CUBLAS routines using Portland under Centos6 and Cuda 5

Accelerated Computing GPU-Accelerated Libraries
1

I want to use the cublas routines from Portland fortran or in some caes intel. The fotran.c fail to compile. I do not have g95 available. I have gcc suite 4.8.0 (gcc/gfortran), Portland 13.3 and Intel 2013.3 available and new Kepler architecture cards. The SDK examples work fine, Portland accelerator work fine. However, I would like to be able to call sgemm and dgemm (c and z too) from Fortran compiled with the Portland compiler.

How do I build the fortran.c wrapper so I can get a library that I can link in and call ?gemm from fortran using Portland or Intel ?

The performance in the matrix multiply examples in the SDK yield over one Tflops/s which is quite good.
Hence my wish to call the highly optimized blas routines.

Ole

2

fortran.c should compile fine using gfortran

What is your compilation problem?

3

It is not clear to me which “Portland Fortran” product you have. Your statement “SDK examples work fine” presumably refers to the CUDA SDK, and I take that as an indication that you are working with PGI CUDA Fortran.

If so, this compiler comes with a handy module that interfaces to CUBLAS. Simply add “use cublas” to your host program. A worked example of CUBLAS calls via this module is on page 56 of the CUDA Fortran Programming Guide and Reference, an online copy of which can be found at [url]http://www.pgroup.com/lit/whitepapers/pgicudaforug.pdf[/url]