ECM for CUDA GPUs in latest GMP-ECM ? - Page 2

Batalov · 2012-02-11, 00:23

For Windows, it would have been nice if someone could downgrade the build/solution files vs2010 to vs2008 and post. For comparison, CUDA toolkit contains scripts for 2005, 2008, 2010 -- that is kinda user friendly. Not everyone has 2010. (Well, temporarily one can get the trial license.)

xilman · 2012-02-11, 00:27

Quote:

Originally Posted by Batalov

For Windows, it would have been nice if someone could downgrade the build/solution files vs2010 to vs2008 and post. For comparison, CUDA toolkit contains scripts for 2005, 2008, 2010 -- that is kinda user friendly. Not everyone has 2010. (Well, temporarily one can get the trial license.)

If you want it, why don't you do it?

Prime95 · 2012-02-11, 01:21

Quote:

Originally Posted by Batalov

Not everyone has 2010. (Well, temporarily one can get the trial license.)

VS 2010 is free (the express edition). Yes, it is 32-bit only, but you can get the Windows SDK for free which contains the 64-bit compiler.

Batalov · 2012-02-11, 02:51

Quote:

Originally Posted by xilman

If you want it, why don't you do it?

Because it is ugly! I am not talking about .sln files; for the .vcxproj -> .vcproj conversion most of the internet based advices amount to 'you might be best served by using the "New project from existing code" wizard to build a new VC2008 project for the code rather than trying to convert the existing project.' It is best done by the authors who know their source and dependencies.

I'll try the free version.

frmky · 2012-02-11, 08:03

Quote:

Originally Posted by R.D. Silverman

Is the code specific to a particular GPU? How portable is it?

It uses CUDA, thus requires an nVidia GPU supporting CC 1.3 or higher.

frmky · 2012-02-11, 08:09

Quote:

Originally Posted by xilman

Each stage one took 70 seconds on a GTX-460.

The GTX 460 has 7 MP's. Did you do 224 curves in parallel or more? Are there memory latencies that are hidden by doing more?

debrouxl · 2012-02-11, 08:59

On my GT540M (admittedly a fairly low-end model, with 2 MPs), under Debian unstable x86_64, gpu_ecm with 64 curves in parallel seems to be somewhat slower than CPU-based stage 1 (tuned GMP-ECM binary) running on Core i7-2670QM @ 2.2 GHz.
I have tested B1 bounds from 5e4 to 16e6, and 32, 64 or 128 parallel curves. 32 curves has throughput markedly slower than 64, but 128 is hardly better than 64 for throughput.

xilman · 2012-02-11, 09:14

Quote:

Originally Posted by frmky

The GTX 460 has 7 MP's. Did you do 224 curves in parallel or more? Are there memory latencies that are hidden by doing more?

The first 896 curves were done in four batches of 224 --- the default. The second ran all in parallel with a block of 32x32 and grid of 70x1x1. I seem to have lost the detailed timing information for the earlier curves

As best I recall, running 896 took slightly less than running 224 four times but I could be quite wrong.

xilman · 2012-02-11, 09:23

I screwed up computing the time per curve

Quote:

Originally Posted by xilman

Each stage one took 70 seconds on a GTX-460. The latest ECM takes 679 seconds per stage 1 on a single core of a 1090T clocked at 3264MHz, so the GPU version is close to 10 times faster in this situation.

1792 curves took 141 hours to run. I evaluated (1792 * 141 / 3600) to obtain the quoted figure of 70 seconds per curve.

The correct expression is (141 * 3600 / 1792), which evaluates to 283 seconds per curve.
Although this is four times worse than the initial figure, it is still 2.4 times faster than a singe core.

Sorry about that.

Karl M Johnson · 2012-02-11, 11:23

Quote:

Originally Posted by frmky

It uses CUDA, thus requires an nVidia GPU supporting CC 1.3 or higher.

DP FP calculations are being performed ?

jasonp · 2012-02-11, 12:30

No, it looks to be integer only. Compute Capability 1.3 also has advanced synchronization primitives that the CC 1.3 branch uses. The code is here for anyone who's interested.

2012-02-11, 08:59	#18
debrouxl Sep 2009 2×3×163 Posts	On my GT540M (admittedly a fairly low-end model, with 2 MPs), under Debian unstable x86_64, gpu_ecm with 64 curves in parallel seems to be somewhat slower than CPU-based stage 1 (tuned GMP-ECM binary) running on Core i7-2670QM @ 2.2 GHz. I have tested B1 bounds from 5e4 to 16e6, and 32, 64 or 128 parallel curves. 32 curves has throughput markedly slower than 64, but 128 is hardly better than 64 for throughput. Last fiddled with by debrouxl on 2012-02-11 at 09:00

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Running CUDA on non-Nvidia GPUs	Rodrigo	GPU Computing	3	2016-05-17 05:43
Error in GMP-ECM 6.4.3 and latest svn	ATH	GMP-ECM	10	2012-07-29 17:15
latest SVN 1677	ATH	GMP-ECM	7	2012-01-07 18:34
Has anyone seen my latest treatise?	davieddy	Lounge	0	2011-01-21 19:29
Latest version?	[CZ]Pegas	Software	3	2002-08-23 17:05

2012-02-11, 00:23	#12
Batalov "Serge" Mar 2008 Phi(4,2^7658614+1)/2 2⁷×7×11 Posts	For Windows, it would have been nice if someone could downgrade the build/solution files vs2010 to vs2008 and post. For comparison, CUDA toolkit contains scripts for 2005, 2008, 2010 -- that is kinda user friendly. Not everyone has 2010. (Well, temporarily one can get the trial license.)

2012-02-11, 12:30	#22
jasonp Tribal Bullet Oct 2004 5×709 Posts	No, it looks to be integer only. Compute Capability 1.3 also has advanced synchronization primitives that the CC 1.3 branch uses. The code is here for anyone who's interested.