Remove MAX_DEVICE_MEMORY_SIZE_DIVISOR from min_max_mem_alloc test while quering maxAllocSize

[joselopeqti]

Removed MAX_DEVICE_MEMORY_SIZE_DIVISOR from get_device_info_max_mem_alloc_size() call so the test uses the actual reported max allocation size instead of a divided value.
This fixes an issue with devices with low memory as the test expects to allocate at least 128mb.

[rjodinchr]

I think we talked about MAX_DEVICE_MEMORY_SIZE_DIVISOR during the teleconference a while ago. What I remember from it is that CL_DEVICE_MAX_MEM_ALLOC_SIZE is not the maximum size that one can allocate. It is the size above which you should not expect to be able to allocate. Failing to allocate CL_DEVICE_MAX_MEM_ALLOC_SIZE -1 is compliant with the OpenCL specification if I remember correctly.

Thus removing MAX_DEVICE_MEMORY_SIZE_DIVISOR does not seem a proper fix. Instead we could make sure that we report at least what OpenCL expects as its minimum?

[joselopeqti]

Is there any reason for the requiredAllocSize of 128mb used by the test?

[rjodinchr]

Is there any reason for the requiredAllocSize of 128mb used by the test?

I think 128mb comes from CL_DEVICE_GLOBAL_MEM_SIZE for the full profile.

I agree that the current implementation of this test is weird (wrong?).

We should at least apply the same formula as we find in the specification:

Max size of memory object allocation in bytes. The
minimum value is max(min(1024 × 1024 × 1024, 1/4th of
CL_DEVICE_GLOBAL_MEM_SIZE), 32 × 1024 × 1024).

But for the real allocation (using clCreateBuffer) in the test, we should keep MAX_DEVICE_MEMORY_SIZE_DIVISOR.