← Back to events
ActiveAIRelease

trunk/769b13f46a24d108e74a807bc3bdb5cc8b478758: Add tiled CUDA kernel for dense 2D transpose copies (#194310)

Photo: PyTorch Releases

What happened

Addresses #113560 . t().contiguous() on CUDA goes through the generic TensorIterator elementwise kernel, which can't coalesce both the read and the write of a transpose. This adds a shared-memory tiled kernel for copies that TensorIterator reduces to a 2D transpose, including sources with a row pitch larger than the row (narrowed or strided outer dims). Kernel 128-bit global loads and stores for aligned 1-, 2- and 4…

Summary assembled by rule from the sources below

Why it's spreading

Sources