I am a PGI and CUDA Fortran newbie.
I am using a third party software (so no source code, only binary is available). This software allows me to link to a BLAS library of my choice at runtime i.e. I have to supply the path to the blas library on the command line at invocation, e.g - blaspath /lib/gotoblas.so. The software will recognise any BLAS library that supports the standard Fortran BLAS interface.
I would like to have a library based on CUDA Fortran and CuBLAS so that I can supply that library to this program and run the BLAS calls on GPU. I suppose therefore that I need some *.so library that is a wrapper around CuBLAS so that the external program does not know the difference from say MKL BLAS, but the BLAS calls are actually run on the GPU.
I am wondering if this is possible and if so, how should I go about setting this up?
Thanks for any help on this,