fix: conda cuda detection

[DavidLapous]

Fixes cuda file detection for linux cuda users.
Without this patch, pykeops fails to find cuda.h and switches back to cpu only mode.

From conda-nvcc's activate.sh file, the relevent variables are the following:

#...

[[ "@cross_target_platform@" == "linux-64" ]] && targetsDir="targets/x86_64-linux"
[[ "@cross_target_platform@" == "linux-ppc64le" ]] && targetsDir="targets/ppc64le-linux"
[[ "@cross_target_platform@" == "linux-aarch64" ]] && targetsDir="targets/sbsa-linux"


CUDA_CFLAGS=""
CUDA_LDFLAGS=""
if [ "${CONDA_BUILD:-0}" = "1" ]; then
    CUDA_CFLAGS="${CUDA_CFLAGS} -I${PREFIX}/${targetsDir}/include"
    CUDA_CFLAGS="${CUDA_CFLAGS} -I${BUILD_PREFIX}/${targetsDir}/include"
    CUDA_LDFLAGS="${CUDA_LDFLAGS} -L${PREFIX}/${targetsDir}/lib"
    CUDA_LDFLAGS="${CUDA_LDFLAGS} -L${PREFIX}/${targetsDir}/lib/stubs"
    CUDA_LDFLAGS="${CUDA_LDFLAGS} -L${BUILD_PREFIX}/${targetsDir}/lib"
    CUDA_LDFLAGS="${CUDA_LDFLAGS} -L${BUILD_PREFIX}/${targetsDir}/lib/stubs"
    # Needed to fix cross compilation.
    # $PREFIX and $PREFIX/${targetsDir} are needed to properly find all host components
    # $BUILD_PREFIX/$HOST/sysroot is needed to find compiler bits and is placed before any `targetsDir` for priority.
    # $BUILD_PREFIX/${targetsDir} is needed for projects that don't enable the CUDA language but use FindCUDAToolkit
    export CMAKE_ARGS="${CMAKE_ARGS} -DCMAKE_FIND_ROOT_PATH=$PREFIX;$BUILD_PREFIX/$HOST/sysroot;$PREFIX/${targetsDir};$BUILD_PREFIX/${targetsDir}"
else
    CUDA_CFLAGS="${CUDA_CFLAGS} -I${CONDA_PREFIX}/${targetsDir}/include"
    CUDA_LDFLAGS="${CUDA_LDFLAGS} -L${CONDA_PREFIX}/${targetsDir}/lib"
    CUDA_LDFLAGS="${CUDA_LDFLAGS} -L${CONDA_PREFIX}/${targetsDir}/lib/stubs"
fi
export CFLAGS="${CFLAGS} ${CUDA_CFLAGS}"
export CPPFLAGS="${CPPFLAGS} ${CUDA_CFLAGS}"
export CXXFLAGS="${CXXFLAGS} ${CUDA_CFLAGS}"
export LDFLAGS="${LDFLAGS} ${CUDA_LDFLAGS}"

# ...

https://github.com/conda-forge/cuda-nvcc-feedstock/blob/e4330bbc991b31e6f3e775524667b806f5495dcf/recipe/activate.sh#L5-L31

So I just added the different possible targetDirs to the keopscore file detection pipeline, and it worked on my laptop.

[Nizben]

Thank you for the PR!
Yes indeed, you are right, i guess things were working for me because I was using cudatoolkit-dev instead of cudatoolkit, which needs the patch you added (as opposed to cudatoolkit-dev which automatically puts the cuda headers in the env/include path).
For instance this is my current cuda_include_path: /home/nizarben/miniconda3/envs/pykeops_dev/include in which I can find cuda.h, no need for adding/specifiying the arch in the path.
Thanks for pointing that out, will fix very soon!

[DavidLapous]

Ah yes, I'm not using cudatoolkit-dev indeed, thanks for the info.
TBH I just noticed this now since the upgrade from fedora 41 to 42 killed cuda on my system (but the conda version seems to work fine with keops indeed).
I also see a link to lib/stubs that I didn't see in your cuda parser (only the lib part) ; it works fine on my system / code, but this may be problematic in another situation.

[Nizben]

Can you give me more context on the 'lib/stubs' please? Which cuda parser are you talking about? What situation could it be problematic in?
Thanks

[DavidLapous]

Sorry this wasn't very clear ; I was talking about this specific line in the conda recipe that I shared in the first message

    CUDA_LDFLAGS="${CUDA_LDFLAGS} -L${CONDA_PREFIX}/${targetsDir}/lib/stubs"

which contains "stubs" that I didn't check

and the cuda parser I was thinking about are lines similar to :

keops/keopscore/keopscore/config/cuda.py

Lines 366 to 369 in 0152630

	self.nvrtc_flags = (
	compile_options
	+ f" -fpermissive -L{libcuda_folder} -L{libnvrtc_folder} -lcuda -lnvrtc"
	)