Windows build for GPU CUDA code

Brian Gladman · 2012-02-16, 16:59

I have just added a Visual Studio 2010 build for the GPU stage 1 code on Windows x64 to the SVN (it might work on win32 but I have not tested it).

As Paul L has already said, this code is only for those who are brave (or foolhardy) since it requires that the project build properties in Visual Studio are properly set up for your GPU configuration before it is built.

Please seek advice, discuss or report issues here.

Brian Gladman

NBtarheel_33 · 2012-02-17, 06:43

Quote:

Originally Posted by Brian Gladman

I have just added a Visual Studio 2010 build for the GPU stage 1 code on Windows x64 to the SVN (it might work on win32 but I have not tested it).

As Paul L has already said, this code is only for those who are brave (or foolhardy) since it requires that the project build properties in Visual Studio are properly set up for your GPU configuration before it is built.

Please seek advice, discuss or report issues here.

Brian Gladman

"GPU Stage 1" - is this going to be GPU code for P-1?

Karl M Johnson · 2012-02-18, 05:55

It's for ecm from GMP-ECM package.

henryzz · 2012-02-19, 19:17

Quote:

Originally Posted by NBtarheel_33

"GPU Stage 1" - is this going to be GPU code for P-1?

While it would be possible to code p-1 for gpu the way gpu-ecm gains its speed is by running multiple curves at once. P-1 is one curve so that wouldn't work.

TimSorbet · 2012-02-20, 03:47

Quote:

Originally Posted by henryzz

While it would be possible to code p-1 for gpu the way gpu-ecm gains its speed is by running multiple curves at once. P-1 is one curve so that wouldn't work.

Would it be possible to simply run P-1 on multiple numbers?

henryzz · 2012-02-20, 15:46

Quote:

Originally Posted by Mini-Geek

Would it be possible to simply run P-1 on multiple numbers?

Dunno.

jasonp · 2012-02-20, 20:48

Yes, if you have to do batch factorization of many different numbers then a GPU would work nicely, but you will need separate state for each number. In fact there's a cottage industry within the number field sieve community to work out a method where many bazillions of small numbers, maybe 80-150 bits each, can get factored with ECM and P+-1. This is the 'NFS cofactorization problem' and graphics cards look like an ideal way to solve it. They're much cheaper and possibly much faster than FPGAs, and programming them is a snap with modern toolkits so you don't need to know any hardware design.

Robert Holmes · 2012-02-24, 11:39

Some interesting developments as of yesterday:

Improved addition chains for Edward's curve ECM: http://eprint.iacr.org/2012/089

and

More ECM-friendly curve families: http://eprint.iacr.org/2012/070

mklasson · 2013-04-27, 16:05

The build.vc11 gpu project file seems outdated and no longer works. I did some changes to it and managed to get gmp-ecm svn2518 to build and run with gpu support using VS 2012 Express and CUDA 5.0 on win7 x64.

I've put the updated files and what I hope are sufficient instructions to get things working at http://mklasson.com/gmp-ecm_70dev_gpu_vc11.zip.

If you're only looking for binaries you can try my x64 i7 executables:
http://mklasson.com/gmp-ecm_gpu_70de...7_sm20_x64.zip
http://mklasson.com/gmp-ecm_gpu_70de...7_sm30_x64.zip

My 660 Ti card seems about as fast as all four cores of my 3.6GHz i7.

Many thanks to Brian, Cyril, and the other fellows behind the scenes!

LaurV · 2013-04-29, 06:08

Quote:

Originally Posted by mklasson

I've put the updated files ..

Thanks a lot for this work! Very useful. I will give it a try to see how fast a gtx580 can ECM.

mklasson · 2013-04-29, 22:22

Quote:

Originally Posted by LaurV

Thanks a lot for this work! Very useful. I will give it a try to see how fast a gtx580 can ECM.

Be sure to factor 300-digit numbers to maximize the gpu speedup

From readme.gpu: "The modular arithmetic on GPU can only handle integers smaller that 2^1018. The execution time does not depend on the size of the interger that is factored as the same code is used for all integers."

My 660 Ti is about 12x the speed of one 3.6GHz i7 core on a c295 input.

And see http://www.mersenneforum.org/showpos...5&postcount=89 for a stage2 bug workaround.

2012-02-16, 16:59	#1
Brian Gladman May 2008 Worcester, United Kingdom 7²·11 Posts	Windows build for GPU CUDA code I have just added a Visual Studio 2010 build for the GPU stage 1 code on Windows x64 to the SVN (it might work on win32 but I have not tested it). As Paul L has already said, this code is only for those who are brave (or foolhardy) since it requires that the project build properties in Visual Studio are properly set up for your GPU configuration before it is built. Please seek advice, discuss or report issues here. Brian Gladman

2012-02-18, 05:55	#3
Karl M Johnson Mar 2010 19B₁₆ Posts	It's for ecm from GMP-ECM package. Last fiddled with by Karl M Johnson on 2012-02-18 at 05:56

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
I'm trying to speedup AES MSIEVE factoring using CUDA build but...	loopdemack	Msieve	11	2016-01-18 13:44
32-bit CUDA build?	f1pokerspeed	Msieve	2	2013-12-30 01:14
Windows x64 CUDA Build	Brian Gladman	Msieve	12	2013-08-30 12:40
CUDA executable code	nucleon	GPU Computing	4	2013-02-04 22:44
need C/C++ modular arithmetic code for Windows	ixfd64	Programming	15	2008-07-30 03:52

2012-02-20, 20:48	#7
jasonp Tribal Bullet Oct 2004 3²·5·79 Posts	Yes, if you have to do batch factorization of many different numbers then a GPU would work nicely, but you will need separate state for each number. In fact there's a cottage industry within the number field sieve community to work out a method where many bazillions of small numbers, maybe 80-150 bits each, can get factored with ECM and P+-1. This is the 'NFS cofactorization problem' and graphics cards look like an ideal way to solve it. They're much cheaper and possibly much faster than FPGAs, and programming them is a snap with modern toolkits so you don't need to know any hardware design.

2012-02-24, 11:39	#8
Robert Holmes Oct 2007 6A₁₆ Posts	Some interesting developments as of yesterday: Improved addition chains for Edward's curve ECM: http://eprint.iacr.org/2012/089 and More ECM-friendly curve families: http://eprint.iacr.org/2012/070

2013-04-27, 16:05	#9
mklasson Feb 2004 2×3×43 Posts	The build.vc11 gpu project file seems outdated and no longer works. I did some changes to it and managed to get gmp-ecm svn2518 to build and run with gpu support using VS 2012 Express and CUDA 5.0 on win7 x64. I've put the updated files and what I hope are sufficient instructions to get things working at http://mklasson.com/gmp-ecm_70dev_gpu_vc11.zip. If you're only looking for binaries you can try my x64 i7 executables: http://mklasson.com/gmp-ecm_gpu_70de...7_sm20_x64.zip http://mklasson.com/gmp-ecm_gpu_70de...7_sm30_x64.zip My 660 Ti card seems about as fast as all four cores of my 3.6GHz i7. Many thanks to Brian, Cyril, and the other fellows behind the scenes!