perf(algebra): ⚡ decide the structural cutoff from a (k, d) digest - #234
perf(algebra): ⚡ decide the structural cutoff from a (k, d) digest#234diagonal-hamiltonian wants to merge 1 commit into
Conversation
The structural cutoffs only ever needed two integers about a monomial: its popcount k and the number of modes holding both of their Majorana bits, d. Computing them meant a call out to cutoff_sums, which walks the dense words to build a sum it then throws away. `paired_mode_count` computes d inline from the dense words and `passes_with_popcount` decides from (k, d) directly, dropping the popcount<=cutoff early-out. The dense representation is untouched: this changes no layout and moves no data, which is why it is a pure decision change and why its counters must stay bit-identical between arms. Default ON. Measured emit_s 0.9200x, 11/12 paired reps over two nodes, sign-test p=0.0063; pooled with 128- and 512-mode shapes 0.9214x, 19/20, p=4.0e-5. Counters (emit/reject/push) are bit-identical between arms in all 20 pairs and memory is flat. Worth ~0.8% of build_graph, NOT the ~3% the microbenchmark projected -- in isolation the kernel measures far better than it delivers, because in the real scan the loop is bound by the latency of the random row read rather than by the arithmetic beside it. Quoted at what it delivers. Measure this on emit_s, not on the wall: at the observed variance the end-to-end benchmark needs ~1590 paired reps to resolve 3%, and the phase timer needs ~5. `env_config_member_defaults_match_parsed_defaults` is the other half of this change and matters more than the perf. Every knob declares its default twice -- once as a Settings member initialiser, once as parse_flag's fallback -- and get() unconditionally overwrites the first with the second, so changing only the member silently does nothing. That shipped a measured regression enabled by default once already, on a knob that had been "turned off" in the member, in the header comment and in the docs, but not in the fallback. Any knob whose two declarations disagree now fails a test. Assisted-by: ClaudeCode:claude-opus-5
|
Docs preview: https://pr-234.monoprop-docs.pages.dev |
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## pr/mpi-lease-staging #234 +/- ##
=====================================================
Coverage 97.70% 97.70%
=====================================================
Files 14 14
Lines 742 742
Branches 98 98
=====================================================
Hits 725 725
Misses 12 12
Partials 5 5
Flags with carried forward coverage won't be shown. Click here to find out more. |
|
|
🤖 AI text below 🤖 Closing this stack to rebuild it around a single-path design. Not abandoned — the work returns as fresh PRs once it is ready. Why. The stack accumulated runtime knobs as a way to A/B competing paths, and an A/B knob is not a shipping knob. Against
What the measurements decided. Several paths were being kept behind default-off knobs on the strength of stale or never-run comparisons:
What returns. The sparse query record together with the positions-based resolve (this PR's successor), the self-resolve simplification, and the layer profiler. Roughly 1,800 lines are deleted rather than added, and every PR in the new stack is a pure addition — the intermediate record format is never introduced, so nothing is added and then removed. Reopening as new PRs once the local work is verified. |
1 participant
🤖 AI text below 🤖
Summary
The structural cutoffs only ever needed two integers about a monomial: its popcount
k, and the number of modes holding both of their Majorana bits,d. Computing them meant calling out tocutoff_sums, which walks the dense words to build a sum it then throws away.paired_mode_countcomputesdinline from the dense words andpasses_with_popcountdecides from(k, d)directly, dropping thepopcount <= cutoffearly-out. The dense representation is untouched: this changes no layout and moves no data, which is why it is a pure decision change and why its counters must stay bit-identical between arms.Measured
Default ON.
emit_s0.9200×, 11/12 paired reps over two nodes, sign-test p=0.0063; pooled with 128- and 512-mode shapes 0.9214×, 19/20, p=4.0e-5. Counters (emit/reject/push) bit-identical between arms in all 20 pairs, memory flat.Worth ~0.8% of
build_graph— not the ~3% the microbenchmark projected. In isolation the kernel measures far better than it delivers, because in the real scan the loop is bound by the latency of the random row read rather than by the arithmetic beside it. Quoted at what it delivers.Measure this on
emit_s, not on the wall: at the observed variance the end-to-end benchmark needs ~1590 paired reps to resolve 3%, and the phase timer needs ~5.The test that matters more than the perf
env_config_member_defaults_match_parsed_defaults.Every knob declares its default twice — once as a
Settingsmember initialiser, once asparse_flag's fallback inget()— andget()unconditionally overwrites the first with the second. So changing only the member silently does nothing.That is not hypothetical. It shipped a measured 7.2% regression enabled by default, on a knob that had been "turned off" in the member, in the header comment and in the docs — but not in the fallback. Any knob whose two declarations disagree now fails a test.
Verification
ctest -L serialgreen;pytest tests --with-mpigreen at 1/2/4 ranks.Checklist
docs/,CONTRIBUTING.md) if neededCHANGELOG/ release notes updated if applicable — release notes are generated from the PR titleAI/LLM disclosure