[aalabsi-amd]

JIRA ID : ALMIOPEN-3097

Motivation

Under ASAN, miopen_gtest exits 1 because LeakSanitizer reports leaks at exit from the find-2.0 batch-norm +
activation tests. Each gpu() verifier calls miopenFindSolutions, runs the returned solutions, and never calls
miopenDestroySolution. Every direct leak points at miopenFindSolutions (src/api/find2_0_commons.cpp:278,
new miopen::Solution{...}). The assertions pass; the leak is in the tests, not the library.

Because of this, #13107 excluded both files from ASAN builds (SKIP_TESTS in test/gtest/CMakeLists.txt), which
drops all 558 cases of */GPU_na_train_find2_* and */GPU_na_inference_find2_* from every ASAN build.

Technical Details

• test/gtest/na_train_find2.cpp (4 verifiers) and test/gtest/na_inference_find2.cpp (1 verifier): a FindSolutions() helper wraps every solution returned by miopenFindSolutions in MIOPEN_MANAGE_PTR(miopenSolution_t, miopenDestroySolution), so each is destroyed when gpu() returns, including on an early fatal assertion. solutions_found is now initialized (it was read uninitialized if find failed).
• test/gtest/CMakeLists.txt: remove the ALMIOPEN-3097 SKIP_TESTS entries, so ASAN builds compile and run both
  files again.

find_2_conv.cpp stays in the ASAN skip list; its leak has not been investigated and is out of scope here.

Test Plan

Local ASAN build of this branch against the rockrel Multi-Arch Release ASAN (nightly) #139 artifacts
(run 40425965688, TheRock e73fd19), -fsanitize=address -shared-libasan, THEROCK_SANITIZER=ASAN,
GPU_TARGETS=gfx942:xnack+; test environment as TheRock CI configure_asan_env.py
(ASAN_OPTIONS=detect_odr_violation=0:quarantine_size_mb=600:verify_asan_link_order=0:symbolize=1, HSA_XNACK=1,
leak detection on):

miopen_gtest --gtest_filter='*/GPU_na_train_find2_*:*/GPU_na_inference_find2_*'

Test Result

	Before (develop, ASAN, PR #13107 canary run)	After (this branch)
na_*find2 cases run under ASAN	none in ASAN builds (excluded via SKIP_TESTS); the Smoke cases passed in the #13107 canary run, but the process failed at exit	558 passed, 0 failed, 0 skipped (12 suites, Smoke + Full)
LeakSanitizer at exit	47556 byte(s) leaked in 452 allocation(s): direct leaks from miopenFindSolutions <- na_train_find2.cpp:534/710, na_inference_find2.cpp:164	no report, process exit code 0
AddressSanitizer errors	0	0

Run: Slurm job 68262003 on banff-cyxtera-s83-2 (MI300X, gfx942:sramecc+:xnack+, amdgpu.noretry=-1), 73 min;
LSan suppressions file loaded but unused (Suppressions used never printed).

Second run:

• Slurm job 68298368 on banff-cyxtera-s70-4 (gfx942:sramecc+:xnack+): 558 passed, 0 failed, exit code 0, and no
  LeakSanitizer or AddressSanitizer output

Submission Checklist

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

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[JonathanLichtnerAMD]

LGTM. Each of the 5 verifiers now destroys the solution returned by miopenFindSolutions, and nothing between find and the run loop can return early. Optional nit: an RAII deleter would keep this leak-proof if someone later adds an ASSERT_* before the loop. No suggested fixes.

Co-authored by Claude (AI).

[aalabsi-amd]

LGTM. Each of the 5 verifiers now destroys the solution returned by miopenFindSolutions, and nothing between find and the run loop can return early. Optional nit: an RAII deleter would keep this leak-proof if someone later adds an ASSERT_* before the loop. No suggested fixes.

Co-authored by Claude (AI).

Done in 0dc7dfc: solutions are now owned by MIOPEN_MANAGE_PTR(miopenSolution_t, miopenDestroySolution) via a
FindSolutions() helper that wraps them right after miopenFindSolutions, so an early ASSERT_*/return can't leak
them. Also initialized solutions_found. Re-ran under ASAN: 558/558 passed, no LSan report.