[Test] Add support for Hecbench_spirv test

[pranavprakash-amd]

Motivation

HeCBench is a large, widely-used collection of HIP/CUDA benchmarks that exercises a broad cross-section of the compiler and HIP runtime stack. The
legacy suite (tests/COMPILERS/hecbench/hecbench.py, the HecBench class) validated these benchmarks but predates the rocm-tests framework — it
cloned the repo into an ad-hoc scratch dir, iterated src/-hip folders imperatively, and decided SPIR-V mode by sniffing an executor's
compiler string for the substring "spirv", reporting into a separate "Hecbench_spirv" suite.

This PR ports HeCBench into the rocm-tests framework as a marker-compliant, parametrized e2e test. The goals:

• Bring HeCBench compile+run coverage under the standard framework (executors, fixtures, markers, Allure metrics, remote-fleet support).
• Promote SPIR-V from an out-of-band compiler-string hack to a first-class mode parametrization axis, so every benchmark produces both a -hip
  (plain hipcc) and a -spirv (--offload-arch=amdgcnspirv) test — giving explicit, per-benchmark SPIR-V codegen signal.
• Provide a fast nightly smoke gate plus a full weekly soak run over the entire catalog.

Technical Details

Adds tests/e2e/compiler/hecbench/ with four files:

• test_hecbench.py — the ported test. A shared _run_benchmark() helper: skips (rather than fails) when a src/-hip folder is absent from the
  checkout; puts --rock-dir's bin/ on PATH and sets LD_LIBRARY_PATH (via ld_path) so the Makefile's hipcc and the built binary resolve TheRock libs;
  runs make clean/make/make run through target_executor so compile and run land on the same node/filesystem.
  • SPIR-V mode: patches the Makefile with sed -i 's/hipcc/hipcc --offload-arch=amdgcnspirv/g'. Before every run it restores the Makefile via git
    checkout -- Makefile, so the shared session checkout is never left in a patched state regardless of run order.
  • Pass criterion (preserves legacy semantics): compile succeeds and the benchmark's subset.json regex extracts a nonzero numeric total
    (_extract_total sums all float-castable captures, robust to multi-group regexes). Benchmarks with an empty regex are validated on clean exit only
    — a deliberate, documented deviation from the legacy behavior, which marked those "failed" purely due to a test-data gap.
  • Metrics reported to Allure via report_metric("HECBENCH__", total).
  • Two test functions: test_hecbench_smoke (curated 5-benchmark subset, runtime.medium, keeps the profile-injected ci.nightly) and
    test_hecbench_full_suite (entire catalog, overrides to ci.weekly + runtime.soak). Both parametrized over benchmark × mode.
• conftest.py — a session-scoped hecbench_repo fixture that clones https://github.com/zjin-lcf/HeCBench (ref master) exactly once via the
  remote-aware external_build.clone_repo helper, replacing the legacy one-off clone. URL/ref overridable via ROCM_TEST_HECBENCH_URL /
  ROCM_TEST_HECBENCH_REF; assert_license_present acts as a provenance guard.
• subset.json — the benchmark → capture-regex catalog (~498 benchmarks), copied verbatim from the legacy suite.
• init.py — package marker.

Markers hw.gpu, layer.runtime, ci.nightly, e2e.stack, os.linux are auto-injected by CATEGORY_PROFILES for tests/e2e/compiler/; explicit runtime.*
/ ci.weekly markers are declared per function. No framework code changed — this is purely additive under tests/.

Prereqs: --rock-dir/ROCK_DIR pointing at a ROCm/TheRock install with bin/hipcc, an AMD GPU, and network access to clone HeCBench.

Test Plan

• Marker/collection validation without hardware:
  pytest tests/e2e/compiler/hecbench/test_hecbench.py --collect-only -q --no-gpu
• Nightly smoke on a real GPU:
  pytest tests/e2e/compiler/hecbench/test_hecbench.py::test_hecbench_smoke -v
  --rock-dir /path/to/therock --gpu-arch gfx942
• Full weekly suite (soak):
  pytest tests/e2e/compiler/hecbench/test_hecbench.py::test_hecbench_full_suite -v
  -m "ci.weekly" --rock-dir /path/to/therock --gpu-arch gfx942
• Lint gate (ruff, black, mypy, pylint) and the marker-lint hook run on the new files.

Test Result

====================================================== ROCm Test Suite Summary ======================================================
Test Directory PASS FAIL SKIP ERROR Duration
───────────────────────────────────────────────────────────────────
tests/e2e/compiler/hecbench 7 3 0 0 147.8 s
───────────────────────────────────────────────────────────────────
TOTAL 10 tests │ 7 passed │ 3 failed │ 0 skipped │ 0 error │ 147.8 s

Submission Checklist

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

[mparamas-amd]

Add a NOTICES.md

[mparamas-amd]

Update regex entry from ["(?:PASS)"] to [""] — an empty string — so the test validates it on exit code alone, aligned with Original.

[mparamas-amd]

Could you check with Original assert ? flatten of group required?

[pranavprakash-amd]

Validated with the original logic. The same logic of regex is been used

[mparamas-amd]

@pranavprakash-amd did you root cause on the 3 test failures ?

[pranavprakash-amd]

@pranavprakash-amd did you root cause on the 3 test failures ?

Yes @mparamas-amd there are JIRA's tracking the failures with respect to the above. Have track of them.