[build] Update IMEX version

[Menooker]

The IMEX upstream has rebased its LLVM version to a recent commit (early August). In this PR, we update the LLVM commit for IMEX build and the IMEX commit as well. The IMEX commit points to the newest main-branch commit of the repo https://github.com/intel/mlir-extensions
For now, We don't need self-hosted IMEX fork. We still need a patch on IMEX at Menooker/mlir-extensions#1 from @dchigarev . Currently still using a forked IMEX, but the patch seems a bug fix and might be able to be pushed upstream.

[Menooker]

The IMEX has originally a single function like:

TensorDescType get(ArrayRef<int64_t> shape, Type type, int aaa = 1, bool bbb = true);

And they added another similar overload function:

TensorDescType get(ArrayRef<int64_t> shape, Type type, int aaa = 1, SomeOtherType bbb = {});

This makes the default arguments unusable, xegpu::TensorDescType::get(shape, elementType) will end up with ambitious overloading error.

[Menooker]

@leshikus @WangJialei-A It seems that updating llvm-imex commit hash will not trigger re-build of LLVM in CI? Would you please fix that?

[dchigarev]

@leshikus @WangJialei-A It seems that updating llvm-imex commit hash will not trigger re-build of LLVM in CI? Would you please fix that?

I think you just simply need to add cmake/llvm-version-imex.txt here:

  push:
    paths:
      - cmake/llvm-version.txt
+     - cmake/llvm-version-imex.txt
      - .github/workflows/build-llvm.yml

[Menooker]

@leshikus @WangJialei-A It seems that updating llvm-imex commit hash will not trigger re-build of LLVM in CI? Would you please fix that?

I think you just simply need to add cmake/llvm-version-imex.txt here:

  push:
    paths:
      - cmake/llvm-version.txt
+     - cmake/llvm-version-imex.txt
      - .github/workflows/build-llvm.yml

Thanks for your suggestions, @dchigarev !

[Menooker]

Current status: Test failed on f16_matmul_32x32.mlir
On my local test env, I got out of resource when calling clFinish to wait for the GPU queue, and this might indicate that GPU hangs after running the kernel. In the CI env, the execution succeeds with wrong results (all zero).

[dchigarev]

Current status: Test failed on f16_matmul_32x32.mlir On my local test env, I got out of resource when calling clFinish to wait for the GPU queue, and this might indicate that GPU hangs after running the kernel. In the CI env, the execution succeeds with wrong results (all zero).

There was a patch in the dev branch of IMEX that fixes that (Menooker/mlir-extensions#1), it seems that it wasn't propagated to the new main branch

[Menooker]

Current status: Test failed on f16_matmul_32x32.mlir On my local test env, I got out of resource when calling clFinish to wait for the GPU queue, and this might indicate that GPU hangs after running the kernel. In the CI env, the execution succeeds with wrong results (all zero).

There was a patch in the dev branch of IMEX that fixes that (Menooker/mlir-extensions#1), it seems that it wasn't propagated to the new main branch

Many thanks, @dchigarev ! I have rebased that to the current HEAD of main branch of IMEX and it works. Are you interested in submitting your PR to IMEX upstream?

[dchigarev]

Are you interested in submitting your PR to IMEX upstream?

Yes, will do