add nvhpc to test suite github runners by sbryngelson · Pull Request #1317 · MFlowCode/MFC

sbryngelson · 2026-03-17T00:59:27Z

Adds github runners for nvhpc builds on cpu and gpu

github-actions · 2026-03-17T01:01:12Z

Claude Code Review

Head SHA: 5dec31f
Files changed: 1

.github/workflows/test.yml

Summary

Adds 15 new matrix entries to the github job covering 5 NVHPC versions (23.11, 24.5, 24.11, 25.1, 25.3) × 3 targets (cpu, gpu-acc, gpu-omp).
NVHPC builds run inside the official nvcr.io/nvidia/nvhpc container on ubuntu-22.04 runners with nvfortran/nvc/nvc++ compilers and root-enabled OpenMPI.
GPU targets (gpu-acc, gpu-omp) exercise a build-only dry-run (--dry-run); actual test execution is restricted to cpu targets, which is appropriate since github-hosted runners lack GPU hardware.
All NVHPC jobs set continue-on-error: true (inherited from the job level), so failures are informational and do not block merges — matching the intended non-gating posture stated in the PR description.
Existing non-NVHPC steps are correctly guarded with !matrix.nvhpc conditions, and Intel/macOS setup steps are unaffected.

Findings

Observation (non-blocking): Missing MPI installation for NVHPC container
File: .github/workflows/test.yml, "Setup NVHPC" step (~line 178)

The NVHPC container image (nvcr.io/nvidia/nvhpc:*-devel-cuda_multi-ubuntu22.04) ships with NVHPC and CUDA toolchains but does not include a system OpenMPI installation. The CPU test step runs mfc.sh test ... --test-all without a --no-mpi flag, which will invoke MPI. The environment sets OMPI_ALLOW_RUN_AS_ROOT (suggesting MPI is expected to be used), but libopenmpi-dev / openmpi-bin are absent from the apt-get install list in the NVHPC setup step. This may cause CPU test failures at runtime if MFC's test runner tries to launch mpirun. Consider either adding openmpi-bin libopenmpi-dev to the NVHPC apt-get install line, or passing --no-mpi to the NVHPC Test step.

Minor style note: HDF5 not installed for NVHPC
File: .github/workflows/test.yml, "Setup NVHPC" step (~line 178)

The standard Ubuntu setup installs libhdf5-dev and hdf5-tools, but the NVHPC setup omits them. If any test case exercises HDF5 I/O (post_process output), the test run may fail. The standard non-NVHPC apt-get install list could serve as a reference for parity.

No issues found in the conditional logic, container image references, GPU flag mapping (gpu-acc → --gpu acc, gpu-omp → --gpu mp), or the CC/CXX/FC overrides. The ternary expression for runs-on and container is correct.

github-actions · 2026-03-17T01:13:36Z

Claude Code Review

Incremental review from: 5dec31f20ec2b50075f91eecae9f1caa98db638d
Head SHA: cd0d691df73cfa4e1528ffc02a3f8310c5a4cde9

New findings since last Claude review:

The two changes in this update are:

nvhpc: [""] and target: [""] added to the base matrix so non-NVHPC entries have these keys defined with empty string values.
The non-NVHPC job name updated from the bare string "Github" to format("Github ({0}, {1}, {2}, intel={3})", matrix.os, matrix.mpi, matrix.debug, matrix.intel).

Both changes are correctness improvements — no new issues introduced.

Observation (carry-over, still unaddressed): MPI and HDF5 missing in NVHPC container setup

The Setup NVHPC step still does not install openmpi-bin libopenmpi-dev or libhdf5-dev hdf5-tools. The Test (NVHPC) step runs --test-all without --no-mpi, so MPI invocations will fail at runtime. This was noted in the prior review and remains unresolved.

No other new high-confidence findings.

Test 5 NVHPC versions (23.11, 24.5, 24.11, 25.1, 25.3) on GitHub-hosted runners using official NVIDIA container images. Each version runs both a CPU build+test and a GPU (OpenACC) build-only target. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…efaults GitHub Actions include entries with new-only keys get merged into all existing combos instead of creating new ones. Adding nvhpc and target as main matrix axes with empty defaults means the include entries overwrite original values, forcing creation of standalone combos.

github-actions · 2026-03-18T02:48:15Z

… build - Add h_iL, h_iR, h_avg_2 and other chemistry locals to private clauses of two GPU_PARALLEL_LOOP regions in s_hllc_riemann_solver (lines 2201, 2421) where they were declared at subroutine scope but not privatized, causing nvfortran 25.1 gpu-omp to fail with 'Could not find allocated-variable index for symbol - h_il' - Disable ZFP compression in silo (-DSILO_ENABLE_ZFP=OFF) since MFC never uses it and nvc crashes compiling zfp/decode1d.c on 23.11/25.3 - Pin silo to 4.12.0 tag with GIT_SHALLOW; clean up HDF5 build flags

…arnings

Change v_vf argument from dimension(sys_size) to dimension(:) to match the caller which passes an allocatable array. The explicit-size vs assumed-shape mismatch caused nvfortran 24.11's IPO inliner to create partial inline temporaries that the OpenACC backend could not map.

This reverts commit dc12849.

…equations" This reverts commit ff30d8a.

github-actions · 2026-03-29T03:11:13Z

Claude Code Review

Incremental review from: 20288ec
Head SHA: f64894f

New findings since last Claude review:

!dir$ noinline directive silently neutered (src/simulation/m_time_steppers.fpp:628): Changing !dir$ noinline to ! dir$ noinline (adding a space after !) converts the NVIDIA nvfortran compiler directive into a plain Fortran comment. !dir$ with no space is the required syntax for nvfortran inline directives; with a space it is ignored entirely. The noinline hint was presumably added to work around a compiler issue with s_compute_dt; removing it silently while keeping the dead #ifdef __NVCOMPILER block may reintroduce whatever problem the directive was guarding against. If the intent is truly to remove the directive, the entire #ifdef block should be deleted; if the intent is to keep noinline, the space must be removed.

…versions - Remove openmpi-bin/libopenmpi-dev from NVHPC container apt install. The system OpenMPI (gfortran-compiled) was overriding the NVHPC-bundled MPI at runtime via ldconfig, causing ABI mismatch segfaults in all CPU tests (rayleigh_taylor_muscl crash in s_interface_compression). - Remove NVHPC version exclusion range for two-pass IPO. Binary search identified s_compute_dt as the single function causing ICE on 24.11; add noinline directive to prevent its extraction into the inline library.

…ailure

sbryngelson and others added 7 commits March 17, 2026 19:25

ci: consolidate NVHPC build steps using matrix.target

790dfbd

ci: add OpenMP target offload (gpu-omp) to NVHPC matrix

e72b8fc

ci: add matrix details to standard GitHub runner job names

fed16a2

ci: add OpenMPI and HDF5 to NVHPC container setup

00857fe

ci: reduce NVHPC GPU build parallelism to -j 2 to avoid OOM

f38b5a2

sbryngelson force-pushed the nvhpc branch from c290881 to f38b5a2 Compare March 17, 2026 23:26

sbryngelson added 2 commits March 18, 2026 02:00

Merge branch 'master' into nvhpc

581645b

Merge branch 'master' into nvhpc

5bfaed3

sbryngelson added 10 commits March 18, 2026 05:29

Merge branch 'master' into nvhpc

23a0bd9

Merge branch 'master' into nvhpc

5ef74e3

ci: add env diagnostics to NVHPC CPU test step

5e8b22b

style: format m_riemann_solvers.fpp

066089f

ci: fix git safe.directory in NVHPC containers to silence ownership w…

9472f4d

…arnings

style: format m_start_up.fpp

dc12849

Revert "style: format m_start_up.fpp"

627d01a

This reverts commit dc12849.

Revert "fix: use assumed-shape array in s_initialize_internal_energy_…

4ad88a9

…equations" This reverts commit ff30d8a.

sbryngelson force-pushed the nvhpc branch from 03703a4 to 20288ec Compare March 29, 2026 16:24

sbryngelson added 2 commits March 29, 2026 12:26

style: format m_time_steppers.fpp

f64894f

ci: move git safe.directory before checkout to fix NVHPC 25.3 clone f…

1a60cbe

…ailure

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add nvhpc to test suite github runners#1317

add nvhpc to test suite github runners#1317
sbryngelson wants to merge 22 commits intoMFlowCode:masterfrom
sbryngelson:nvhpc

sbryngelson commented Mar 17, 2026

Uh oh!

github-actions bot commented Mar 17, 2026

Uh oh!

github-actions bot commented Mar 17, 2026

Uh oh!

github-actions bot commented Mar 18, 2026 •

edited

Loading

Uh oh!

github-actions bot commented Mar 29, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

1 participant

Conversation

sbryngelson commented Mar 17, 2026

Uh oh!

github-actions bot commented Mar 17, 2026

Claude Code Review

Summary

Findings

Uh oh!

github-actions bot commented Mar 17, 2026

Claude Code Review

Uh oh!

github-actions bot commented Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Mar 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

1 participant

github-actions bot commented Mar 18, 2026 •

edited

Loading

github-actions bot commented Mar 29, 2026 •

edited

Loading