[ROCm] CMake and lit changes to enable AMDGPU support.

[whchung]

• Introduce GPUToROCDL to map thread/block/grid device functions to corresponding intrinsics and device functions on AMD GPU.
• Introduce GPUToROCm to drive MLIR->LLVM->HSA code object lowering process. Some passes are pretty similar with those in GPUToCUDA and may deserve to be merged.
• Introduce mlir-rocm-runner in mlir/tools.

Authors:

• Jack Chung (@whchung)
• Deven Desai (@deven-amd)

[googlebot]

All (the pull request submitter and all commit authors) CLAs are signed, but one or more commits were authored or co-authored by someone other than the pull request submitter.

We need to confirm that all authors are ok with their commits being contributed to this project. Please have them confirm that by leaving a comment that contains only @googlebot I consent. in this pull request.

Note to project maintainer: There may be cases where the author cannot leave a comment, or the comment is not properly detected as consent. In those cases, you can manually confirm consent of the commit author(s), and set the cla label to yes (if enabled on your project).

ℹ️ Googlers: Go here for more info.

[googlebot]

CLAs look good, thanks!

ℹ️ Googlers: Go here for more info.

[joker-eph]

Hi,

Thanks for the contribution! It is nice to see another GPU target :)

What is the status of this pull-request right now? Are you seeking early feedback or are you looking for a quick path to integrate this upstream?
I saw with a quick scan things like git clone -b exp-mlir-rocm-runner https://github.com/whchung/mlir added to the README which lead me to be unsure about the current state.

[whchung]

@joker-eph Thanks for your prompt reply.

All tests are passing on this PR, which means we can run MLIR core end-to-end on AMD ROCm platform.

$ cmake --build . --target check-mlir         
[0/1] Running the MLIR regression tests
Testing Time: 5.13s
  Expected Passes    : 223
  Unsupported Tests  : 4

4 unsupported tests are those on CUDA platform.

We are starting building internal tools based on work derived from this PR, so we'd like to understand the steps to get the PR revised and merged.

Granted there are quite a few duplication between GPUToCUDA and GPUToROCm at this moment, plus there are some hard-coded logic at this moment so I think some discussion would be surely needed.

I'll remove my change to README.md. They should stay downstream.

[joker-eph]

All tests are passing on this PR, which means we can run MLIR core end-to-end on AMD ROCm platform.

Fantastic!

Granted there are quite a few duplication between GPUToCUDA and GPUToROCm at this moment, plus there are some hard-coded logic at this moment so I think some discussion would be surely needed.

Right, I suspect we will want to integrate this piece by piece, in PRs as small as possible.
We will also look into avoiding as much duplication as possible and instead refactor and reuse the logic in the repo.

I believe it would be best to start with an RFC on the public mailing-list to discuss this and involve the folks actively working on GPU targets already (SPIR-V and Cuda).

Thanks!

[whchung]

@joker-eph Thanks. I'll revise this PR and break it into smaller patches, probably tablegen for ROCDL dialect to begin with. Also I'll send an RFC to the mailing list.

[whchung]

@joker-eph I've reduced this PR to a minimal size. Will file subsequent ones moving forward, plus submitting RFCs to the public mailing list.

[joker-eph]

Thanks!

From what I see in the diff now, these are mostly only build configuration changes left? I expect these to land with the actual runner in one of the later PRs.

[whchung]

@joker-eph Yeah, come to think of it there is really no meat in this PR now.

Let me keep this PR open for awhile, and use this PR as a placeholder for mlir-rocm-runner then.

[joker-eph]

Closing for now (cleanup), feel free to re-open as needed though!