projects / platform / upstream / llvm.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: ac1bedd) | patch
[mlir][nvpu] Prevent F32ToTF32 pattern to generate illegal IR
authorThomas Raoux <thomasraoux@google.com>
Mon, 15 Aug 2022 16:16:46 +0000 (16:16 +0000)
committerThomas Raoux <thomasraoux@google.com>
Mon, 15 Aug 2022 16:46:18 +0000 (16:46 +0000)
commit91594b5b985cf144f9cde64cddc534684d644665
tree25b3e7dbc091d57f57a606c170a60283cfe4417atree | snapshot
parentac1bedd584359567266b1c41ac26734556031b9fcommit | diff
[mlir][nvpu] Prevent F32ToTF32 pattern to generate illegal IR

We shouldn't apply this pattern to non F32->F32 mma.sync operations.

Differential Revision: https://reviews.llvm.org/D131902
mlir/lib/Dialect/NVGPU/Transforms/MmaSyncTF32Transform.cpp diff | blob | history
mlir/test/Dialect/NVGPU/mma-sync-f32-to-tf32.mlir diff | blob | history
Domain: System / Toolchain;