Add GPU-initiated and Host-initiated SDMA executors via Anvil - #343

Draft

nileshnegi wants to merge 35 commits into

candidate-1.70from

users/nileshnegi/feat/rad-sdma-anvil

Commits on Jul 31, 2026

feat: add anvil KFD/SDMA queue library headers

nileshnegi
and
claude
committed
feat: add anvil.cpp — xio-free adaptation of anvil.hip

nileshnegi
and
claude
committed
build: add ENABLE_ANVIL_EXEC cmake option for GPU-initiated SDMA executor

nileshnegi
and
claude
committed
feat: add EXE_GPU_INITIATED_DMA enum value (char code 'S')

nileshnegi
and
claude
committed
build: fix hsakmt IMPORTED target — use STATIC for .a, add drm transitive deps

nileshnegi
and
claude
committed
feat: add anvilQueues field to ExeInfo for SDMA queue handles

nileshnegi
and
claude
committed
feat: add TransfersHaveErrors validation for EXE_GPU_INITIATED_DMA

nileshnegi
and
claude
committed
feat: implement PrepareAnvilExecutor and extend Teardown for anvil

nileshnegi
and
claude
committed
feat: implement AnvilTransferKernel and RunAnvilExecutor

nileshnegi
and
claude
committed
fix: add gfx1250 s_waitcnt compat in sdma-ep.h; remove kernel no-op guard

nileshnegi
and
claude
committed
fix: make IsGpuExeType unconditional to prevent segfault on gfx950

nileshnegi
and
claude
committed
build: add ENABLE_ANVIL_EXEC flag to Makefile

nileshnegi
and
claude
committed
build: pass libhsakmt.a via -Wl, to avoid -x hip parsing it as source

nileshnegi
and
claude
committed
feat: add TB_VERBOSE=1 logging for anvil alloc/dealloc lifecycle

nileshnegi
and
claude
committed
fix: add EXE_GPU_INITIATED_DMA to ExeTypeToStr as "GISDMA"

nileshnegi
and
claude
committed
feat: extend USE_DMA_EXEC in AllToAll preset to support BMA and GISDMA

nileshnegi
and
claude
committed
fix: remove ANVIL_EXEC_ENABLED guard from RecursiveWildcardTransferExpansion switch

nileshnegi
and
claude
committed
fix: remove ANVIL_EXEC_ENABLED guard from GetHsaAgent switch case

nileshnegi
and
claude
committed
anvil: suppress nodiscard warnings on hipFree in SdmaQueue dtor

nileshnegi
and
claude
committed
anvil: add HSA preferred-engine selection + SDMA affinity topology table

nileshnegi
and
claude
committed
topology: fix SDMA affinity table column alignment

nileshnegi
and
claude
committed
fix: honor byteOffset in DMA and Anvil writes

nileshnegi
and
cursoragent
committed
anvil: run GISDMA transfers concurrently with per-transfer timing

nileshnegi
and
cursoragent
committed
anvil: add opt-in round-robin SDMA engine selection

nileshnegi
and
cursoragent
committed
anvil: use hipExtLaunchKernelGGL for GISDMA kernel timing

nileshnegi
and
cursoragent
committed
anvil: use engine-ordered signal for GISDMA completion

nileshnegi
and
cursoragent
committed
anvil: add runtime toggles for GISDMA doorbell placement and completion wait

nileshnegi
and
cursoragent
committed
anvil: add ANVIL_LEGACY toggle for legacy put/quiet GISDMA completion

nileshnegi
and
cursoragent
committed
anvil: split GISDMA kernel into four non-branching variants

nileshnegi
and
cursoragent
committed
anvil: chunk GISDMA copies to lift the 1 GiB per-transfer limit

nileshnegi
and
cursoragent
committed
anvil: add host-initiated (CPU-driven) SDMA queue path

nileshnegi
and
cursoragent
committed
anvil: add OSS7 fused copy+signal SDMA packet (gfx1250)

nileshnegi
and
cursoragent
committed
build: enable XIO_SDMA_OSS7 fused packets by default (all-arch safe)

nileshnegi
and
cursoragent
committed
anvil: rename GISDMA executor to GMA and split host path into HMA

nileshnegi
and
cursoragent
committed
docs: document GMA/HMA anvil SDMA executors in CHANGELOG

nileshnegi
and
cursoragent
committed