mfaktc: a CUDA program for Mersenne prefactoring

[ET_]

Here it is

http://www.moregimps.it/billion/expo_f.php

Let me know if you need it in some particular format.

BTW, 5 minutes from 2^50 to 2^71 is amazing!

Luigi

[TheJudger]

Hi Luigi,

that's exactly what I was looking for. :)

I had this one in my bookmarks: http://home.earthlink.net/~elevensmooth/Billion.html

Oliver

[RichD]

TheJudger,

Does this require a 64-bit CUDA library? I'm on a 64-bit OS (Mac - Snow Leopard) but Nvidia does not have the Mac libraries at 64-bit (still at 32-bit). I noticed some old post of yours in the Math forum asking questions about 32-bit integers.

RichD.

[TheJudger]

Hi RichD,

I'm running in 64bit, but I think that msft did some tests on a 32bit Ubuntu.
The size of the adress room doesn't affect the size of the datatypes (usually ;)).
E.g. on a 32bit OS there are 64bit ints aswell. And on x86 you have 80 bit floats, too.

It _SHOULD_ run on 32bit OS but I haven't checked. If I remember correctly the siever runs a bit slower in 32bit mode. :/

Oliver

[ldesnogu]

Originally Posted by TheJudger

The size of the adress room doesn't affect the size of the datatypes (usually ;)).

long is considered as 32-bit on both Windows 32- and 64-bit (LLP64) while it is considered as 32-bit on Linux 32-bit and 64-bit on Linux 64-bit (LP64).
OTOH, I can't say if your program uses long :)

[TheJudger]

Hi!

ldesnogu: right, I've forgotten "long". :/
-----
The (not yet released) 0.05 is faster again. Raw speed on my GTX 275: ~73M candidates per second for M66362159 above 64 bits.

Single Process
THREADS_PER_GRID: 30 * 2^15 # this is specific for my GTX 275 since it has 30 multiprocessors
THREADS_PER_BLOCK: 256
SIEVE_PRIMES: 22500 # siever becomes limiting on my system... again

M66362159 from 2^ 1 to 2^64: 98386msec
M66362159 from 2^64 to 2^65: 90711msec
M66362159 from 2^65 to 2^66: 177915msec
M66362159 from 2^66 to 2^67: 353126msec
-----
No more ptx hacking needed!

__device__ unsigned int __umul24hi(unsigned int a, unsigned int b)
{
  unsigned int r;
  asm("mul24.hi.u32 %0, %1, %2;" : "=r" (r) : "r" (a) , "r" (b));
/* _SLOW_ workaround if inline assembly above doesn't work (e.g. device emulation)*/
  //  r = (__umul24(a,b) >> 16) + (__umulhi(a&0xFFFFFF,b&0xFFFFFF)<<16);
  return r;
}

This works fine on Linux - CUDA toolkit 2.3. Hopefully this is possible on Windows, too.
AFAIK inline assembly is a unsupported feature of nvcc but I think it is better this way.

Oliver

[ET_]

I hope you will include a short HOWTO_compile.txt to your next release...

Luigi

[TheJudger]

Hi Luigi,

for the upcomming 0.05 on Linux with a proper installed CUDA Toolkit:

./compile.sh

;)

Oliver

[henryzz]

Originally Posted by TheJudger

Hi Luigi,

for the upcomming 0.05 on Linux with a proper installed CUDA Toolkit:

./compile.sh

;)

Oliver

have you changed the path to the more standard destination then?

[TheJudger]

Hi,

good news:
Yesterday I've added more than 200 known factors to the selftest.
Every single factor was verified using my code. :)
In some cases it misses factors when there are mutliple factors in one class close together but this is not critical. The is a known problem since the first version... This has nothing to do with the calculations itself, it is just how the results are returned from the GPU to the CPU.
-----
Raw speed on my GTX 275 for M66362159 above 64 bits: ~74M candidates per second.

Siever received a nice performace improvement for free by adding "-funroll-all-loops" to the gcc options. :) (only useful for CPU-limited scenarios)

Oliver

[ET_]

Luigi