trunk/67af8e140dce108c4c2fe6ffdec8317f2b93deb1: Fix AVX512 grid sampler mask gather (#197635)
- PyTorch: 1578 events in the last 90 days
- PyTorch: 1559th Release in the last 90 days
- Previous: earlier the same day · trunk/536fcc5eeabf50519f9173c7bf8410993d5b2440
What happened
This PR fixes an incorrect OP AI-assisted summary, reviewed by the author: Summary The AVX512 mask_gather converted an all-ones mask to a bitmask with floating-point equality. The all-ones pattern is NaN as float or double, so every comparison was false and grid_sample returned zeros when the AVX512 kernel ran. Extract the sign bit with _mm512_movepi{64,32}_mask , matching AVX2. Also make the sign bit the active-lan…
Summary assembled by rule from the sources below