GH-46165: [C++] Add cuda option to Meson configuration

[WillAyd]

Rationale for this change

This continues building out support for Meson as a build system gnerator

What changes are included in this PR?

Adds a cuda option that pulls in the gpu directory

Are these changes tested?

No

Are there any user-facing changes?

No

• GitHub Issue: [C++] Add gpu directory to Meson #46165

[github-actions]

⚠️ GitHub issue #46165 has been automatically assigned in GitHub to PR creator.

[WillAyd]

I am unable to fully test this locally as I do not have a CUDA-enabled device. I can configure the project and compile, but at link time I get errors like:

/usr/bin/ld: src/arrow/gpu/libarrow-cuda.so.p/cuda_context.cc.o: in function `arrow::cuda::CudaDevice::AllocateHostBuffer(long)':
/home/willayd/clones/arrow/cpp/builddir/../src/arrow/gpu/cuda_context.cc:266:(.text+0xa88): undefined reference to `cuMemHostAlloc'
/usr/bin/ld: src/arrow/gpu/libarrow-cuda.so.p/cuda_context.cc.o: in function `arrow::cuda::CudaDevice::MakeStream(unsigned int)::{lambda(void*)#1}::operator()(void*) const':
/home/willayd/clones/arrow/cpp/builddir/../src/arrow/gpu/cuda_context.cc:286:(.text+0xc29): undefined reference to `cuStreamDestroy_v2'
/usr/bin/ld: src/arrow/gpu/libarrow-cuda.so.p/cuda_context.cc.o: in function `arrow::cuda::CudaDevice::MakeStream(unsigned int)':
/home/willayd/clones/arrow/cpp/builddir/../src/arrow/gpu/cuda_context.cc:280:(.text+0xf1f): undefined reference to `cuStreamCreate'

I believe those missing symbols are from drivers that I cannot install, although my familiarity with building CUDA applications is admittedly low

[kou]

We can build cpp/src/arrow/gpu/ without CUDA-enabled device. For example, our deb packages do it: https://github.com/ursacomputing/crossbow/actions/runs/14590422628/job/40924210051#step:8:8349

-- Installing: /build/apache-arrow-20.0.0.dev326/debian/tmp/usr/lib/x86_64-linux-gnu/libarrow_cuda.so.2000.0.0

FYI: We can use nvidia-cuda-toolkit on Debian for it:

arrow/dev/tasks/linux-packages/apache-arrow/apt/debian-bookworm/Dockerfile

Line 85 in a583e6a

apt install -y -V ${quiet} nvidia-cuda-toolkit; \

[WillAyd]

@github-actions crossbow submit *meson

[github-actions]

Revision: 5b4a4be4ff81439f3a017aae5583dd6303b76b10

Submitted crossbow builds: ursacomputing/crossbow @ actions-c4ce1f69f0

Task	Status
test-conda-cpp-meson

[WillAyd]

@github-actions crossbow submit *meson

[github-actions]

Revision: e21c24fbe237c1e6bf123f1e69cf6ce817e57210

Submitted crossbow builds: ursacomputing/crossbow @ actions-13cedcf855

Task	Status
test-conda-cpp-meson

[WillAyd]

It looks like detecting CUDA installed via Python may need the upcoming Meson 1.9 release to build correctly. We could also try to install on the system instead for current Meson versions, although I think waiting for 1.9 is easiest (should be released soonish)