Autonomous engineering built, validated ROCm Windows runtime fix on Strix Halo

ACE Validation of a Source-Level Fix for ROCm Issue #12571: Windows PAL/SDMA Stream-Ordering Fault on gfx1151

Description

ACE K2, the governed autonomous engineering system developed at NH Applied, independently built and physically validated a source-level correction for ROCm/rocm-systems issue #12571 on Windows 11 using an AMD Radeon 8060S (gfx1151).

The issue concerns a PAL/SDMA stream-ordering defect in which an SDMA copy issued after hipEventRecord can race earlier compute on the same stream, potentially returning stale data. ACE recovered the matching source/runtime provenance, built the Windows HIP runtime, generated a distinct durable-compute-event source modification, and performed a controlled physical A/B validation.

A forced-pristine same-toolchain build reproduced the targeted failure as STALE 50/50, while a forced-recompiled patched build returned clean 50/50 results across all six test cells with SDMA enabled. The evidence was then submitted to an independent deterministic evaluator, which returned ACCEPT / PATCH_VALIDATED_SDMA_ENABLED.

The ACE implementation is not source-identical to AMD PR #12952. Both approaches preserve compute-event state across the relevant flush, but ACE uses the retained lastComputeEvent_ directly for the compute-idle fence, while the AMD implementation preserves the existing live-event path and uses the retained event as a fallback.

This report documents the source/runtime provenance, causal validation method, physical gfx1151 results, implementation comparison, attribution timeline, cryptographic hashes, limitations, and relationship to AMD's subsequent matched-stack validation on gfx1201.

Claim boundary: the result is bounded to the tested Windows gfx1151 environment, matched source/runtime revision, and governed #12571 reproducer matrix. It is not a claim of universal correctness across ROCm platforms or releases.

Files

ACE_ROCm_12571_White_Paper_Updated_REV3.pdf

Files

(238.5 kB)

Additional details

Related works

- Is supplement to

- Technical note: 10.5281/zenodo.22869831 (DOI)

- References

- Other: https://github.com/ROCm/rocm-systems/issues/12571 (URL)

- Other: https://github.com/ROCm/rocm-systems/pull/12952 (URL)

- Other: https://github.com/ROCm/rocm-systems/issues/12571#issuecomment-6062023767 (URL)