tree 92f26848315aad0cb30bc0358022dc6fab5060bb
parent ae6f01883fcdcf90956af4180ce725ba04a5e1e7
author Sean Silva <silvasean@google.com> 1588205859 -0700
committer Copybara-Service <copybara-worker@google.com> 1588205948 -0700

[vmla] Dynamic shape lowering for dot_general

- Cleanly separate VMLA/Transforms/PreConversionLowering.cpp which does preparatory tensor->tensor rewrites.

- restructure conversion so xla_hlo.broadcast_in_dim and xla_hlo.dynamic_broadcast_in_dim are lowered via shapex.ranked_broadcast_in_dim

- add shape transfer function plugin for VMLA (only vmla.batch.matmul.pseudo for now); Also fixup docs on vmla.batch.matmul.pseudo.

- shape transfer function for xla_hlo.dynamic_reshape

- augment matrix_ops_test.py to test the cases all the way from Python!!! This is the end of the dynamic matmul! (though there are still some major issues, like error handling :/)

Also, fix up CMake build for convert-shape-to-shapex pass.

PiperOrigin-RevId: 309129583