viable/strict/1790368010: [MPS] Port norm to the shared reduction kernels (#198611)
- PyTorch: 1480 events in the last 90 days
- PyTorch: 1461th Release in the last 90 days
- Previous: earlier the same day · trunk/8540373d3cd96255ab87ee2f2c91353a363690cb
What happened
Port norm / linalg.vector_norm on MPS from its standalone Metal kernel (one threadgroup per output, so full reductions ran on one threadgroup) onto the shared reduction kernels. Op p bf16 before (us) bf16 after (us) bf16 before/after f32 before/after [8192, 1024] , dim=-1 2 35.4 35.3 1.00x 1.00x [8192, 1024] , dim=0 2 509 37.4 13.6x 5.13x [64, 1024, 64] , dim=1 2 357 20.3 17.6x 10.8x [1000000, 4] , dim=0 2 989 27.9…
Summary assembled by rule from the sources below