trunk/19243ab6d82ea8f9ec13dbc610c4915dcbb40ed6: [MPS] Migrate pixel shuffle/unshuffle to metal (#201724)
- PyTorch 近 90 天出现 2671 次
- PyTorch 近 90 天第 2629 次发版
- 上一次:同一天稍早 · trunk/933b351ff11fbce371597a984b0740606597bfdf: [AOTI] Fix consts cpp build on macOS (#201698)
发生了什么
Replaces the MPSGraph depthToSpace/spaceToDepth path with Metal kernel. Perf: input (NCHW) dtype shuffle before shuffle after speedup unshuffle before unshuffle after speedup (2,3,12,8,8) f=2 bf16 31.2 7.0 4.46x 27.2 6.4 4.25x (2,3,12,8,8) f=2 fp32 29.8 7.1 4.20x 28.7 6.8 4.22x (16,64,32,32) f=4 bf16 24.7 12.8 1.93x 24.4 11.8 2.07x (16,64,32,32) f=4 fp32 26.5 14.8 1.79x 25.3 13.7 1.85x (1,27,256,256) f=3 bf16 28.5 1…
摘要按规则整理自下方来源原文