GNU and GPU programming

[aihaike]

Dear all,

I wonder if there is any chance to one day be able to code a graphic card (i.e NVedia and CUDA) with GNU software (i.e gfortran) like PGI dos with its fortran compiler.
Thanks.

[Rubberman]

Yes. Visit the nVidia web site. They have a lot of documentation and such for Linux/CUDA programming. I'd post it all here, but all the docs, drivers, and such come to almost 700MB...

[aihaike]

Thanks for the reply,

Well, so far it seems to me that only PGI proposes a CUDA fortran compiler which include OpenMP like directives. From my knowledge no free compiler does that.
I may be wrong ...

[Rubberman]

There is a CUDA driver and SDT toolkit that allows parallel C programming with GNU toolchain compilers. All of these are available on the nVidia web site. There is also a commercial 3rd party Fortran compiler available that has a free for personal/educational use license that you can use if you REALLY want to use Fortran. However, that is becoming more and more rare these days. Even the HEP (High Energy Physics) research community that used to use a lot of Fortran is using C++ for almost all current and new software for HEP research purposes such as data acquisition, reduction, analysis, etc. Fortran is only used for legacy code bases that haven't been migrated to C++ as yet. So, you should be able to use the MPI libraries in conjunction with the CUDA SDK and Linux system drivers in order to get serious parallelization of most floating point math code.

[aihaike]

thank you for your and sorry for my slow reply.
Even I'm not agree with you about Fortran, your answer is very helpful.
Thanks,

Eric.