(fix:cuda) CUDA portability improved\nbuild system now checks for valid compiler compatible with nvcc and fails if that is not found and cuda is selected. Further, compiler choices are propegated throughout build along with CFLAGS and CXXFLAGS

tools/check_cuda_toolchain.py

@@ -0,0 +1,72 @@
+#!/usr/bin/env python3
+"""Probe the installed toolkit instead of maintaining a CUDA/compiler version table."""
+from __future__ import annotations
+
+import argparse
+import os
+from pathlib import Path
+import shlex
+import subprocess
+import tempfile
+
+
+def check_toolchain(nvcc: str, cc: str, cxx: str, *, standard: str = '17',
+                    c_args: list[str] | None = None,
+                    cxx_args: list[str] | None = None) -> str:
+    version = subprocess.run([nvcc, '--version'], text=True, capture_output=True, check=True).stdout.strip()
+    # Do not permit ambient flags to bypass the installed toolkit's version checks.
+    env = dict(os.environ)
+    for key in ('NVCC_PREPEND_FLAGS', 'NVCC_APPEND_FLAGS', 'NVCC_CCBIN'):
+        env.pop(key, None)
+    with tempfile.TemporaryDirectory(prefix='mfem-cuda-check-') as directory:
+        root = Path(directory)
+        source = root / 'probe.cu'
+        source.write_text('''#include <cuda_runtime.h>
+#include <thrust/device_vector.h>
+#include <cmath>
+__global__ void probe(int *p) { *p = 1; }
+int main() { int *p = nullptr; cudaMalloc(&p, sizeof(int));
+probe<<<1, 1>>>(p); return (int)cudaFree(p); }
+''')
+        for label, host, flags in [('C', cc, c_args or []), ('C++', cxx, cxx_args or [])]:
+            command = [nvcc, '--compiler-bindir', host, '--std=c++' + standard,
+                       str(source), '-o', str(root / ('probe.exe' if os.name == 'nt' else 'probe'))]
+            if label == 'C':
+                command.append('-c')
+            for flag in flags:
+                command += ['-Xcompiler', flag]
+            result = subprocess.run(command, env=env, text=True, capture_output=True)
+            if result.returncode:
+                raise RuntimeError(
+                    f'CUDA compatibility check failed for the selected {label} compiler: {host}\n'
+                    f'{version}\nCommand: {shlex.join(command)}\n'
+                    f'{result.stdout}{result.stderr}\n'
+                    'Select a C/C++ toolchain supported by this CUDA toolkit using CC and CXX '
+                    '(for example CC=gcc-15 CXX=g++-15), or install a compatible toolkit. '
+                    'Use a fresh Meson build directory when changing compilers; --reconfigure '
+                    'does not change cached compilers. For a CPU build use -Dmfem_cuda=disabled. '
+                    'The probe compiles and links but does not require a running GPU.'
+                )
+    return f'CUDA compatibility checks passed: C={cc}, C++={cxx}, CUDA C++{standard}\n{version}'
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('--nvcc', required=True)
+    parser.add_argument('--cc', required=True)
+    parser.add_argument('--cxx', required=True)
+    parser.add_argument('--standard', default='17')
+    parser.add_argument('--c-arg', action='append', default=[])
+    parser.add_argument('--cxx-arg', action='append', default=[])
+    args = parser.parse_args()
+    try:
+        print(check_toolchain(args.nvcc, args.cc, args.cxx, standard=args.standard,
+                              c_args=args.c_arg, cxx_args=args.cxx_arg))
+    except (RuntimeError, OSError, subprocess.SubprocessError) as error:
+        print(error)
+        return 1
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())