Fail to download CUDA using the python in $EPREFIX

[ocaisa]

Ran into an issue when trying to install CUDA using the Python from $EPREFIX:

== FAILED: Installation ended unsuccessfully (build directory: /tmp/CUDAcore/11.1.1/system-system): build failed (first 300 chars): Unexpected error occurred when trying to download https://developer.download.nvidia.com/compute/cuda/11.1.1/local_installers/cuda_11.1.1_455.32.00_linux.run to /project/def-sponsor00/easybuild/sources/c/CUDAcore/cuda_11.1.1_455.32.00_linux.run: signed integer is greater than maximum (took 0 sec)

It looks like this is a bug in Python 3.8 (from https://bugs.python.org/issue42853)

[boegel]

See also easybuilders/easybuild-framework#3455

[ocaisa]

Building some CUDA examples also fails, I think this might be due to the fact that the prefix GCC compiler is ahead of the foss GCC (GCC10 versus GCC9):

[ocaisa@gpu-node1 reduction]$ make HOST_COMPILER=$(which g++)
>>> GCC Version is greater or equal to 4.7.0 <<<
/project/def-sponsor00/easybuild/software/CUDAcore/11.1.1/bin/nvcc -ccbin /cvmfs/pilot.eessi-hpc.org/2020.12/software/x86_64/amd/zen2/software/GCCcore/9.3.0/bin/g++ -I../../Common -m64 --std=c++17 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_86,code=compute_86 -o reduction_kernel.o -c reduction_kernel.cu
/cvmfs/pilot.eessi-hpc.org/2020.12/compat/linux/x86_64/usr/include/math.h: In function 'int iseqsig(_T1, _T2)':
/cvmfs/pilot.eessi-hpc.org/2020.12/compat/linux/x86_64/usr/include/math.h:1374:39: error: expected ')' before numeric constant
 1374 |   typedef decltype (__MATH_EVAL_FMT2 (__x, __y)) _T3;
      |                                   ~   ^~~
      |                                       )

[ocaisa]

I guess it is a problem with the CUDA samples package, it is picking up the math.h from the prefix rather than from GCCcore, not sure why

[ocaisa]

This no longer seems to be an issue