[Issue] Windows: `rocBLAS full-suite tests crash (SEH divide-by-zero/access violation) and end with heap corruption (0xc0000374)`

[chiranjeevipattigidi]

Summary

rocBLAS client full-suite tests crash on Windows in Multi-Arch CI (PR #5072 run), showing repeated SEH exceptions in AXPY batched tests (divide-by-zero), then TRSV strided-batched failures (access violation + HIP invalid value), and finally the overall test process terminates with a heap-corruption style exit code.

Failing CI link

• https://github.com/ROCm/TheRock/actions/runs/25459048352/job/74751466996?pr=5072
• https://github.com/ROCm/TheRock/actions/runs/25459050326/job/74753445059?pr=5073
• https://github.com/ROCm/TheRock/actions/runs/25459050326/job/74753445030?pr=5073

Tests Failed

1. Multiple AXPY batched/strided-batched parameterized tests crash with:

• unknown file: error: SEH exception with code 0xc0000094 thrown in the test body.

Examples from the log:

• _ /axpy_batched_ex ... nightly_blas1_batched_with_alpha_..._400000_...
• _ /axpy_strided_batched_ex ... nightly_blas1_strided_batched_with_alpha_..._300001_...
• also some pre_checkin ... blas1_graph_check ... variants

(SEH 0xc0000094 = integer divide-by-zero on Windows.)

2. TRSV strided-batched shows access violations and a rocBLAS/HIP error sequence:

• unknown file: error: SEH exception with code 0xc0000005 thrown in the test body.
• .../rocblas/clients/common/gtest_helpers.cpp(217): error: Failed
• Received uncaught exception: rocblas_status_internal_error
• hipGetLastError at end of test function call: ...
• rocBLAS error from hip error code: 'hipErrorInvalidValue':1

(SEH 0xc0000005 = access violation on Windows.)

3. The overall CTest target fails with process exit code:

• rocblas-test_full_suite (Exit code 0xc0000374)

(0xc0000374 is commonly associated with heap corruption / corrupted heap detected on Windows.)

Environment

• Repo: ROCm/TheRock
• Workflow: "Multi-Arch CI"
• Target Archs: gfx110x-all
• Platform: Windows runner (SEH + Windows-style paths in log)
• Commit/ref: bb544cdc1731adb5a1d16afc42027242bf1cd9eb

Log excerpts

• unknown file: error: SEH exception with code 0xc0000094 thrown in the test body.
• unknown file: error: SEH exception with code 0xc0000005 thrown in the test body.
• .../rocblas/clients/common/gtest_helpers.cpp(217): error: Failed
• Received uncaught exception: rocblas_status_internal_error
• rocBLAS error from hip error code: 'hipErrorInvalidValue':1
• rocblas-test_full_suite (Exit code 0xc0000374)

Impact:

Blocker for promotion of rocm-systems ROCm/TheRock#5072 , rocm-libraries ROCm/TheRock#5073

[chiranjeevipattigidi]

@amcamd / @TorreZuk
Can you take a look at these failures ?

[TorreZuk]

1. The first exception:
   Exception thrown at 0x00007FFD237EC547 (rocm-openblas64.dll) in rocblas-test.exe: 0xC0000094: Integer division by zero.

@chiranjeevipattigidi the crash is in OpenBLAS, these all pass with AOCL so we need to analyze that.
Are you able to work on fixing OpenBLAS?

[chiranjeevipattigidi]

1. The first exception:
   Exception thrown at 0x00007FFD237EC547 (rocm-openblas64.dll) in rocblas-test.exe: 0xC0000094: Integer division by zero.

@chiranjeevipattigidi the crash is in OpenBLAS, these all pass with AOCL so we need to analyze that. Are you able to work on fixing OpenBLAS?

Do you know the point of contact for openblas ? Can we add them to this thread ?

[TorreZuk]

I was asking if you were willing to look into fixing OpenBLAS, as it is open source. @chiranjeevipattigidi is your team getting assigned to test dependencies to certify them for use? I will see if I can bypass the crashing function as a workaround.

[JeniferC99]

@TorreZuk I don't think our team is familiar with OpenBLAS. do you know which commit is causing this failure on the first place? maybe seeing that we can try to see how to fix it.

[TorreZuk]

This has nothing to do with large-batch hipErrorInvalidConfig, see ROCM-23872 for details on this one. The tests weren't enabled so this is not a regression but looks like a bug in the reference code. I am trying to get these tests to run on my PR with workaround #7184

[chiranjeevipattigidi]

#7184 workaround PR merged and this following rocm-libraries bump included that fix commit ROCm/TheRock#5247, after this merge, we should not see these failures.