[AIROCMLIR-1030] Port recent enhancements in tuning scripts from rocMLIR to rocmlirTriton

[erizheng-amd]

Motivation

This PR examines the new changes on tuning scripts from rocMLIR side since the last porting in PR #192, this includes tuningRunner.py, perfRunner.py, and quickTuningGen.py. This PR ports the new changes and also add new tests while keeping the rocmlirTriton specific changes.

Technical Details

This PR includes several major changes on the rocMLIR side:

• [AIROCMLIR-673] Canonicalize tuningRunner output rocMLIR#2350 introduces the canonicalization in tuningRunner and perfRunner. This was partially done with [AIROCMLIR-1022] Normalize configs in tuningRunner #303. Now we are having a complete backport
• [AIROCMLIR-217] Replace MITuna with tuningRunner in CI rocMLIR#2310 introduces a NumaNodeLock CPU verification lock
• [AIROCMLIR-71] Add gemm+gemm and conv+gemm support to quickTuningGen.py rocMLIR#2262 introduces GemmGemm and ConvGemm support in quickTuningGen
• Resolve invalid escape sequence in perf utils rocMLIR#2308 removes the warnings in pytest

Also includes the two new pytests in MLIR test_tuningRunner.py and test_perfRunner.py, rewired and refactored them into LIT test:

Original class	New file	Tests
TestParseTuningDbLine, TestReadTuningDb, TestParseDataTypes, TestLayoutHelpers, TestGetNanoseconds, TestGetProfilerOutputPath, TestGetMetricArgsForRocprof, TestGetMiliseconds	perfRunner-test.py	27
TestFormatError, TestGetConfigClass, TestParseArguments, TestNumaTopologyParseCpuList	tuningRunner-arg-validation.py	16
TestGetStateFilepath, TestTuningState, TestTuningStateFile, TestTunedConfigsCache, TestDebugFileWriter, TestGetGitCommitHash	tuningRunner-state.py	21
TestCanonicalizeTestVector	tuningRunner-canonicalization.py	9
TestNumaNodeLock	tuningRunner-numa-lock.py	6
TestFindBestPerfconfig, TestTuneConfig	tuningRunner-driver-results.py	3

Changes are intentionally left out:

• [AIROCMLIR-80] Utilize amd_arch_db python bindings where applicable rocMLIR#2352 is skipped because the amd_arch_db is very different between the two repos.
• [AIROCMLIR-426] Translate rocMLIR conv layouts to MIOpen or skip rocMLIR#2422 already gets a PR for backporting it [AIROCMLIR-1078] Port MIOpen conv layout translation to rocmlirTriton #359

Divergence still presenting due to specific rocmlirTriton work:

• --verify-mode is cpu only on rocmlirTriton introduced by Delete GPU verification (New) #94
• Tooling difference using rocm-run introduced by [AIROCMLIR-819] perf scripts: route mlir-runner through rocm-run wrapper #191
• Difference flag usage in rocmlirTriton

Test Plan

• ninja check-rocmlir
• PR CI passed
• nightly CI running
• Tuning on a few configs and make sure it is working

Test Result

• ninja check-rocmlir passes
• PR CI and nightly CI running
• Tuning and perfRunner works

Submission Checklist

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

[rocmlir-pr-reviewer]

⚠️ This PR modifies the Claude-review CI security perimeter

The following files in this PR control whether and how the
claude-review workflow protects its LLM Gateway secrets at
runtime (see .github/workflows/CLAUDE_AUTO_REVIEW.md):

• .github/workflows/ci.yml

Before applying the claude-review label on this PR, please:

1. Audit the diff in these paths line-by-line. A malicious or
   accidental change could disable the --allowedTools
   restriction, the overlay step, the sanitizer, or the
   review/post job split -- any of which would expose the
   secrets in env to the PR-modified workflow.
2. If the changes are legitimate and you want a Claude review,
   do NOT apply the claude-review label. Instead, run via
   Actions → Claude Auto Review → Run workflow and enter
   this PR's number. The dispatch path runs from the trusted,
   code-owner-approved version of the workflow on
   develop, so a malicious PR-side
   modification cannot affect the run.

(This banner is automated. The modifies-ci-paths label
will be removed automatically if a future push removes the
perimeter modifications. The Layer-3 in-workflow guard will
additionally fail the claude-review label-triggered run
on this PR if the label is applied while perimeter changes
are present.)

[Copilot]

Pull request overview

This PR ports recent rocMLIR performance/tuning script enhancements into rocmlirTriton, including more consistent config canonicalization, improved verification behavior (timeouts + NUMA-aware locking), expanded quick-tuning generation for additional ops, and adds CI-ready Python tests for these scripts.

Changes:

• Add NUMA-node reader/writer locking to reduce CPU-verification interference during parallel tuning, and introduce a configurable verification timeout.
• Canonicalize and validate configs more consistently across tuningRunner.py and perfRunner.py, including stricter parsing/canonicalization in DB/config loaders.
• Add new pytest-based unit tests (GPU-less via HIP mocking) and wire them into GitHub Actions CI; update CLI flag spellings and related docs/tests.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
pip_requirements.txt	Adds pytest dependency for new test suite.
mlir/utils/performance/tuningRunner.py	Ports tuning enhancements: verification timeout, NUMA locking, canonicalization/state handling, debug schema checks, CLI tweaks.
mlir/utils/performance/perfRunner.py	Ports canonicalization + DB parsing improvements; updates CLI flags (with backward-compatible aliases).
mlir/utils/performance/perfCommonUtils.py	Minor regex literal fix (raw string).
mlir/utils/performance/analysis/quickTuningGen.py	Adds GemmGemm/ConvGemm support and kernel-type mapping updates.
mlir/utils/jenkins/Jenkinsfile	Updates perfRunner invocation flags to new dashed forms.
mlir/test/perf-scripts/runtime/perfRunner-help.py	Updates help-output checks for new dashed flags.
mlir/test/perf-scripts/attention-tuning-db-compat.py	Updates read_tuning_db call signature to include config class and fallback params.
.github/workflows/ci.yml	Adds a new containerized pytest job for performance scripts.
mlir/utils/performance/tests/test_tuningRunner.py	New unit tests for tuningRunner (mock HIP/GPU).
mlir/utils/performance/tests/test_perfRunner.py	New unit tests for perfRunner (mock HIP/GPU).
mlir/utils/performance/tests/mock_hip.py	New HIP + amd_arch_db mock module for CI environments without ROCm.
mlir/utils/performance/tests/init.py	New test package marker and brief description.

---

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

[erizheng-amd]

Note for reviewers: I will rebase onto the develop branch after #306 merged.

[Copilot]

Pull request overview

Copilot reviewed 14 out of 14 changed files in this pull request and generated no new comments.

Suppressed comments (1)

mlir/utils/performance/tuningRunner.py:1112

• DebugFileWriter currently rejects appending when the column order differs from the existing header, even if the schemas are identical. This can happen when the existing file was produced by a different pandas/Python version (or any code path that emits the same columns in a different order), and it prevents safe resumption even though the data could be re-ordered to match the existing header.

Instead of requiring new_columns == existing_columns, consider treating it as a mismatch only when the column sets differ, and when they match, reindex the DataFrame to existing_columns before writing so appended rows align with the on-disk header.

        df = pd.DataFrame(result.entries)
        new_columns = list(df.columns)
        if self._existing_columns is not None and new_columns != self._existing_columns:
            raise ValueError(
                f"Debug file '{self.filepath}' has a schema that does not match the current "

[bogdan-petkovic]

LGTM now