I have a code which compiles (and executes) correctly with de 11.1 version of the cuda fortran compiler. When I try not compile it with the 11.6 version, I get the following message:
/opt/pgi/linux86-64/11.6/libso/libcudafor.so: undefined reference to cublasAlloc' /opt/pgi/linux86-64/11.6/libso/libcudafor.so: undefined reference to cublasFree’
I installed the cuda 4.0 toolkit, before AND after the 11.6 version, to no avail.
We were able to build your example when we removed the ‘-mcmodel=medium’
from your Makefile. We have filed TPR 17972 to address -mcmodel=medium.
If you do not need -mcmodel=medium to compile and link your program
(because the statically declared data is <2Gb collectively), you are better off
without it. Better optimizations and more flexible linkage follows.