trunk/7e1c35ebdb7015a89ecdb99030893a767550eef1: [CUDA] Add non-overlapping fast path for avg_pool2d backward (#191086)
- PyTorch: 239 events in the last 90 days
- PyTorch: 236th Release in the last 90 days
- Previous: earlier the same day · trunk/10a2f3aa6abef3a84ad0ba37007fbce4d4681a6b: [TokenSwitch] Add symmetric-memory zero-copy dispatch/combine (#187785)
What happened
Summary Adds a dedicated CUDA backward kernel for avg_pool2d when the pooling windows tile the input exactly ( stride == kernel_size , no padding, no ceil_mode ) -- the common downsampling case such as avg_pool2d(2, 2) . In that regime each input element maps to exactly one output window with a constant divisor, so the general kernel's per-element window scan and per-window bound/divisor recomputation collapse to a…
Summary assembled by rule from the sources below