mfaktc: a CUDA program for Mersenne prefactoring

[msft]

Hi,

On my GTX260

a) compiles
OK
b) runs correctly and
correct
c) is faster or slower?
slower

[BigBrother]

Originally Posted by axn

Presumably, you've figured out what to do already, if not check this out: http://msdn.microsoft.com/en-us/library/hd0hzyf8.aspx

Yes, I replaced the two sieve_clear_bit() calls with the _bittestandreset intrinsic, but it didn't improve performance. Actually, the program was slightly slower.

[axn]

Bummer

[BigBrother]

Using a method described on http://code.cheesydesign.com/ I managed to uncover the commands issued by nvcc without using the defective --dryrun command line option.

The program now processes 4.2M candidates/second on my 9600M GS, which is an increase of ~33%

[ldesnogu]

Originally Posted by TheJudger

I'm not 100% sure if I got the point of your post.
The smallest primes (3, 5, 7, 11) are removed from the sieve totally.

I quickly read your code, and it seems that it's not using a dense representation for the sieve array. You don't need to have bits for even numbers for instance.

If I misread your code, sorry.

[TheJudger]

ldesnogu: there are no even numbers.

The sieve represents the k-values of the factor candidates. The factor candidates are 2*k*p+1, so they are allways odd.

[TheJudger]

Hi!

axn: the asm code works but is slower (as msft reported allready) :(

I don't remeber the specific settings for this run but here are the numbers.
My C code as inline function: ~33M/s
My C code as macro: ~32M/s
Your asm code as inline function: ~28M/s
Your asm code as macro: ~30M/s
-----

Anyway I've improved the siever a little bit. :)

-----
New Benchmarks on my system (openSUSE 11.1 x86-64, CUDA 2.3, GTX 275, 4GHz C2D)

Single Process
THREADS_PER_GRID: 2^20
THREADS_PER_BLOCK: 256
SIEVE_PRIMES: 15000

M66362159 from 2^ 1 to 2^64: 142.4s
M66362159 from 2^64 to 2^65: 142.2s
M66362159 from 2^65 to 2^66: 282.4s
M66362159 from 2^66 to 2^67: 559.4s
---
Single Process
THREADS_PER_GRID: 2^20
THREADS_PER_BLOCK: 256
SIEVE_PRIMES: 15000
MORE_CLASSES

M66362159 from 2^65 to 2^66: 284.1s
M66362159 from 2^66 to 2^67: 545.5s
M66362159 from 2^67 to 2^68: 1068.6s
---
Two Processes at the same time
THREADS_PER_GRID: 2^20
THREADS_PER_BLOCK: 256
SIEVE_PRIMES: 150000 (10 fold increase compared to Single Process)

M66362159 from 2^64 to 2^65: 206.3s
M66362159 from 2^65 to 2^66: 401.0s
M66362159 from 2^66 to 2^67: 794.0s
-----
find attached the new version. :)
version 0.02 (2010-01-13)
- fixed some printf's
- allocate and free arrays only ONCE (was per class before)
- added check of return values of most *alloc()
- siever: improved the loop which creates the candidate list


Oliver

[moebius]

Originally Posted by TheJudger

Hi henryzz,

I forgot to mention: I have run it only on Linux right now (openSUSE 11.1 x86-64).
If you still want a binary I can create one on my system. Let me known which CPU you have, I'll make some settings than.
You need to install the CUDA software aswell.

I'm interested in an exe for 32-bit Windows XP SP3,CUDA 2.3, Athlon 64 (single core), Geforce 8600GT (256 MB). It would be nice if that were possible, because I haven't installed MSVC-compiler on this machine.

[BigBrother]

Here are the executables I compiled, they work on my 32-bit Vista system, so I guess they also work on 32-bit XP. Both versions are included, the original and the one with the ptx hack.

[moebius]

Very nice, I try it out later, because msieve_gpu currently running on the nvidia.

bedankt..

[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've improved the sieve code again. I've unrolled the loop which creates the ktab and use a precalculated table to get check 8 bits at once. This minimized the effect of "MORE_CLASSES" (becomes benefical at 2^69 to 2^70 for M66362159).

The GPU-code is untouched.

Single Process
THREADS_PER_GRID: 2^20
THREADS_PER_BLOCK: 256
SIEVE_PRIMES: 45000 (with this value my CPU can keep the GPU busy with on core)

M66362159 from 2^ 1 to 2^64: 113.4s
M66362159 from 2^64 to 2^65: 113.4s
M66362159 from 2^65 to 2^66: 223.6s
M66362159 from 2^66 to 2^67: 442.3s
M66362159 from 2^67 to 2^68: 879.7s

Sorry, no new code release yet. You have to wait, I want to try some things before. ;)

Oliver