feat(hipsr): add compute and compute_yield ops

[zpye]

Summary

Adds hipsr.compute and hipsr.compute_yield. hipsr.compute is a DPS op whose isolated single-block region groups the operations that implement one high-level computation, so an ONNX op needing several tensor or shape operations stays a single hipsr op.

Ops only: nothing emits one yet, so no existing behavior changes.

Related issue or design

None.

Why

hipsr.placeholder accepts only block arguments and constant or placeholder results as inputs. Without a grouping op, a multi-operation lowering leaves intermediates in the enclosing block, and the placeholder of a later op would have to take one as an input, which its verifier rejects.

The body is IsolatedFromAbove, so the context, inputs, and inits enter it as entry-block arguments in that order.

There is no hand-written verifier. SizedRegion<1>, SingleBlock, SingleBlockImplicitTerminator, RegionBranchOpInterface, DestinationStyleOpInterface, and IsolatedFromAbove already produce every check, and the tests assert the message each one emits.

What

• hipsr.compute: !hipsr.context first, then variadic ins and outs inits, variadic tensor results, one $body region.
• Inits are Hipsr_TensorOrDeviceMemRef, so the op survives bufferization to device memrefs. Memref inits mean zero tensor results, which the optional result-type group allows.
• Inputs are Variadic<AnyType>, so the body can also take values like an index extent.
• hipsr.compute_yield mirrors hipsr.pool_domain_yield.
• No shape region: output shapes stay the hipsr.placeholder job, and -hipsr-populate-shape-region skips this op.

Test plan

New test/lit/Dialect/Hipsr/compute.mlir: three positive cases (tensor init, multiple inits with a non-tensor input, device-memref inits with no results) and ten negative cases, one per verification mechanism. Expected diagnostics were copied from real hip-mlir-opt output.

ctest -R MorphizenMLIRLitTests
366 tests: 361 passed, 3 unsupported, 2 failed

Both failures are pre-existing on main and unrelated: test_gather_block_quantized.mlir (--convert-onnx-to-hip) and test_compress_model.mlir (--hipdnn-pipeline). pre-commit run --all-files passes.

Notes for reviewers

• No bufferization model, so a hipsr.compute reaching one-shot bufferize would fail. Unreachable today because -hipsr-bufferize does not exist.
• No custom builder. TableGen's (resultTypes, ctx, inputs, inits) builder leaves the body blockless, so the first conversion pattern must create the entry block itself; @empty_body and @missing_context_argument pin the diagnostics it sees otherwise.
• Nothing computes output shapes for compute-backed ONNX ops yet. A follow-up must decide where reshape-like shape math lives.

Checklist

• The change is focused, or links a design/series explaining its scope.
• Relevant tests were added or updated and the results are documented.
• User-facing or design documentation was updated when needed: no in-repo doc covers hipsr, so the ODS description blocks carry the contract.
• Substantial AI assistance is disclosed, and I reviewed and understand the result.

AI assistance: drafted with Cursor. Diagnostics were captured from real tool output and the test numbers come from local build and CTest runs. The commit carries Made-with and Co-authored-by trailers.

[github-actions]

Thanks for opening a PR!

This project follows LLVM's incremental-development and AI-tool-use
guidance. See CONTRIBUTING.md
for the project workflow.

Before requesting review, please check that:

1. The change is focused. Substantial work links the relevant issue
   or design discussion.
2. The PR documents relevant test results and updates affected
   documentation.
3. If AI tools provided substantial assistance, the description
   explains what was assisted and how it was validated, and commit
   trailers identify the tool. The contributor has reviewed and
   understands the result.

Reviewers are assigned through
CODEOWNERS where ownership
is configured.

[zpye]

Superseded by #615, which carries the same commit from a branch in this repository so that CI runs with full repository access. Closing this one.