nvcc package should constrain c / cxx compiler versions based on cuda version

[kkraus14]

Compatibility matrix: https://gist.github.com/ax3l/9489132

Latest versions supported:

CUDA Version	GCC Version
9.2	7
10.0	7
10.1	8
10.2	8
11.0	9
11.1	10

[jaimergp]

I was thinking about this too. Is this to be addressed as part of the global pinnings at conda-forge-ci-pinning via zipped_keys?

[jakirkham]

IIRC Keith we are only worried about this for packages that actually contain CUDA kernel code right? Not packages that merely link to CUDA libraries. Is that correct or am I missing something?

[kkraus14]

Any code that compiled with nvcc, so even if they don't have device code if they use a .cu extension it will go through nvcc.

[leofang]

I think this should be fixed asap, as Conda-Forge has updated the global pinning to gcc9, and cupy-feedstock starts to see tons of build errors due to incompatible host compiler versions: conda-forge/cupy-feedstock#80 (comment).

[leofang]

Any code that compiled with nvcc, so even if they don't have device code if they use a .cu extension it will go through nvcc.

Even worse, any host code that is compiled by host compilers and includes cuda_runtime.h will also be affected (see the cupy error linked above).

[kkraus14]

Looks like it's more than just libraries with device code based on the above.

[jakirkham]

What happens if we build against GCC 7 and the runtime library from GCC 9 is pulled in by the compiled packages?

[kkraus14]

What happens if we build against GCC 7 and the runtime library from GCC 9 is pulled in by the compiled packages?

As far as I'm aware that should be fine, there's just compiler checks in the headers that prevent using too new of a compiler.

[leofang]

Update: with conda-forge/conda-forge-pinning-feedstock#1052 all CUDA builds fall back to gcc 7 now, so it works for all CUDA versions (9.2 - 11.0). It'd be great to enable gcc 9 for CUDA 11.0+ only.

[jakirkham]

Does PR ( conda-forge/conda-forge-pinning-feedstock#2010 ) sufficiently address this concern @kkraus14? Or are there other things still needed?

[kkraus14]

Does PR ( conda-forge/conda-forge-pinning-feedstock#2010 ) sufficiently address this concern @kkraus14? Or are there other things still needed?

I believe it generally handles things for building packages within conda-forge, but if someone was to use the nvcc and gcc packages themselves they can still land themselves in an incompatible situation as far as I can tell.

[jakirkham]

Suppose we could add run_constrained to the package

[robertmaynard]

The compatiblity matrix listed above and the one on the CUDA Linux installation guide ( https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html ) list the maxium supported gcc version, not the minimum.

This is clarified in footnote 2 of the CUDA docs where CUDA 11.X requires at least gcc 6.