pgf77 support for Intel PHI

I have some legacy fortran 77 code that is not parallel, not special in any way just f77 code. Without out making ANY modifications to my code is there a way for me to take advantage of the Intel PHI cores? Does PGI support this in any capacity? I would to be able to recompile my code with some special phi or gpu magic option and viola take advantage of some additional cores on the system? Maybe this is just a pipe dream.

Hi Carl,

No, sorry, we don’t support Phi as of yet. Maybe in a few years when the next generation, Knight’s Landing, is available.

I would to be able to recompile my code with some special phi or gpu magic option and viola take advantage of some additional cores on the system?

You can try auto-parallelization (-Mconcur) when targeting a multi-core x86 processor to see if there’s any hope.

However, to really take full advantage of the new parallel architectures, you will need to modify your code. Some methods such as OpenACC are easier than others but parallel programming is still difficult.

Both NVIDIA and Intel are changing their future architectures to make things easier, NVIDIA with unified memory and NVLINK and Intel with self-hosting, but the program needs to be parallel.

  • Mat

Thanks Mat. We will be looking forward to it. We are not so much interested in taking advantage of making code parallel as we are the promise of using more cores on a single system. We really need lots of cores cheap. PHI could be a good way for many of the old f77 programs to achieve exactly that. TBD

-Carl

Possible. Though, if you’re just looking for a bunch of low powered cores, you might want to look at some of the current ARM server offerings. I don’t know enough about them to offer any insight or recommendation, but if you don’t want to wait a few years, it something to consider.

  • Mat

Interesting subject concerning ARM. Parallela has just announced some new offerings: http://www.parallella.org/2014/07/14/new-parallella-product-offerings/. Btw. are you guys aware of any development towards OpenCL support for Fortran? I think that might make quite a few people rethink a code portation since they fear vendor lock-in.