trunk/7fe528830802c15f416786ae4c9947ac925a018a: Stop instantiating unreachable cat kernels (#194184)
- PyTorch: 1512 events in the last 90 days
- PyTorch: 1493th Release in the last 90 days
- Previous: earlier the same day · trunk/4aefab0c58cad9034fb12e208b18ec3d1a6d6500
What happened
parallel_cat cleared isAligned and isInOutAligned at runtime under USE_ROCM , because CatArrayBatchedCopy_contig is faster there ( #118685 ). That makes CatArrayBatchedCopy_alignedK_contig and CatArrayBatchedCopy_vectorized unreachable on ROCm, but a runtime flag cannot drop them: __global__ functions are DCE roots, so both families are still instantiated and emitted. A constexpr bool does, as in SoftMax.cu:1109 . T…
Summary assembled by rule from the sources below