Request: support add perf testing of matrix multiply "accelerator units" in benchmark.. #29
Description
Hi,
seems after dot product (int8) support next "hot" feature could add is "matrix multiply accelerator" perf testing support..
aka tensor cores on nvidia etc...
searching on opencl.gpuinfo.org I see ARM and Intel extensions (not Nvidia nor AMD nor Qualcomm which is sad to see):
*cl_arm_matrix_multiply 0.4%
*cl_intel_subgroup_matrix_multiply_accumulate 2.2%
*cl_intel_subgroup_matrix_multiply_accumulate_tf32 0.5%
*cl_intel_subgroup_split_matrix_multiply_accumulate 1.6%
of intel the one to test seems "cl_intel_subgroup_matrix_multiply_accumulate" (as tf32 is the tensorfloat32 variant) supported on Intel ARC on Windows and Linux and also..
the ARM one is only supported on Pixel products which is sad:
https://opencl.gpuinfo.org/listreports.php?extension=cl_arm_matrix_multiply
more sad is that no ext seems published altough seems to be supported in ARM Compute library which is open source I think so maybe there is code to learn from..:
https://android.googlesource.com/platform/external/ComputeLibrary/+/refs/heads/android14-qpr1-s2-release%5E1..refs/heads/android14-qpr1-s2-release/
so in brief.. do you plan on investigating adding a cross vendor "matrix multiply accelerator units" test at least for Intel and ARM GPUs..
NOTE: vkpeak does the same with equivalent cooperative matrix ext, but there HW support is broader (AMD,NV,Intel, etc..)
thanks..