Meson CUDA dependency ignores target-specific CUDA library paths during cross-compilation (aarch64)

[MaxandreOgeret]

Describe the bug

When cross-compiling CUDA-based projects for aarch64 platforms, Meson's built-in CUDA dependency
dependency('cuda', modules: ['cudart'])
always provides the host link directory /usr/local/cuda/lib64 instead of the target architecture's /usr/local/cuda/targets/aarch64-linux/lib

To Reproduce

Any cross compiled project which uses dependency('cuda')

Expected behavior

dependency('cuda') should use the target's include and link dir when available.

--

Issues I noticed in the cuda dependency:

• cuda dependency will always include the build machine libdir:
  

  meson/mesonbuild/dependencies/cuda.py

  Lines 213 to 214 in 25854d7

  	def _detect_arch_libdir(self) -> str:
  	arch = detect_cpu_family(self.env.coredata.compilers.host)

  
  

  meson/mesonbuild/dependencies/cuda.py

  Line 226 in 25854d7

  return libdirs[arch]

  
  Arch should be looked up via the compilers "for_machine" entry instead.
• lib64 is hardcoded.
  

  meson/mesonbuild/dependencies/cuda.py

  Lines 222 to 223 in 25854d7

  	elif machine.is_linux():
  	libdirs = {'x86_64': 'lib64', 'ppc64': 'lib', 'aarch64': 'lib64', 'loongarch64': 'lib64'}

  
  When the actual NVIDIA cross-compilation toolchain uses lib
  ( /usr/local/cuda/targets/aarch64-linux/lib)
  This prevents setting CUDA_PATH=/usr/local/cuda/targets/aarch64-linux as a workaround.
• The host machine target is totally ignored, and only the build machine lib dir is used.
  

  meson/mesonbuild/dependencies/cuda.py

  Lines 59 to 60 in 25854d7

  	arch_libdir = self._detect_arch_libdir()
  	self.libdir = os.path.join(self.cuda_path, arch_libdir)

Approaches to fix the issue

With some guidance I'd be honored to make a PR, I have the following thoughts on how to solve the issues:

• _detect_arch_libdir should use "for_machine" entry instead of detect_cpu_family.
• When cross compiling the lib directory is forced to lib (or make configurable ?)
• When cross-compiling, the dependency detector should search for target-specific CUDA toolkit paths in /usr/local/cuda/targets/{target_arch}-linux/.

--

system parameters

• cross-build (build machine: x86_64; host_machine: aarch64)
• OS: Ubuntu 18.04 x86_64
• Python version: Python 3.8.0
• Meson version: 1.8.0
• Ninja version: 1.11.1.git.kitware.jobserver-1

---

Thanks to @eli-schwartz for guidance on Matrix chat about this issue! 🙂