oneDNN icon indicating copy to clipboard operation
oneDNN copied to clipboard
verified publisher icon oneapi-src

xehpg: jit: gemm: reduce m tile size for int4 * int8 gemv strategies

Open petercad opened this issue 8 months ago • 4 comments

Addresses MFDNN-13752. Some of the new strategies from #2788 run out of registers -- this PR reduces the m tile size, which avoids this and also seems to improve performance.

petercad avatar May 28 '25 21:05 petercad

make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb disable arch_gpu_xe-hpc enable arch_gpu_xe-hpg-atsm enable arch_gpu_xe-hpg-dg2 disable arch_gpu_xe-lp enable arch_gpu_xe-lpg enable arch_gpu_xe-lpg+ disable arch_gpu_xe2-hpg-bmg disable arch_gpu_xe2-lpg disable arch_gpu_xe3-lpg disable benchdnn_all enable benchdnn_matmul

petercad avatar May 28 '25 21:05 petercad

make test perf-gpu set primitive=matmul disable arch_gpu_xe-hpc enable arch_gpu_xe-hpg-atsm enable arch_gpu_xe-hpg-dg2 disable arch_gpu_xe-lp enable arch_gpu_xe-lpg enable arch_gpu_xe-lpg+ disable arch_gpu_xe2-hpg-bmg disable arch_gpu_xe2-lpg disable arch_gpu_xe3-lpg

petercad avatar May 28 '25 21:05 petercad

make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb disable arch_gpu_xe-hpc enable arch_gpu_xe-hpg-atsm enable arch_gpu_xe-hpg-dg2 disable arch_gpu_xe-lp enable arch_gpu_xe-lpg enable arch_gpu_xe-lpg+ disable arch_gpu_xe2-hpg-bmg disable arch_gpu_xe2-lpg disable arch_gpu_xe3-lpg disable benchdnn_all enable benchdnn_matmul

petercad avatar May 28 '25 22:05 petercad

CI failures are unrelated.

The large perf regressions are real but somewhat tangential to this PR as they will disappear when MFDNN-13651 is resolved.

petercad avatar May 29 '25 15:05 petercad

Addressed as part of #3390.

petercad avatar Sep 23 '25 22:09 petercad