INT4 GEMV OpenCL Kernel for Adreno GPU Compatibility

[djeong20]

This PR updates the INT4 GEMV OpenCL kernel to ensure compatibility with the Adreno GPU while maintaining the Intel intrinsics.

Self-evaluation:

1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

[myungjoo]

Be careful with such #define.

Someone may do...:

if (a == b)
  BLOCK_READ_IMPL_2;
else
  BLOCK_READ_IMPL_4;

which will break everything.

https://stackoverflow.com/questions/923822/whats-the-use-of-do-while0-when-we-define-a-macro

[piotrrak]

I don't get it:

shouldn't

if (get_subgroup_local_id() != 0)
{
   for (uint i = 1; i <SUBGROUP_SIZE; ++i)
   {
       buf_line[0] = buf_line[0] + buf_line[i];
   }
}

as we would want to reduce all no just some.
Simply don't get it what's going on with this stride.

[piotrrak]

Oh think I get it, to make it work the way you want the reduction accumulation local buf_line[0] + to strided element should be atomic (local atomit fetch add) otherwise it races between subgroup threads. (race between is on buf_line[0] and modifiaction of buf_line[0])

Have you tested correctness of this reduction?

[piotrrak]

According to this document qualcomm gpus support subgroup operations, as both cl_khr_subgroup_* as part of OpenCL 3.0 and cl_qcom_subgroups_*
Only part that should be reimplemented is cl_intel_subgroup_local_block_io

https://docs.qualcomm.com/bundle/publicresource/80-NB295-11_REV_C_Qualcomm_Snapdragon_Mobile_Platform_Opencl_General_Programming_and_Optimization.pdf

Please refer to chapter 8.9 and 9.2 iof above document. I don't see the point in reimplementing those, as they are supported by non-intel extensions. Reusing those should simplify debugging code.

In any case subgroup_reduce_add could be implemented as shuffle_down operation correctly.

[djeong20]

Hello @piotrrak, I agree that only cl_intel_subgroup_local_block_io should be implemented.

#define SLM_BLOCK_READ_FLOAT(ptr_)                                             \
  as_float(intel_sub_group_block_read((const __local uint *)(ptr_)))

Do you have any suggestions for implementing this in non-Intel extensions?

[piotrrak]

https://github.com/pocl/pocl/blob/6842d4a2de5f568d17cbfc30657016662ab7d6c9/lib/kernel/subgroups.cl#L443 might refer to that, that's the other opensource implementation of block_read block_write and might be good starting point. As for supporting very old adreno targets, subgroup reduction and broadcast are basically correspondingly shuffle_down and shuffle_up and those primitives were present by qcom_subgroups extension on even 5 yrs old targets.