[AIROCMLIR-1010] Separate --debug flag into --debug and --debug-quick-tune-data to fix ~50GB .debug file

[erizheng-amd]

Motivation

This PR backports PR https://github.com/ROCm/rocmlirTriton/pull/297 in rocmlirTriton to rocMLIR.

After the PR #2407 change in timing method, the .debug file generated with tuningRunner.py with --debug flag now has a huge array of measurements for every tested config. Previously, we only had something like {'total_cpu_time': 40.19872, 'iterations': 31} for small kernels, and this change increased the .debug file size from ~500MB to ~50GB.

Technical Details

This PR fixes this issue by separate --debug flag into --debug and --debug-quick-tune-data as quickTuningGen.py requires the .debug file but not using the detail measurements at all. Now with --debug-quick-tune-data, we still generate a .debug file but we drop the MeasurementsMs column entirely. We only print that if we want to do some actual debug work and require the MeasurementsMs data with --debug flag.

Test Plan

-- verify the QuickTuningGen works fine without MeasurementsMs
-- verify the .debug file size shrinks
-- ninja check-rocmlir
-- PR CI

Test Result

-- I have tried QuickTuningGen works fine without MeasurementsMs on a .debug file with only 2 problem configs with exhaustive tuning
-- For the 2 problem configs with exhaustive tuning, the .debug file size drops from 341MB to 3MB
-- ninja check-rocmlir passed
-- PR CI running

Submission Checklist

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

[Copilot]

Pull request overview

Backports a tuningRunner debug-output change from rocmlirTriton to rocMLIR to prevent .debug TSV files from exploding in size after the timing-method change (by making per-iteration measurement arrays opt-in).

Changes:

• Adds a new --debug-quick-tune-data flag to emit .debug files without the heavy MeasurementsMs per-iteration arrays.
• Keeps --debug as the “full debug” mode that includes detailed per-iteration measurement arrays.
• Adds a lit test that checks .debug output is produced and omits MeasurementsMs when using --debug-quick-tune-data.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
mlir/utils/performance/tuningRunner.py	Adds --debug-quick-tune-data and enables .debug file generation for either debug mode while only emitting MeasurementsMs for full --debug.
mlir/test/perf-scripts/runtime/tuningRunner-gemm.py	Adds a runtime lit test validating the new quick-debug mode’s .debug schema (no MeasurementsMs).

---

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2414      +/-   ##
===========================================
+ Coverage    82.57%   82.68%   +0.11%     
===========================================
  Files          120      120              
  Lines        42852    42828      -24     
  Branches      7110     7106       -4     
===========================================
+ Hits         35381    35410      +29     
+ Misses        4815     4791      -24     
+ Partials      2656     2627      -29     

Flag	Coverage Δ
gfx120x	82.53% <ø> (+0.01%)	⬆️
gfx950	82.49% <ø> (+0.14%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 33 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.