pgfortran compiling error

Accelerated Computing HPC Compilers Legacy PGI Compilers
#1

Hy Guys ,

I have installed the latest PGI version and I have a NVIDA Geforce GT 230M with kernel driver version 340.107.

when I compile the deviceQueery.cuf in the examples directory I have the following errors:


[/code]


/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseCcsrgemm2”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseDcsrgemm2_bufferSizeExt”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseZcsrgemm2_bufferSizeExt”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cudaLaunchKernel”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseScsrgemm2”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseZcsrgemm2”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseCcsrgemm2_bufferSizeExt”
./libcudaforwrapblas.so: riferimento non definito a “cublasDgetrsBatched”
./libcudaforwrapblas.so: riferimento non definito a “cublasSgetrsBatched”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseScsrgemm2_bufferSizeExt”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseZcsrcolor”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseDcsrgemm2”
./libcudaforwrapblas.so: riferimento non definito a “cublasCgetrsBatched”
./libcudaforwrapblas.so: riferimento non definito a “cublasZgetrsBatched”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseCcsrcolor”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseScsrcolor”
/usr/local/pgi/linux86-64/18.4/lib/libcudafor.so: riferimento non definito a “cusparseDcsrcolor”
pgacclnk: child process exit status 1: /bin/ld


I use pgfortran deviceQuery.cuf -lcublas -lblas -lcusparce

I suppose it is linkend to the fact that he cannot find the library path. Any Idea?
Could you help me to solve this problem?

Thank you in advance

#2

Hi Emrys,

Do you have the CUDA libs installed somewhere locally? If so you can specify the path to it directly with -L/path/to/cuda/lib64 on your compile line. If you’ve installed the CUDA libs as part of the installation script, I think you can specify the libs with -Mcudalib=cusparse,cublas.

The bigger problem here is that the GT 230M has CC 1.2 (https://developer.nvidia.com/cuda-legacy-gpus) which we no longer support. CC 3.0 is the minimum. I think around version 15 CC 1X was deprecated, so if you are targeting an older GPU version 14.10 is probably going to be your best bet (or seek out an accelerator with a higher CC)