Any recent work with CUDA / OpenCL?

[Krzysztof Kamieniecki]

Hi All,

Has anyone recently been doing work with CUDA / OpenCL? The planets are aligning again and I will have time to work on this once more, but if someone else has developed something I would be more than happy to use it for my HPEC needs. 

Regards,

Krys

[Tim Holy]

Hi Krys,

That's great! I don't know of anything. In my group we may be getting to the
point in the next month or two where we might want to play with this a bit
ourselves, but we haven't done anything yet.

Best,
--Tim

[Tracy Wadleigh]

I haven't started anything myself, but am interested in seeing this happen. I've been getting my data parallelism fix with OpenMP lately, but would like to bring my graphics card to the party, too. I'm volunteering to help incubate/test the design.

--

[Diego Javier Zea]

Idem too Tracy!  I would like to help on a OpenCL/Julia binding. I dont't have enough knowledge for doit it myself, but I can offer testing and help trying to find bugs or with the API ;)

[david tweed]

On Wednesday, January 16, 2013 2:52:33 PM UTC, Krzysztof Kamieniecki wrote:

Hi All,

Has anyone recently been doing work with CUDA / OpenCL? The planets are aligning again and I will have time to work on this once more, but if someone else has developed something I would be more than happy to use it for my HPEC needs.

Hi,

no work from me on OpenCL. However, I'm still working on my array engine stuff (this isn't vaporware,honest! literally a case that I just need to actually find the off hours to code up the algorithms) and although the "code-emission" functions are currently emitting LLVM it might be interesting to add emitters for OpenCL. The only thing is that targetting OpenCL has the same problems you get with targetting C: there are things that you may know about the program behaviour that you can't express in C/OpenCL and while OpenCL is _not_ guaranteed to do any optimizations to your code to make it perform better on a given piece of hardware, it always has the ability to do what it thinks are optimizations based on it's understanding of the code you've written, even though that may mean it's actually undoing an optimization the front-end has carefully done. So I wouldn't necessarily expect targetting OpenCL would acheive close to peak performance, but it might well be effective enough that it would be better than running the code on the CPU.

Regards,
Dave

[Krzysztof Kamieniecki]

Hi David,

We may have talked about this before, but at some point we should compare notes and maybe factor out common code for building the input expressions. 

Krys

[Krzysztof Kamieniecki]

Also ...

My next steps are:

1. Make my code into a package

2. Make it work with the latest Julia and CUDA 5.0

3. Investigate the CUDA LLVM backend, and see if it is worth while to change my Julia -> PTX -> kernel workflow.

[Diego Javier Zea]

https://github.com/ddemidov/vexcl ...maybe an interesting library using OpenCL for wrapping with Julia ?