Activate script in conda package contains hardcoded /usr/local/cuda

[pearu]

Issue:
The activate script in conda package, say in https://anaconda.org/conda-forge/nvcc_linux-64/10.1/download/linux-64/nvcc_linux-64-10.1-h3d80acd_0.tar.bz2, contains hardcoded /usr/local/cuda and ignores CUDA_HOME environment variable:

#!/bin/bash

if [[ ! -d "/usr/local/cuda" ]]
then
    echo "Cannot find: /usr/local/cuda"
    exit 1
fi
if [[ ! -f "/usr/local/cuda/lib64/stubs/libcuda.so" ]]
then
    echo "Cannot find: /usr/local/cuda/lib64/stubs/libcuda.so"
    exit 1
fi
grep "CUDA Version 10.1" /usr/local/cuda/version.txt &>/dev/null
if [[ 0 -ne 0 ]]
then
    echo "Version of installed CUDA didn't match package"
    exit 1
fi

export CUDA_HOME="/usr/local/cuda"
export CFLAGS="${CFLAGS} -I${CUDA_HOME}/include"
export CPPFLAGS="${CPPFLAGS} -I${CUDA_HOME}/include"
export CXXFLAGS="${CXXFLAGS} -I${CUDA_HOME}/include"

It looks like install_nvcc.sh (that creates the activate script) should escape using CUDA_HOME when building the conda package.

[kkraus14]

Those values are currently generated from the install_nvcc.sh script, but aren't very friendly to a user changing their environment / CUDA version which is userspace. @jakirkham @mike-wendt @raydouglass any thoughts on how we should best handle this?

[kkraus14]

As someone who bounces between multiple installed CUDA versions, I would be in favor of moving the CUDA_HOME handling into the activate.d script so that if I change my user environment to switch from say CUDA 9.2 to 10.1 that it's nicely handled.

[isuruf]

I raised this issue in conda-forge/staged-recipes#8229 and @jakirkham mentioned that he did not want to support building outside of the conda-forge docker image. But we should support that in my opinion.

[jakirkham]

If there are other people willing to handle issues that come up from this addition, I wouldn't be opposed to seeing things become more flexible.