ciflow/torchtitan/200699: [inductor] Split persistent and looped reductions in combo kernels
- PyTorch: 2235 events in the last 90 days
- PyTorch: 2206th Release in the last 90 days
- Previous: earlier the same day · ciflow/trunk/200591
What happened
Summary: Without combo_kernel_per_subkernel_blocks , the sub-kernels of a combo kernel share one XBLOCK and one num_warps , and select_combo_heuristics takes the autotune configs from a single sub-kernel, a looped reduction whenever the combo has one. A persistent reduction keeps a whole row in registers and wants a small XBLOCK , while a looped reduction wants a large XBLOCK over a small R0_BLOCK , so in a mixed co…
Summary assembled by rule from the sources below