Should cudatoolkit set NUMBAPRO_CUDALIB on Windows?

[MattWenham]

Not quite sure exactly who this is an issue for, but here goes.

I finally got numba -s to correctly recognise my GPU by using cudatoolkit (current version 10.0) and not the nVidia CUDA SDK (currently at version 10.1). I got the following output:

__CUDA Information__
Found 1 CUDA devices
id 0        b'GeForce MX130'                              [SUPPORTED]
                      compute capability: 5.0
                           pci device id: 0
                              pci bus id: 3
Summary:
        1/1 devices are supported
CUDA driver version                           : 10010
CUDA libraries:
Finding cublas
        ERROR: can't locate lib
Finding cusparse
        ERROR: can't locate lib
Finding cufft
        ERROR: can't locate lib
Finding curand
        ERROR: can't locate lib
Finding nvvm
        named  nvvm64_33_0.dll
        finding libdevice for compute_20...     ok
        finding libdevice for compute_30...     ok
        finding libdevice for compute_35...     ok
        finding libdevice for compute_50...     ok

This page helped me realise that I needed to manually set NUMBAPRO_CUDALIB. I searched for cublas64_100.dll and set NUMBAPRO_CUDALIB manually, I then got output which showed that the libraries had been found:

__CUDA Information__
Found 1 CUDA devices
id 0        b'GeForce MX130'                              [SUPPORTED]
                      compute capability: 5.0
                           pci device id: 0
                              pci bus id: 3
Summary:
        1/1 devices are supported
CUDA driver version                           : 10010
CUDA libraries:
Finding cublas
        named  cublas64_100.dll
        trying to open library...       ok
Finding cusparse
        named  cusparse64_100.dll
        trying to open library...       ok
Finding cufft
        named  cufft64_100.dll
        trying to open library...       ok
Finding curand
        named  curand64_100.dll
        trying to open library...       ok
Finding nvvm
        named  nvvm64_33_0.dll
        trying to open library...       ok
        finding libdevice for compute_20...     ok
        finding libdevice for compute_30...     ok
        finding libdevice for compute_35...     ok
        finding libdevice for compute_50...     ok

Should cudatoolkit be setting NUMBAPRO_CUDALIB in Windows (and perhaps other OSes too?) to save others this pain? Or is this a matter for numba itself?

[seibert]

Numba should be searching for the cudatoolkit where it is installed in the conda environment and not relying on the environment variable, which is intended to override the default behavior. If Numba is failing to find the required libraries, that is a Numba bug.

[MattWenham]

Thanks, I'll bounce this over to numba.

[stuartarchibald]

Think this got fixed here numba/numba#3773!?