viable/strict/1792428192: [MPS] Migrate std_mean to metal (#201310)
- PyTorch: 2585 events in the last 90 days
- PyTorch: 2550th Release in the last 90 days
- Previous: earlier the same day · ciflow/xpu/201034: Update torch-xpu-ops commit pin
What happened
Migrates std_mean to metal by reusing existing reduction primitives. Perf on m5 pro (no regressions): Shape Reduction dims Dtype Layout Strides std_mean speedup var_mean speedup (32,) all fp32 contiguous (1,) 3.36x 3.46x (32,) all bf16 contiguous (1,) 3.42x 3.46x (32,) all fp32 stride 2 (offset 1) (2,) 3.97x 4.01x (32,) all bf16 stride 2 (offset 1) (2,) 3.97x 3.99x (1024,) all fp32 contiguous (1,) 2.67x 2.60x (1024,…
Summary assembled by rule from the sources below