[Kernel] conv3d: parameterize tile size + add autotuner - #820

Merged

coderfeli merged 17 commits into

jiacao-amd:jiacao/conv3d-multi-tile-autotune-v2

Jul 24, 2026

Commits on Jul 22, 2026

conv3d: parameterize tile size + add autotuner
jiacao-amd
committed
conv3d: fix large-tensor (>2^31 elem) input addressing for bf16 and fp8
jiacao-amd
committed
conv3d: vectorized epilogue store + 2D/1D entry points

jiacao-amd
authored and
jiacao-amd
committed
conv3d: support >2^31-element inputs on the 3x1x1 fast path

jiacao-amd
authored and
jiacao-amd
committed
conv3d fp8: port vectorized epilogue, 2D/1D entry points, and >2^31 support

jiacao-amd
authored and
jiacao-amd
committed
conv3d: remove bench_conv3d_tiles.py

jiacao-amd
authored and
jiacao-amd
committed
conv3d: unify 1D/2D/3D behind one entry + merge split-K helpers

authored and
jiacao-amd
committed
conv3d: async global->LDS copy + 4-stage pipeline (+10-14%)

jiacao-amd
authored and
jiacao-amd
committed
conv3d: extend async copy to large (>2^31 elem) tensors

jiacao-amd
authored and
jiacao-amd
committed
conv3d: rename kernels to drop _8wave; sync bf16 to gemm-port

jiacao-amd
authored and
jiacao-amd
committed
conv3d: add wgm L2-swizzle + XOR bank-swizzle to bf16 implicit-GEMM kernel

jiacao-amd
and
claude
committed
conv3d: support BIG_IN (>2^31-elem) with n>1 batch size

jiacao-amd
and
claude
committed
conv3d: remove comments added in BIG_IN n>1 commit

jiacao-amd
and
claude
committed
conv3d fp8: fix epilogue and style after rebase conflict resolution
jiacao-amd
committed
conv3d: apply black formatting to conv3d_implicit_autotune
jiacao-amd
committed

Commits on Jul 23, 2026