tree 06ade9e40f5ec3356aa610ab5469a70591f67a7a
parent 9ffd9eee7247adebd0f48d53a8d593fc6b9b7bf6
author Benjamin Kramer <kramerb@google.com> 1589965652 -0700
committer TensorFlower Gardener <gardener@tensorflow.org> 1589966070 -0700

[XLA:CPU] Plumb through a minimal emitter for matmuls using the mlir linalg dialect

This is just the most basic lowering and will generate linalg.matmul for small
matmuls and then convert to loops. The result is fairly slow, but we can
iterate on that.

To make XLA use it set XLA_FLAGS=--xla_backend_extra_options=xla_use_linalg_for_dot

PiperOrigin-RevId: 312443993
Change-Id: Icaf20764c954803f5b1bacfaa9456839b28ba52c