[Kernel] Add submanifold sparse 3D convolution (bf16 implicit GEMM, gfx950) - #942

Draft

jiacao-amd wants to merge 10 commits into

jiacao-amd:spconv-bf16-implicit-gemm

Commits on Jul 31, 2026

[Kernel] Add submanifold sparse 3D convolution (bf16 implicit GEMM, gfx950)
jiacao-amd
committed
d4161ef
[Kernel] spconv: migrate the map kernels to the fx operator surface
jiacao-amd
committed
3ebc28d
[Kernel] spconv: document why the Z-Delta kernel stays on raw arith
jiacao-amd
committed
a58ae56
[Kernel] spconv: migrate the GEMM prologue and hot loop to the fx surface
jiacao-amd
committed
591ce58
[Kernel] spconv: take kernel_size, not kv_total, in the dense map compiler
jiacao-amd
committed
9708a6d
[Kernel] spconv: use Python `if` instead of explicit scf.IfOp
jiacao-amd
committed
71d8614
[Kernel] spconv: size the compile caches like the rest of kernels/
jiacao-amd
committed
5ba6669
[Kernel] spconv: move pair compaction from torch onto three fx kernels
jiacao-amd
committed
6ee20cd
[Kernel] spconv: derive the rulebook cache key from the coords tensor
jiacao-amd
committed
706a0bf
[Kernel] spconv: scan the pair counts on the device
jiacao-amd
committed
f9cb4f1