Re: NCL fortran->Nvidia GPU porting

From: Rick Brownrigg <brownrig_at_nyahnyahspammersnyahnyah>
Date: Wed Aug 17 2011 - 09:34:09 MDT

Hi Elmer,

Those are some impressive gains, and yes, I have to think people would want more! ;-)

I believe Parvis uses MPI and targets distributed compute nodes. In an orthogonal effort, this summer we began looking into how we might leverage GPUs found in commodity machines. We examined just a handful of potential operations that could be parallelized, and the results were positive (not nearly as dramatic as yours, however).

One issue of concern for us is portability. To address that, our implementation leveraged OpenCL, which can target multiple vendors' GPUs as well as multicore CPUs.

What are your thoughts in terms of sharing your efforts? Do you have other functions/algorithms in mind that might be ripe for parallelizing?

Thanks for letting us know about your efforts.


On Aug 16, 2011, at 10:05 AM, Elmer Joandi wrote:

> Hi,
> I ported in springtime for fun cape 2d calculus from NCL fortran to GPU, got it about 100x faster (dual versus single precision, on singe corei7 CPU core versus casual Nvidia GPU, 300 secs versus 3 secs on 500x500 domain ).
> I wonder -
> a) what is the legal status of this code ?
> b) would anyone want it or more of such porting ?
> c) would ParVis have MPI or GPU or both ?
> --
> Elmer Joandi
> +3725185428
> _______________________________________________
> ncl-talk mailing list
> List instructions, subscriber options, unsubscribe:

ncl-talk mailing list
List instructions, subscriber options, unsubscribe:
Received on Wed Aug 17 09:34:15 2011

This archive was generated by hypermail 2.1.8 : Wed Sep 07 2011 - 10:58:58 MDT