From 40e2c33e48ee810c8062a779915c7c79b3ef9685 Mon Sep 17 00:00:00 2001
From: Raunaq Morarka <raunaqmorarka@gmail.com>
Date: Mon, 22 Jun 2026 11:53:39 +0530
Subject: [PATCH 1/5] Fix VectorFastPFOR corruption on reused output buffers

slowpack OR-accumulates into the output, so a reused buffer kept stale
bits. Zero the target words first.
---
 pom.xml                                       | 12 +++++
 .../vector/VectorBitPacker.java               |  6 +++
 .../vector/VectorFastPFORTest.java            | 49 +++++++++++++++++++
 3 files changed, 67 insertions(+)
 create mode 100644 src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java

diff --git a/pom.xml b/pom.xml
index 180a49d..ccdc909 100644
--- a/pom.xml
+++ b/pom.xml
@@ -105,6 +105,18 @@
               </excludes>
             </configuration>
           </execution>
+          <execution>
+            <id>default-testCompile</id>
+            <phase>test-compile</phase>
+            <goals>
+              <goal>testCompile</goal>
+            </goals>
+            <configuration>
+              <testExcludes>
+                <testExclude>me/lemire/integercompression/vector/*</testExclude>
+              </testExcludes>
+            </configuration>
+          </execution>
           <!-- The vector module is experimental, currently only for advanced users. -->
           <!--<execution>
             <id>vector-fastpfor</id>
diff --git a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
index 9b2e1ca..052499c 100644
--- a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
+++ b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
@@ -428,6 +428,11 @@ public static void fastunpack(final int[] in, int inpos, final int[] out,
     }
   }
 
+  /**
+   * Packs an arbitrary count of integers, OR-accumulating into {@code out}.
+   * Zeroes its target words first so a reused output buffer carries no stale
+   * bits.
+   */
   public static int slowpack(final int[] in, int inpos, int inlen,
                              final int[] out, int outpos, int b) {
     if (inlen == 0)
@@ -437,6 +442,7 @@ public static int slowpack(final int[] in, int inpos, int inlen,
       return outpos + inlen;
     }
     int mask = (1 << b) - 1;
+    Arrays.fill(out, outpos, outpos + (inlen * b + 31) / 32, 0);
     int c = 0;
     int l = 0;
     int r = 0;
diff --git a/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java b/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
new file mode 100644
index 0000000..04d67d1
--- /dev/null
+++ b/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
@@ -0,0 +1,49 @@
+/**
+ * This code is released under the
+ * Apache License Version 2.0 http://www.apache.org/licenses/.
+ *
+ * (c) Daniel Lemire, http://lemire.me/en/
+ */
+package me.lemire.integercompression.vector;
+
+import static org.junit.Assert.assertArrayEquals;
+
+import java.util.Arrays;
+
+import org.junit.Test;
+
+import me.lemire.integercompression.IntWrapper;
+
+/**
+ * Tests for the vectorized FastPFOR codec.
+ */
+public class VectorFastPFORTest {
+
+  /**
+   * A few exceptions in a single block leave a sub-block remainder that is
+   * packed with slowpack, which OR-accumulates into the output. Compressing
+   * into a non-zero buffer must still produce a clean roundtrip.
+   */
+  @Test
+  public void dirtyOutputBufferRoundTrip() {
+    int[] data = new int[VectorFastPFOR.BLOCK_SIZE];
+    for (int i = 0; i < data.length; i++) {
+      data[i] = i % 8; // base values fit in 3 bits
+    }
+    data[5] = 1 << 20; // exceptions sharing one width, count not a multiple
+    data[200] = 1 << 20; // of BLOCK_SIZE, so the remainder goes through slowpack
+
+    VectorFastPFOR codec = new VectorFastPFOR();
+    int[] compressed = new int[2 * data.length];
+    Arrays.fill(compressed, -1); // stale bits the slowpack remainder must overwrite
+    IntWrapper inpos = new IntWrapper(0);
+    IntWrapper outpos = new IntWrapper(0);
+    codec.headlessCompress(data, inpos, data.length, compressed, outpos);
+
+    int[] recovered = new int[data.length];
+    codec.headlessUncompress(compressed, new IntWrapper(0), outpos.get(),
+        recovered, new IntWrapper(0), data.length);
+
+    assertArrayEquals(data, recovered);
+  }
+}

From 2a9e75033b34de1804df38eb040121abfaf1bd91 Mon Sep 17 00:00:00 2001
From: Raunaq Morarka <raunaqmorarka@gmail.com>
Date: Mon, 22 Jun 2026 11:55:25 +0530
Subject: [PATCH 2/5] Add 128-bit vector bit-packing with width-tagged streams

Introduce a VectorBitPackerKernels interface with a 128-lane
VectorBitPacker128 for Arm NEON and other 128-bit hardware, alongside
the existing 512-bit kernel. A LaneWidth enum selects the encode kernel
from the preferred vector width and tags each stream; decode dispatches
to the tagging width's kernel and fails loud when the host runs only
narrower lanes natively. slowpack/slowunpack move into VectorFastPFOR.

Decode speedup over the scalar FastPFOR codec:
  Graviton2 (Neoverse N1, NEON 128): 2.0-2.9x (~2.7x), encode ~15% faster
  Graviton4 (Neoverse V2, SVE 128): 2.3-4.1x (~3.0x), encode ~8% faster
---
 .../vector/VectorBitPacker.java               |    91 +-
 .../vector/VectorBitPacker128.java            | 31953 ++++++++++++++++
 .../vector/VectorBitPackerKernels.java        |    77 +
 .../vector/VectorFastPFOR.java                |   140 +-
 .../vector/VectorBitPackerTest.java           |    68 +
 .../vector/VectorFastPFORTest.java            |    36 +
 6 files changed, 32270 insertions(+), 95 deletions(-)
 create mode 100644 src/main/java/me/lemire/integercompression/vector/VectorBitPacker128.java
 create mode 100644 src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java
 create mode 100644 src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java

diff --git a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
index 052499c..cd19344 100644
--- a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
+++ b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
@@ -15,7 +15,7 @@
  * classes.
  *
  */
-public class VectorBitPacker {
+public class VectorBitPacker implements VectorBitPackerKernels {
   private static final VectorSpecies<Integer> SPECIES_512 =
       IntVector.SPECIES_512;
   private static final VectorSpecies<Integer> SPECIES_256 =
@@ -101,8 +101,9 @@ public class VectorBitPacker {
    * @param b
    *                number of bits to use per integer
    */
-  public static void fastpack(final int[] in, int inpos, final int[] out,
-                              int outpos, int b) {
+  @Override
+  public void fastpack(final int[] in, int inpos, final int[] out,
+                       int outpos, int b) {
     switch (b) {
     case 0:
       break;
@@ -205,7 +206,8 @@ public static void fastpack(final int[] in, int inpos, final int[] out,
     }
   }
 
-  static void fastpackNoMask(final int[] in, int inpos, final int[] out,
+  @Override
+  public void fastpackNoMask(final int[] in, int inpos, final int[] out,
                              int outpos, int b) {
     switch (b) {
     case 0:
@@ -323,8 +325,9 @@ static void fastpackNoMask(final int[] in, int inpos, final int[] out,
    * @param b
    *                number of bits to use per integer
    */
-  public static void fastunpack(final int[] in, int inpos, final int[] out,
-                                int outpos, int b) {
+  @Override
+  public void fastunpack(final int[] in, int inpos, final int[] out,
+                         int outpos, int b) {
     switch (b) {
     case 0:
       Arrays.fill(out, outpos, outpos + 256, 0);
@@ -428,82 +431,6 @@ public static void fastunpack(final int[] in, int inpos, final int[] out,
     }
   }
 
-  /**
-   * Packs an arbitrary count of integers, OR-accumulating into {@code out}.
-   * Zeroes its target words first so a reused output buffer carries no stale
-   * bits.
-   */
-  public static int slowpack(final int[] in, int inpos, int inlen,
-                             final int[] out, int outpos, int b) {
-    if (inlen == 0)
-      return outpos;
-    if (b == 32) {
-      System.arraycopy(in, inpos, out, outpos, inlen);
-      return outpos + inlen;
-    }
-    int mask = (1 << b) - 1;
-    Arrays.fill(out, outpos, outpos + (inlen * b + 31) / 32, 0);
-    int c = 0;
-    int l = 0;
-    int r = 0;
-    int val = 0;
-    for (int i = 0; i < inlen; i++) {
-      val = in[inpos + i] & mask;
-      out[outpos] |= val << (c + r);
-      c += b;
-      l = (32 - r) % b;
-      if (c + r >= 32) {
-        if (i < inlen - 1 || l != 0)
-          outpos++;
-        r = l == 0 ? 0 : b - l;
-        if (l != 0)
-          out[outpos] = val >> (b - r);
-        c = 0;
-      }
-    }
-    return outpos;
-  }
-
-  public static int slowunpack(final int[] in, int inpos, final int[] out,
-                               int outpos, int outlen, int b) {
-    if (outlen == 0) {
-      return inpos;
-    }
-    if (b == 32) {
-      System.arraycopy(in, inpos, out, outpos, outlen);
-      return inpos + outlen;
-    }
-    int mask = (1 << b) - 1;
-    int limit = outpos + outlen;
-    int r = 0;
-    int val = 0;
-    int i = 0;
-    for (; outpos < limit; i++) {
-      if (r > 0)
-        out[outpos++] =
-            (val >>> (32 - (b - r))) | ((in[inpos + i] << (b - r)) & mask);
-      val = in[inpos + i];
-      int j = 0;
-      int l = 32 - r;
-      int ll = l % b == 0 ? l : l - b;
-      while (j < ll && outpos < limit) {
-        out[outpos++] = (val >> (j + r)) & mask;
-        j += b;
-      }
-      r = l % b == 0 ? 0 : b - (l % b);
-    }
-    return inpos + i;
-  }
-
-  public static int numCompressedInts(int n, int b) {
-    int width = b % 2 == 0 ? VLEN_512 : VLEN_256;
-    if (n <= width)
-      return n;
-    int intsPerVec = (32 / b) * width;
-    int q = (n + intsPerVec - 1) / intsPerVec;
-    return q * width;
-  }
-
   private static void fastpack1(final int[] in, int inpos, final int[] out,
                                 int outpos) {
     var iV = IntVector.fromArray(SPECIES_256, in, inpos);
diff --git a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker128.java b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker128.java
new file mode 100644
index 0000000..6e08546
--- /dev/null
+++ b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker128.java
@@ -0,0 +1,31953 @@
+/**
+ * This code is released under the
+ * Apache License Version 2.0 http://www.apache.org/licenses/.
+ *
+ * (c) Daniel Lemire, http://lemire.me/en/
+ */
+package me.lemire.integercompression.vector;
+
+import java.util.Arrays;
+import jdk.incubator.vector.*;
+
+/**
+ * Vectorized bit-packing routines using 128-bit (4 x int32) vectors.
+ *
+ * A 256-integer block is packed across 4 SIMD lanes, each lane packing 64
+ * values into 2*b 32-bit words. Selected at runtime by VectorBitPackerKernels
+ * when the preferred hardware vector width is 128 bits (e.g. Arm NEON, Graviton).
+ */
+public class VectorBitPacker128 implements VectorBitPackerKernels {
+  private static final VectorSpecies<Integer> SPECIES_128 =
+      IntVector.SPECIES_128;
+  private static final int VLEN_128 = 4;
+  private static final int BLOCK_SIZE = 256;
+
+  private static final IntVector MASK_1 =
+      IntVector.broadcast(SPECIES_128, (1 << 1) - 1);
+  private static final IntVector MASK_2 =
+      IntVector.broadcast(SPECIES_128, (1 << 2) - 1);
+  private static final IntVector MASK_3 =
+      IntVector.broadcast(SPECIES_128, (1 << 3) - 1);
+  private static final IntVector MASK_4 =
+      IntVector.broadcast(SPECIES_128, (1 << 4) - 1);
+  private static final IntVector MASK_5 =
+      IntVector.broadcast(SPECIES_128, (1 << 5) - 1);
+  private static final IntVector MASK_6 =
+      IntVector.broadcast(SPECIES_128, (1 << 6) - 1);
+  private static final IntVector MASK_7 =
+      IntVector.broadcast(SPECIES_128, (1 << 7) - 1);
+  private static final IntVector MASK_8 =
+      IntVector.broadcast(SPECIES_128, (1 << 8) - 1);
+  private static final IntVector MASK_9 =
+      IntVector.broadcast(SPECIES_128, (1 << 9) - 1);
+  private static final IntVector MASK_10 =
+      IntVector.broadcast(SPECIES_128, (1 << 10) - 1);
+  private static final IntVector MASK_11 =
+      IntVector.broadcast(SPECIES_128, (1 << 11) - 1);
+  private static final IntVector MASK_12 =
+      IntVector.broadcast(SPECIES_128, (1 << 12) - 1);
+  private static final IntVector MASK_13 =
+      IntVector.broadcast(SPECIES_128, (1 << 13) - 1);
+  private static final IntVector MASK_14 =
+      IntVector.broadcast(SPECIES_128, (1 << 14) - 1);
+  private static final IntVector MASK_15 =
+      IntVector.broadcast(SPECIES_128, (1 << 15) - 1);
+  private static final IntVector MASK_16 =
+      IntVector.broadcast(SPECIES_128, (1 << 16) - 1);
+  private static final IntVector MASK_17 =
+      IntVector.broadcast(SPECIES_128, (1 << 17) - 1);
+  private static final IntVector MASK_18 =
+      IntVector.broadcast(SPECIES_128, (1 << 18) - 1);
+  private static final IntVector MASK_19 =
+      IntVector.broadcast(SPECIES_128, (1 << 19) - 1);
+  private static final IntVector MASK_20 =
+      IntVector.broadcast(SPECIES_128, (1 << 20) - 1);
+  private static final IntVector MASK_21 =
+      IntVector.broadcast(SPECIES_128, (1 << 21) - 1);
+  private static final IntVector MASK_22 =
+      IntVector.broadcast(SPECIES_128, (1 << 22) - 1);
+  private static final IntVector MASK_23 =
+      IntVector.broadcast(SPECIES_128, (1 << 23) - 1);
+  private static final IntVector MASK_24 =
+      IntVector.broadcast(SPECIES_128, (1 << 24) - 1);
+  private static final IntVector MASK_25 =
+      IntVector.broadcast(SPECIES_128, (1 << 25) - 1);
+  private static final IntVector MASK_26 =
+      IntVector.broadcast(SPECIES_128, (1 << 26) - 1);
+  private static final IntVector MASK_27 =
+      IntVector.broadcast(SPECIES_128, (1 << 27) - 1);
+  private static final IntVector MASK_28 =
+      IntVector.broadcast(SPECIES_128, (1 << 28) - 1);
+  private static final IntVector MASK_29 =
+      IntVector.broadcast(SPECIES_128, (1 << 29) - 1);
+  private static final IntVector MASK_30 =
+      IntVector.broadcast(SPECIES_128, (1 << 30) - 1);
+  private static final IntVector MASK_31 =
+      IntVector.broadcast(SPECIES_128, (1 << 31) - 1);
+
+  @Override
+  public void fastpack(final int[] in, int inpos, final int[] out,
+                       int outpos, int b) {
+    switch (b) {
+    case 0:
+      break;
+    case 1:
+      fastpack1(in, inpos, out, outpos);
+      break;
+    case 2:
+      fastpack2(in, inpos, out, outpos);
+      break;
+    case 3:
+      fastpack3(in, inpos, out, outpos);
+      break;
+    case 4:
+      fastpack4(in, inpos, out, outpos);
+      break;
+    case 5:
+      fastpack5(in, inpos, out, outpos);
+      break;
+    case 6:
+      fastpack6(in, inpos, out, outpos);
+      break;
+    case 7:
+      fastpack7(in, inpos, out, outpos);
+      break;
+    case 8:
+      fastpack8(in, inpos, out, outpos);
+      break;
+    case 9:
+      fastpack9(in, inpos, out, outpos);
+      break;
+    case 10:
+      fastpack10(in, inpos, out, outpos);
+      break;
+    case 11:
+      fastpack11(in, inpos, out, outpos);
+      break;
+    case 12:
+      fastpack12(in, inpos, out, outpos);
+      break;
+    case 13:
+      fastpack13(in, inpos, out, outpos);
+      break;
+    case 14:
+      fastpack14(in, inpos, out, outpos);
+      break;
+    case 15:
+      fastpack15(in, inpos, out, outpos);
+      break;
+    case 16:
+      fastpack16(in, inpos, out, outpos);
+      break;
+    case 17:
+      fastpack17(in, inpos, out, outpos);
+      break;
+    case 18:
+      fastpack18(in, inpos, out, outpos);
+      break;
+    case 19:
+      fastpack19(in, inpos, out, outpos);
+      break;
+    case 20:
+      fastpack20(in, inpos, out, outpos);
+      break;
+    case 21:
+      fastpack21(in, inpos, out, outpos);
+      break;
+    case 22:
+      fastpack22(in, inpos, out, outpos);
+      break;
+    case 23:
+      fastpack23(in, inpos, out, outpos);
+      break;
+    case 24:
+      fastpack24(in, inpos, out, outpos);
+      break;
+    case 25:
+      fastpack25(in, inpos, out, outpos);
+      break;
+    case 26:
+      fastpack26(in, inpos, out, outpos);
+      break;
+    case 27:
+      fastpack27(in, inpos, out, outpos);
+      break;
+    case 28:
+      fastpack28(in, inpos, out, outpos);
+      break;
+    case 29:
+      fastpack29(in, inpos, out, outpos);
+      break;
+    case 30:
+      fastpack30(in, inpos, out, outpos);
+      break;
+    case 31:
+      fastpack31(in, inpos, out, outpos);
+      break;
+    case 32:
+      System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
+      break;
+    }
+  }
+
+  @Override
+  public void fastpackNoMask(final int[] in, int inpos, final int[] out,
+                             int outpos, int b) {
+    switch (b) {
+    case 0:
+      break;
+    case 1:
+      fastpackNoMask1(in, inpos, out, outpos);
+      break;
+    case 2:
+      fastpackNoMask2(in, inpos, out, outpos);
+      break;
+    case 3:
+      fastpackNoMask3(in, inpos, out, outpos);
+      break;
+    case 4:
+      fastpackNoMask4(in, inpos, out, outpos);
+      break;
+    case 5:
+      fastpackNoMask5(in, inpos, out, outpos);
+      break;
+    case 6:
+      fastpackNoMask6(in, inpos, out, outpos);
+      break;
+    case 7:
+      fastpackNoMask7(in, inpos, out, outpos);
+      break;
+    case 8:
+      fastpackNoMask8(in, inpos, out, outpos);
+      break;
+    case 9:
+      fastpackNoMask9(in, inpos, out, outpos);
+      break;
+    case 10:
+      fastpackNoMask10(in, inpos, out, outpos);
+      break;
+    case 11:
+      fastpackNoMask11(in, inpos, out, outpos);
+      break;
+    case 12:
+      fastpackNoMask12(in, inpos, out, outpos);
+      break;
+    case 13:
+      fastpackNoMask13(in, inpos, out, outpos);
+      break;
+    case 14:
+      fastpackNoMask14(in, inpos, out, outpos);
+      break;
+    case 15:
+      fastpackNoMask15(in, inpos, out, outpos);
+      break;
+    case 16:
+      fastpackNoMask16(in, inpos, out, outpos);
+      break;
+    case 17:
+      fastpackNoMask17(in, inpos, out, outpos);
+      break;
+    case 18:
+      fastpackNoMask18(in, inpos, out, outpos);
+      break;
+    case 19:
+      fastpackNoMask19(in, inpos, out, outpos);
+      break;
+    case 20:
+      fastpackNoMask20(in, inpos, out, outpos);
+      break;
+    case 21:
+      fastpackNoMask21(in, inpos, out, outpos);
+      break;
+    case 22:
+      fastpackNoMask22(in, inpos, out, outpos);
+      break;
+    case 23:
+      fastpackNoMask23(in, inpos, out, outpos);
+      break;
+    case 24:
+      fastpackNoMask24(in, inpos, out, outpos);
+      break;
+    case 25:
+      fastpackNoMask25(in, inpos, out, outpos);
+      break;
+    case 26:
+      fastpackNoMask26(in, inpos, out, outpos);
+      break;
+    case 27:
+      fastpackNoMask27(in, inpos, out, outpos);
+      break;
+    case 28:
+      fastpackNoMask28(in, inpos, out, outpos);
+      break;
+    case 29:
+      fastpackNoMask29(in, inpos, out, outpos);
+      break;
+    case 30:
+      fastpackNoMask30(in, inpos, out, outpos);
+      break;
+    case 31:
+      fastpackNoMask31(in, inpos, out, outpos);
+      break;
+    case 32:
+      System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
+      break;
+    }
+  }
+
+  @Override
+  public void fastunpack(final int[] in, int inpos, final int[] out,
+                         int outpos, int b) {
+    switch (b) {
+    case 0:
+      Arrays.fill(out, outpos, outpos + 256, 0);
+      break;
+    case 1:
+      fastunpack1(in, inpos, out, outpos);
+      break;
+    case 2:
+      fastunpack2(in, inpos, out, outpos);
+      break;
+    case 3:
+      fastunpack3(in, inpos, out, outpos);
+      break;
+    case 4:
+      fastunpack4(in, inpos, out, outpos);
+      break;
+    case 5:
+      fastunpack5(in, inpos, out, outpos);
+      break;
+    case 6:
+      fastunpack6(in, inpos, out, outpos);
+      break;
+    case 7:
+      fastunpack7(in, inpos, out, outpos);
+      break;
+    case 8:
+      fastunpack8(in, inpos, out, outpos);
+      break;
+    case 9:
+      fastunpack9(in, inpos, out, outpos);
+      break;
+    case 10:
+      fastunpack10(in, inpos, out, outpos);
+      break;
+    case 11:
+      fastunpack11(in, inpos, out, outpos);
+      break;
+    case 12:
+      fastunpack12(in, inpos, out, outpos);
+      break;
+    case 13:
+      fastunpack13(in, inpos, out, outpos);
+      break;
+    case 14:
+      fastunpack14(in, inpos, out, outpos);
+      break;
+    case 15:
+      fastunpack15(in, inpos, out, outpos);
+      break;
+    case 16:
+      fastunpack16(in, inpos, out, outpos);
+      break;
+    case 17:
+      fastunpack17(in, inpos, out, outpos);
+      break;
+    case 18:
+      fastunpack18(in, inpos, out, outpos);
+      break;
+    case 19:
+      fastunpack19(in, inpos, out, outpos);
+      break;
+    case 20:
+      fastunpack20(in, inpos, out, outpos);
+      break;
+    case 21:
+      fastunpack21(in, inpos, out, outpos);
+      break;
+    case 22:
+      fastunpack22(in, inpos, out, outpos);
+      break;
+    case 23:
+      fastunpack23(in, inpos, out, outpos);
+      break;
+    case 24:
+      fastunpack24(in, inpos, out, outpos);
+      break;
+    case 25:
+      fastunpack25(in, inpos, out, outpos);
+      break;
+    case 26:
+      fastunpack26(in, inpos, out, outpos);
+      break;
+    case 27:
+      fastunpack27(in, inpos, out, outpos);
+      break;
+    case 28:
+      fastunpack28(in, inpos, out, outpos);
+      break;
+    case 29:
+      fastunpack29(in, inpos, out, outpos);
+      break;
+    case 30:
+      fastunpack30(in, inpos, out, outpos);
+      break;
+    case 31:
+      fastunpack31(in, inpos, out, outpos);
+      break;
+    case 32:
+      System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
+      break;
+    }
+  }
+
+  private static void fastpack1(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask1(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack1(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 31).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    iV.and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 31).and(MASK_1).intoArray(out, outpos);
+  }
+
+  private static void fastpack2(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask2(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack2(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    iV.and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    iV.and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
+  }
+
+  private static void fastpack3(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask3(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack3(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_3).intoArray(out, outpos);
+  }
+
+  private static void fastpack4(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask4(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack4(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+  }
+
+  private static void fastpack5(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask5(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack5(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    iV.and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_5).intoArray(out, outpos);
+  }
+
+  private static void fastpack6(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask6(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack6(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    iV.and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    iV.and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
+  }
+
+  private static void fastpack7(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask7(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack7(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    iV.and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_7).intoArray(out, outpos);
+  }
+
+  private static void fastpack8(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask8(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack8(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+  }
+
+  private static void fastpack9(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask9(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack9(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    iV.and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_9).intoArray(out, outpos);
+  }
+
+  private static void fastpack10(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask10(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack10(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    iV.and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    iV.and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+  }
+
+  private static void fastpack11(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask11(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack11(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    iV.and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_11).intoArray(out, outpos);
+  }
+
+  private static void fastpack12(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask12(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack12(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+  }
+
+  private static void fastpack13(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask13(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack13(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    iV.and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_13).intoArray(out, outpos);
+  }
+
+  private static void fastpack14(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask14(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack14(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    iV.and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    iV.and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    iV.and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+  }
+
+  private static void fastpack15(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask15(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack15(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_15).intoArray(out, outpos);
+  }
+
+  private static void fastpack16(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask16(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack16(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+  }
+
+  private static void fastpack17(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask17(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack17(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    iV.and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_17).intoArray(out, outpos);
+  }
+
+  private static void fastpack18(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask18(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack18(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    iV.and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    iV.and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    iV.and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+  }
+
+  private static void fastpack19(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask19(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack19(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    iV.and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_19).intoArray(out, outpos);
+  }
+
+  private static void fastpack20(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask20(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack20(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+  }
+
+  private static void fastpack21(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask21(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack21(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    iV.and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_21).intoArray(out, outpos);
+  }
+
+  private static void fastpack22(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask22(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack22(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    iV.and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    iV.and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    iV.and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
+  }
+
+  private static void fastpack23(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask23(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack23(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    iV.and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_23).intoArray(out, outpos);
+  }
+
+  private static void fastpack24(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask24(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack24(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+  }
+
+  private static void fastpack25(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask25(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack25(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    iV.and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_25).intoArray(out, outpos);
+  }
+
+  private static void fastpack26(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask26(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack26(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    iV.and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    iV.and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    iV.and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
+  }
+
+  private static void fastpack27(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask27(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack27(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    iV.and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_27).intoArray(out, outpos);
+  }
+
+  private static void fastpack28(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask28(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack28(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+  }
+
+  private static void fastpack29(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask29(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack29(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(134217727).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    iV.and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(134217727).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_29).intoArray(out, outpos);
+  }
+
+  private static void fastpack30(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask30(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack30(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    iV.and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    iV.and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    iV.and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_30).intoArray(out, outpos);
+  }
+
+  private static void fastpack31(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV.and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask31(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 252);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastunpack31(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_128, in, inpos);
+    iV.and(MASK_31).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 4);
+    oV = iV.and(1073741823).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 8);
+    oV = iV.and(536870911).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 12);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 16);
+    oV = iV.and(134217727).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 20);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 24);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 28);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 32);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 36);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 40);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 44);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 48);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 52);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 56);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 60);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 64);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 68);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 72);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 76);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 80);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 84);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 88);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 92);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 96);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 100);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 104);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 108);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 112);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 116);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 120);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_31).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 124);
+    iV.and(MASK_31).intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 128);
+    oV = iV.and(1073741823).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 132);
+    oV = iV.and(536870911).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 136);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 140);
+    oV = iV.and(134217727).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 144);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 148);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 152);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 156);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 160);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 164);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 168);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 172);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 176);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 180);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 184);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 188);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 192);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 196);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 200);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 204);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 208);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 212);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 216);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 220);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 224);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 228);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 232);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 236);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 240);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_128, in, inpos + 244);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_128;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_31).intoArray(out, outpos);
+  }
+
+}
diff --git a/src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java b/src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java
new file mode 100644
index 0000000..da7b7c2
--- /dev/null
+++ b/src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java
@@ -0,0 +1,77 @@
+/**
+ * This code is released under the
+ * Apache License Version 2.0 http://www.apache.org/licenses/.
+ *
+ * (c) Daniel Lemire, http://lemire.me/en/
+ */
+package me.lemire.integercompression.vector;
+
+import jdk.incubator.vector.IntVector;
+
+/**
+ * Width-specific vectorized bit-packing kernels for a 256-integer block.
+ * Implemented by VectorBitPacker (512-bit lanes) and VectorBitPacker128
+ * (128-bit lanes). The packed layout differs per width, so a stream is decoded
+ * by the same kernel that packed it.
+ */
+public interface VectorBitPackerKernels {
+
+  void fastpack(int[] in, int inpos, int[] out, int outpos, int b);
+
+  void fastpackNoMask(int[] in, int inpos, int[] out, int outpos, int b);
+
+  void fastunpack(int[] in, int inpos, int[] out, int outpos, int b);
+
+  /**
+   * Hardware vector lane width a stream was packed for. The packed byte layout
+   * differs per width and is not interchangeable across widths, so the stream
+   * carries its width (as {@link #code}) and is decoded by {@link #kernel}.
+   */
+  enum LaneWidth {
+    BITS_128(0, 128, new VectorBitPacker128()),
+    BITS_512(2, 512, new VectorBitPacker());
+
+    /** Compact wire tag stored in the stream (fits in 2 bits). */
+    public final int code;
+    /** Native vector lane width in bits. */
+    public final int bits;
+    /** Kernel that packs and unpacks at this width. */
+    public final VectorBitPackerKernels kernel;
+
+    LaneWidth(int code, int bits, VectorBitPackerKernels kernel) {
+      this.code = code;
+      this.bits = bits;
+      this.kernel = kernel;
+    }
+
+    /** Width whose kernel runs natively on this machine (encode default). */
+    public static final LaneWidth PREFERRED =
+        forHost(IntVector.SPECIES_PREFERRED.vectorBitSize());
+
+    /** Largest kernel width that runs natively on a machine of {@code hostBits}. */
+    public static LaneWidth forHost(int hostBits) {
+      LaneWidth best = null;
+      for (LaneWidth width : values()) {
+        if (width.bits <= hostBits && (best == null || width.bits > best.bits)) {
+          best = width;
+        }
+      }
+      if (best == null) {
+        throw new IllegalStateException(
+            "no vector bit-packing kernel fits this machine's preferred vector width of "
+            + hostBits + " bits");
+      }
+      return best;
+    }
+
+    /** Maps a stream wire tag back to its width. */
+    public static LaneWidth fromCode(int code) {
+      for (LaneWidth width : values()) {
+        if (width.code == code) {
+          return width;
+        }
+      }
+      throw new IllegalArgumentException("unknown vector lane-width tag " + code);
+    }
+  }
+}
diff --git a/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java b/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java
index 7374fa5..adbe432 100644
--- a/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java
+++ b/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java
@@ -12,6 +12,7 @@
 import me.lemire.integercompression.IntegerCODEC;
 import me.lemire.integercompression.SkippableIntegerCODEC;
 import me.lemire.integercompression.IntWrapper;
+import me.lemire.integercompression.vector.VectorBitPackerKernels.LaneWidth;
 
 /**
  * This is a patching scheme designed for speed.
@@ -41,6 +42,14 @@
  * For multi-threaded applications, each thread should use its own FastPFOR
  * object.
  *
+ * Blocks are packed in a vectorized layout that differs by hardware vector
+ * lane width, so each stream is tagged with the width it was packed for and is
+ * decoded by the matching kernel. Decoding requires a machine whose preferred
+ * vector width is at least the stream's; a narrower machine fails fast rather
+ * than emulating. The default constructor packs at this machine's preferred
+ * width; the {@code (int, LaneWidth)} constructor pins a width so a
+ * heterogeneous cluster can decode on its narrowest node.
+ *
  * @author Daniel Lemire
  */
 public class VectorFastPFOR implements IntegerCODEC, SkippableIntegerCODEC {
@@ -50,7 +59,14 @@ public class VectorFastPFOR implements IntegerCODEC, SkippableIntegerCODEC {
   public final static int BLOCK_SIZE = 256;
   private final static int INTS_PER_BLOCK = BLOCK_SIZE >>> 5;
 
+  // The page header word holds the metadata offset in its low 30 bits and the
+  // packing lane-width tag in its top 2 bits.
+  private final static int WIDTH_SHIFT = 30;
+  private final static int WHEREMETA_MASK = (1 << WIDTH_SHIFT) - 1;
+
   private final int pageSize;
+  private final LaneWidth encodeWidth;
+  private final VectorBitPackerKernels encoder;
   private final int[][] dataTobePacked = new int[33][];
   private int[] exceptData = null;
 
@@ -64,9 +80,18 @@ public class VectorFastPFOR implements IntegerCODEC, SkippableIntegerCODEC {
    * @param pagesize
    *                the desired page size (recommended value is
    * FastPFOR.DEFAULT_PAGE_SIZE)
+   * @param encodeWidth
+   *                the vector lane width to pack with. Use
+   * {@link LaneWidth#PREFERRED} for this machine's fastest layout, or pin a
+   * cluster to its narrowest node's width so every node can decode the stream.
    */
-  private VectorFastPFOR(int pagesize) {
+  public VectorFastPFOR(int pagesize, LaneWidth encodeWidth) {
+    if (pagesize >= (1 << WIDTH_SHIFT))
+      throw new IllegalArgumentException("page size must be smaller than "
+                                         + (1 << WIDTH_SHIFT));
     pageSize = pagesize;
+    this.encodeWidth = encodeWidth;
+    this.encoder = encodeWidth.kernel;
     // Initiate arrrays.
     bem = new byte[3 * pageSize / BLOCK_SIZE + pagesize];
     for (int k = 1; k < dataTobePacked.length; ++k)
@@ -75,9 +100,10 @@ private VectorFastPFOR(int pagesize) {
   }
 
   /**
-   * Construct the fastPFOR CODEC with default parameters.
+   * Construct the fastPFOR CODEC with default parameters, packing with this
+   * machine's preferred vector lane width.
    */
-  public VectorFastPFOR() { this(DEFAULT_PAGE_SIZE); }
+  public VectorFastPFOR() { this(DEFAULT_PAGE_SIZE, LaneWidth.PREFERRED); }
 
   /**
    * Compress data in blocks of BLOCK_SIZE integers (if fewer than BLOCK_SIZE
@@ -165,11 +191,11 @@ private void encodePage(int[] in, IntWrapper inpos, int thissize, int[] out,
       } else {
         bindex += 2;
       }
-      VectorBitPacker.fastpack(in, tmpinpos, out, tmpoutpos, tmpbestb);
+      encoder.fastpack(in, tmpinpos, out, tmpoutpos, tmpbestb);
       tmpoutpos += INTS_PER_BLOCK * tmpbestb;
     }
     inpos.set(tmpinpos);
-    out[headerpos] = tmpoutpos - headerpos;
+    out[headerpos] = (tmpoutpos - headerpos) | (encodeWidth.code << WIDTH_SHIFT);
 
     int bytesize = bindex;
     out[tmpoutpos++] = bytesize;
@@ -196,13 +222,13 @@ private void encodePage(int[] in, IntWrapper inpos, int thissize, int[] out,
         int j = 0;
         int n = (dataPointers[k] / BLOCK_SIZE) * BLOCK_SIZE;
         for (; j < n; j += BLOCK_SIZE) {
-          VectorBitPacker.fastpackNoMask(dataTobePacked[k], j, out, tmpoutpos,
+          encoder.fastpackNoMask(dataTobePacked[k], j, out, tmpoutpos,
                                          k);
           tmpoutpos += INTS_PER_BLOCK * k;
         }
         int r = dataPointers[k] % BLOCK_SIZE;
         if (r != 0) {
-          tmpoutpos = VectorBitPacker.slowpack(dataTobePacked[k], j, r, out,
+          tmpoutpos = slowpack(dataTobePacked[k], j, r, out,
                                                tmpoutpos, k);
           tmpoutpos++;
         }
@@ -246,10 +272,29 @@ private void loadMetaData(int[] in, int inexcept, int bytesize) {
     }
   }
 
+  /**
+   * Rejects a stream packed for wider lanes than this machine runs natively.
+   * Decoding it would silently fall back to scalar emulation; failing loud lets
+   * the caller re-encode the cluster at a lower width instead.
+   */
+  static void checkDecodable(LaneWidth streamWidth, LaneWidth hostWidth) {
+    if (streamWidth.bits > hostWidth.bits) {
+      throw new IllegalStateException(
+          "VectorFastPFOR stream was packed for " + streamWidth.bits
+          + "-bit vector lanes, but this machine runs at most " + hostWidth.bits
+          + "-bit lanes natively. Re-encode with lanes <= " + hostWidth.bits
+          + " bits.");
+    }
+  }
+
   private void decodePage(int[] in, IntWrapper inpos, int[] out,
                           IntWrapper outpos, int thissize) {
     final int initpos = inpos.get();
-    final int wheremeta = in[inpos.get()];
+    final int header = in[inpos.get()];
+    final int wheremeta = header & WHEREMETA_MASK;
+    final LaneWidth streamWidth = LaneWidth.fromCode(header >>> WIDTH_SHIFT);
+    checkDecodable(streamWidth, LaneWidth.PREFERRED);
+    final VectorBitPackerKernels decoder = streamWidth.kernel;
     inpos.increment();
     int inexcept = initpos + wheremeta;
 
@@ -268,11 +313,11 @@ private void decodePage(int[] in, IntWrapper inpos, int[] out,
           int j = 0;
           int len = (size / BLOCK_SIZE) * BLOCK_SIZE;
           for (; j < len; j += BLOCK_SIZE) {
-            VectorBitPacker.fastunpack(in, inexcept, dataTobePacked[k], j, k);
+            decoder.fastunpack(in, inexcept, dataTobePacked[k], j, k);
             inexcept += INTS_PER_BLOCK * k;
           }
           int r = size % BLOCK_SIZE;
-          inexcept = VectorBitPacker.slowunpack(in, inexcept, dataTobePacked[k],
+          inexcept = slowunpack(in, inexcept, dataTobePacked[k],
                                                 j, r, k);
         } else {
           int j = 0;
@@ -282,12 +327,12 @@ private void decodePage(int[] in, IntWrapper inpos, int[] out,
           System.arraycopy(in, inexcept, buf, 0, in.length - inexcept);
           int l = (size / BLOCK_SIZE) * BLOCK_SIZE;
           for (; j < l; j += BLOCK_SIZE) {
-            VectorBitPacker.fastunpack(buf, inexcept - initinexcept,
+            decoder.fastunpack(buf, inexcept - initinexcept,
                                        dataTobePacked[k], j, k);
             inexcept += INTS_PER_BLOCK * k;
           }
           int r = size % BLOCK_SIZE;
-          inexcept = VectorBitPacker.slowunpack(in, inexcept, dataTobePacked[k],
+          inexcept = slowunpack(in, inexcept, dataTobePacked[k],
                                                 j, r, k);
         }
       }
@@ -300,7 +345,7 @@ private void decodePage(int[] in, IntWrapper inpos, int[] out,
          ++run, tmpoutpos += BLOCK_SIZE) {
       final int b = bem[idx];                  // byteContainer.get();
       final int cexcept = bem[idx + 1] & 0xFF; // byteContainer.get() & 0xFF;
-      VectorBitPacker.fastunpack(in, tmpinpos, out, tmpoutpos, b);
+      decoder.fastunpack(in, tmpinpos, out, tmpoutpos, b);
       tmpinpos += INTS_PER_BLOCK * b;
       if (cexcept > 0) {
         final int maxbits = bem[idx + 2]; // byteContainer.get();
@@ -363,4 +408,73 @@ public String toString() {
   protected ByteBuffer makeBuffer(int sizeInBytes) {
     return ByteBuffer.allocateDirect(sizeInBytes);
   }
+
+  /**
+   * Packs the sub-block exception remainder, which is not a multiple of the
+   * vector block size, into the sequential scalar layout read back by
+   * {@link #slowunpack}. Zeroes its target words first, then OR-accumulates the
+   * packed bits, so a reused output buffer carries no stale bits.
+   */
+  private static int slowpack(final int[] in, int inpos, int inlen,
+                              final int[] out, int outpos, int b) {
+    if (inlen == 0)
+      return outpos;
+    if (b == 32) {
+      System.arraycopy(in, inpos, out, outpos, inlen);
+      return outpos + inlen;
+    }
+    int mask = (1 << b) - 1;
+    Arrays.fill(out, outpos, outpos + (inlen * b + 31) / 32, 0);
+    int c = 0;
+    int l = 0;
+    int r = 0;
+    int val = 0;
+    for (int i = 0; i < inlen; i++) {
+      val = in[inpos + i] & mask;
+      out[outpos] |= val << (c + r);
+      c += b;
+      l = (32 - r) % b;
+      if (c + r >= 32) {
+        if (i < inlen - 1 || l != 0)
+          outpos++;
+        r = l == 0 ? 0 : b - l;
+        if (l != 0)
+          out[outpos] = val >> (b - r);
+        c = 0;
+      }
+    }
+    return outpos;
+  }
+
+  /** Reverses {@link #slowpack}. */
+  private static int slowunpack(final int[] in, int inpos, final int[] out,
+                                int outpos, int outlen, int b) {
+    if (outlen == 0) {
+      return inpos;
+    }
+    if (b == 32) {
+      System.arraycopy(in, inpos, out, outpos, outlen);
+      return inpos + outlen;
+    }
+    int mask = (1 << b) - 1;
+    int limit = outpos + outlen;
+    int r = 0;
+    int val = 0;
+    int i = 0;
+    for (; outpos < limit; i++) {
+      if (r > 0)
+        out[outpos++] =
+            (val >>> (32 - (b - r))) | ((in[inpos + i] << (b - r)) & mask);
+      val = in[inpos + i];
+      int j = 0;
+      int l = 32 - r;
+      int ll = l % b == 0 ? l : l - b;
+      while (j < ll && outpos < limit) {
+        out[outpos++] = (val >> (j + r)) & mask;
+        j += b;
+      }
+      r = l % b == 0 ? 0 : b - (l % b);
+    }
+    return inpos + i;
+  }
 }
diff --git a/src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java b/src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java
new file mode 100644
index 0000000..0322c77
--- /dev/null
+++ b/src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java
@@ -0,0 +1,68 @@
+/**
+ * This code is released under the
+ * Apache License Version 2.0 http://www.apache.org/licenses/.
+ *
+ * (c) Daniel Lemire, http://lemire.me/en/
+ */
+package me.lemire.integercompression.vector;
+
+import static org.junit.Assert.assertArrayEquals;
+
+import java.util.Random;
+
+import org.junit.Test;
+
+/**
+ * Tests for the width-specific bit-packing kernels.
+ *
+ * VectorBitPacker (256/512-bit lanes) and VectorBitPacker128 (128-bit lanes)
+ * use different lane strides, so their packed layouts differ and are not
+ * wire-compatible. Both must satisfy the same roundtrip contract: packing a
+ * 256-integer block whose values fit in b bits and unpacking it recovers the
+ * input for every width b.
+ */
+public class VectorBitPackerTest {
+
+  private static final int BLOCK_SIZE = 256;
+
+  private static int[] randomBlock(Random random, int b) {
+    int mask = b == 32 ? -1 : (1 << b) - 1;
+    int[] in = new int[BLOCK_SIZE];
+    for (int i = 0; i < BLOCK_SIZE; i++) {
+      in[i] = random.nextInt() & mask;
+    }
+    return in;
+  }
+
+  private static void roundTrip(VectorBitPackerKernels packer) {
+    Random random = new Random(42);
+    for (int b = 1; b <= 32; b++) {
+      int[] in = randomBlock(random, b);
+
+      int[] packed = new int[8 * b];
+      packer.fastpack(in, 0, packed, 0, b);
+      int[] recovered = new int[BLOCK_SIZE];
+      packer.fastunpack(packed, 0, recovered, 0, b);
+      assertArrayEquals("fastpack b=" + b, in, recovered);
+
+      int[] packedNoMask = new int[8 * b];
+      packer.fastpackNoMask(in, 0, packedNoMask, 0, b);
+      int[] recoveredNoMask = new int[BLOCK_SIZE];
+      packer.fastunpack(packedNoMask, 0, recoveredNoMask, 0, b);
+      assertArrayEquals("fastpackNoMask b=" + b, in, recoveredNoMask);
+
+      assertArrayEquals("fastpack vs fastpackNoMask b=" + b, packed,
+          packedNoMask);
+    }
+  }
+
+  @Test
+  public void vectorBitPacker128RoundTrip() {
+    roundTrip(new VectorBitPacker128());
+  }
+
+  @Test
+  public void vectorBitPackerRoundTrip() {
+    roundTrip(new VectorBitPacker());
+  }
+}
diff --git a/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java b/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
index 04d67d1..aaf5ac3 100644
--- a/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
+++ b/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
@@ -7,12 +7,14 @@
 package me.lemire.integercompression.vector;
 
 import static org.junit.Assert.assertArrayEquals;
+import static org.junit.Assert.assertThrows;
 
 import java.util.Arrays;
 
 import org.junit.Test;
 
 import me.lemire.integercompression.IntWrapper;
+import me.lemire.integercompression.vector.VectorBitPackerKernels.LaneWidth;
 
 /**
  * Tests for the vectorized FastPFOR codec.
@@ -46,4 +48,38 @@ public void dirtyOutputBufferRoundTrip() {
 
     assertArrayEquals(data, recovered);
   }
+
+  /** A stream packed for wider lanes than the host runs natively is refused. */
+  @Test
+  public void checkDecodableRejectsWiderStream() {
+    assertThrows(IllegalStateException.class,
+        () -> VectorFastPFOR.checkDecodable(LaneWidth.BITS_512, LaneWidth.BITS_128));
+    // equal or narrower stream decodes natively
+    VectorFastPFOR.checkDecodable(LaneWidth.BITS_128, LaneWidth.BITS_128);
+    VectorFastPFOR.checkDecodable(LaneWidth.BITS_128, LaneWidth.BITS_512);
+  }
+
+  /** 128-bit lanes are the universal floor, so such a stream decodes on any host. */
+  @Test
+  public void lowestCommonWidthRoundTripsOnAnyHost() {
+    int[] data = new int[3 * VectorFastPFOR.BLOCK_SIZE];
+    for (int i = 0; i < data.length; i++) {
+      data[i] = i % 8;
+    }
+    data[5] = 1 << 20;
+    data[600] = 1 << 25;
+
+    VectorFastPFOR codec =
+        new VectorFastPFOR(VectorFastPFOR.DEFAULT_PAGE_SIZE, LaneWidth.BITS_128);
+    int[] compressed = new int[2 * data.length];
+    IntWrapper inpos = new IntWrapper(0);
+    IntWrapper outpos = new IntWrapper(0);
+    codec.headlessCompress(data, inpos, data.length, compressed, outpos);
+
+    int[] recovered = new int[data.length];
+    codec.headlessUncompress(compressed, new IntWrapper(0), outpos.get(),
+        recovered, new IntWrapper(0), data.length);
+
+    assertArrayEquals(data, recovered);
+  }
 }

From bfd93b6d7beaa8736b4674bc263a404e146f45c2 Mon Sep 17 00:00:00 2001
From: Raunaq Morarka <raunaqmorarka@gmail.com>
Date: Mon, 22 Jun 2026 11:57:54 +0530
Subject: [PATCH 3/5] Add 256-bit vector bit-packing kernel

Add VectorBitPacker256 (AVX2, 256-bit SVE) and register it in the
LaneWidth enum, so 256-bit hosts pack natively instead of stepping down
to the 128-bit kernel.

Decode speedup over the scalar FastPFOR codec:
  Graviton3 (SVE 256): 2.8-4.5x (~3.4x), encode ~13% faster
  AMD EPYC Zen 1 (AVX2 256): 2.4-3.1x (~2.8x), encode ~15% faster
---
 .../vector/VectorBitPacker.java               | 11536 ++---------
 .../vector/VectorBitPacker256.java            | 16226 ++++++++++++++++
 .../vector/VectorBitPackerKernels.java        |     7 +-
 .../vector/VectorIntBitPackBenchmark.java     |   103 +
 .../vector/VectorBitPackerTest.java           |    15 +-
 .../vector/VectorFastPFORTest.java            |     5 +-
 6 files changed, 17898 insertions(+), 9994 deletions(-)
 create mode 100644 src/main/java/me/lemire/integercompression/vector/VectorBitPacker256.java
 create mode 100644 src/main/java/me/lemire/integercompression/vector/VectorIntBitPackBenchmark.java

diff --git a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
index cd19344..1b0df04 100644
--- a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
+++ b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker.java
@@ -18,74 +18,39 @@
 public class VectorBitPacker implements VectorBitPackerKernels {
   private static final VectorSpecies<Integer> SPECIES_512 =
       IntVector.SPECIES_512;
-  private static final VectorSpecies<Integer> SPECIES_256 =
-      IntVector.SPECIES_256;
   private static final int VLEN_512 = 16;
-  private static final int VLEN_256 = 8;
   private static final int BLOCK_SIZE = 256;
 
-  private static final IntVector MASK_1 =
-      IntVector.broadcast(SPECIES_256, (1 << 1) - 1);
   private static final IntVector MASK_2 =
       IntVector.broadcast(SPECIES_512, (1 << 2) - 1);
-  private static final IntVector MASK_3 =
-      IntVector.broadcast(SPECIES_256, (1 << 3) - 1);
   private static final IntVector MASK_4 =
       IntVector.broadcast(SPECIES_512, (1 << 4) - 1);
-  private static final IntVector MASK_5 =
-      IntVector.broadcast(SPECIES_256, (1 << 5) - 1);
   private static final IntVector MASK_6 =
       IntVector.broadcast(SPECIES_512, (1 << 6) - 1);
-  private static final IntVector MASK_7 =
-      IntVector.broadcast(SPECIES_256, (1 << 7) - 1);
   private static final IntVector MASK_8 =
       IntVector.broadcast(SPECIES_512, (1 << 8) - 1);
-  private static final IntVector MASK_9 =
-      IntVector.broadcast(SPECIES_256, (1 << 9) - 1);
   private static final IntVector MASK_10 =
       IntVector.broadcast(SPECIES_512, (1 << 10) - 1);
-  private static final IntVector MASK_11 =
-      IntVector.broadcast(SPECIES_256, (1 << 11) - 1);
   private static final IntVector MASK_12 =
       IntVector.broadcast(SPECIES_512, (1 << 12) - 1);
-  private static final IntVector MASK_13 =
-      IntVector.broadcast(SPECIES_256, (1 << 13) - 1);
   private static final IntVector MASK_14 =
       IntVector.broadcast(SPECIES_512, (1 << 14) - 1);
-  private static final IntVector MASK_15 =
-      IntVector.broadcast(SPECIES_256, (1 << 15) - 1);
   private static final IntVector MASK_16 =
       IntVector.broadcast(SPECIES_512, (1 << 16) - 1);
-  private static final IntVector MASK_17 =
-      IntVector.broadcast(SPECIES_256, (1 << 17) - 1);
   private static final IntVector MASK_18 =
       IntVector.broadcast(SPECIES_512, (1 << 18) - 1);
-  private static final IntVector MASK_19 =
-      IntVector.broadcast(SPECIES_256, (1 << 19) - 1);
   private static final IntVector MASK_20 =
       IntVector.broadcast(SPECIES_512, (1 << 20) - 1);
-  private static final IntVector MASK_21 =
-      IntVector.broadcast(SPECIES_256, (1 << 21) - 1);
   private static final IntVector MASK_22 =
       IntVector.broadcast(SPECIES_512, (1 << 22) - 1);
-  private static final IntVector MASK_23 =
-      IntVector.broadcast(SPECIES_256, (1 << 23) - 1);
   private static final IntVector MASK_24 =
       IntVector.broadcast(SPECIES_512, (1 << 24) - 1);
-  private static final IntVector MASK_25 =
-      IntVector.broadcast(SPECIES_256, (1 << 25) - 1);
   private static final IntVector MASK_26 =
       IntVector.broadcast(SPECIES_512, (1 << 26) - 1);
-  private static final IntVector MASK_27 =
-      IntVector.broadcast(SPECIES_256, (1 << 27) - 1);
   private static final IntVector MASK_28 =
       IntVector.broadcast(SPECIES_512, (1 << 28) - 1);
-  private static final IntVector MASK_29 =
-      IntVector.broadcast(SPECIES_256, (1 << 29) - 1);
   private static final IntVector MASK_30 =
       IntVector.broadcast(SPECIES_512, (1 << 30) - 1);
-  private static final IntVector MASK_31 =
-      IntVector.broadcast(SPECIES_256, (1 << 31) - 1);
 
   /**
    * Pack 32 integers
@@ -108,97 +73,97 @@ public void fastpack(final int[] in, int inpos, final int[] out,
     case 0:
       break;
     case 1:
-      fastpack1(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack1(in, inpos, out, outpos);
       break;
     case 2:
       fastpack2(in, inpos, out, outpos);
       break;
     case 3:
-      fastpack3(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack3(in, inpos, out, outpos);
       break;
     case 4:
       fastpack4(in, inpos, out, outpos);
       break;
     case 5:
-      fastpack5(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack5(in, inpos, out, outpos);
       break;
     case 6:
       fastpack6(in, inpos, out, outpos);
       break;
     case 7:
-      fastpack7(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack7(in, inpos, out, outpos);
       break;
     case 8:
       fastpack8(in, inpos, out, outpos);
       break;
     case 9:
-      fastpack9(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack9(in, inpos, out, outpos);
       break;
     case 10:
       fastpack10(in, inpos, out, outpos);
       break;
     case 11:
-      fastpack11(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack11(in, inpos, out, outpos);
       break;
     case 12:
       fastpack12(in, inpos, out, outpos);
       break;
     case 13:
-      fastpack13(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack13(in, inpos, out, outpos);
       break;
     case 14:
       fastpack14(in, inpos, out, outpos);
       break;
     case 15:
-      fastpack15(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack15(in, inpos, out, outpos);
       break;
     case 16:
       fastpack16(in, inpos, out, outpos);
       break;
     case 17:
-      fastpack17(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack17(in, inpos, out, outpos);
       break;
     case 18:
       fastpack18(in, inpos, out, outpos);
       break;
     case 19:
-      fastpack19(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack19(in, inpos, out, outpos);
       break;
     case 20:
       fastpack20(in, inpos, out, outpos);
       break;
     case 21:
-      fastpack21(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack21(in, inpos, out, outpos);
       break;
     case 22:
       fastpack22(in, inpos, out, outpos);
       break;
     case 23:
-      fastpack23(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack23(in, inpos, out, outpos);
       break;
     case 24:
       fastpack24(in, inpos, out, outpos);
       break;
     case 25:
-      fastpack25(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack25(in, inpos, out, outpos);
       break;
     case 26:
       fastpack26(in, inpos, out, outpos);
       break;
     case 27:
-      fastpack27(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack27(in, inpos, out, outpos);
       break;
     case 28:
       fastpack28(in, inpos, out, outpos);
       break;
     case 29:
-      fastpack29(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack29(in, inpos, out, outpos);
       break;
     case 30:
       fastpack30(in, inpos, out, outpos);
       break;
     case 31:
-      fastpack31(in, inpos, out, outpos);
+      VectorBitPacker256.fastpack31(in, inpos, out, outpos);
       break;
     case 32:
       System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
@@ -213,97 +178,97 @@ public void fastpackNoMask(final int[] in, int inpos, final int[] out,
     case 0:
       break;
     case 1:
-      fastpackNoMask1(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask1(in, inpos, out, outpos);
       break;
     case 2:
       fastpackNoMask2(in, inpos, out, outpos);
       break;
     case 3:
-      fastpackNoMask3(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask3(in, inpos, out, outpos);
       break;
     case 4:
       fastpackNoMask4(in, inpos, out, outpos);
       break;
     case 5:
-      fastpackNoMask5(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask5(in, inpos, out, outpos);
       break;
     case 6:
       fastpackNoMask6(in, inpos, out, outpos);
       break;
     case 7:
-      fastpackNoMask7(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask7(in, inpos, out, outpos);
       break;
     case 8:
       fastpackNoMask8(in, inpos, out, outpos);
       break;
     case 9:
-      fastpackNoMask9(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask9(in, inpos, out, outpos);
       break;
     case 10:
       fastpackNoMask10(in, inpos, out, outpos);
       break;
     case 11:
-      fastpackNoMask11(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask11(in, inpos, out, outpos);
       break;
     case 12:
       fastpackNoMask12(in, inpos, out, outpos);
       break;
     case 13:
-      fastpackNoMask13(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask13(in, inpos, out, outpos);
       break;
     case 14:
       fastpackNoMask14(in, inpos, out, outpos);
       break;
     case 15:
-      fastpackNoMask15(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask15(in, inpos, out, outpos);
       break;
     case 16:
       fastpackNoMask16(in, inpos, out, outpos);
       break;
     case 17:
-      fastpackNoMask17(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask17(in, inpos, out, outpos);
       break;
     case 18:
       fastpackNoMask18(in, inpos, out, outpos);
       break;
     case 19:
-      fastpackNoMask19(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask19(in, inpos, out, outpos);
       break;
     case 20:
       fastpackNoMask20(in, inpos, out, outpos);
       break;
     case 21:
-      fastpackNoMask21(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask21(in, inpos, out, outpos);
       break;
     case 22:
       fastpackNoMask22(in, inpos, out, outpos);
       break;
     case 23:
-      fastpackNoMask23(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask23(in, inpos, out, outpos);
       break;
     case 24:
       fastpackNoMask24(in, inpos, out, outpos);
       break;
     case 25:
-      fastpackNoMask25(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask25(in, inpos, out, outpos);
       break;
     case 26:
       fastpackNoMask26(in, inpos, out, outpos);
       break;
     case 27:
-      fastpackNoMask27(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask27(in, inpos, out, outpos);
       break;
     case 28:
       fastpackNoMask28(in, inpos, out, outpos);
       break;
     case 29:
-      fastpackNoMask29(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask29(in, inpos, out, outpos);
       break;
     case 30:
       fastpackNoMask30(in, inpos, out, outpos);
       break;
     case 31:
-      fastpackNoMask31(in, inpos, out, outpos);
+      VectorBitPacker256.fastpackNoMask31(in, inpos, out, outpos);
       break;
     case 32:
       System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
@@ -333,97 +298,97 @@ public void fastunpack(final int[] in, int inpos, final int[] out,
       Arrays.fill(out, outpos, outpos + 256, 0);
       break;
     case 1:
-      fastunpack1(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack1(in, inpos, out, outpos);
       break;
     case 2:
       fastunpack2(in, inpos, out, outpos);
       break;
     case 3:
-      fastunpack3(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack3(in, inpos, out, outpos);
       break;
     case 4:
       fastunpack4(in, inpos, out, outpos);
       break;
     case 5:
-      fastunpack5(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack5(in, inpos, out, outpos);
       break;
     case 6:
       fastunpack6(in, inpos, out, outpos);
       break;
     case 7:
-      fastunpack7(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack7(in, inpos, out, outpos);
       break;
     case 8:
       fastunpack8(in, inpos, out, outpos);
       break;
     case 9:
-      fastunpack9(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack9(in, inpos, out, outpos);
       break;
     case 10:
       fastunpack10(in, inpos, out, outpos);
       break;
     case 11:
-      fastunpack11(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack11(in, inpos, out, outpos);
       break;
     case 12:
       fastunpack12(in, inpos, out, outpos);
       break;
     case 13:
-      fastunpack13(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack13(in, inpos, out, outpos);
       break;
     case 14:
       fastunpack14(in, inpos, out, outpos);
       break;
     case 15:
-      fastunpack15(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack15(in, inpos, out, outpos);
       break;
     case 16:
       fastunpack16(in, inpos, out, outpos);
       break;
     case 17:
-      fastunpack17(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack17(in, inpos, out, outpos);
       break;
     case 18:
       fastunpack18(in, inpos, out, outpos);
       break;
     case 19:
-      fastunpack19(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack19(in, inpos, out, outpos);
       break;
     case 20:
       fastunpack20(in, inpos, out, outpos);
       break;
     case 21:
-      fastunpack21(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack21(in, inpos, out, outpos);
       break;
     case 22:
       fastunpack22(in, inpos, out, outpos);
       break;
     case 23:
-      fastunpack23(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack23(in, inpos, out, outpos);
       break;
     case 24:
       fastunpack24(in, inpos, out, outpos);
       break;
     case 25:
-      fastunpack25(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack25(in, inpos, out, outpos);
       break;
     case 26:
       fastunpack26(in, inpos, out, outpos);
       break;
     case 27:
-      fastunpack27(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack27(in, inpos, out, outpos);
       break;
     case 28:
       fastunpack28(in, inpos, out, outpos);
       break;
     case 29:
-      fastunpack29(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack29(in, inpos, out, outpos);
       break;
     case 30:
       fastunpack30(in, inpos, out, outpos);
       break;
     case 31:
-      fastunpack31(in, inpos, out, outpos);
+      VectorBitPacker256.fastunpack31(in, inpos, out, outpos);
       break;
     case 32:
       System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
@@ -431,106 +396,6 @@ public void fastunpack(final int[] in, int inpos, final int[] out,
     }
   }
 
-  private static void fastpack1(final int[] in, int inpos, final int[] out,
-                                int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_1);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
   private static void fastpack2(final int[] in, int inpos, final int[] out,
                                 int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -583,116 +448,6 @@ private static void fastpack2(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack3(final int[] in, int inpos, final int[] out,
-                                int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_3);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
   private static void fastpack4(final int[] in, int inpos, final int[] out,
                                 int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -749,126 +504,6 @@ private static void fastpack4(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack5(final int[] in, int inpos, final int[] out,
-                                int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_5);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
   private static void fastpack6(final int[] in, int inpos, final int[] out,
                                 int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -930,136 +565,6 @@ private static void fastpack6(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack7(final int[] in, int inpos, final int[] out,
-                                int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_7);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
   private static void fastpack8(final int[] in, int inpos, final int[] out,
                                 int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -1126,146 +631,6 @@ private static void fastpack8(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack9(final int[] in, int inpos, final int[] out,
-                                int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_9);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
   private static void fastpack10(final int[] in, int inpos, final int[] out,
                                  int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -1337,186 +702,36 @@ private static void fastpack10(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack11(final int[] in, int inpos, final int[] out,
+  private static void fastpack12(final int[] in, int inpos, final int[] out,
                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_11);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 11).or(oV);
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV.and(MASK_12);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 22).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 10);
+    outpos += VLEN_512;
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 23).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 9);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpack12(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_12);
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
     oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
@@ -1563,166 +778,6 @@ private static void fastpack12(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack13(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_13);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
   private static void fastpack14(final int[] in, int inpos, final int[] out,
                                  int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -1804,176 +859,6 @@ private static void fastpack14(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack15(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_15);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
   private static void fastpack16(final int[] in, int inpos, final int[] out,
                                  int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -2060,510 +945,564 @@ private static void fastpack16(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack17(final int[] in, int inpos, final int[] out,
+  private static void fastpack18(final int[] in, int inpos, final int[] out,
                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_17);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 17).or(oV);
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV.and(MASK_18);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 15);
+    outpos += VLEN_512;
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 19).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 13);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 21).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 11);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 6).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 23).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 9);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 25).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 7);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 27).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 5);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 29).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 3);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 31).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 1);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 16).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+  }
+
+  private static void fastpack20(final int[] in, int inpos, final int[] out,
+                                 int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV.and(MASK_20);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 16);
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 18).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 14);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 3).or(oV);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 20).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 12);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 22).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 10);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 7).or(oV);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 24).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 8);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 9).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.and(MASK_20).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 26).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 6);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 11).or(oV);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 28).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 13).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 30).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 2);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 15).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack18(final int[] in, int inpos, final int[] out,
+  private static void fastpack22(final int[] in, int inpos, final int[] out,
                                  int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_18);
+    var oV = iV.and(MASK_22);
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack19(final int[] in, int inpos, final int[] out,
+  private static void fastpack24(final int[] in, int inpos, final int[] out,
                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_19);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 19).or(oV);
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV.and(MASK_24);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 13);
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 6).or(oV);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 25).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 7);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(MASK_24).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 31).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 1);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 18).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 14);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 5).or(oV);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 24).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 8);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 11).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.and(MASK_24).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 30).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 2);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 17).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 15);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 4).or(oV);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 23).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 9);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.and(MASK_24).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 29).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 3);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 16).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 16);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 3).or(oV);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 22).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+  }
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 10);
+  private static void fastpack26(final int[] in, int inpos, final int[] out,
+                                 int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV.and(MASK_26);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 9).or(oV);
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 28).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 15).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 17);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 21).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 11);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 27).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 5);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 18);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 1).or(oV);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 20).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 12);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 7).or(oV);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 26).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 6);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 13).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack20(final int[] in, int inpos, final int[] out,
+  private static void fastpack28(final int[] in, int inpos, final int[] out,
                                  int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_20);
+    var oV = iV.and(MASK_28);
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
@@ -2571,611 +1510,522 @@ private static void fastpack20(final int[] in, int inpos, final int[] out,
     oV = oV.zero(SPECIES_512);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(MASK_20).or(oV);
+    oV = iV.and(MASK_28).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack21(final int[] in, int inpos, final int[] out,
+  private static void fastpack30(final int[] in, int inpos, final int[] out,
                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_21);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 31).or(oV);
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV.and(MASK_30);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 1);
+    outpos += VLEN_512;
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 20).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 12);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 9).or(oV);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 30).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 2);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 19).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 13);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 29).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 3);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 18).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 14);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 7).or(oV);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 28).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 17).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 15);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 6).or(oV);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 27).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 5);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 16).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 16);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 5).or(oV);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 26).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 6);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 15).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 17);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 4).or(oV);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 25).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 7);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 18);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 3).or(oV);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 24).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 8);
+  }
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 13).or(oV);
+  private static void fastpackNoMask2(final int[] in, int inpos,
+                                      final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 19);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 23).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 9);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 20);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 22).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
 
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 10);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 11).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack22(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
+  private static void fastpackNoMask4(final int[] in, int inpos,
+                                      final int[] out, int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_22);
+    var oV = iV;
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
+    oV = oV.zero(SPECIES_512);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack23(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_23);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 14).or(oV);
+  private static void fastpackNoMask6(final int[] in, int inpos,
+                                      final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 18);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 28).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 4);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 19).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 13);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 22);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 24).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 8);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 15).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 29).or(oV);
+  }
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+  private static void fastpackNoMask8(final int[] in, int inpos,
+                                      final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 3);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 20).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 11).or(oV);
+    outpos += VLEN_512;
+    oV = oV.zero(SPECIES_512);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 21);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 25).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 7);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 16).or(oV);
+    oV = oV.zero(SPECIES_512);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 16);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 7).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 30).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 2);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 21).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 11);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 26).or(oV);
+  }
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+  private static void fastpackNoMask10(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 6);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 17).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 15);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 31).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 1);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 22).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 10);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 13).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 19);
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 27).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 5);
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 18).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
 
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 14);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 9).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack24(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
+  private static void fastpackNoMask12(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_24);
+    var oV = iV;
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
-
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_24).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
@@ -3183,7535 +2033,1600 @@ private static void fastpack24(final int[] in, int inpos, final int[] out,
     oV = oV.zero(SPECIES_512);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(MASK_24).or(oV);
+    oV = iV.or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(MASK_24).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpack25(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_25);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 7).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack26(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
+  private static void fastpackNoMask14(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_26);
+    var oV = iV;
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
-
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
 
     oV.intoArray(out, outpos);
   }
 
-  private static void fastpack27(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_27);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 27).or(oV);
+  private static void fastpackNoMask16(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 5);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 22).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 17).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 15);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 20);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 7).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 25);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 29).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 24).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 8);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 19).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 13);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpack28(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_28);
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(MASK_28).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpack29(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_29);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 28);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpack30(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV.and(MASK_30);
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpack31(final int[] in, int inpos, final int[] out,
-                                 int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV.and(MASK_31);
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 28);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 30);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask1(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask2(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask3(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask4(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask5(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask6(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask7(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask8(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask9(final int[] in, int inpos,
-                                      final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask10(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask11(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask12(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask13(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask14(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask15(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask16(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask17(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask18(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask19(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask20(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask21(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask22(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask23(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask24(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask25(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask26(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask27(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask28(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.or(oV);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask29(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask30(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastpackNoMask31(final int[] in, int inpos,
-                                       final int[] out, int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    var oV = iV;
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 1);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
-    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-  }
-
-  private static void fastunpack1(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 31).and(MASK_1).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
-
-  private static void fastunpack2(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
-    outpos += VLEN_512;
-  }
-
-  private static void fastunpack3(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_3);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_3).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
-
-  private static void fastunpack4(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    var oV = iV.and(MASK_4);
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0xf).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
-    outpos += VLEN_512;
-  }
-
-  private static void fastunpack5(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_5);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_5).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
-
-  private static void fastunpack6(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
-    outpos += VLEN_512;
-  }
-
-  private static void fastunpack7(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_7);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_7).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
-
-  private static void fastunpack8(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    var oV = iV.and(MASK_8);
-    oV = oV.zero(SPECIES_512);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0xff).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
     oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(0xff).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV.intoArray(out, outpos);
+  }
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
+  private static void fastpackNoMask18(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
 
-    oV = oV.zero(SPECIES_512);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(0xff).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
 
-    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
-    outpos += VLEN_512;
-  }
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-  private static void fastunpack9(final int[] in, int inpos, final int[] out,
-                                  int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_9);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
 
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_9);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_9);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 6).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_9);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_9);
+    oV.intoArray(out, outpos);
+  }
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 7).or(oV);
+  private static void fastpackNoMask20(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_9);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 3).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_9);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_9);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_9).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
-  private static void fastunpack10(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_10).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
+    oV.intoArray(out, outpos);
+  }
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 4).or(oV);
+  private static void fastpackNoMask22(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
 
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 6).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
-  }
-
-  private static void fastunpack11(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
 
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_11);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_11);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 9).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_11);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_11);
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 7).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_11);
+    oV.intoArray(out, outpos);
+  }
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 6).or(oV);
+  private static void fastpackNoMask24(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_11);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_11);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_11);
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 3).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_11);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_11);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = oV.zero(SPECIES_512);
 
-    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_11).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.or(oV);
 
-  private static void fastunpack12(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+    oV.intoArray(out, outpos);
+  }
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 4).or(oV);
+  private static void fastpackNoMask26(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = oV.zero(SPECIES_512);
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(0xfff).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 8).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 4).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
-    outpos += VLEN_512;
-  }
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-  private static void fastunpack13(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 6).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_13);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 11).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 3).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+  }
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+  private static void fastpackNoMask28(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_13);
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 9).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_13);
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = oV.zero(SPECIES_512);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_13);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.or(oV);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 7).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
 
-    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_13).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
-  private static void fastunpack14(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_14).intoArray(out, outpos);
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
 
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
+    oV.intoArray(out, outpos);
+  }
+
+  private static void fastpackNoMask30(final int[] in, int inpos,
+                                       final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    var oV = iV;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
 
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 6).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 10).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
-    outpos += VLEN_512;
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
 
-    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
     outpos += VLEN_512;
-  }
 
-  private static void fastunpack15(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
 
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_15);
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
 
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_15);
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_15);
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 6).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_15);
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_15);
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_15);
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_15);
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+  }
 
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+  private static void fastunpack2(final int[] in, int inpos, final int[] out,
+                                  int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    iV.and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_15);
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_15);
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 3).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_15);
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_15);
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_512;
+  }
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 7).or(oV);
+  private static void fastunpack4(final int[] in, int inpos, final int[] out,
+                                  int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_15);
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 9).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_15);
+    var oV = iV.and(MASK_4);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 11).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(0xf).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_15);
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 13).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_15).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_512;
   }
 
-  private static void fastunpack16(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
+  private static void fastunpack6(final int[] in, int inpos, final int[] out,
+                                  int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_16).intoArray(out, outpos);
+    iV.and(MASK_6).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    var oV = iV.and(MASK_16);
-    oV = oV.zero(SPECIES_512);
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0xffff).or(oV);
+    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = oV.zero(SPECIES_512);
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(0xffff).or(oV);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = oV.zero(SPECIES_512);
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(0xffff).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
     outpos += VLEN_512;
+  }
 
-    oV = oV.zero(SPECIES_512);
+  private static void fastunpack8(final int[] in, int inpos, final int[] out,
+                                  int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(0xffff).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
     outpos += VLEN_512;
 
+    var oV = iV.and(MASK_8);
     oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(0xffff).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(0xff).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
     outpos += VLEN_512;
 
     oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(0xffff).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(0xff).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
     outpos += VLEN_512;
 
     oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(0xffff).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(0xff).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
     outpos += VLEN_512;
   }
 
-  private static void fastunpack17(final int[] in, int inpos, final int[] out,
+  private static void fastunpack10(final int[] in, int inpos, final int[] out,
                                    int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 15).or(oV);
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    iV.and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_17);
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 13).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 11).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_17);
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 9).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_17);
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 7).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 6).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_17);
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_17);
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 3).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_512;
+  }
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+  private static void fastunpack12(final int[] in, int inpos, final int[] out,
+                                   int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_17);
+    var oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 1).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 16).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_17);
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 12).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_17);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(0xfff).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_17);
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 6).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_17);
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_17);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_17).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_512;
   }
 
-  private static void fastunpack18(final int[] in, int inpos, final int[] out,
+  private static void fastunpack14(final int[] in, int inpos, final int[] out,
                                    int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
-    iV.and(MASK_18).intoArray(out, outpos);
+    iV.and(MASK_14).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    var oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 14).or(oV);
+    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 10).or(oV);
+    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 6).or(oV);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
+    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 6).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
 
     iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 12).or(oV);
+    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 10).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_512;
+  }
 
-    oV.intoArray(out, outpos);
+  private static void fastunpack16(final int[] in, int inpos, final int[] out,
+                                   int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    iV.and(MASK_16).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
     outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
+    var oV = iV.and(MASK_16);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(0xffff).or(oV);
 
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
     outpos += VLEN_512;
-  }
-
-  private static void fastunpack19(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 13).or(oV);
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_19);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 7).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(0xffff).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 1).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_19);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(0xffff).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_19);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 8).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(0xffff).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0x1ffff).lanewise(VectorOperators.LSHL, 2).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_19);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 15).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(0xffff).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_19);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 9).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(0xffff).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_19);
+    oV = oV.zero(SPECIES_512);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 3).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(0xffff).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_19);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 16).or(oV);
+    outpos += VLEN_512;
 
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_512;
+  }
 
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+  private static void fastunpack18(final int[] in, int inpos, final int[] out,
+                                   int outpos) {
+    var iV = IntVector.fromArray(SPECIES_512, in, inpos);
+    iV.and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_19);
+    var oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 16);
+    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_19);
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 4).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 32);
+    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 10).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_19);
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 17).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 48);
+    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 6).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_19);
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 11).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 64);
+    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_19);
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 80);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_19);
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 18).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 96);
+    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_19);
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_19);
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 6).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_19).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_512;
   }
 
   private static void fastunpack20(final int[] in, int inpos, final int[] out,
@@ -10811,206 +3726,6 @@ private static void fastunpack20(final int[] in, int inpos, final int[] out,
     outpos += VLEN_512;
   }
 
-  private static void fastunpack21(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x7ffff).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x1ffff).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_21);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_21).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
-
   private static void fastunpack22(final int[] in, int inpos, final int[] out,
                                    int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -11071,256 +3786,46 @@ private static void fastunpack22(final int[] in, int inpos, final int[] out,
     oV.intoArray(out, outpos);
     outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
-    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
-    outpos += VLEN_512;
-  }
-
-  private static void fastunpack23(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0x7ffff).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0x1fffff).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(0x1ffff).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_23);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(0x3fffff).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_23);
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 10).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 112);
+    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 4).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_23);
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 19).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 128);
+    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_23);
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 5).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 144);
+    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_23);
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 14).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 12).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_23).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_512;
   }
 
   private static void fastunpack24(final int[] in, int inpos, final int[] out,
@@ -11430,226 +3935,6 @@ private static void fastunpack24(final int[] in, int inpos, final int[] out,
     outpos += VLEN_512;
   }
 
-  private static void fastunpack25(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x3fffff).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x7ffff).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0x7fffff).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(0xffffff).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(0x1ffff).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(0x1fffff).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_25);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_25).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
-
   private static void fastunpack26(final int[] in, int inpos, final int[] out,
                                    int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -11732,264 +4017,34 @@ private static void fastunpack26(final int[] in, int inpos, final int[] out,
     oV = iV.and(0xffffff).lanewise(VectorOperators.LSHL, 2).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
-
-    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_512;
-
-    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
-    outpos += VLEN_512;
-  }
-
-  private static void fastunpack27(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_27).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x3fffff).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0x1ffff).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_27).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0xffffff).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0x7ffff).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_27).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0x3ffffff).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x1fffff).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_27).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(0x7fffff).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_27).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(0x1ffffff).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_27);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_27);
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 12).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 160);
+    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 8).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_27);
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 17).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 176);
+    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 14).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_27);
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
 
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 22).or(oV);
+    iV = IntVector.fromArray(SPECIES_512, in, inpos + 192);
+    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 20).or(oV);
 
     oV.intoArray(out, outpos);
-    outpos += VLEN_256;
+    outpos += VLEN_512;
 
-    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_27).intoArray(out, outpos);
-    outpos += VLEN_256;
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_512;
   }
 
   private static void fastunpack28(final int[] in, int inpos, final int[] out,
@@ -12109,246 +4164,6 @@ private static void fastunpack28(final int[] in, int inpos, final int[] out,
     outpos += VLEN_512;
   }
 
-  private static void fastunpack29(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_29).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x3ffffff).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0x7fffff).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x1ffff).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_29).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0xfffffff).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0x1ffffff).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x3fffff).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0x7ffff).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_29).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(0x7ffffff).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(0xffffff).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(0x1fffff).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_29);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_29).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
-
   private static void fastunpack30(final int[] in, int inpos, final int[] out,
                                    int outpos) {
     var iV = IntVector.fromArray(SPECIES_512, in, inpos);
@@ -12471,253 +4286,4 @@ private static void fastunpack30(final int[] in, int inpos, final int[] out,
     outpos += VLEN_512;
   }
 
-  private static void fastunpack31(final int[] in, int inpos, final int[] out,
-                                   int outpos) {
-    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
-    iV.and(MASK_31).intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    var oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
-    oV = iV.and(0x3fffffff).lanewise(VectorOperators.LSHL, 1).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
-    oV = iV.and(0x1fffffff).lanewise(VectorOperators.LSHL, 2).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
-    oV = iV.and(0xfffffff).lanewise(VectorOperators.LSHL, 3).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
-    oV = iV.and(0x7ffffff).lanewise(VectorOperators.LSHL, 4).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
-    oV = iV.and(0x3ffffff).lanewise(VectorOperators.LSHL, 5).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
-    oV = iV.and(0x1ffffff).lanewise(VectorOperators.LSHL, 6).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
-    oV = iV.and(0xffffff).lanewise(VectorOperators.LSHL, 7).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
-    oV = iV.and(0x7fffff).lanewise(VectorOperators.LSHL, 8).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
-    oV = iV.and(0x3fffff).lanewise(VectorOperators.LSHL, 9).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
-    oV = iV.and(0x1fffff).lanewise(VectorOperators.LSHL, 10).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
-    oV = iV.and(0xfffff).lanewise(VectorOperators.LSHL, 11).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
-    oV = iV.and(0x7ffff).lanewise(VectorOperators.LSHL, 12).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
-    oV = iV.and(0x3ffff).lanewise(VectorOperators.LSHL, 13).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
-    oV = iV.and(0x1ffff).lanewise(VectorOperators.LSHL, 14).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
-    oV = iV.and(0xffff).lanewise(VectorOperators.LSHL, 15).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
-    oV = iV.and(0x7fff).lanewise(VectorOperators.LSHL, 16).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
-    oV = iV.and(0x3fff).lanewise(VectorOperators.LSHL, 17).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
-    oV = iV.and(0x1fff).lanewise(VectorOperators.LSHL, 18).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
-    oV = iV.and(0xfff).lanewise(VectorOperators.LSHL, 19).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
-    oV = iV.and(0x7ff).lanewise(VectorOperators.LSHL, 20).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
-    oV = iV.and(0x3ff).lanewise(VectorOperators.LSHL, 21).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
-    oV = iV.and(0x1ff).lanewise(VectorOperators.LSHL, 22).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
-    oV = iV.and(0xff).lanewise(VectorOperators.LSHL, 23).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
-    oV = iV.and(0x7f).lanewise(VectorOperators.LSHL, 24).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
-    oV = iV.and(0x3f).lanewise(VectorOperators.LSHL, 25).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
-    oV = iV.and(0x1f).lanewise(VectorOperators.LSHL, 26).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
-    oV = iV.and(0xf).lanewise(VectorOperators.LSHL, 27).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
-    oV = iV.and(7).lanewise(VectorOperators.LSHL, 28).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 3).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
-    oV = iV.and(3).lanewise(VectorOperators.LSHL, 29).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    oV = iV.lanewise(VectorOperators.LSHR, 2).and(MASK_31);
-
-    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
-    oV = iV.and(1).lanewise(VectorOperators.LSHL, 30).or(oV);
-
-    oV.intoArray(out, outpos);
-    outpos += VLEN_256;
-
-    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_31).intoArray(out, outpos);
-    outpos += VLEN_256;
-  }
 }
diff --git a/src/main/java/me/lemire/integercompression/vector/VectorBitPacker256.java b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker256.java
new file mode 100644
index 0000000..02596f5
--- /dev/null
+++ b/src/main/java/me/lemire/integercompression/vector/VectorBitPacker256.java
@@ -0,0 +1,16226 @@
+/**
+ * This code is released under the
+ * Apache License Version 2.0 http://www.apache.org/licenses/.
+ *
+ * (c) Daniel Lemire, http://lemire.me/en/
+ */
+package me.lemire.integercompression.vector;
+
+import java.util.Arrays;
+import jdk.incubator.vector.*;
+
+/**
+ * Vectorized bit-packing routines using 256-bit (8 x int32) vectors.
+ *
+ * A 256-integer block is packed across 8 SIMD lanes, each lane packing 32
+ * values into b 32-bit words. Selected at runtime by VectorBitPackerKernels
+ * when the preferred hardware vector width is 256 bits (e.g. AVX2, Graviton3),
+ * where the 512-bit paths of VectorBitPacker fall back to slow emulation.
+ */
+public class VectorBitPacker256 implements VectorBitPackerKernels {
+  private static final VectorSpecies<Integer> SPECIES_256 =
+      IntVector.SPECIES_256;
+  private static final int VLEN_256 = 8;
+  private static final int BLOCK_SIZE = 256;
+
+  private static final IntVector MASK_1 =
+      IntVector.broadcast(SPECIES_256, (1 << 1) - 1);
+  private static final IntVector MASK_2 =
+      IntVector.broadcast(SPECIES_256, (1 << 2) - 1);
+  private static final IntVector MASK_3 =
+      IntVector.broadcast(SPECIES_256, (1 << 3) - 1);
+  private static final IntVector MASK_4 =
+      IntVector.broadcast(SPECIES_256, (1 << 4) - 1);
+  private static final IntVector MASK_5 =
+      IntVector.broadcast(SPECIES_256, (1 << 5) - 1);
+  private static final IntVector MASK_6 =
+      IntVector.broadcast(SPECIES_256, (1 << 6) - 1);
+  private static final IntVector MASK_7 =
+      IntVector.broadcast(SPECIES_256, (1 << 7) - 1);
+  private static final IntVector MASK_8 =
+      IntVector.broadcast(SPECIES_256, (1 << 8) - 1);
+  private static final IntVector MASK_9 =
+      IntVector.broadcast(SPECIES_256, (1 << 9) - 1);
+  private static final IntVector MASK_10 =
+      IntVector.broadcast(SPECIES_256, (1 << 10) - 1);
+  private static final IntVector MASK_11 =
+      IntVector.broadcast(SPECIES_256, (1 << 11) - 1);
+  private static final IntVector MASK_12 =
+      IntVector.broadcast(SPECIES_256, (1 << 12) - 1);
+  private static final IntVector MASK_13 =
+      IntVector.broadcast(SPECIES_256, (1 << 13) - 1);
+  private static final IntVector MASK_14 =
+      IntVector.broadcast(SPECIES_256, (1 << 14) - 1);
+  private static final IntVector MASK_15 =
+      IntVector.broadcast(SPECIES_256, (1 << 15) - 1);
+  private static final IntVector MASK_16 =
+      IntVector.broadcast(SPECIES_256, (1 << 16) - 1);
+  private static final IntVector MASK_17 =
+      IntVector.broadcast(SPECIES_256, (1 << 17) - 1);
+  private static final IntVector MASK_18 =
+      IntVector.broadcast(SPECIES_256, (1 << 18) - 1);
+  private static final IntVector MASK_19 =
+      IntVector.broadcast(SPECIES_256, (1 << 19) - 1);
+  private static final IntVector MASK_20 =
+      IntVector.broadcast(SPECIES_256, (1 << 20) - 1);
+  private static final IntVector MASK_21 =
+      IntVector.broadcast(SPECIES_256, (1 << 21) - 1);
+  private static final IntVector MASK_22 =
+      IntVector.broadcast(SPECIES_256, (1 << 22) - 1);
+  private static final IntVector MASK_23 =
+      IntVector.broadcast(SPECIES_256, (1 << 23) - 1);
+  private static final IntVector MASK_24 =
+      IntVector.broadcast(SPECIES_256, (1 << 24) - 1);
+  private static final IntVector MASK_25 =
+      IntVector.broadcast(SPECIES_256, (1 << 25) - 1);
+  private static final IntVector MASK_26 =
+      IntVector.broadcast(SPECIES_256, (1 << 26) - 1);
+  private static final IntVector MASK_27 =
+      IntVector.broadcast(SPECIES_256, (1 << 27) - 1);
+  private static final IntVector MASK_28 =
+      IntVector.broadcast(SPECIES_256, (1 << 28) - 1);
+  private static final IntVector MASK_29 =
+      IntVector.broadcast(SPECIES_256, (1 << 29) - 1);
+  private static final IntVector MASK_30 =
+      IntVector.broadcast(SPECIES_256, (1 << 30) - 1);
+  private static final IntVector MASK_31 =
+      IntVector.broadcast(SPECIES_256, (1 << 31) - 1);
+
+  @Override
+  public void fastpack(final int[] in, int inpos, final int[] out,
+                       int outpos, int b) {
+    switch (b) {
+    case 0:
+      break;
+    case 1:
+      fastpack1(in, inpos, out, outpos);
+      break;
+    case 2:
+      fastpack2(in, inpos, out, outpos);
+      break;
+    case 3:
+      fastpack3(in, inpos, out, outpos);
+      break;
+    case 4:
+      fastpack4(in, inpos, out, outpos);
+      break;
+    case 5:
+      fastpack5(in, inpos, out, outpos);
+      break;
+    case 6:
+      fastpack6(in, inpos, out, outpos);
+      break;
+    case 7:
+      fastpack7(in, inpos, out, outpos);
+      break;
+    case 8:
+      fastpack8(in, inpos, out, outpos);
+      break;
+    case 9:
+      fastpack9(in, inpos, out, outpos);
+      break;
+    case 10:
+      fastpack10(in, inpos, out, outpos);
+      break;
+    case 11:
+      fastpack11(in, inpos, out, outpos);
+      break;
+    case 12:
+      fastpack12(in, inpos, out, outpos);
+      break;
+    case 13:
+      fastpack13(in, inpos, out, outpos);
+      break;
+    case 14:
+      fastpack14(in, inpos, out, outpos);
+      break;
+    case 15:
+      fastpack15(in, inpos, out, outpos);
+      break;
+    case 16:
+      fastpack16(in, inpos, out, outpos);
+      break;
+    case 17:
+      fastpack17(in, inpos, out, outpos);
+      break;
+    case 18:
+      fastpack18(in, inpos, out, outpos);
+      break;
+    case 19:
+      fastpack19(in, inpos, out, outpos);
+      break;
+    case 20:
+      fastpack20(in, inpos, out, outpos);
+      break;
+    case 21:
+      fastpack21(in, inpos, out, outpos);
+      break;
+    case 22:
+      fastpack22(in, inpos, out, outpos);
+      break;
+    case 23:
+      fastpack23(in, inpos, out, outpos);
+      break;
+    case 24:
+      fastpack24(in, inpos, out, outpos);
+      break;
+    case 25:
+      fastpack25(in, inpos, out, outpos);
+      break;
+    case 26:
+      fastpack26(in, inpos, out, outpos);
+      break;
+    case 27:
+      fastpack27(in, inpos, out, outpos);
+      break;
+    case 28:
+      fastpack28(in, inpos, out, outpos);
+      break;
+    case 29:
+      fastpack29(in, inpos, out, outpos);
+      break;
+    case 30:
+      fastpack30(in, inpos, out, outpos);
+      break;
+    case 31:
+      fastpack31(in, inpos, out, outpos);
+      break;
+    case 32:
+      System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
+      break;
+    }
+  }
+
+  @Override
+  public void fastpackNoMask(final int[] in, int inpos, final int[] out,
+                             int outpos, int b) {
+    switch (b) {
+    case 0:
+      break;
+    case 1:
+      fastpackNoMask1(in, inpos, out, outpos);
+      break;
+    case 2:
+      fastpackNoMask2(in, inpos, out, outpos);
+      break;
+    case 3:
+      fastpackNoMask3(in, inpos, out, outpos);
+      break;
+    case 4:
+      fastpackNoMask4(in, inpos, out, outpos);
+      break;
+    case 5:
+      fastpackNoMask5(in, inpos, out, outpos);
+      break;
+    case 6:
+      fastpackNoMask6(in, inpos, out, outpos);
+      break;
+    case 7:
+      fastpackNoMask7(in, inpos, out, outpos);
+      break;
+    case 8:
+      fastpackNoMask8(in, inpos, out, outpos);
+      break;
+    case 9:
+      fastpackNoMask9(in, inpos, out, outpos);
+      break;
+    case 10:
+      fastpackNoMask10(in, inpos, out, outpos);
+      break;
+    case 11:
+      fastpackNoMask11(in, inpos, out, outpos);
+      break;
+    case 12:
+      fastpackNoMask12(in, inpos, out, outpos);
+      break;
+    case 13:
+      fastpackNoMask13(in, inpos, out, outpos);
+      break;
+    case 14:
+      fastpackNoMask14(in, inpos, out, outpos);
+      break;
+    case 15:
+      fastpackNoMask15(in, inpos, out, outpos);
+      break;
+    case 16:
+      fastpackNoMask16(in, inpos, out, outpos);
+      break;
+    case 17:
+      fastpackNoMask17(in, inpos, out, outpos);
+      break;
+    case 18:
+      fastpackNoMask18(in, inpos, out, outpos);
+      break;
+    case 19:
+      fastpackNoMask19(in, inpos, out, outpos);
+      break;
+    case 20:
+      fastpackNoMask20(in, inpos, out, outpos);
+      break;
+    case 21:
+      fastpackNoMask21(in, inpos, out, outpos);
+      break;
+    case 22:
+      fastpackNoMask22(in, inpos, out, outpos);
+      break;
+    case 23:
+      fastpackNoMask23(in, inpos, out, outpos);
+      break;
+    case 24:
+      fastpackNoMask24(in, inpos, out, outpos);
+      break;
+    case 25:
+      fastpackNoMask25(in, inpos, out, outpos);
+      break;
+    case 26:
+      fastpackNoMask26(in, inpos, out, outpos);
+      break;
+    case 27:
+      fastpackNoMask27(in, inpos, out, outpos);
+      break;
+    case 28:
+      fastpackNoMask28(in, inpos, out, outpos);
+      break;
+    case 29:
+      fastpackNoMask29(in, inpos, out, outpos);
+      break;
+    case 30:
+      fastpackNoMask30(in, inpos, out, outpos);
+      break;
+    case 31:
+      fastpackNoMask31(in, inpos, out, outpos);
+      break;
+    case 32:
+      System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
+      break;
+    }
+  }
+
+  @Override
+  public void fastunpack(final int[] in, int inpos, final int[] out,
+                         int outpos, int b) {
+    switch (b) {
+    case 0:
+      Arrays.fill(out, outpos, outpos + 256, 0);
+      break;
+    case 1:
+      fastunpack1(in, inpos, out, outpos);
+      break;
+    case 2:
+      fastunpack2(in, inpos, out, outpos);
+      break;
+    case 3:
+      fastunpack3(in, inpos, out, outpos);
+      break;
+    case 4:
+      fastunpack4(in, inpos, out, outpos);
+      break;
+    case 5:
+      fastunpack5(in, inpos, out, outpos);
+      break;
+    case 6:
+      fastunpack6(in, inpos, out, outpos);
+      break;
+    case 7:
+      fastunpack7(in, inpos, out, outpos);
+      break;
+    case 8:
+      fastunpack8(in, inpos, out, outpos);
+      break;
+    case 9:
+      fastunpack9(in, inpos, out, outpos);
+      break;
+    case 10:
+      fastunpack10(in, inpos, out, outpos);
+      break;
+    case 11:
+      fastunpack11(in, inpos, out, outpos);
+      break;
+    case 12:
+      fastunpack12(in, inpos, out, outpos);
+      break;
+    case 13:
+      fastunpack13(in, inpos, out, outpos);
+      break;
+    case 14:
+      fastunpack14(in, inpos, out, outpos);
+      break;
+    case 15:
+      fastunpack15(in, inpos, out, outpos);
+      break;
+    case 16:
+      fastunpack16(in, inpos, out, outpos);
+      break;
+    case 17:
+      fastunpack17(in, inpos, out, outpos);
+      break;
+    case 18:
+      fastunpack18(in, inpos, out, outpos);
+      break;
+    case 19:
+      fastunpack19(in, inpos, out, outpos);
+      break;
+    case 20:
+      fastunpack20(in, inpos, out, outpos);
+      break;
+    case 21:
+      fastunpack21(in, inpos, out, outpos);
+      break;
+    case 22:
+      fastunpack22(in, inpos, out, outpos);
+      break;
+    case 23:
+      fastunpack23(in, inpos, out, outpos);
+      break;
+    case 24:
+      fastunpack24(in, inpos, out, outpos);
+      break;
+    case 25:
+      fastunpack25(in, inpos, out, outpos);
+      break;
+    case 26:
+      fastunpack26(in, inpos, out, outpos);
+      break;
+    case 27:
+      fastunpack27(in, inpos, out, outpos);
+      break;
+    case 28:
+      fastunpack28(in, inpos, out, outpos);
+      break;
+    case 29:
+      fastunpack29(in, inpos, out, outpos);
+      break;
+    case 30:
+      fastunpack30(in, inpos, out, outpos);
+      break;
+    case 31:
+      fastunpack31(in, inpos, out, outpos);
+      break;
+    case 32:
+      System.arraycopy(in, inpos, out, outpos, BLOCK_SIZE);
+      break;
+    }
+  }
+
+  static void fastpack1(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_1).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask1(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack1(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_1).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 31).and(MASK_1).intoArray(out, outpos);
+  }
+
+  static void fastpack2(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_2).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask2(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack2(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    iV.and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_2).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 30).and(MASK_2).intoArray(out, outpos);
+  }
+
+  static void fastpack3(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_3).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask3(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack3(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_3).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 29).and(MASK_3).intoArray(out, outpos);
+  }
+
+  static void fastpack4(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_4).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask4(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack4(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    iV.and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_4).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 28).and(MASK_4).intoArray(out, outpos);
+  }
+
+  static void fastpack5(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_5).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask5(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack5(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_5).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 27).and(MASK_5).intoArray(out, outpos);
+  }
+
+  static void fastpack6(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_6).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask6(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack6(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    iV.and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_6).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 26).and(MASK_6).intoArray(out, outpos);
+  }
+
+  static void fastpack7(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_7).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask7(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack7(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_7).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 25).and(MASK_7).intoArray(out, outpos);
+  }
+
+  static void fastpack8(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_8).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask8(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack8(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    iV.and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_8).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 24).and(MASK_8).intoArray(out, outpos);
+  }
+
+  static void fastpack9(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_9).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask9(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack9(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_9).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 23).and(MASK_9).intoArray(out, outpos);
+  }
+
+  static void fastpack10(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_10).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask10(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack10(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    iV.and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_10).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 22).and(MASK_10).intoArray(out, outpos);
+  }
+
+  static void fastpack11(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_11).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask11(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack11(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_11).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 21).and(MASK_11).intoArray(out, outpos);
+  }
+
+  static void fastpack12(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_12).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask12(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack12(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    iV.and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_12).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 20).and(MASK_12).intoArray(out, outpos);
+  }
+
+  static void fastpack13(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_13).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask13(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack13(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_13).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 19).and(MASK_13).intoArray(out, outpos);
+  }
+
+  static void fastpack14(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_14).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask14(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack14(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    iV.and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_14).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 18).and(MASK_14).intoArray(out, outpos);
+  }
+
+  static void fastpack15(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_15).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask15(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack15(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_15).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 17).and(MASK_15).intoArray(out, outpos);
+  }
+
+  static void fastpack16(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_16).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask16(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack16(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    iV.and(MASK_16).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 16).and(MASK_16).intoArray(out, outpos);
+  }
+
+  static void fastpack17(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_17).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask17(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack17(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_17).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 15).and(MASK_17).intoArray(out, outpos);
+  }
+
+  static void fastpack18(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_18).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask18(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack18(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    iV.and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_18).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 14).and(MASK_18).intoArray(out, outpos);
+  }
+
+  static void fastpack19(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_19).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask19(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack19(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_19).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 13).and(MASK_19).intoArray(out, outpos);
+  }
+
+  static void fastpack20(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_20).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask20(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack20(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    iV.and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_20).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 12).and(MASK_20).intoArray(out, outpos);
+  }
+
+  static void fastpack21(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_21).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask21(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack21(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_21).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 11).and(MASK_21).intoArray(out, outpos);
+  }
+
+  static void fastpack22(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_22).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask22(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack22(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    iV.and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_22).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 10).and(MASK_22).intoArray(out, outpos);
+  }
+
+  static void fastpack23(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_23).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask23(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack23(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_23).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 9).and(MASK_23).intoArray(out, outpos);
+  }
+
+  static void fastpack24(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_24).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask24(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack24(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    iV.and(MASK_24).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 8).and(MASK_24).intoArray(out, outpos);
+  }
+
+  static void fastpack25(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_25).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask25(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack25(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_25).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 7).and(MASK_25).intoArray(out, outpos);
+  }
+
+  static void fastpack26(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_26).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask26(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack26(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    iV.and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_26).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 6).and(MASK_26).intoArray(out, outpos);
+  }
+
+  static void fastpack27(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_27).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask27(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack27(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_27).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 5).and(MASK_27).intoArray(out, outpos);
+  }
+
+  static void fastpack28(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_28).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask28(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack28(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    iV.and(MASK_28).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 4).and(MASK_28).intoArray(out, outpos);
+  }
+
+  static void fastpack29(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_29).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask29(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack29(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_29).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(134217727).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 3).and(MASK_29).intoArray(out, outpos);
+  }
+
+  static void fastpack30(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_30).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask30(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack30(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    iV.and(MASK_30).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 2).and(MASK_30).intoArray(out, outpos);
+  }
+
+  static void fastpack31(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV.and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHR, 30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.and(MASK_31).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastpackNoMask31(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    var oV = iV;
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.lanewise(VectorOperators.LSHL, 31).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 1);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 248);
+    oV = iV.lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+  }
+
+  static void fastunpack31(final int[] in, int inpos, final int[] out, int outpos) {
+    var iV = IntVector.fromArray(SPECIES_256, in, inpos);
+    iV.and(MASK_31).intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    var oV = iV.lanewise(VectorOperators.LSHR, 31).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 8);
+    oV = iV.and(1073741823).lanewise(VectorOperators.LSHL, 1).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 30).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 16);
+    oV = iV.and(536870911).lanewise(VectorOperators.LSHL, 2).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 29).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 24);
+    oV = iV.and(268435455).lanewise(VectorOperators.LSHL, 3).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 28).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 32);
+    oV = iV.and(134217727).lanewise(VectorOperators.LSHL, 4).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 27).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 40);
+    oV = iV.and(67108863).lanewise(VectorOperators.LSHL, 5).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 26).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 48);
+    oV = iV.and(33554431).lanewise(VectorOperators.LSHL, 6).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 25).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 56);
+    oV = iV.and(16777215).lanewise(VectorOperators.LSHL, 7).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 24).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 64);
+    oV = iV.and(8388607).lanewise(VectorOperators.LSHL, 8).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 23).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 72);
+    oV = iV.and(4194303).lanewise(VectorOperators.LSHL, 9).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 22).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 80);
+    oV = iV.and(2097151).lanewise(VectorOperators.LSHL, 10).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 21).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 88);
+    oV = iV.and(1048575).lanewise(VectorOperators.LSHL, 11).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 20).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 96);
+    oV = iV.and(524287).lanewise(VectorOperators.LSHL, 12).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 19).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 104);
+    oV = iV.and(262143).lanewise(VectorOperators.LSHL, 13).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 18).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 112);
+    oV = iV.and(131071).lanewise(VectorOperators.LSHL, 14).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 17).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 120);
+    oV = iV.and(65535).lanewise(VectorOperators.LSHL, 15).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 16).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 128);
+    oV = iV.and(32767).lanewise(VectorOperators.LSHL, 16).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 15).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 136);
+    oV = iV.and(16383).lanewise(VectorOperators.LSHL, 17).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 14).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 144);
+    oV = iV.and(8191).lanewise(VectorOperators.LSHL, 18).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 13).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 152);
+    oV = iV.and(4095).lanewise(VectorOperators.LSHL, 19).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 12).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 160);
+    oV = iV.and(2047).lanewise(VectorOperators.LSHL, 20).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 11).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 168);
+    oV = iV.and(1023).lanewise(VectorOperators.LSHL, 21).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 10).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 176);
+    oV = iV.and(511).lanewise(VectorOperators.LSHL, 22).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 9).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 184);
+    oV = iV.and(255).lanewise(VectorOperators.LSHL, 23).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 8).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 192);
+    oV = iV.and(127).lanewise(VectorOperators.LSHL, 24).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 7).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 200);
+    oV = iV.and(63).lanewise(VectorOperators.LSHL, 25).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 6).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 208);
+    oV = iV.and(31).lanewise(VectorOperators.LSHL, 26).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 5).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 216);
+    oV = iV.and(15).lanewise(VectorOperators.LSHL, 27).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 4).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 224);
+    oV = iV.and(7).lanewise(VectorOperators.LSHL, 28).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 3).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 232);
+    oV = iV.and(3).lanewise(VectorOperators.LSHL, 29).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    oV = iV.lanewise(VectorOperators.LSHR, 2).and(MASK_31);
+
+    iV = IntVector.fromArray(SPECIES_256, in, inpos + 240);
+    oV = iV.and(1).lanewise(VectorOperators.LSHL, 30).or(oV);
+
+    oV.intoArray(out, outpos);
+    outpos += VLEN_256;
+
+    iV.lanewise(VectorOperators.LSHR, 1).and(MASK_31).intoArray(out, outpos);
+  }
+
+}
diff --git a/src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java b/src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java
index da7b7c2..9179827 100644
--- a/src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java
+++ b/src/main/java/me/lemire/integercompression/vector/VectorBitPackerKernels.java
@@ -10,9 +10,9 @@
 
 /**
  * Width-specific vectorized bit-packing kernels for a 256-integer block.
- * Implemented by VectorBitPacker (512-bit lanes) and VectorBitPacker128
- * (128-bit lanes). The packed layout differs per width, so a stream is decoded
- * by the same kernel that packed it.
+ * Implemented by VectorBitPacker (512-bit lanes), VectorBitPacker256 (256-bit
+ * lanes) and VectorBitPacker128 (128-bit lanes). The packed layout differs per
+ * width, so a stream is decoded by the same kernel that packed it.
  */
 public interface VectorBitPackerKernels {
 
@@ -29,6 +29,7 @@ public interface VectorBitPackerKernels {
    */
   enum LaneWidth {
     BITS_128(0, 128, new VectorBitPacker128()),
+    BITS_256(1, 256, new VectorBitPacker256()),
     BITS_512(2, 512, new VectorBitPacker());
 
     /** Compact wire tag stored in the stream (fits in 2 bits). */
diff --git a/src/main/java/me/lemire/integercompression/vector/VectorIntBitPackBenchmark.java b/src/main/java/me/lemire/integercompression/vector/VectorIntBitPackBenchmark.java
new file mode 100644
index 0000000..ee69c74
--- /dev/null
+++ b/src/main/java/me/lemire/integercompression/vector/VectorIntBitPackBenchmark.java
@@ -0,0 +1,103 @@
+/**
+ * This code is released under the
+ * Apache License Version 2.0 http://www.apache.org/licenses/.
+ *
+ * (c) Daniel Lemire, http://lemire.me/en/
+ */
+package me.lemire.integercompression.vector;
+
+import java.text.DecimalFormat;
+import java.util.Random;
+
+import me.lemire.integercompression.BitPacking;
+
+/**
+ * Benchmarks the vectorized int bit-packing kernels against the scalar
+ * unrolled {@link BitPacking}, packing 256-integer blocks of each width.
+ * Speeds are millions of integers per second. (For expert use; requires
+ * --add-modules jdk.incubator.vector.)
+ */
+public class VectorIntBitPackBenchmark {
+
+  private static final int BLOCK = 256;
+  private static final int SUBBLOCKS = BLOCK / 32;
+
+  private static void scalarPack(int[] in, int[] out, int bit) {
+    for (int blk = 0; blk < SUBBLOCKS; blk++) {
+      BitPacking.fastpackwithoutmask(in, blk * 32, out, blk * bit, bit);
+    }
+  }
+
+  private static void scalarUnpack(int[] in, int[] out, int bit) {
+    for (int blk = 0; blk < SUBBLOCKS; blk++) {
+      BitPacking.fastunpack(in, blk * bit, out, blk * 32, bit);
+    }
+  }
+
+  private static void test(boolean verbose) {
+    DecimalFormat df = new DecimalFormat("0");
+    final int times = 100000;
+    Random r = new Random(0);
+    int[] data = new int[BLOCK];
+    int[] compressed = new int[8 * 32];
+    int[] uncompressed = new int[BLOCK];
+    VectorBitPacker256 vec256 = new VectorBitPacker256();
+    VectorBitPacker vec = new VectorBitPacker();
+
+    for (int bit = 1; bit <= 32; ++bit) {
+      int mask = bit == 32 ? -1 : (1 << bit) - 1;
+      long scalarComp = 0;
+      long scalarDecomp = 0;
+      long vec256Comp = 0;
+      long vec256Decomp = 0;
+      long vecComp = 0;
+      long vecDecomp = 0;
+      for (int t = 0; t < times; ++t) {
+        for (int k = 0; k < BLOCK; ++k) {
+          data[k] = r.nextInt() & mask;
+        }
+        long time1 = System.nanoTime();
+        scalarPack(data, compressed, bit);
+        long time2 = System.nanoTime();
+        scalarUnpack(compressed, uncompressed, bit);
+        long time3 = System.nanoTime();
+        vec256.fastpackNoMask(data, 0, compressed, 0, bit);
+        long time4 = System.nanoTime();
+        vec256.fastunpack(compressed, 0, uncompressed, 0, bit);
+        long time5 = System.nanoTime();
+        vec.fastpackNoMask(data, 0, compressed, 0, bit);
+        long time6 = System.nanoTime();
+        vec.fastunpack(compressed, 0, uncompressed, 0, bit);
+        long time7 = System.nanoTime();
+        scalarComp += time2 - time1;
+        scalarDecomp += time3 - time2;
+        vec256Comp += time4 - time3;
+        vec256Decomp += time5 - time4;
+        vecComp += time6 - time5;
+        vecDecomp += time7 - time6;
+      }
+      if (verbose) {
+        double sc = BLOCK * times * 1000.0;
+        System.out.println("bit = " + bit
+            + " | scalar comp = " + df.format(sc / scalarComp)
+            + " vec256 comp = " + df.format(sc / vec256Comp)
+            + " vec comp = " + df.format(sc / vecComp)
+            + " | scalar decomp = " + df.format(sc / scalarDecomp)
+            + " vec256 decomp = " + df.format(sc / vec256Decomp)
+            + " vec decomp = " + df.format(sc / vecDecomp));
+      }
+    }
+  }
+
+  /**
+   * Main method.
+   *
+   * @param args command-line arguments
+   */
+  public static void main(String[] args) {
+    System.out.println("Testing int packing (scalar vs VectorBitPacker256 vs "
+        + "VectorBitPacker), 256-int blocks, speeds in millions of ints/s");
+    test(false);
+    test(true);
+  }
+}
diff --git a/src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java b/src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java
index 0322c77..6640890 100644
--- a/src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java
+++ b/src/test/java/me/lemire/integercompression/vector/VectorBitPackerTest.java
@@ -15,11 +15,11 @@
 /**
  * Tests for the width-specific bit-packing kernels.
  *
- * VectorBitPacker (256/512-bit lanes) and VectorBitPacker128 (128-bit lanes)
- * use different lane strides, so their packed layouts differ and are not
- * wire-compatible. Both must satisfy the same roundtrip contract: packing a
- * 256-integer block whose values fit in b bits and unpacking it recovers the
- * input for every width b.
+ * VectorBitPacker (256/512-bit lanes), VectorBitPacker256 (256-bit lanes) and
+ * VectorBitPacker128 (128-bit lanes) use different lane strides, so their packed
+ * layouts differ and are not wire-compatible. Each must satisfy the same
+ * roundtrip contract: packing a 256-integer block whose values fit in b bits and
+ * unpacking it recovers the input for every width b.
  */
 public class VectorBitPackerTest {
 
@@ -61,6 +61,11 @@ public void vectorBitPacker128RoundTrip() {
     roundTrip(new VectorBitPacker128());
   }
 
+  @Test
+  public void vectorBitPacker256RoundTrip() {
+    roundTrip(new VectorBitPacker256());
+  }
+
   @Test
   public void vectorBitPackerRoundTrip() {
     roundTrip(new VectorBitPacker());
diff --git a/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java b/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
index aaf5ac3..e222c09 100644
--- a/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
+++ b/src/test/java/me/lemire/integercompression/vector/VectorFastPFORTest.java
@@ -53,10 +53,13 @@ public void dirtyOutputBufferRoundTrip() {
   @Test
   public void checkDecodableRejectsWiderStream() {
     assertThrows(IllegalStateException.class,
-        () -> VectorFastPFOR.checkDecodable(LaneWidth.BITS_512, LaneWidth.BITS_128));
+        () -> VectorFastPFOR.checkDecodable(LaneWidth.BITS_256, LaneWidth.BITS_128));
+    assertThrows(IllegalStateException.class,
+        () -> VectorFastPFOR.checkDecodable(LaneWidth.BITS_512, LaneWidth.BITS_256));
     // equal or narrower stream decodes natively
     VectorFastPFOR.checkDecodable(LaneWidth.BITS_128, LaneWidth.BITS_128);
     VectorFastPFOR.checkDecodable(LaneWidth.BITS_128, LaneWidth.BITS_512);
+    VectorFastPFOR.checkDecodable(LaneWidth.BITS_256, LaneWidth.BITS_512);
   }
 
   /** 128-bit lanes are the universal floor, so such a stream decodes on any host. */

From 0f4d6590ae7606d565a62084260bc8ea773076bb Mon Sep 17 00:00:00 2001
From: Raunaq Morarka <raunaqmorarka@gmail.com>
Date: Fri, 19 Jun 2026 20:50:11 +0530
Subject: [PATCH 4/5] Build and test the vector module by default

Declare jdk.incubator.vector as requires static so scalar consumers
resolve the module without --add-modules; only VectorFastPFOR users need it.
---
 README.md                      |  5 ++--
 pom.xml                        | 48 +++++++---------------------------
 src/main/java/module-info.java |  8 +++---
 3 files changed, 18 insertions(+), 43 deletions(-)

diff --git a/README.md b/README.md
index 0246789..5888603 100644
--- a/README.md
+++ b/README.md
@@ -101,8 +101,9 @@ Most others do not.
 The Java Team at Intel (R) introduced the vector implementation for FastPFOR
 based on the Java Vector API that showed significant gains over the
 non-vectorized implementation. For an example usage, see
-examples/vector/Example.java. The feature requires JDK 19+ and is currently for 
-advanced users.
+examples/vector/Example.java. On aarch64 (e.g. Graviton3), use JDK 24 or
+later: earlier releases lack the Vector API SVE intrinsics and run a fallback
+that is slower than the scalar codec.
 
 JavaFastPFOR as a dependency
 ------------------------
diff --git a/pom.xml b/pom.xml
index ccdc909..5594a43 100644
--- a/pom.xml
+++ b/pom.xml
@@ -90,45 +90,17 @@
         <configuration>
             <source>21</source>
             <target>21</target>
+            <compilerArgs>
+              <arg>--add-modules</arg>
+              <arg>jdk.incubator.vector</arg>
+            </compilerArgs>
+        </configuration>
+      </plugin>
+      <plugin>
+        <artifactId>maven-surefire-plugin</artifactId>
+        <configuration>
+          <argLine>--add-modules jdk.incubator.vector</argLine>
         </configuration>
-        <executions>
-          <execution>
-            <id>default-compile</id>
-            <phase>compile</phase>
-            <goals>
-              <goal>compile</goal>
-            </goals>
-            <configuration>
-              <excludes>
-                <exclude>me/lemire/integercompression/vector/*</exclude>
-                <exclude>module-info.java</exclude>
-              </excludes>
-            </configuration>
-          </execution>
-          <execution>
-            <id>default-testCompile</id>
-            <phase>test-compile</phase>
-            <goals>
-              <goal>testCompile</goal>
-            </goals>
-            <configuration>
-              <testExcludes>
-                <testExclude>me/lemire/integercompression/vector/*</testExclude>
-              </testExcludes>
-            </configuration>
-          </execution>
-          <!-- The vector module is experimental, currently only for advanced users. -->
-          <!--<execution>
-            <id>vector-fastpfor</id>
-            <phase>compile</phase>
-            <goals>
-              <goal>compile</goal>
-            </goals>
-             <configuration>
-              <release>29</release>
-            </configuration>
-          </execution>       -->
-        </executions>
       </plugin>
       <plugin>
         <groupId>org.apache.felix</groupId>
diff --git a/src/main/java/module-info.java b/src/main/java/module-info.java
index f134601..4d652dd 100644
--- a/src/main/java/module-info.java
+++ b/src/main/java/module-info.java
@@ -2,11 +2,13 @@
 
 // SPDX-License-Identifier: Apache-2.0
 module me.lemire.integercompression {
-  // This is currently only for advanced users:
-  // requires jdk.incubator.vector;
+  // Optional at runtime: only consumers of the vector package (VectorFastPFOR)
+  // need jdk.incubator.vector resolved (e.g. --add-modules jdk.incubator.vector).
+  // Scalar consumers resolve without it.
+  requires static jdk.incubator.vector;
   exports me.lemire.integercompression;
   exports me.lemire.longcompression;
   exports me.lemire.longcompression.differential;
   exports me.lemire.integercompression.differential;
-  // exports me.lemire.integercompression.vector;
+  exports me.lemire.integercompression.vector;
 }

From 9093b1ad4e3a526a333825b2de1392f064251227 Mon Sep 17 00:00:00 2001
From: Raunaq Morarka <raunaqmorarka@gmail.com>
Date: Fri, 19 Jun 2026 22:58:51 +0530
Subject: [PATCH 5/5] Cover VectorFastPFOR with the shared codec test suites

SkippableBasicTest exercises maxHeadlessCompressedLength, so implement it
(mirroring FastPFOR) rather than throwing.
---
 .../lemire/integercompression/vector/VectorFastPFOR.java  | 8 +++++++-
 src/test/java/me/lemire/integercompression/BasicTest.java | 2 ++
 .../me/lemire/integercompression/SkippableBasicTest.java  | 2 ++
 3 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java b/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java
index adbe432..cc755c8 100644
--- a/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java
+++ b/src/main/java/me/lemire/integercompression/vector/VectorFastPFOR.java
@@ -54,6 +54,8 @@
  */
 public class VectorFastPFOR implements IntegerCODEC, SkippableIntegerCODEC {
   private final static int OVERHEAD_OF_EACH_EXCEPT = 8;
+  private static final int OVERHEAD_OF_EACH_PAGE_IN_INTS = 36;
+  private static final int OVERHEAD_OF_EACH_BLOCK_IN_INTS = 1;
   public final static int DEFAULT_PAGE_SIZE = 64 << 10;
 
   public final static int BLOCK_SIZE = 256;
@@ -257,7 +259,11 @@ public void headlessUncompress(int[] in, IntWrapper inpos, int inlength,
 
   @Override
   public int maxHeadlessCompressedLength(IntWrapper compressedPositions, int inlength) {
-    throw new UnsupportedOperationException("Calculating the max compressed length is not supported yet.");
+    inlength = inlength - inlength % BLOCK_SIZE;
+    int pageCount = (inlength + pageSize - 1) / pageSize;
+    int blockCount = inlength / BLOCK_SIZE;
+    int blockSizeInInts = OVERHEAD_OF_EACH_BLOCK_IN_INTS + BLOCK_SIZE;
+    return OVERHEAD_OF_EACH_PAGE_IN_INTS * pageCount + blockSizeInInts * blockCount + 24;
   }
 
   private void loadMetaData(int[] in, int inexcept, int bytesize) {
diff --git a/src/test/java/me/lemire/integercompression/BasicTest.java b/src/test/java/me/lemire/integercompression/BasicTest.java
index b29ae0d..6743017 100644
--- a/src/test/java/me/lemire/integercompression/BasicTest.java
+++ b/src/test/java/me/lemire/integercompression/BasicTest.java
@@ -17,6 +17,7 @@
 import me.lemire.integercompression.differential.IntegratedVariableByte;
 import me.lemire.integercompression.differential.XorBinaryPacking;
 import me.lemire.integercompression.synth.ClusteredDataGenerator;
+import me.lemire.integercompression.vector.VectorFastPFOR;
 
 import org.junit.Test;
 
@@ -43,6 +44,7 @@ public class BasicTest {
             new Composition(new OptPFDS16(), new VariableByte()),
             new Composition(new FastPFOR128(), new VariableByte()),
             new Composition(new FastPFOR(), new VariableByte()),
+            new Composition(new VectorFastPFOR(), new VariableByte()),
             new Simple9(),
             new Simple16(),
             new GroupSimple9(),
diff --git a/src/test/java/me/lemire/integercompression/SkippableBasicTest.java b/src/test/java/me/lemire/integercompression/SkippableBasicTest.java
index 881dada..ca919d4 100644
--- a/src/test/java/me/lemire/integercompression/SkippableBasicTest.java
+++ b/src/test/java/me/lemire/integercompression/SkippableBasicTest.java
@@ -13,6 +13,7 @@
 import me.lemire.integercompression.differential.IntegratedVariableByte;
 import me.lemire.integercompression.differential.SkippableIntegratedComposition;
 import me.lemire.integercompression.differential.SkippableIntegratedIntegerCODEC;
+import me.lemire.integercompression.vector.VectorFastPFOR;
 import org.junit.Test;
 
 import static org.junit.Assert.assertArrayEquals;
@@ -37,6 +38,7 @@ public class SkippableBasicTest {
             new SkippableComposition(new OptPFDS16(), new VariableByte()),
             new SkippableComposition(new FastPFOR128(), new VariableByte()),
             new SkippableComposition(new FastPFOR(), new VariableByte()),
+            new SkippableComposition(new VectorFastPFOR(), new VariableByte()),
             new Simple9(),
             new Simple16() };