Add '' to cmake common_paths

[zasdfgbnm]

PyTorch is a project built with cmake, and provides cmake packages for applications to link with it. The cmake prefix of PyTorch can be found by running:

python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)'

you will see something like below from the above command:

/home/gaoxiang/.virtualenvs/nvfuser/lib/python3.11/site-packages/torch/share/cmake

Inspecting this directory:

❯ tree /home/gaoxiang/.virtualenvs/nvfuser/lib/python3.11/site-packages/torch/share/cmake
/home/gaoxiang/.virtualenvs/nvfuser/lib/python3.11/site-packages/torch/share/cmake
├── ATen
│   └── ATenConfig.cmake
├── Caffe2
│   ├── Caffe2Config.cmake
│   ├── Caffe2Targets.cmake
│   ├── Caffe2Targets-release.cmake
│   ├── FindCUDAToolkit.cmake
│   ├── FindCUSPARSELT.cmake
│   ├── Modules_CUDA_fix
│   │   ├── FindCUDA.cmake
│   │   ├── FindCUDNN.cmake
│   │   └── upstream
│   │       ├── CMakeInitializeConfigs.cmake
│   │       ├── FindCUDA
│   │       │   ├── make2cmake.cmake
│   │       │   ├── parse_cubin.cmake
│   │       │   ├── run_nvcc.cmake
│   │       │   └── select_compute_arch.cmake
│   │       ├── FindCUDA.cmake
│   │       ├── FindPackageHandleStandardArgs.cmake
│   │       └── FindPackageMessage.cmake
│   └── public
│       ├── cuda.cmake
│       ├── gflags.cmake
│       ├── glog.cmake
│       ├── LoadHIP.cmake
│       ├── mkl.cmake
│       ├── mkldnn.cmake
│       ├── protobuf.cmake
│       └── utils.cmake
├── Tensorpipe
│   ├── TensorpipeTargets.cmake
│   └── TensorpipeTargets-release.cmake
└── Torch
    ├── TorchConfig.cmake
    └── TorchConfigVersion.cmake

9 directories, 28 files

However, meson currently filters this directory out by _preliminary_find_check. As a result, doing

torch_dep = dependency('Torch')

will fail, even if you set cmake_prefix_path with the value returned by PyTorch.

Possibly related issues:
https://stackoverflow.com/questions/68884434/libtorch-c-meson-dependency
#9740
https://discuss.pytorch.org/t/libtorch-meson-build/139648

[eli-schwartz]

The current commit message states the mechanical nature of the change: "add such and such string to XXX". It would be best to explain in the commit message why this change is useful.

I'm no cmake expert but I think what it's doing is allowing you to pass the value of the share/cmake/ directory as cmake_prefix_path although normally that directory is supposed to be set to the base directory which contains "lib" and "include" and "share/cmake" subdirectories.

Am I understanding correctly?

[zasdfgbnm]

@eli-schwartz Yes, your understanding is correct. I will update commit message later.

[zasdfgbnm]

Commit message updated

[tristan957]

Meson has a cmake_prefix_path option or --cmake-prefix-path on the command line. I feel like setting that to the output of python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)' makes more sense to me. This could be achieved with a machine file pretty easily.

[tristan957]

Alternatively writing a custom dependency lookup for pytorch might be even better?

[eli-schwartz]

@tristan957 afaict that's what this PR does. It allows you to set the meson option to that value and have it work -- previously, the value of torch.utils.cmake_prefix_path was a value that cmake would search but meson would reject.

[zasdfgbnm]

Meson has a cmake_prefix_path option or --cmake-prefix-path on the command line. I feel like setting that to the output of python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)' makes more sense to me. This could be achieved with a machine file pretty easily.

I am actually setting --cmake-prefix-path to python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)' and it does not work. And this is exactly why I write this PR.

[tristan957]

Ah ok. Now I get it. Thanks for explaining a little bit more.

[tristan957]

@eli-schwartz are we good to merge it?