[r2.2:Cherryppick] Split unroll_batch_matmul pass into own build target

[lgeiger]

This is a cherry-pick of #37356

@jpienaar Since this doesn't include functional changes it would be amazing if it could be cherry-picked onto the 2.2 release branch.

[mihaimaruseac]

I think you are missing the removal of these lines

tensorflow/tensorflow/compiler/mlir/tensorflow/transforms/unroll_batch_matmul.cc

Lines 28 to 29 in b720399

	#include "mlir/Dialect/QuantOps/FakeQuantSupport.h" // TF:llvm-project
	#include "mlir/Dialect/QuantOps/UniformSupport.h" // TF:llvm-project

It was in #37356 but it is not here. Is it possible the cherrypick was wrong?

[lgeiger]

@mihaimaruseac If I compare 0e86d51 and this PR side by side the only difference are the changes in tensorflow/compiler/mlir/tensorflow/transforms/batchmatmul_to_einsum.cc which are not relevant on the r2.2 branch since the file doesn't exist, or am I missing something?

[mihaimaruseac]

Oh, sorry, I was looking at the wrong file.

However, there are still differences between the PR on master and the PR on this branch: On master https://github.com/tensorflow/tensorflow/pull/37356/files#diff-f20d136d02b72175cd9891df563f8d62R267-R280. On this branch https://github.com/tensorflow/tensorflow/pull/37362/files#diff-f20d136d02b72175cd9891df563f8d62R267-R287 There are more deps added on this branch than on master.

[lgeiger]

However, there are still differences between the PR on master and the PR on this branch

This is intentional since during the merge of #37356 those dependencies where added to resolve internal build errors mentioned in #37356 (comment)

See 0e86d51 for the diff of the merged PR into master.

[mihaimaruseac]

Makes sense. Thank you