tree 594ea51db738758299be7da638545ee8f9440730
parent ed8ff5e79a164943789c165cb06c5bcd1cc5a98a
author Han-Chung Wang <hanchung@google.com> 1613677418 +0800
committer GitHub <noreply@github.com> 1613677418 +0800
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJgLsNqCRBK7hj4Ov3rIwAAdHIIAK3vmGto/YA+2BnSTY+c51Wu
 Y8ZrDcoGx6f3MlGSthDTkhSKoCp1rb/NJL6qZLnCKSrLV9LoKyjY9HXnBmJmz8Dy
 BcnijS0GGKrUosQ3dS0ARiQaAQd7uKDk+PyldjpF+ZUumaioALRY9OiYdqytOLSO
 AsP8k+EB6Y5kK+QNKoJ//Zt9w+ypwvPcvjQS4cFuAQvzXAAxbNiNt/DwBPBOb+lj
 HdYCnGeeL/i3NfhtiEPrZsMqXsiEdnWpi1RZXuc3H5ooZhN01OxrISsZHjLXnsCv
 XgPy/z8cfLFY8d69zRyw0xk1B2RDjsmeD+Kb1Xz8JSSrp+veiZnYwOoiq8PLFCw=
 =Lssn
 -----END PGP SIGNATURE-----
 

Adapt mhlo.conv to Linalg on buffers patterns to be on tensors. (#4748)

Lowers a conv op to a static rank conv op. The general conv op only
supports on memrefs. This limits the support for conv op to specific
data layout and only work for few ranks, ie 1d, 2d, and 3d. This
separates the lowering to normal conv and depthwise conv. The former one
is lowering to Linalg on tensors, and the latter one is leaving with the
same behavior, ie, lowering to Linalg on buffers.

Add support of TC conv ops to later GPU passes.

Depends on https://reviews.llvm.org/D96038