[Triton] GEMM tunning script

[k50112113]

This PR adds GEMM tunning scripts that

1. screen through parameter space and select the most performant config
2. generate JSON config files
3. verify performance results

[fxmarty-amd]

This is very useful and works! I am able to get significant speedups in certain configurations that are not by default tuned.

Do you have a recommendation to verify correctness compared to using the default triton config? e.g. as compared to the default https://github.com/ROCm/aiter/blob/main/aiter/ops/triton/configs/gemm/gfx950-GEMM-AFP4WFP4_PRESHUFFLED.json for mxfp4 preshuffled?

edit: using https://github.com/ROCm/aiter/blob/main/op_tests/triton_tests/gemm/basic/test_gemm_afp4wfp4.py works

[k50112113]

Do you have a recommendation to verify correctness compared to using the default triton config? e.g. as compared to the default https://github.com/ROCm/aiter/blob/main/aiter/ops/triton/configs/gemm/gfx950-GEMM-AFP4WFP4_PRESHUFFLED.json for mxfp4 preshuffled?

For new shapes that you tune, the JSON file name would be gfx950-GEMM-AFP4WFP4_PRESHUFFLED-N={N}-K={2*K}.json in the case of mxfp4 GEMM, so we don't have to replace the default JSON file (the one without N and K sufix), and of course, once you have generated those JSON files, you can add them to aiter/ops/triton/configs/gemm/ and push a PR for us to review.

edit: using https://github.com/ROCm/aiter/blob/main/op_tests/triton_tests/gemm/basic/test_gemm_afp4wfp4.py works

Yes, we recommend that you use the test_gemm_afp4wfp4.py scripts once you have added/modified your JSON file to aiter/ops/triton/configs/gemm/ if test_gemm_afp4wfp4 does not contain the shape you tunned, please add your shape at test_gemm_afp4wfp4::get_x_vals() and run the test script locally before pushing the PR. If you see that the test fails, please push your changes to a branch and contact us.