Message boards :
Number crunching :
OpenCL/GPU app?
Message board moderation
Author | Message |
---|---|
Send message Joined: 19 Aug 11 Posts: 31 Credit: 63,078,601 RAC: 22,236 |
Any chance of an app for our GPUs? I know some math does not port well to GPUs. But maybe this one would? Reno, NV Team: SETI.USA |
Send message Joined: 8 Jul 11 Posts: 1318 Credit: 404,002,318 RAC: 289,341 |
Any chance of an app for our GPUs? I know some math does not port well to GPUs. But maybe this one would? Greg and I discussed this a while back, and came to the conclusion that it might be possible but would require alot of work. We are using the pari/gp library to do the number theoretic calculations, and although this is open source code, it would require non-trivial modifications. We are also using the gmp library, which may not port well. But I am not an expert on GPUs, so maybe someone else can chime in here with some better insight. Eric |
Send message Joined: 19 Aug 11 Posts: 76 Credit: 2,002,860 RAC: 0 |
Any chance of an app for our GPUs? I know some math does not port well to GPUs. But maybe this one would? well, at least GMP-ECM has a working CUDA-branch, you don't happen to be using that lib? primegrid recently started a world-record hunt for Generalized Fermat-primes using CUDA-apps. current runtimes are up into several days on fastest GPU's, but would take months most likely on current high-end CPU's. that app allready owns the top 4 spots. ;) |
Send message Joined: 8 Jul 11 Posts: 1318 Credit: 404,002,318 RAC: 289,341 |
No, we need the standard GMP. And even if GMP had cuda support, we still have the pari/gp library to deal with. We use pari for computing integral bases, checking irreducibility, computing field discriminants, and also factoring large integers. It's a large amount of code that would need to be "cudified". It would be a fun undertaking, if only I had the time. |
Send message Joined: 13 Sep 11 Posts: 5 Credit: 5,004,016 RAC: 0 |
Well... i'm not sure why you don't upload a copy of the code somewhere for users to have a look at it. Maybe one will have time to look at it an port it to use a GPU. Actually, since you're using GPL'ed GMP and PARI, you'll have to release your source code upone request (cosider this post a request ;) ), or otherwhise you're violating GPL. TIA |
Send message Joined: 8 Jul 11 Posts: 1318 Credit: 404,002,318 RAC: 289,341 |
I may just do that. I'd like to clean it up first, before I open it up for the whole world to see. Maybe someone will even see a way to make it more efficient; after all I'm really a mathematician/engineer and not a programmer. |
Send message Joined: 8 Jul 11 Posts: 46 Credit: 7,144,042 RAC: 0 |
You can link userspace apps to GPL libraries and not encumber under the GPL. I'm all for posting the source though. |
Send message Joined: 19 Aug 11 Posts: 76 Credit: 2,002,860 RAC: 0 |
I'm all for posting the source though. where can i sign? ;) |
Send message Joined: 1 Jul 12 Posts: 13 Credit: 2,099,843 RAC: 0 |
Eric, Any news on the GPU front? I would dearly love to have my GPU crunch away on this project. S. |
Send message Joined: 8 Jul 11 Posts: 1318 Credit: 404,002,318 RAC: 289,341 |
Eric, Well, it would be a ton of work and right now I just don't have the time. But I'll find some time this weekend to post the source code and see if any expert GPU programmers want to take a crack at it. |
Send message Joined: 3 May 13 Posts: 13 Credit: 102,701 RAC: 0 |
Has anyone rethought the GPU tasking issue in the last year? Thanks |
Send message Joined: 8 Jul 11 Posts: 1318 Credit: 404,002,318 RAC: 289,341 |
We had a volunteer named Erno who actually did the OpenCL port and got something working. If I recall, a portion of the computation was still being done on the cpu, and the overall speed improvement was something like 20%. We've been meaning to look at it some more to see if we could find ways to improve the speed. |
Send message Joined: 3 May 13 Posts: 13 Credit: 102,701 RAC: 0 |
We had a volunteer named Erno who actually did the OpenCL port and got something working. If I recall, a portion of the computation was still being done on the cpu, and the overall speed improvement was something like 20%. We've been meaning to look at it some more to see if we could find ways to improve the speed. With WCG's completion of their HCC1 project, crunchers will be looking for a home for their GPUs. A WU that can function with an app_config.xml configuration of 0.2C + 0.2ATI (as the Hcc1 did) would surely expedite the completion of your projects as well. It is my understanding that WCG realized their project's completion two or three years ahead of time. That fact by itself should stimulate the interest of any Project coordinator. I am not a programmer, but IMO the GPU trick is to keep the WUs small. Respectfully, Gandalf |
Send message Joined: 25 Feb 13 Posts: 216 Credit: 9,899,302 RAC: 0 |
Hi Eric, Any news on the GPU front? I found this:http://geraldine.fjfi.cvut.cz/mmg/index.php?page=cuda CUDA-expert from the "Department of Mathematics at the Czech Technical University in Prague" can help us. They have an CUDA-Research-Center. The group-site by NVIDIA: http://research.nvidia.com/content/ctu-crc-summary Take a look at it, if you want help. ;) I think they will help this project. |
Send message Joined: 8 Jul 11 Posts: 1318 Credit: 404,002,318 RAC: 289,341 |
Hi Eric, Hi Aurel, Sorry, no news. I've just been too busy to take a look at it. A while back, one of our volunteers (Erno) got something that worked but it offered little speed improvement. If I recall, some of the number field calculations were still being done on the cpu side due to the complication of CUDA-fying the pari libraries. When I get a chance later, I will take a look at the link. Thanks! |
Send message Joined: 13 Sep 11 Posts: 5 Credit: 5,004,016 RAC: 0 |
... A while back, one of our volunteers (Erno) got something that worked but it offered little speed improvement. ... How about realeasing the apps source for the CPU app ? |
Send message Joined: 25 Feb 13 Posts: 216 Credit: 9,899,302 RAC: 0 |
Hi all, I was today an the Hannover Fair 2014. I meet some interesting persons, I´ll explain it on a single thread. ;) To the GPU_App: We all know, MATLAB supports GPU_Computing and Pari not. [Thats the problem...] We only need an little part from the MATLAB-programm, the mathematical part. With look I´ll become an one-year-testing-lizenz for COMSOL. [COSMOL is an physical_mathematical computing programm; one lizenz costs ~50.000€] COMSOL is written in C++ and a little bit Java. [I understand C++] I can copy the code and write an single version. Some other ways? Yes. Using "R". [An spezial programming language] No experince about "R", but I have an open-source programm for "R". [I don´t know who to programm it...] Thats just some small informations. My test-version will come in 2-3 weeks. ;) Greets, Aurel |
Send message Joined: 13 Sep 11 Posts: 5 Credit: 5,004,016 RAC: 0 |
Hi all, I don't care about Pari NOR MATLAB at ALL! I just need the source code of the cpu app that's all.. |
Send message Joined: 25 Feb 13 Posts: 216 Credit: 9,899,302 RAC: 0 |
Hi Crunch3r, please read the PM I´ve written to you. ;) Pari has an compilation problem, they can´t fix it. This problem is only at Windows OS. I don´t know, if an GPU-App would run with an Pari libary. |
Send message Joined: 13 Sep 11 Posts: 5 Credit: 5,004,016 RAC: 0 |
Hi Crunch3r, Yes, that's caused by some stupid pointer arithmetic... Ten years ago that was a common issue, but the NumberFields app tops it,with what i'd call "code retards in practice".... Anyway, i have not had the time to fix this garbage code(getting rid of the pari crap etc.)... Had to fix DistRTGens OCL MD5 code for my own AMD GPUs first...(Yes, credits matter to me!!! Fuck yourself, science whores!!!) Soooooo... to be continued. |