Fix element type mismatch in attention preSoftmax fusion

[justinrosner]

Motivation

This PR fixes a crash that MIGraphX was seeing when compiling an attention kernel with fusion: https://amd-hub.atlassian.net/browse/AIROCMLIR-438

Technical Details

When lowering gridwise_attention_accel ops with preSoftmax fusion, the gemm0 output buffer element type was unconditionally set to elemTypeV (the values input element type). This caused a type mismatch when the preSoftmax body's linalg.generic expected a different element type for it's gemm0 based input (e.g., when the linalg.generic was truncating/extending).

Test Plan

• PR CI
• Original kernel from MIGraphX is passing (and turned into a LIT test)

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 fixes a crash in MIGraphX when compiling attention kernels with preSoftmax fusion. The issue occurred when lowering gridwise_attention_accel operations where the gemm0 output buffer element type was incorrectly set to the values input element type (elemTypeV), causing a type mismatch when the preSoftmax body's linalg.generic operation expected a different element type (e.g., when truncating or extending).

Changes:

• Modified element type determination logic to walk the preSoftmax body and extract the correct type from the first linalg.generic operation's gemm0-based input
• Added a comprehensive LIT test that reproduces the original MIGraphX failure scenario with type conversions in the preSoftmax fusion body

Reviewed changes

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

File	Description
mlir/lib/Dialect/Rock/Transforms/GridwiseGemmToBlockwise.cpp	Added logic to walk preSoftmax body and determine gemmOutElemType from the first generic's gemm0-based input, and fusionOutElemType from the last generic's output, fixing the element type mismatch
mlir/test/Dialect/Rock/gridwise-gemm-linalg-failure.mlir	New test file verifying correct handling of attention operations with preSoftmax fusion that performs f16 to f32 extension, ensuring the lowering produces correct buffer types

---

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

[pabloantoniom]

I second both of Daniel's comments