CUDA?
 

Hi,

I've got a GeForce GTX 285, and 'd like to know, if this can be used for the CRUS project? Is there any sieving or PRP software available, which supports CUDA? Or, is anything in development?

Regards,
Xentar

None that I know of.
Someone has made an [URL="http://www.mersenneforum.org/showthread.php?t=12576"]FFT library for CUDA[/URL], (still somewhat experimental, although many DCs on GIMPS candidates have successfully been run) so it should be possible for that to be implemented into a program to run k*b^n+c (instead of just 2^p-1, like it currently is...I think) numbers. But AFAIK nothing that'd work for CRUS exists in a ready-to-use form.
(I don't know of any CUDA sieving software, closest is [URL="http://www.mersenneforum.org/showthread.php?t=12827"]TF for Mersenne numbers[/URL])

[quote=Mini-Geek;209816]None that I know of.
Someone has made an [URL="http://www.mersenneforum.org/showthread.php?t=12576"]FFT library for CUDA[/URL], (still somewhat experimental, although many DCs on GIMPS candidates have successfully been run) so it should be possible for that to be implemented into a program to run k*b^n+c (instead of just 2^p-1, like it currently is...I think) numbers. But AFAIK nothing that'd work for CRUS exists in a ready-to-use form.
(I don't know of any CUDA sieving software, closest is [URL="http://www.mersenneforum.org/showthread.php?t=12827"]TF for Mersenne numbers[/URL])[/quote]
I imagine that the GIMPS LL test CUDA application wouldn't be too hard to modify for LLR tests (for the uninitiated, that takes care of Riesel base 2 and Riesel power-of-2 bases but nothing else) since the LLR test is based on the LL test and I think there's only a couple differences in the initialization. Proth tests (Sierp. base 2 and power-of-2 bases) and PRP tests (everything else) might require a bit more work since a wholly new algorithm would need to be coded, though I think Proth and PRP tests are somewhat related algorithmically so once one is done the other shouldn't be as hard.

If I had a GPU I'd be all over the LL test development effort, helping with testing and the like so that it can be moved forward into an LLR version (which would not only help CRUS but NPLB and numerous other projects as well--PrimeGrid for one would surely be quite interested). But alas, when I built my computer I didn't see a nice GPU as much of a priority since I don't play 3D games. :smile: Sometime or other I may invest in one though for the purpose of crunching; since all my computers have integrated graphics it wouldn't even interfere with normal use of the computer.

Ok, thank you for the answers.
This means, I have to wait, till there is a new software available.

Or do I have to learn how to program CUDA myself? :devil:

[quote=Xentar;209961]Or do I have to learn how to program CUDA myself? :devil:[/quote]Sounds like a good idea to me.

Paul

Have you any tips for making CUDA programming less frustrating? I found that, once I'd once written outside the bounds of an array, my machine was (unsurprisingly, I suppose) prone to X-related misbehaviour including horrible crashes unless I rebooted it; this was sufficiently troublesome to have happen on my workstation that I've not written a line of CUDA for six months.

I'm not really prepared to get a sacrificial machine that I can ssh to and run CUDA with a reboot after every segfault; maybe the answer is to spend most of my time developing on the simulator and only using the hardware on stuff that I'm confident is correct, but this is still a pretty convoluted way to go when my goal's to get the code fast on the hardware.

[quote=fivemack;210120]Have you any tips for making CUDA programming less frustrating? I found that, once I'd once written outside the bounds of an array, my machine was (unsurprisingly, I suppose) prone to X-related misbehaviour including horrible crashes unless I rebooted it; this was sufficiently troublesome to have happen on my workstation that I've not written a line of CUDA for six months.

I'm not really prepared to get a sacrificial machine that I can ssh to and run CUDA with a reboot after every segfault; maybe the answer is to spend most of my time developing on the simulator and only using the hardware on stuff that I'm confident is correct, but this is still a pretty convoluted way to go when my goal's to get the code fast on the hardware.[/quote]CUDA 3.0 has a memory bouds checker, or so it appears. I only downloaded 3.0 a couple of days ago and haven't yet used it enough to see whether it addresses your situation.

Device emulation is now deprecated and on-device debugging is the way forward. Not sure when the emulator will be discontinued.


Paul