Support default configs for BATCH_MATMUL_* in MaterializeEncoding pass

[pzread]

In #14705 we added BATCH_MATMUL_* encoding users and handled the materialization in IREE codegen path by providing a custom MaterializeEncodingFn. However MaterializeEncoding pass also can be called standalone, this change updates the default chooseEncodingInfo function to support new encoding users.

This is done by moving chooseEncodingInfoForMatmul under LinalgExt to reuse the logic of deciding materialization based on matmul encoding.

The problem was found when adding tests for MaterializeEncoding with BATCH_MATMUL_* encoding users, since MaterializeEncoding pass is called standalone by --iree-linalg-ext-materialize-encoding

#14431

[pzread]

I was trying to add this code in to LinalgExt/Utils/Utils.h, but LinalgExt/Utils/Utils.h is used in LinalgExt/IR/LinalgExtOps.cpp. And we need to use linalg_ext ops, which results in cyclic library dependencies.

I'm not sure what the better way is to solve this. Here I simply creates a separate utils library.

If I put this code into LinalgExt/Utils/Utils.h and add IREELinalgExtDialect to the dependencies of IREELinalgExtUtils, I'll get the error:

CMake Error: The inter-target dependency graph contains the following strongly connected component (cycle):
  "obj.IREELinalgExtDialect" of type OBJECT_LIBRARY
  depends on "IREELinalgExtUtils" (strong)
  "IREELinalgExtDialect" of type STATIC_LIBRARY
  depends on "obj.IREELinalgExtDialect" (strong)
  depends on "IREELinalgExtUtils" (strong)
  "obj.IREELinalgExtUtils" of type OBJECT_LIBRARY
  depends on "IREELinalgExtDialect" (strong)
  "IREELinalgExtUtils" of type STATIC_LIBRARY
  depends on "IREELinalgExtDialect" (strong)
  depends on "obj.IREELinalgExtUtils" (strong)
  
At least one of these targets is not a STATIC_LIBRARY.  Cyclic dependencies are allowed only among static libraries.
CMake Generate step failed.  Build files cannot be regenerated correctly.

[MaheshRavishankar]

WHy does Utils needs IR ?

[pzread]

The new code added here needs IREE::LinalgExt::EncodingUser, which is defined in LinalgExt/IR/LinalgExtOps.h. So if we add the code into LinalgExt/Utils/Utils.h instead, IIUC IREELinalgExtUtils will have a dependency to IREELinalgExtDialect.

On the other hand, LinalgExtOps.cpp is using getDimValue and computeInterchangeFromDimPos defined in Utils.cpp, so it currently has a dependecy to IREELinalgExtUtils

https://github.com/openxla/iree/blob/5142828f9d6fb761c98900e45f4e398c6da9af6d/llvm-external-projects/iree-dialects/lib/Dialect/LinalgExt/IR/CMakeLists.txt#L12-L13

[pzread]

To clarify, here I'm trying to move chooseEncodingInfoForMatmul from IREE codegen into LinalgExt so we can reuse the same logic to choose tiling and permutations for materialization based on LinExt::EncodingUser.

[hanhanW]

I've seen similar issues in MLIR repo. I think the solution is that *IR should not depends on *Util. The common methods are moved to *IR.h in upstream, and we can probably follow the convention. Dependency issue looks fine to me, and we can refactor it later.

[pzread]

@MaheshRavishankar Kindly ping : ) Do you have further comments on this PR?

[pzread]

@RoboTux FYI

[MaheshRavishankar]

Looks good to me! Sorry I didnt know you were waiting for me to approve. I saw Hanhan approve so I just left it at that.