[marifamd]

Summary

Follow-up to #7102, which added -DENABLE_WSL_BACKEND=ON to the amdsmi subproject unconditionally.

TheRock already has a switch for this component, THEROCK_ENABLE_WSL_ROCDXG (feature group WSL, off in default/CI builds). With the backend forced on, the two disagree:

• Default builds ship an amdsmi WSL backend that dlopens librocdxg.so, while the wsl-rocdxg subproject that produces that library is not built. The dlopen just fails at runtime, so nothing breaks, but the code is dead weight in every default build.
• ENABLE_WSL_BACKEND=ON is not dependency-free on the amdsmi side. It has a FATAL_ERROR gate requiring rocr-runtime/libhsakmt/include/hsakmt/rocdxg_smi.h, resolved through a hardcoded sibling path. That works today only because rocm-systems is checked out whole, and it is a build-time dependency default builds should not take on.

This makes the amdsmi flag follow TheRock's existing switch, matching the -DBUILD_TESTS=${THEROCK_BUILD_TESTING} pattern on the adjacent line. Turning on THEROCK_ENABLE_WSL_ROCDXG now builds both the bridge library and the amdsmi backend that consumes it.

JIRA ID

ROCM-28065

Test plan

Verified THEROCK_ENABLE_WSL_ROCDXG is a CACHE BOOL created by therock_add_feature, and that therock_finalize_features() (top-level CMakeLists.txt:277) runs well before add_subdirectory(core) (CMakeLists.txt:503), so the variable is in scope at the point of use.

Substitution checked for every state:

THEROCK_ENABLE_WSL_ROCDXG	Emitted argument
OFF	-DENABLE_WSL_BACKEND=OFF
ON	-DENABLE_WSL_BACKEND=ON
undefined	-DENABLE_WSL_BACKEND= (falsey; amdsmi's option(... OFF) default preserved, no error)

• Default Linux build no longer compiles the WSL backend (restores pre-core: enable WSL backend for amdsmi on non-Windows #7102 artifact content)
• THEROCK_ENABLE_WSL_ROCDXG=ON propagates ENABLE_WSL_BACKEND=ON to amdsmi
• Windows unaffected (change stays inside the existing if(NOT WIN32) guard)
• pre-commit run --files core/CMakeLists.txt docs/development/wsl_rocdxg.md passes

For reference, in run 32498518713 the compiler-runtime stage built amdsmi with the backend on and core-amdsmi_lib_generic grew from 29,896,091 to 30,051,269 bytes (+151 KB) versus main. That delta is what default builds stop carrying with this change.

[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.

[marifamd]

This PR only needs to be merged if this PR doesn't make it in: ROCm/rocm-systems#10436

[marifamd]

Doesn't look like my build time dependency removal changes will make it in 10.1, would be wise to merge this

[JeniferC99]

lgtm