Remove strict stride multiple enforcement for expanded tensor support

[justinrosner]

Motivation

This PR removes a strict check in MIGraphXToTosa that was preventing MIGraphX from compiling certain concat kernels. This implements: https://amd-hub.atlassian.net/browse/AIROCMLIR-472

Technical Details

The AsUnderlyingShapeConverter was incorrectly rejecting valid non-contiguous output buffer strides. Specifically, it required that memory layout dimensions be integer multiples of the logical tensor dimensions. This failed for cases like <4x5x24xf16, 288x24x1> where the logical shape is 4x5x24 but the memory layout is 4x12x24. This pattern could occur when MIGraphX concatenates tensors (e.g., a 4x5x24 and 4x7x24 into a 4x12x24 buffer), where each kernel writes to a different slice of the final buffer.

Test Plan

• PR CI

Test Result

• PR CI

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

This PR removes an overly restrictive stride validation check that was preventing MIGraphX from compiling valid concat kernels. The check required memory layout dimensions to be integer multiples of logical tensor dimensions, which failed for cases where a kernel writes to a non-contiguous slice of a concatenated output buffer.

Changes:

• Removed the modulo check in AsUnderlyingShapeConverter that enforced integer-multiple relationship between memory layout and logical dimensions
• Added a positive test case demonstrating 4x5x24 logical tensor with 4x12x24 memory layout (non-multiple expansion)
• Removed the negative test case that was validating the now-removed error condition

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp	Removes 4-line modulo validation check that was rejecting valid non-contiguous output buffer strides
mlir/test/Conversion/MIGraphXToTosa/migraphx-to-tosa-non-contiguous-strides.mlir	Adds positive test for non-multiple expansion (4x5x24→4x12x24) and removes obsolete negative test case

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[pfultz2]

I still get a similar error:

Error: 'rock.expand_strides' op output dimension 12 is not a multiple of input dimension 7

Which might come from here:

rocMLIR/mlir/lib/Dialect/Rock/IR/RockDialect.cpp

Line 899 in 4462152

return emitOpError("output dimension ")

[justinrosner]

I still get a similar error:

Error: 'rock.expand_strides' op output dimension 12 is not a multiple of input dimension 7

Which might come from here:

rocMLIR/mlir/lib/Dialect/Rock/IR/RockDialect.cpp

Line 899 in 4462152

return emitOpError("output dimension ")

Removed that check as well. I'm assuming this is from running the second kernel?

[justinrosner]

Passing PR CI: https://ml-ci-internal.amd.com/job/MLIR/job/mlir/job/PR-2226/5/pipeline-overview