[nunnikri]

Motivation

Nightly ASAN builds have failed two nights running when ld.lld links libMIOpen.so, with relocation R_X86_64_PC32 out of range errors. libMIOpen.so statically links hipconv and MIOpen's own ai_heuristics.cpp directly, and ASAN's stack-use-after-return instrumentation on their dense, per-arch/per-kernel templated code has grown that link past the linker's +/-2GB PC-relative relocation limit. This is cumulative image bloat from ordinary kernel/feature growth (e.g. #13718, #14627), not a defect in any single PR, and will keep recurring as more kernels land.

ISSUE ID : ROCm/TheRock#9812

Technical Details

Adds MIOPEN_HIPCONV_SANITIZER and MIOPEN_CKIMPL_SANITIZER CACHE options, consumed in hipconv's and ck_impl's own CMakeLists.txt. When set to OFF, each strips -fsanitize=address from just that component:

• hipconv: compile-option override only, on both the main hipconv target and each per-arch object-library target in _arch_targets (where the oversized template instantiations actually live). hipconv contributes object files/a static archive into libMIOpen.so's own link, not a link of its own, so there's no link step to override.
• ck_impl: both compile- and link-option overrides, applied per MIOpenCKGroupedConv_<arch> plugin target — each is its own SHARED library with its own independent link step, mirroring the existing MIOPEN_DISABLE_DEVICE_UBSAN pattern already in this file. Unlike that UBSan precedent, this isn't -Xarch_device-scoped: the failures we're seeing are in host-side kernel-launcher code (under this build's unscoped, full-ASAN mode), so both host and device instrumentation are dropped here.

Neither option affects the rest of MIOpen, which keeps full ASAN/TSAN coverage. Both default to unset (no override, today's behavior unchanged) and are driven by matching preset knobs in ROCm/TheRock's CMakePresets.json (companion PR).

Mixing instrumented and uninstrumented code in one process is a standard, supported ASAN configuration — allocation/deallocation is intercepted process-wide regardless of which TU allocated or freed, and the sanitizer flag doesn't change a function's ABI. The tradeoff is narrower detection coverage: a memory bug originating purely inside hipconv/ck_impl's own code won't be caught at the point of fault anymore (though resulting heap corruption can still sometimes surface later via instrumented code touching the same memory).

Test Plan

• Confirmed libMIOpen.so's ASAN link succeeds with feat(ck): TDM and named barriers in wavelet pipeline conv fwd #13718 reverted (ROCm/TheRock run 42324389102)
• Verify the ci:asan build passes with this change and feat(ck): TDM and named barriers in wavelet pipeline conv fwd #13718/fix(miopen): pick the FP8 exponent bias from the device at run time #14627 both present

[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.