[amd-eochoalo]

Motivation

Making rocjitsu tests more portable across different versions of amdclang.

Technical Details

When switching between different nightly builds, tests compiled with amdclang may not be recompiled unless there is a strict dependency.

From the CMake documentation:

DEPENDS
Specify files on which the command depends. Each argument is converted to a dependency as follows:

[...] if the target is an executable or library, a file-level dependency is created to cause the custom command to re-run whenever the target is recompiled.

Issue Tracking

N/A.

Test Plan

Test on CI.

Test Result

Tests pass on CI.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.

[therock-pr-bot]

❌ PR Check — Action Required

Check	Status	Details
📝 PR Description	❌ Fail	Error: PR description must reference a JIRA ID, ISSUE ID, or a GitHub closing keyword. Expected: include a JIRA ID / ISSUE ID line (separator : or -, or omitted; value may be a JIRA key, a number with/without #, or a link), OR a closing keyword + issue reference. Accepted examples: • JIRA ID : TESTAUTO-6039 • JIRA ID - #330 • JIRA ID #330 • JIRA ID (on separate line) ROCM-25757 • ISSUE ID : TESTUTO-3334 • ISSUE ID #3334 • ISSUE ID - TESTAUTO-3433 • ISSUE ID (on separate line) AIRUNTIME-2352 • ISSUE ID : https://github.com/<org_name>/<repo_name>/issues/1234 • Closes #10 • Fixes octo-org/octo-repo#100 • Resolves: #123 • #123 • https://github.com/<org_name>/<repo_name>/issues/123 Current: no valid JIRA/ISSUE/closing-keyword reference found
⛔ 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	—

⚠️ 1 policy check(s) failed. Please address the issues above before this PR can be Reviewed.

🚫 Please fix the failed policies

• ❌ PR Description

The Not ready to Review label was added to this PR. Once all policies pass, the label is removed automatically.

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

🙋 Wish to Override Policy?

[therock-pr-bot]

🚫 Please fix the failed policies before requesting reviews.

The following policy checks failed:

• ❌ PR Description

The Not ready to Review label has been added to this PR.
Once all policies pass, the label will be removed automatically.

[atgutier]

[MAJOR] emulation/rocjitsu/tests/CMakeLists.txt:1815 — Configure-time KFD permissions can
permanently remove hardware tests

Reasoning: The new AND RJ_HAS_KFD_ACCESS prevents rocm_agent_enumerator from running whenever the
machine performing CMake configuration cannot access /dev/kfd. That conflates target discovery with
permission to execute tests on the configure host.

rocm_agent_enumerator intentionally supports several discovery paths that do not require KFD access:
ROCM_TARGET_LST, a colocated target.lst, sysfs topology, and PCI enumeration. The first two are
explicitly intended for sandboxed and containerized build environments.

When enumeration is skipped, every HAS_*_GPU variable remains false, so the conditional
rj_add_test() calls around lines 1910, 1921, 1941, 2084, 2250, 2442, and 2923 are never executed.
Because rj_add_test() also creates installed CTest entries only when called (tests/
CMakeLists.txt:216-308), a package configured without KFD access permanently omits those tests—even
if it is later installed and tested on a GPU worker with full access.

Evidence:

• projects/rocminfo/rocm_agent_enumerator:127-144,263-303 performs non-KFD discovery before falling
  back to rocminfo.

• projects/rocminfo/README.md:34-44 documents ROCM_TARGET_LST and target.lst specifically for
  sandbox/container environments.

• The new condition at tests/CMakeLists.txt:1815 bypasses all those paths.

• Many affected registrations supply INSTALL_COMMAND, so this changes the installed test manifest as
  well as the current build tree.

Suggested fix: Keep architecture discovery independent of /dev/kfd accessibility. Register the
applicable hardware tests based on discovered or explicitly supplied targets, then check KFD access
at test execution time and return a CTest skip result when it is unavailable. If build-tree and
installed-test behavior must differ, separate their registration paths.

Confidence: high. The remaining environmental question is which current CI/package flows configure
and execute on different hosts, but the supported installed-test and ROCM_TARGET_LST workflows are
unambiguously affected.

[MINOR] emulation/rocjitsu/tests/CMakeLists.txt:1065 — The access check misses the earlier gfx1201
hardware-test gate

Reasoning: RJ_HAS_GFX1201_GPU is still derived solely from readable sysfs topology and does not
incorporate the newly computed RJ_HAS_KFD_ACCESS. On a gfx1201 system where topology is visible
but /dev/kfd is not accessible, the six DbtGuestRocminfoTest.* cases under line 1680 remain
registered.

Those tests use guest_gfx950_on_gfx1201.json, which enables hardware-backed DBT guest mode. GuestKfd
then opens the real device with openat(..., "/dev/kfd", O_RDWR | O_CLOEXEC, ...) at
guest_kfd.cpp:587; an access failure propagates through initialization. rocminfo_test.cpp:156-159
treats the resulting nonzero process exit as a test failure rather than a skip.

Suggested fix: Include these earlier gfx1201 registrations in the runtime-access gating described
above. As a smaller immediate fix, require both RJ_HAS_GFX1201_GPU and RJ_HAS_KFD_ACCESS at line
1680.

Confidence: high. The trigger is the common configuration where KFD topology is world-readable but
the device node is unavailable to the current user.

[RattataKing]

Adding missing deps of ${AMDCXX} and ${CLANG_OFFLOAD_BUNDLER} looks correct to me

[amd-eochoalo]

@atgutier I removed the code about kfd access, just adding missing dependencies.

[systems-assistant]

TheRock Submodule Bump Activity

Newest first

• 2026-09-08 — Included in TheRock via ROCm/TheRock#7974.