Check for feedstock use of compiler-bindir as well as -ccbin

[mabraham]

Solution to issue cannot be found in the documentation.

• I checked the documentation.

Issue

recipe/linux/nvcc.sh tries to avoid adding a second -ccbin argument to nvcc, but it does not check for the use of --compiler-bindir. The latter synonym was introduced in recent years.

I think nvcc.sh should check for both, per https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#compiler-bindir-directory-ccbin

This is currently causing issues with the gromacs-feedstock at conda-forge/gromacs-feedstock#17, whose CMake tries to check nvcc for compatible flags and experiences warnings about the use of multiple such flags.

Installed packages

not applicable

Environment info

not applicable

[mabraham]

The case I'm working on also complains about cicc not being found. What does that mean, please? https://dev.azure.com/conda-forge/feedstock-builds/_build/results?buildId=719647&view=logs&j=bb1c2637-64c6-57bd-9ea6-93823b2df951&t=350df31b-3291-5209-0bb7-031395f0baa1&l=1356

[jakirkham]

There is this line that could be tweaked somehow

nvcc-feedstock/recipe/linux/nvcc.sh

Line 4 in 04b48c9

case ${arg} in -ccbin)

Though note nvcc is used on CUDA 11. It is a wrapper script that calls an existing nvcc (in conda-forge the nvcc binary comes from the Docker image)

The CUDA 12 compiler package is in the cuda-nvcc feedstock. It currently lacks this logic for either flag. So it looks like we need to add this there. Have filed issue ( conda-forge/cuda-nvcc-feedstock#20 ) as well