mfaktc: a CUDA program for Mersenne prefactoring

[BigBrother]

Originally Posted by TheJudger

Thank you, BigBrother.
Actually I've no CUDA-capable compiler environment installed under Windows, is MSVC the first choice for CUDA on Windows? Is it available for free for non-commercial usage?

I don't know if MSVC is the first choice for CUDA on Windows, it's the only one i've got here. I'm using a student license for MSVC.

[moebius]

Microsoft Visual Studio Express and MSVC++ Express Edition is free-of-charge, MFC, ATL, OpenPM require the Standard Edition or higher.

[moebius]

E:\mfactc>mfaktc-hack 65255629 68 69 0
mfaktc v0.02 Copyright (C) 2009, 2010 Oliver Weihe
THREADS_PER_GRID 1048576
THREADS_PER_BLOCK 256
SIEVE_SIZE_LIMIT 32kiB
SIEVE_SIZE 230945bits
SIEVE_PRIMES 50000
USE_PINNED_MEMORY enabled
USE_ASYNC_COPY enabled
VERBOSE_TIMING disabled
SELFTEST disabled
MORE_CLASSES disabled
sieve_init(): sieving factor candidates with small primes up to 611957
tf(65255629, 68, 69);
k_min = 2261474677980

no factor for M65255629 from 2^68 to 2^69
tf(): total time spent: 12785031msec

too bad that I can't upload the results to primenet

[TheJudger]

Hello!

moebius: If you want to help: rerun allready factored exponents and help proving that my code finds the factor(s) aswell.

False positives are not the problem (if there are some)
But we need to be sure that my code doesn't miss the factors.

On your hardware "SIEVE_PRIMES 200000" might be benifical, to bad that you can't check easily (needs recompile).

Can you add one or two lines starting with "class" aswell so I can see the actual speed?

-----
Find attached version 0.03

- allow exponents up to 2^32 -1
(tested with some exponents around M3321xxxxxx)
- siever: improved the loop which creates the candidate list (again)
- loop unrolled
- use a lookup table to parse 8 bits at once
- added 40 known factors from ElevenSmooth "Operation Billion Digits" in
M3321xxxxxx range to the selftest
- added another timer which helps to adjust SIEVE_PRIMES (needs to be
enabled with VERBOSE_TIMING)

The bigger exponents need further testing!

You'll see the speedups only in configurations where the CPU was limiting!

-----
Uncwilly: M3312xxxxxx from 2^1 to 2^71 needs ~5m 35s on my system :)


Oliver

[ET_]

Originally Posted by TheJudger

Hello!

moebius: If you want to help: rerun allready factored exponents and help proving that my code finds the factor(s) aswell.

False positives are not the problem (if there are some)
But we need to be sure that my code doesn't miss the factors.

On your hardware "SIEVE_PRIMES 200000" might be benifical, to bad that you can't check easily (needs recompile).

Can you add one or two lines starting with "class" aswell so I can see the actual speed?

-----
Find attached version 0.03

- allow exponents up to 2^32 -1
(tested with some exponents around M3321xxxxxx)
- siever: improved the loop which creates the candidate list (again)
- loop unrolled
- use a lookup table to parse 8 bits at once
- added 40 known factors from ElevenSmooth "Operation Billion Digits" in
M3321xxxxxx range to the selftest
- added another timer which helps to adjust SIEVE_PRIMES (needs to be
enabled with VERBOSE_TIMING)

The bigger exponents need further testing!

You'll see the speedups only in configurations where the CPU was limiting!

-----
Uncwilly: M3312xxxxxx from 2^1 to 2^71 needs ~5m 35s on my system :)


Oliver

As a owner of a Nvidia card on a Linux OS and a manager of Operation Billion Digits, I will soon ask for help testing the range 3321xxxxxx

Luigi

[TheJudger]

Hi Luigi,

keep in mind that it is still limited to 71 bits factor size.

Oliver

[ET_]

Originally Posted by TheJudger

Hi Luigi,

keep in mind that it is still limited to 71 bits factor size.

Oliver

Keep in mind that I still need to install Linux and learn how to use CUDA...

We may test our factors below 71 bits in less than a day (note that we didn't use Prime95...)

I keep on waiting for some 80 bits version to put on OBD page

Luigi

[ixfd64]

Wow, this could be a turning point for GIMPS. I hope your code gets implemented in Prime95!

[Uncwilly]

Originally Posted by TheJudger

Uncwilly: M3312xxxxxx from 2^1 to 2^71 needs ~5m 35s on my system :)

I may have to invest in a box for home then. I currently only have a laptop.

[nucleon]

Originally Posted by TheJudger

Uncwilly: M3312xxxxxx from 2^1 to 2^71 needs ~5m 35s on my system :)

How is that comparing with CPUs? (core i7 etc...)


-- Craig

[Uncwilly]

Originally Posted by nucleon

How is that comparing with CPUs? (core i7 etc...)

70->71 on a single core ~8h on my laptop. 0.72 GHz Days credit.