[bazel] Add a bazel flag to enable building MLIR with CUDA support

[apaszke]

This makes it possible to specify --@llvm-project//mlir:enable_cuda=true on the bazel command line and get a build that includes NVIDIA GPU support in MLIR.

[chsigg]

Does this flag correspond to one of the CMake flags? Should it have a similar name, or be referenced in a comment?

I assume this flag will not be tested in CI. Have you tested it locally? Does it trigger a dependency on the CUDA toolkit?

[apaszke]

Does this flag correspond to one of the CMake flags?

It corresponds to this part of the CMake config.

Have you tested it locally? Does it trigger a dependency on the CUDA toolkit?

Yes, I have tested the flag locally. It doesn't actually add a dependency on the CUDA toolkit, but only enables a number of MLIR passes that are otherwise #ifdefed away by MLIR_ENABLE_CUDA_CONVERSIONS.

[chsigg]

Thanks Adam.

Ok, it seems that if_cuda_available doesn't trigger MLIR_ENABLE_NVPTXCOMPILER (yet), which would require a CUDA toolkit to be available.

What about converting to NVVM binary though, doesn't this call expect to find CUDA tools?

[apaszke]

It does, but it's a run-time dependency. Which shouldn't affect the build significantly and after all... I'd expect that if someone enables this flag then it's on them to provide all the tools necessary.

[chsigg]

I'd expect that if someone enables this flag then it's on them to provide all the tools necessary.

I would expect that people will use a flag if it's there. I'm not saying we shouldn't land this as is and in the current state it looks clean. I was just trying to understand what is already supported. E.g., it doesn't trigger any new tests that would require a CUDA toolkit to be found.

I would expect though that proper CUDA support with bazel would pull the toolkit as an external dependency. But that's clearly a separate next step.