tree 05c419129705e35638cf95fb1224eac1b82d245b
parent 66216656acb831d954dada6001fc02735c98ee6d
author Yuanzhong Xu <yuanzx@google.com> 1583458934 -0800
committer TensorFlower Gardener <gardener@tensorflow.org> 1583459258 -0800

[MLIR:TF/XLA] Shape inference of variant types

1) use the same handling as resource when calling to registry's shape functions.
2) create some pass-through cases, because the shape functions often skip subshapes, which is now a problem for tensor list. Since a while loop's tensorlist output is usually passed to other ops, while's shape function skipping the handle shapes causes cascading problems.

PiperOrigin-RevId: 299238548
Change-Id: I63ced0bde726a04ddd2aa204eef636646626c4eb