[ROCM] Fix buffer_load_lds support for gfx950#2248

Draft

benenzhu wants to merge 36 commits into

tile-ai:mainfrom

benenzhu:feat-c-remove_vmcnt0

Commits on May 17, 2026

M2: add cp_async_gs_lds_with_rsrc<N> device templates for gfx950

benenzhu
and
claude
committed
M3: register ptx_cp_async_lds / ptx_make_buffer_resource / ptx_cp_async_lds_rsrc TIR ops

benenzhu
and
claude
committed
M4: HIP codegen handlers for buffer_load_lds ops + hoisted-resource AttrStmt

benenzhu
and
claude
committed
M5: route eligible 16B copies to ptx_cp_async_lds on gfx950

benenzhu
and
claude
committed
M6: HoistBufferResource Python pass + pipeline wiring (gfx950 only)

benenzhu
and
claude
committed
M7: layout SwizzleDelta API + ROCm swizzle-swap in lower_tile_op

benenzhu
and
claude
committed
M8a: re-enable MergeSharedMemoryAllocations in OptimizeForTarget

benenzhu
and
claude
committed
M8b: integrate ptx_cp_async_lds with vec-loop folding + buf-merge + codegen num_elems convention

benenzhu
and
claude
committed
M5-harden: replace XOR-call gate with real affine lane-contiguity proof

benenzhu
and
claude
committed
M9: VisitExpr_(CallNode) swizzle-swap on tl::ptx_cp_async_lds (real fix for AC-2/AC-1)

benenzhu
and
claude
committed
M6.5: AMD vmcnt wait-count scaling in HoistBufferResource

benenzhu
and
claude
committed
add a flag to save temp files
benenzhu
committed
M9-safe: downgrade ptx_cp_async_lds to ptx_cp_async when swap can't produce linear LDS

benenzhu
and
claude
committed

Commits on May 18, 2026

Commits on May 21, 2026

Commits on May 22, 2026

Commits on May 23, 2026

Commits on May 27, 2026