projects / platform / upstream / llvm.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: 4a96b2e) | patch
[mlir][Linalg] Temporarily circumvent TensorConstant bufferize bug
authorNicolas Vasilache <nicolas.vasilache@gmail.com>
Mon, 12 Oct 2020 20:18:28 +0000 (20:18 +0000)
committerNicolas Vasilache <nicolas.vasilache@gmail.com>
Mon, 12 Oct 2020 20:23:57 +0000 (20:23 +0000)
commit81ead8a53525abe1f149b3c5ad1bb44ee258d10f
tree8409a0f2a168ea88fdd9ecd6681f568d37110804tree | snapshot
parent4a96b2e75f6db5b6f51febdd2ee559bacb7833b7commit | diff
[mlir][Linalg] Temporarily circumvent TensorConstant bufferize bug

The TensorConstantOp bufferize conversion pattern has a bug that
makes it incorrect in the case of vectors whose alignment is not
the natural alignment. Circumvent it temporarily by using a power of 2.

Differential Revision: https://reviews.llvm.org/D89265
mlir/integration_test/Dialect/Linalg/CPU/test-tensor-matmul.mlir diff | blob | history
Domain: System / Toolchain;