[Issue] Windows gfx1151: rocBLAS trmm_strided_batched f64 complex crash (hipErrorLaunchFailure / guard corruption)

[JeniferC99]

Windows gfx1151 rocBLAS shard 6/6 crashes during rocblas-test_standard_suite on a specific pre_checkin TRMM strided-batched inplace f64 complex case. Linux gfx1151 shard 6/6 passes on the same commit/build, and the failure is unrelated to hipBLASLt/TensileLite StreamK changes in #8442 (zero rocBLAS files touched).

Failing CI

• https://github.com/ROCm/rocm-libraries/actions/runs/29951922822/job/89072433187 (PR feat(tensilelite): StreamK single-hop work stealing #8442 attempt 1)
• Runner: wshshark-strixhalo-14-gfx1151-g0-baefd0
• rocBLAS commit in log: 79ebb10ad160c92db6797d55a92ec868016a0727

Passing control (same run, same ctest shard)

• Linux shard 6/6 passed, including all 157 trmm_strided_batched cases: https://github.com/ROCm/rocm-libraries/actions/runs/29951922822/job/89256052854

Log excerpt

SIGNAL raised in: blas3_tensile/pre_checkin_trmm_strided_batched_medium_inplace_f64_c_LLTU_800_900_2r3i_801_640800_901_810900_901_810900_2_inplace
C:/.../testing_trmm_strided_batched.hpp(863): error: Expected equality of these values:
  hC.transfer_from(*dOut)
    Which is: 719   # hipErrorLaunchFailure
  hipSuccess
    Which is: 0
Error: hipMemcpy post-guard copy failure.
Error: hipMemcpy pre-guard copy failure.
rocBLAS error during freeing of allocated memory in handle destructor: rocblas_status_internal_error

Line 863 is CHECK_HIP_ERROR(hC.transfer_from(*dOut)) — not a numerical norm mismatch. The d_vector.hpp guard EXPECT_EQ(memcmp(...), 0) failures are follow-on corruption after the GPU context is poisoned.

Test params (from gtest yaml): M=800, N=900, side=L, uplo=L, transA=T, diag=U, f64 complex, batch_count=2, inplace.

Steps to reproduce

On a Windows gfx1151 runner with TheRock CI artifacts installed:

ctest -L '^standard$' -LE ex_gpu --output-on-failure --parallel 1 --timeout 7200 \
  --test-dir build/bin/rocblas -V --tests-information 6,,6

Or run the single gtest case:

rocblas-test --gtest_filter='blas3_tensile/pre_checkin_trmm_strided_batched_medium_inplace_f64_c_LLTU_800_900_2r3i_801_640800_901_810900_901_810900_2_inplace'

[JeniferC99]

Closing as an intermittent Windows gfx1151 flake, not a reproducible rocBLAS regression.

Attempt 1 (job 89072433187) hit SIGNAL + guard hipMemcpy failures on one trmm_strided_batched f64 test. Attempt 2 ran the same shard on a different runner and all 157 trmm_strided_batched tests passed (failed on an unrelated dotc_batched_ex f16 test instead). PR #8442 touches tensilelite only (0 rocblas files); shards 1–5 passed on the same run.

Reopen if this trmm signature starts failing deterministically on develop.