[MLIR][TORCH] Add TorchToTosa lowering for aten.where.self op

[AmosLewis]

Find this issue when lowering gpt2 to tosa dialect in SHARK with python nod-ai/SHARK#338

If you are boot camp torch-mlir:
Here is useful comments to understand the debug step: #961 (comment)

Here is the link to useful torch_mlir_debug_command.txt that will be helpful for boot camp and refer later in development of torch-mlir.

After you read all the following comments, for ONNX work, refer to this dc9ea08

[AmosLewis]

https://gist.github.com/AmosLewis/fa8662eef03c3379bda6a3974036abd1 for test this op.
nod-ai/SHARK#338 to use this op.

[AmosLewis]

In the test, it converts torch to tosa successfully. But why after that, it converts tosa to linalg? The tosa to linalg conversion arise the bug. @sjarus @eric-k256

> (mlir_venv) nod% python -m e2e_testing.main -f 'ElementwiseAtenWhereSelf' --config=tosa -v 
> Compiling ElementwiseAtenWhereSelfModule_basic...
> FAIL - "ElementwiseAtenWhereSelfModule_basic"
> 
> Unexpected outcome summary:
> 
> ****** Failed tests - 1 tests
>     FAIL - "ElementwiseAtenWhereSelfModule_basic"
>         Compilation error: Traceback (most recent call last):
>           File "/home/chi/src/ubuntu20/shark/torch-mlir/build/tools/torch-mlir/python_packages/torch_mlir/torch_mlir_e2e_test/framework.py", line 290, in compile_and_run_test
>             compiled = config.compile(test.program_factory())
>           File "/home/chi/src/ubuntu20/shark/torch-mlir/build/tools/torch-mlir/python_packages/torch_mlir/torch_mlir_e2e_test/configs/tosa_backend.py", line 35, in compile
>             return self.backend.compile(module)
>           File "/home/chi/src/ubuntu20/shark/torch-mlir/build/tools/torch-mlir/python_packages/torch_mlir/torch_mlir_e2e_test/tosa_backends/linalg_on_tensors.py", line 57, in compile
>             run_pipeline_with_repro_report(
>           File "/home/chi/src/ubuntu20/shark/torch-mlir/build/tools/torch-mlir/python_packages/torch_mlir/torch_mlir/compiler_utils.py", line 73, in run_pipeline_with_repro_report
>             raise TorchMlirCompilerError(trimmed_message) from None
>         torch_mlir.compiler_utils.TorchMlirCompilerError: Lowering TOSA to Linalg-on-Tensors failed with the following diagnostics:
>         error: 'linalg.generic' op expected indexing_map #2 to have 4 dim(s) to match the number of loops
>         note: see current operation: 
>         %2 = "linalg.generic"(%1, %arg1, %arg2, %0) ({
>         ^bb0(%arg3: i1, %arg4: f32, %arg5: f32, %arg6: f32):
>         %3 = "arith.select"(%arg3, %arg4, %arg5) : (i1, f32, f32) -> f32
>         "linalg.yield"(%3) : (f32) -> ()
>         }) {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<() -> ()>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"], operand_segment_sizes = array<i32: 3, 1>} : (tensor<1x5x5xi1>, tensor<1x12x5x5xf32>, tensor<f32>, tensor<1x12x5x5xf32>) -> tensor<1x12x5x5xf32>

[ramiro050]

I think the code that generates the indexing maps for the linalg.generic op is not handling zero-rank tensors correctly.

[ramiro050]

I think this line should be using the result rank variable rank rather than the operand type.getRank():

https://github.com/llvm/llvm-project/blob/86bf43d2ab1334af1ca7cb10d407b5afe19fc65f/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp#L617

[AmosLewis]

Yes. After replacing with rank, everything looks good.

    indexingMaps.push_back(AffineMap::get(
        /*dimCount=*/rank, /*symbolCount=*/0, affineExprs,
        rewriter.getContext()));

[AmosLewis]

llvm to use llvm/llvm-project@main...AmosLewis:llvm-project:tosa-to-linalg

llvm patch waiting for review and merge : https://reviews.llvm.org/D135343

[AmosLewis]

Need this merged first.
build: update llvm tag to 438e5918 #1475
Or this
build: update llvm tag to d325d2b #1483

[AmosLewis]

Need this merged first. build: update llvm tag to 438e5918 #1475 Or this build: update llvm tag to d325d2b #1483

@ramiro050 Could you help uplift llvm to d325d2b?

[AmosLewis]

Finish rebase and uplift the llvm to #1502.
But got a new build-test out-of-tree bug which is not from my patch.

  pip installing Pytorch..
  ERROR: Invalid requirement: '/main_checkout/torch-mlir/build_tools/../build_tools/python_deploy/wheelhouse/*'
  Hint: It looks like a path. File '/main_checkout/torch-mlir/build_tools/../build_tools/python_deploy/wheelhouse/*' does not exist.
  CMake Error at python/CMakeLists.txt:36 (message):
    Failed to run `build_libtorch.sh`

[ramiro050]

Finish rebase and uplift the llvm to #1502. But got a new build-test out-of-tree bug which is not from my patch.

  pip installing Pytorch..
  ERROR: Invalid requirement: '/main_checkout/torch-mlir/build_tools/../build_tools/python_deploy/wheelhouse/*'
  Hint: It looks like a path. File '/main_checkout/torch-mlir/build_tools/../build_tools/python_deploy/wheelhouse/*' does not exist.
  CMake Error at python/CMakeLists.txt:36 (message):
    Failed to run `build_libtorch.sh`

Is this happening locally on your machine? The CI seems to be building fine.

[ashay]

Is this happening locally on your machine? The CI seems to be building fine.

It was happening in CI (because of a bogus cache entry) so I deleted that cache entry and restarted the build, which then passed. All is fine now, but I'll keep an eye out for more failures.

[ramiro050]

It was happening in CI (because of a bogus cache entry) so I deleted that cache entry and restarted the build, which then passed. All is fine now, but I'll keep an eye out for more failures.

Nice! Thanks for the quick fix!

[AmosLewis]

Here is the link to useful torch_mlir_debug_command.txt that will be helpful for boot camp and refer later in development of torch-mlir.