[ruru25254649]

Summary

When the CUDA toolkit is 12.9 or newer, so CUDA::nvToolsExt does not exist, MFCTargets.cmake links NVHPC's NVTX wrapper twice:

target_link_libraries(${a_target} PRIVATE nvhpcwrapnvtx)
target_link_options(${a_target} PRIVATE "-cudalib=nvtx3")

NVHPC 26.9 ships the library as libnvhpcnvtx and no longer ships libnvhpcwrapnvtx. A --gpu acc or --gpu mp build with 26.9 therefore stops at its first link:

/usr/bin/ld: cannot find -lnvhpcwrapnvtx: No such file or directory
gmake[3]: *** [CMakeFiles/syscheck.dir/build.make:100: syscheck] Error 2
Error: Failed to build the syscheck target.

The CI lanes go up to NVHPC 26.3, so none of them hit this.

The library does not need to be named at all. The driver flag -cudalib=nvtx3 already adds the wrapper under whichever name the installed release uses:

nvfortran	nvfortran -cudalib=nvtx3 -dryrun link line
25.9, 26.3	... -lnvhpcwrapnvtx ...
26.9	... -lnvhpcnvtx ...

This PR removes the explicit nvhpcwrapnvtx and keeps -cudalib=nvtx3. The CUDA ≤ 12.8 branch (CUDA::nvToolsExt) is unchanged.

The second commit adds 'cmake/**' to the cmakelist filter in .github/file-filter.yml:

• The filter lists CMakeLists.txt and toolchain/cmake/**, but not the cmake/ directory added in Share the global-parameters core across executables; generate MPI broadcasts; build hygiene #1552.
• So a PR that only changes cmake/*.cmake, like this one, skips the whole build and test matrix.
• It is the same line as in Fix compiler configuration, build-cache identity, and CMake CI coverage #2019.

Verification

On 2× H100 NVL (master 3ef5e1e):

	master	this PR
NVHPC 26.9, --gpu acc	link fails (above)	all targets build
NVHPC 25.9, --gpu acc	builds	builds
NVTX ranges in nsys (./mfc.sh run ... --nsys, 2D_riemann_test, 50 steps)	—	present on 25.9 and 26.9: TIMESTEP ×50, COMPUTE-RHS ×150, RHS-RIEMANN-SOLVER ×300, RHS-RECONSTRUCTION ×300, ...
Full --gpu acc test suite, NVHPC 26.9	cannot build	675 passed, 47 failed. All 47 are hypoelasticity tests (see below)

./mfc.sh precheck passes.

The 47 hypoelasticity failures on NVHPC 26.9 are unrelated to this PR:

• Unpatched master, built on 26.9 with a compat symlink so it links, fails the same way. For example, 34F3999B gives abs: 1.56E-02, rel: 1.56E-05 on both.
• A CPU-only 26.9 build, which has no NVTX at all, fails 41 of the first 51 hypoelasticity tests.
• --reldebug (-O1) and --debug (-O0 -Mbounds -C) builds pass, with no bounds errors. Dropping -Mfprelaxed does not help.

They pass on NVHPC 25.9, and we will report them separately.

Re-checked on master 961cde5, RTX 4060 Ti (sm_89):

	master	this PR
NVHPC 26.9, --gpu acc	ld: cannot find -lnvhpcwrapnvtx	pre/sim/post build
./mfc.sh test --gpu acc -% 3, NVHPC 26.9	—	19 passed, 2 failed. Both failures are hypoelasticity tests, the known 26.9 issue above
nsys NVTX summary, 64² 2D case, 20 steps	—	TIMESTEP ×20, COMPUTE-RHS ×60, RHS-RIEMANN-SOLVER ×120, RHS-RECONSTRUCTION ×120
NVHPC 26.3 (newest CI lane), --gpu acc, clean build dir	builds	syscheck, pre_process build; ldd shows libnvhpcwrapnvtx.so, as before

Side effects

• CUDA ≤ 12.8 (CUDA::nvToolsExt): unchanged.
• NVHPC ≤ 26.3: -cudalib=nvtx3 links the same libnvhpcwrapnvtx that was named explicitly before (checked with -dryrun on 25.9 and 26.3, and with ldd on a 26.3 build), so the link line is equivalent.
• The file-filter.yml line makes CI run the build matrix for any cmake/ change, which is more CI time for such PRs.

Contribution Policy

All contributions are expected to demonstrate:

• A clear understanding of the codebase
• Alignment with product direction
• Thoughtful reasoning behind changes
• Evidence of real-world usage or hands-on experience with the problem

If these expectations are not met, we would prefer to implement the changes ourselves rather than spend time reviewing low-effort submissions.

Acknowledgement

• I confirm this PR meets the above expectations and reflects my own understanding and real-world context.

---

This PR is from Your Team Name participating in the SC26 Student Cluster Competition.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WPhkY8p1EHhGY7jScizJmC

[github-actions]

Claude Code Review

Head SHA: aaaf38a

Files changed:

• 2
• .github/file-filter.yml
• cmake/MFCTargets.cmake

Findings:

• medium cmake/MFCTargets.cmake: The CUDA >= 12.9 branch no longer links nvhpcwrapnvtx. It now relies on -cudalib=nvtx3 to pull in NVHPC's NVTX wrapper library. The new comment says this, but the diff doesn't show it holds for NVHPC releases before 26.9 (where the library is libnvhpcwrapnvtx). If -cudalib=nvtx3 does not link the wrapper there, profiling builds with CUDA 12.9+ and older NVHPC will fail with undefined NVTX symbols. Confidence is moderate, since I can't check the NVHPC behavior from the diff. Please confirm on an NVHPC < 26.9 build with CUDA 12.9+ and --gpu profiling. If it fails, keep the old link conditional on the library existing, for example if (TARGET nvhpcwrapnvtx) or find_library.