xelpg: jit: gemm: additional f16 accumulation strategies

Open petercad opened this issue 7 months ago • 4 comments

Adds some f16 accumulation FMA strategies (opt-in with --attr-acc-mode=f16) for MTL. Theoretical peak is 2x faster than f32 accumulation and actual performance speedup is similar.

Jun 11 '25 17:06 petercad

make test linters

Jun 16 '25 16:06 petercad

make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb disable benchdnn_all enable benchdnn_matmul

Jun 16 '25 16:06 petercad

make test linters

Jun 30 '25 22:06 petercad

make test disable test_device_cpu disable build_cpu_runtime_omp disable build_cpu_runtime_sycl disable build_cpu_runtime_tbb disable benchdnn_all enable benchdnn_matmul

Jun 30 '25 22:06 petercad