[Test] Add kfd test

[pranavprakash-amd]

Motivation

Add end-to-end test coverage for KFD (Kernel Fusion Driver) — the AMD GPU kernel driver / thunk layer (libhsakmt) at the very base of the ROCm
stack. KFD sits below the HIP runtime, so validating it directly closes a gap at the lowest software layer of our e2e coverage. This ports the
legacy ROCmTest tests/KFD/kfd.py into the marker-compliant rocm-tests framework.

Technical Details

Ports ROCmTest's class-based KFD.execute(executeData) into tests/e2e/kfd/, driving the upstream kfdtest GTest suite from the libhsakmt project.

• tests/e2e/kfd/test_kfd.py — four tests mirroring the original's coverage:
  • test_kfd_smoke (runtime.fast) — enumerates KFD suites via --gtest_list_tests; proves the binary links against libhsakmt and launches.
  • test_kfd_full_suite (ci.weekly, runtime.soak) — the full run_kfdtest.sh default run.
  • test_kfd_hmm_svm (runtime.medium) — HMM/SVM path (KFDSVMRangeTest/KFDSVMEvictTest), mirroring the original "HMM" case.
  • test_kfd_multi_gpu_parallel (hw.multi_gpu, gpu_count("ALL"), runtime.medium) — parallel multi-GPU mode via HSA_TEST_GPUS_NUM, mirroring
    --hsa_test_gpus_num.
• tests/e2e/kfd/conftest.py — session-scoped kfdtest_binary fixture that sparse-clones the whole libhsakmt project from the ROCm/rocm-systems
  monorepo (not just tests/kfdtest, since its CMakeLists resolves thunk headers via ../../include) and CMake-builds it against the resolved
  TheRock/ROCm install. Includes fail-fast guards: an actionable skip when libdrm/libnuma pkg-config deps are un-provisioned, and a clear error when
  the sibling include/ tree is missing from a stale checkout.
• framework/builder/binary_builder.py — clone_repo now re-asserts the sparse-checkout spec on a reused cached checkout (both local and remote
  paths), so a narrower checkout from an earlier run is widened rather than silently missing newly-needed files.
• framework/markers/taxonomy.py — new tests/e2e/kfd category profile (hw.gpu, layer.runtime, ci.nightly, e2e.stack, os.linux).
• pyproject.toml — disable pylint R0917 (too-many-positional-arguments), since pytest injects fixtures positionally.

Host provisioning (OS build deps, GCC pinning) is intentionally not ported into test code — nodes are provisioned via --pre-install pkg=libdrm-dev
/ pkg=libnuma-dev. GPU selection/visibility is left to target_executor (which injects ROCR_VISIBLE_DEVICES) rather than the original's fragile
host-topology -n offset.

Test Plan

• Marker-lint hook validated all files under tests/e2e/kfd/ at write time.
• Collection preview: pytest tests/e2e/kfd/ --collect-only -q --no-gpu
• Smoke gate on a gfx942 node: pytest tests/e2e/kfd/test_kfd.py::test_kfd_smoke --gpu-arch gfx942 -v
• Nightly subset (smoke + HMM/SVM): pytest tests/e2e/kfd/ -m "hw.gpu and ci.nightly" --gpu-arch gfx942 -v
• Multi-GPU parallel path on a ≥2-GPU node.

Test Result

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html ====================================================== ROCm Test Suite Summary ====================================================== Test Directory PASS FAIL SKIP ERROR Duration ───────────────────────────────────────────────────── tests/e2e/kfd 2 0 0 0 244.7 s ───────────────────────────────────────────────────── TOTAL 2 tests │ 2 passed │ 0 failed │ 0 skipped │ 0 error │ 244.7 s ================================================================= ================================================================== ====================================== 2 passed, 2 deselected, 1 warning in 245.89s (0:04:05) =======================================

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[pranavprakash-amd]

====================================================== ROCm Test Suite Summary ======================================================
Test Directory PASS FAIL SKIP ERROR Duration
─────────────────────────────────────────────────────
tests/e2e/kfd 1 3 0 0 2389.3 s
─────────────────────────────────────────────────────
TOTAL 4 tests │ 1 passed │ 3 failed │ 0 skipped │ 0 error │ 2389.3 s

============================================= 3 failed, 1 passed in 2403.73s (0:40:03) ==============================================