master

文件	最后提交记录	最后更新时间
CMakeLists.txt	[Huawei][AscendNPU IR] migrate NormalizeVector and FlattenVectorOps from A5 Migrate the A5 Vector NormalizeVector and FlattenVectorOps passes, register them in the Vector transform pass set, wire them into BiShengIRVectorTransforms, and import the associated A5 lit coverage. This includes the NormalizeVector pass option path from A5, including NormalizeVectorOptions, the enable-dot-scaled-compile command-line option, the option-aware createNormalizeVectorPass entry point, and the conditional TransferReadToGatheringLoadPattern registration. Also extract the minimal shared helpers needed by the migrated passes from HIVM/Utils and Dialect/Utils, and update the Vector transform CMake wiring. A5 source/test basis: f9d978b144433ff121048601f315733610c8525d. Local adaptations: keep the existing A3 Vector pass declaration ordering and dependent dialect style; replace one flatten-vector test producer that used an unavailable AVE op with an arith.addf vector producer; guard HIVMAVE conversion tests with REQUIRES: regbase and TODOs because -convert-arith-to-hivmave and -convert-vector-to-hivmave are not registered in this branch. LLVM dependency: update third-party/llvm-project to local commit c2d43da95d9b, which fixes vector transfer drop-unit-dims mask handling so the migrated NormalizeVector tests verify on the current LLVM verifier. Verification: ninja; build/bin/llvm-lit -sv on the 24 migrated/related tests reports 14 passed and 10 unsupported due to regbase-guarded missing pass stacks; build/bin/llvm-lit -sv on Vector + HFusion reports 139 passed and 25 unsupported; touched HIVM/RegBase/HIVMAVE set reports 5 passed and 2 unsupported; bishengir-compile/hfusion reports 2 passed and 27 unsupported.	1 个月前
FlattenVectorOps.cpp	[Vector] preserve strided layout when collapsing to 1D collapseAllDims built the collapsed memref type with MemRefType::get(), which drops the strided layout and dynamic offset of a rank-reduced subview source, producing an invalid collapse_shape the verifier rejects. Use memref::CollapseShapeOp::computeCollapsedType to preserve them, and reject non-contiguous strided sources via isGuaranteedCollapsible to avoid the computeCollapsedType assert.	5 天前
MaterializeVectorWriteToDestination.cpp	[AscendNPU-IR] fix: fix llvm-22	1 个月前
NormalizeVector.cpp	Revert "!2583 merge flatten-perf into master" This reverts commit 6c6c800c7df82ca4e96ba79cd2ab8dc6432f4182, reversing changes made to 31d7b93bc4f71cf05b5592b6b6ce28b687c14799.	6 天前
PeelLoopsContainingTranspose.cpp	[Huawei][AscendNPU IR] migrate PeelLoopsContainingTranspose from A5 Migrate the A5 Vector PeelLoopsContainingTranspose pass, register it in the Vector transform pass set, wire it into BiShengIRVectorTransforms, and add the corresponding lit test. A5 source/test commit: ca90b6e3b Initialize A5 AscendNPUIR. Verification: ninja -C build; build/bin/llvm-lit -sv bishengir/test/Dialect/Vector/peel_loops_containing_transpose.mlir; build/bin/llvm-lit bishengir/test reports 474 total, 376 passed, 97 unsupported, 1 accepted baseline failure: bishengir :: bishengir-compile/save-temps.mlir.	1 个月前
VectorTransferLowering.cpp	[Huawei][AscendNPU IR] migrate VectorTransferLowering from A5 Migrate the A5 VectorTransferLowering pass and register it in the Vector transform library. A5 source commit: ca90b6e3b Initialize A5 AscendNPUIR. A5 test analysis: the pass is exercised from bishengir/test/Dialect/Vector/normalize_vector.mlir in A5; no standalone VectorTransferLowering lit file was found. Local adaptation: normalize VectorTransferLowering.cpp line endings to LF and add a final newline to satisfy diff checks. Verification: ninja -C build; build/bin/bishengir-opt -vector-transfer-lowering on a temporary 1D transfer read/write snippet; build/bin/llvm-lit bishengir/test. Full lit has accepted baseline failure only: bishengir :: bishengir-compile/save-temps.mlir.	1 个月前