ciflow/trunk/195661: [ROCm] Tune MediumRadixSort items_per_thread for small sort sizes
- PyTorch: 402 events in the last 90 days
- PyTorch: 399th Release in the last 90 days
- Previous: earlier the same day · ciflow/xpu/189335
What happened
Use 4 items per thread for 1024/2048 element sorts on ROCm, while keeping 8 items per thread for 4096 element sorts. Benchmarks showed speedups on MI350X (gfx950) and MI450 (gfx1250) for the smaller sizes, but a regression at 4096 due to the larger thread block. CUDA behavior is unchanged.
Summary assembled by rule from the sources below