I am testing a shared library across multiple platforms and it has cuFFT statically linked (CUDA 9.1). Loading the lib and running the functions in C/C++ programs in Linux (Ubuntu 16.04, RHEL 6.x) and Mac (OSX High Sierra) work without any problems, but not in Python (using 3.5).
When creating the plan, cufftPlan1d returns 5, which means “Driver or internal cuFFT library error”. When I switch to dynamic linkage of cuFFT, the thing works fine. It doesn’t really disturb me as all I have to do is ship the runtime cuFFT along with my lib (it also saves seconds during compilation).
Just want to know if you guys have faced something similar, or if there is any action on the CUDA programming side that could make Python accept the static linkage (not a Python specialist here).