You are here

Any info on 1) building for CUDA and 2) with the PGI compiler suite

3 posts / 0 new
Last post
Any info on 1) building for CUDA and 2) with the PGI compiler suite
#1

In looking to try out a build of Rosetta so that a researcher in another
department can try it out on a large memory resource we have here, I
noticed, because I ended up looking at this file, that there's a stanza
in

tools/build/basic.settings

that appears to allow for compilation against a CUDA toolkit.

As I am not a Rosetta user but an implementer, I was wondering if there's
a simple test/demo that I could try out, if I build such a variant, which
would be an Ubuntu 10.10 (GCC 445) with NVIDIA_GPU_Computing_SDK-4.0.

Just to say that I grabbed the full 3.4 bundle but there doesn't appear to
be a demo with anything akin to CUDA in the demo directory names, nor has
a search within the forums for "CUDA" turned up anything useful either.

On a second resource I look after, I have access to the PGI compiler suite
and was wondering if anyone has tried to compile Rosetta using that?

Presumably it is "just a matter of adding a stanza for pgiCC and choosing
the optimisation flags", for some value of "optimisation flags"?

Any pointers welcome,

Kevin M. Buckley
ECS, VUW, NZ.

Post Situation: 
Sun, 2012-07-01 23:07
vuw_ecs_kevin

We don't have much, or maybe any, public GPU compatibility. There are a few algorithms ported to GPU (Luki ported sc, which is a shape complementarity algorithm; I don't know if that's in 3.4) but 99% of Rosetta is still CPU-only. So, yes, you can compile in GPU support, but it only has any effect in one very narrow case. Further GPU work is under development.

I am not aware of anybody working with PGI compilers (it would be in tools/build). Your interpretation of what is necessary is correct. I believe I have access to PGI compilers so if you create a solution I can minimally test it and put it in the codebase. GCC is eternally popular for production work, with some use of ICC. A lot of us use clang for development but not production (which may indicate lack of clang on the supercomputers, not a deficiency of clang).

Mon, 2012-07-02 11:15
smlewis

> We don't have much, or maybe any, public GPU compatibility.

OK. Just thought to ask

> ... not aware of anybody working with PGI compilers ... if you create a solution
> I can minimally test it and put it in the codebase.

OK. We'll see where we can take it

Kevin M. Buckley
ECS, VUW, NZ.

Mon, 2012-07-02 16:34
vuw_ecs_kevin