The P-1 factoring CUDA program - Page 4

ET_ · 2013-03-02, 13:42

Quote:

Originally Posted by firejuggler

1H and 34 min? that's 5 to 7 time faster than a ordinary CPU. nice.

Please notice the B1=580,000...

Is the B1 limit actually a fixed one?

Luigi

firejuggler · 2013-03-02, 13:54

yup,
1H34 is about 5700 second.
Below a run of a similar sized expo , total run for phase 1 is 24400 seconds. Ok.. so speed up is only about 4.2 time

ET_ · 2013-03-02, 14:13

Quote:

Originally Posted by firejuggler

yup,
1H34 is about 5700 second.
Below a run of a similar sized expo , total run for phase 1 is 24400 seconds. Ok.. so speed up is only about 4.2 time

Please note that not everyone here has an high-clocked AVX2 processor available... Not to mention that the code here is still a proof of concept with just few optimizations.

Luigi

firejuggler · 2013-03-02, 14:45

Please note that i'm genuinely impressed by the "proof-of-concept" speed up.

As for my CPU, it is a stock speed i5 2500k, which is pretty 'ordinary' for today computer-enthusiast ( the run was done on one core).

kracker · 2013-03-02, 15:42

Quote:

Originally Posted by chalsall

+1!!!

Ditto. I only have a 560, but it is the 2GB version, so when you need some Stage 2 testing....

Now give us your half-first born child. Naow!

chalsall · 2013-03-02, 15:50

Quote:

Originally Posted by kracker

Now give us your half-first born child. Naow!

Sure. Where should half of my non-existent first-born be delivered?

owftheevil · 2013-03-02, 16:32

Quote:

Originally Posted by henryzz

I am intrigued how fast this would run small numbers upto a much higher B1. How much does the exponent affect the runtime? Could you try a sub 1000-bit exponent? Maybe a range of different size exponents would be appropriate.

It could be coerced into taking exponents that small (I assume you mean exponents p with Mp < 1000 bits), but it wouldn't be very efficient. Toom-Cook multiplication would be better, or even grammar school multiplication if you go small enough. A very rough upper bound on the number of iterations you need for a given B1 is log2(B1) * the number of primes < B1. Iteration times will be close to what CuLu gets for the same fft. This is after all only a slight modification of CuLu. For very large B1 things will be about 5-10% slower for some final segment.

owftheevil · 2013-03-02, 16:40

Quote:

Originally Posted by ET_

Please notice the B1=580,000...

Is the B1 limit actually a fixed one?

Luigi

Yes, at the moment, I have to rebuild it every time I want to use a different B1. If fact the next thing I'm going to do is enable it to get the B1 from the command line.

Quote:

Please note that not everyone here has an high-clocked AVX2 processor available... Not to mention that the code here is still a proof of concept with just few optimizations.

Luigi

It won't get much faster than it is now.

owftheevil · 2013-03-02, 16:45

Quote:

Originally Posted by chalsall

Sure. Where should half of my non-existent first-born be delivered?

e-mail would be fine. But if its half of a boy, don't bother. My wife only wants it if its half of a girl.

And as for the half night in the gap hotel, I presume I have to find my own way to Barbados? Or are you also going to provide transportation for half of the way there?

ET_ · 2013-03-02, 16:55

Quote:

Originally Posted by owftheevil

It could be coerced into taking exponents that small (I assume you mean exponents p with Mp < 1000 bits), but it wouldn't be very efficient. Toom-Cook multiplication would be better, or even grammar school multiplication if you go small enough. A very rough upper bound on the number of iterations you need for a given B1 is log2(B1) * the number of primes < B1. Iteration times will be close to what CuLu gets for the same fft. This is after all only a slight modification of CuLu. For very large B1 things will be about 5-10% slower for some final segment.

Maybe Montgomery/Barrett could fill the gap between Toom-Cook and grammar multiplication, but at the moment I'd rather have a "fully-functional" P-1 on current exponents, leaving the smaller ones to either mfakt or gpu-ecm...

You rock, owftheevil!

Luigi

chalsall · 2013-03-02, 17:00

Quote:

Originally Posted by owftheevil

e-mail would be fine. But if its half of a boy, don't bother. My wife only wants it if its half of a girl.

Actually it's transgendered... MtF...

Quote:

Originally Posted by owftheevil

And as for the half night in the gap hotel, I presume I have to find my own way to Barbados? Or are you also going to provide transportation for half of the way there?

Since I just got a big contact, I'll also provide half a trip here. You'll have to swim the rest of the way....

Sincerely though, thanks very much for your work!

2013-03-02, 13:54	#35
firejuggler "Vincent" Apr 2010 Over the rainbow 2³×5×73 Posts	yup, 1H34 is about 5700 second. Below a run of a similar sized expo , total run for phase 1 is 24400 seconds. Ok.. so speed up is only about 4.2 time Attached Thumbnails

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfaktc: a CUDA program for Mersenne prefactoring	TheJudger	GPU Computing	3628	2023-04-17 22:08
World's second-dumbest CUDA program	fivemack	Programming	112	2015-02-12 22:51
World's dumbest CUDA program?	xilman	Programming	1	2009-11-16 10:26
Factoring program need help	Citrix	Lone Mersenne Hunters	8	2005-09-16 02:31
Factoring program	ET_	Programming	3	2003-11-25 02:57

2013-03-02, 14:45	#37
firejuggler "Vincent" Apr 2010 Over the rainbow 2³·5·73 Posts	Please note that i'm genuinely impressed by the "proof-of-concept" speed up. As for my CPU, it is a stock speed i5 2500k, which is pretty 'ordinary' for today computer-enthusiast ( the run was done on one core).