[Test] Add rocHPL test

[pranavprakash-amd]

Motivation

Add a rocHPL (High-Performance Linpack) end-to-end HPC benchmark to the ROCm test framework. This validates the full ROCm compute stack — HIP
runtime → rocBLAS/DGEMM → MPI multi-rank orchestration — through a real, GPU-saturating LU-factorization workload, one of the canonical HPC
acceptance benchmarks for AMD GPU systems.

It also ports the legacy tests/HPC/perf_rochpl/{runRocHpl.py,build_rocHPL.sh,runRochpl.sh} scripts out of the external AMD BareMetal perf-test
framework and into the marker-compliant, remote-transparent rocm-tests framework — replacing rocminfo ASIC scraping, hand-set
ROCR_VISIBLE_DEVICES, sudo+password build steps, and per-ASIC CSV lookups with executor-injected GPU allocation and env-configurable parameters.

Technical Details

• New test tests/e2e/hpc/rochpl/test_rochpl.py — clones, builds, and runs rocHPL across P*Q MPI ranks (one rank per GPU) via the generated
  mpirun_rochpl launcher, then asserts the HPL residual check PASSED, parses the total GFLOPS from HPL.out, records it as a metric (ROCHPL_GFLOPS),
  and enforces an optional ROCHPL_MIN_GFLOPS floor.
• conftest.py — session-scoped rochpl_build and rochpl_mpi_runtime fixtures. rocHPL is a full CMake/install.sh project (not buildable via the
  single-file compile_binary/hipcc path), so it uses the framework's remote-transparent external-build primitives (clone_repo,
  assert_license_present, detect_mpi_runtime / provision_openmpi_runtime) plus a streaming install.sh runner. MPI + ROCm env is injected via an env
  VAR=... cmd prefix, never by mutating os.environ. Build trees are namespaced by GPU arch and process-grid config to force a clean rebuild on
  --gpu-arch/ROCHPL_NUM_GPUS changes.
• _workload.py — shared, env-configurable run parameters (ROCHPL_NUM_GPUS/P/Q/N/NB/ITERATIONS/REF/MIN_GFLOPS/MPI_EXTRA_ENV) imported by both
  build and run sides so grid size and gpu_count can never drift. Includes an OpenMPI+UCX single-node fallback for nodes without a fast RDMA
  fabric.
• framework/markers/taxonomy.py — registers a CATEGORY_PROFILE for tests/e2e/hpc/rochpl/ (hw.multi_gpu, layer.math_lib, ci.weekly, e2e.stack,
  os.linux). The test overrides hw.* to hw.gpu in single-GPU mode (ROCHPL_NUM_GPUS=1) and declares runtime.soak + gpu_count(N) explicitly.
• NOTICES.md — upstream rocHPL license/provenance.

Test Plan

• Collection / marker-lint validation (no GPU):
  pytest tests/e2e/hpc/rochpl/ --collect-only -q --no-gpu
• Single-GPU run:
  ROCHPL_NUM_GPUS=1 pytest tests/e2e/hpc/rochpl/ -m "hw.gpu" --gpu-arch -v
• Multi-GPU run (default 2-GPU, 2x1 grid):
  pytest tests/e2e/hpc/rochpl/ -m "hw.multi_gpu" --gpu-arch -v

Test Result

====================================================== ROCm Test Suite Summary ======================================================
Test Directory PASS FAIL SKIP ERROR Duration
────────────────────────────────────────────────────────────
tests/e2e/hpc/rochpl 0 1 0 0 5.2 s
────────────────────────────────────────────────────────────
TOTAL 1 tests │ 0 passed │ 1 failed │ 0 skipped │ 0 error │ 5.2 s

Failed tests:
• tests/e2e/hpc/rochpl/test_rochpl.py::test_rochpl_benchmark

Submission Checklist

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

[mparamas-amd]

this may not be required, as rocm_path referred to PATH for build. pls check.

[mparamas-amd]

Check to include all matrix variants, see If a json variant within this file (if small) shall replace the csv combinations. typically run for 1:2:4:8 GPU counts X ASIC specific P/Q/N/NB

Look to parameterize this input to run the test_rochpl.py

[pranavprakash-amd]

Noted.

[mparamas-amd]

is OpenMPI version - '4.1.4' necessary ?, I see the source has different - 5.0.6