mfaktc: a CUDA program for Mersenne prefactoring - Page 96

Christenson · 2011-07-02, 18:32

Oliver:
Exactly when the automated primenet interaction is released depends on me finishing it and how much other work you do. It happens when it happens. I hope to get the necessary changes into the main part of mfaktc on 0.18, but we shall see. I have to do them and Oliver has to fold them in along with the dummy version of mfaktc-primenet.

My separate file needs a termination function so that we can ensure that libcurl is also informed to close connections.

**********
As for estimating assignments, I was thinking that if we ensured there were 5 assignments in worktodo.txt, that would provide enough fodder for figuring out how much to get for next time.
********

apsen · 2011-07-03, 14:24

I guess it does not happen often :-) but you probably should mention in the readme that mfaktc does not support compute capability. I went through getting getting CUDA dlls and installing new video drivers just to find out it does not support v1.0 :-(

BTW is there real problem with using v1.0 or it just hasn't been considered?

Andriy

TheJudger · 2011-07-03, 17:58

Hi Andriy,

sorry, I'll add the information into the README.txt. I just have noticed that this kind of information is only available in the Changelog.txt and in the mfaktc article on the mersennewiki.org (http://mersennewiki.org/index.php/Mfaktc).

Early versions ran fine on my 8800 GTX (G80 GPU, the only CC 1.0 GPU) but newer versions don't. There are (at least) two reasons why it doesn't work on CC 1.0 GPUs:

use of atomic instructions for access to the results array (this needs CC >=1.1)
I don't know what the problem is but at some point it stopped working on CC 1.0.

From mfaktc Changelog.txt

Code:

- officially GPUs with compute capability 1.0 are not supported. AFAIK the
  only GPU affected is the G80 (8800 GTS 320, 8800 GTS 640, 8800 GTX, 8800
  Ultra and their Quadro/Tesla variants (but not a 8800 GTS 512, this one is
  a G92 GPU)). The issue seems to be the synchronisation of the writes to
  *d_RES.
  _PERHAPS_ I'm able to fix this in feature releases. BUT are there really
  many G80 GPUs out there? I think it is not worth the work (and yes,
  personally I own a 8800GTX).

I won't place my money on a fix for CC 1.0 GPUs...

Oliver

Bdot · 2011-07-03, 20:10

Quote:

Originally Posted by TheJudger

use of atomic instructions for access to the results array (this needs CC >=1.1)

That is also the reason why my OpenCL version will not be able to run on ATI GPUs before HD5xxx - the older ones don't have atomics ...

I spent a few thoughts on how to make this work without atomics, but implementing locks on GPUs is a little complicated. Without locks it will only work when we assume that we find only one factor per grid. However, I found hints how to implement the locking. So if there's a big vote for supporting HD4xxx or maybe CUDA CC 1.0 I could add something mfakto (and suggest it for mfaktc

).

Christenson · 2011-07-03, 20:32

Oliver:
Should/is Bdot on the distribution list for pre-release mfaktc?

Just to let you, Bdot, know, the plan for mfaktc is to add a few CPU-side calls to the current code for when factors are found or not or work is needed, and to use as much P95 code as possible (preferably, whole source files unmodified) to do the work.

Christenson · 2011-07-03, 21:45

Hey, what's the preferred exchange rate for spaces to tabs?
And are we using the one true style or the other true style for braces? [I like mine to line up, but I'll follow whatever convention].

Add to list: More robust parsing of worktodo.txt....in parse.c.

apsen · 2011-07-03, 22:25

Quote:

Originally Posted by TheJudger

Early versions ran fine on my 8800 GTX (G80 GPU, the only CC 1.0 GPU) but newer versions don't.

Is it possible to get my hands on one of the earlier builds or they are not good enough to use them?

I have two 8800 GTX and one HD4550. I wonder how much work could they do?..

James Heinrich · 2011-07-04, 01:11

Quote:

Originally Posted by apsen

I have two 8800 GTX and one HD4550. I wonder how much work could they do?..

The 8800GTX should perform fairly close to my 8800GT; I get ~26GHz-days/day; assuming there was a stable version of mfaktc that performed similarly to recent builds, and there's no performance penalty for Compute v1.0 implementation, I'd expect ~25-30 GHz-days/day from an 8800GTX. That was quite a few "ifs", however...

Christenson · 2011-07-04, 02:18

Quote:

Originally Posted by apsen

Is it possible to get my hands on one of the earlier builds or they are not good enough to use them?

I have two 8800 GTX and one HD4550. I wonder how much work could they do?..

If you wade far enough back in this thread (it's 700 posts long!), you will find the early versions of mfaktc. You will want to visit the "Putting it All Together" thread, as the necessary windows DLL is very easy to find there. Finally, if you want to make a post on that thread with links to the posts with the early versions, I think Rodrigo will find that helpful.

Rodrigo · 2011-07-04, 06:01

Quote:

Originally Posted by Christenson

If you wade far enough back in this thread (it's 700 posts long!), you will find the early versions of mfaktc. You will want to visit the "Putting it All Together" thread, as the necessary windows DLL is very easy to find there. Finally, if you want to make a post on that thread with links to the posts with the early versions, I think Rodrigo will find that helpful.

That would be great!!

Rodrigo

Bdot · 2011-07-04, 11:01

Quote:

Originally Posted by Christenson

Oliver:
Should/is Bdot on the distribution list for pre-release mfaktc?

Just to let you, Bdot, know, the plan for mfaktc is to add a few CPU-side calls to the current code for when factors are found or not or work is needed, and to use as much P95 code as possible (preferably, whole source files unmodified) to do the work.

I received one prerel-version in order to get the signal handling into my stuff. However, mfakto is not yet ready to report results to primenet. Currently some GPUs do not correctly execute some kernels and I may need a fix from AMD. Once it is ready, and George agreed, I can diff-merge your and Olivers latest changes to my code so that we have the same type of communication with primenet.

Quote:

Originally Posted by apsen

... and one HD4550. I wonder how much work could they do?..

With my current kernels that would be about 8-10 GHz-days/day. And the code compiles fine for HD4xxx when I just skip the atomic_inc ...

2011-07-03, 14:24	#1047
apsen Jun 2011 10000011₂ Posts	Compute capability I guess it does not happen often :-) but you probably should mention in the readme that mfaktc does not support compute capability. I went through getting getting CUDA dlls and installing new video drivers just to find out it does not support v1.0 :-( BTW is there real problem with using v1.0 or it just hasn't been considered? Andriy

2011-07-03, 17:58	#1048
TheJudger "Oliver" Mar 2005 Germany 11·101 Posts	Hi Andriy, sorry, I'll add the information into the README.txt. I just have noticed that this kind of information is only available in the Changelog.txt and in the mfaktc article on the mersennewiki.org (http://mersennewiki.org/index.php/Mfaktc). Early versions ran fine on my 8800 GTX (G80 GPU, the only CC 1.0 GPU) but newer versions don't. There are (at least) two reasons why it doesn't work on CC 1.0 GPUs: use of atomic instructions for access to the results array (this needs CC >=1.1) I don't know what the problem is but at some point it stopped working on CC 1.0. From mfaktc Changelog.txt Code: - officially GPUs with compute capability 1.0 are not supported. AFAIK the only GPU affected is the G80 (8800 GTS 320, 8800 GTS 640, 8800 GTX, 8800 Ultra and their Quadro/Tesla variants (but not a 8800 GTS 512, this one is a G92 GPU)). The issue seems to be the synchronisation of the writes to *d_RES. _PERHAPS_ I'm able to fix this in feature releases. BUT are there really many G80 GPUs out there? I think it is not worth the work (and yes, personally I own a 8800GTX). I won't place my money on a fix for CC 1.0 GPUs... Oliver

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1676	2021-06-30 21:23
The P-1 factoring CUDA program	firejuggler	GPU Computing	753	2020-12-12 18:07
gr-mfaktc: a CUDA program for generalized repunits prefactoring	MrRepunit	GPU Computing	32	2020-11-11 19:56
mfaktc 0.21 - CUDA runtime wrong	keisentraut	Software	2	2020-08-18 07:03
World's second-dumbest CUDA program	fivemack	Programming	112	2015-02-12 22:51

2011-07-02, 18:32	#1046
Christenson Dec 2010 Monticello 703₁₆ Posts	Oliver: Exactly when the automated primenet interaction is released depends on me finishing it and how much other work you do. It happens when it happens. I hope to get the necessary changes into the main part of mfaktc on 0.18, but we shall see. I have to do them and Oliver has to fold them in along with the dummy version of mfaktc-primenet. My separate file needs a termination function so that we can ensure that libcurl is also informed to close connections. ******** As for estimating assignments, I was thinking that if we ensured there were 5 assignments in worktodo.txt, that would provide enough fodder for figuring out how much to get for next time. ******

2011-07-03, 20:32	#1050
Christenson Dec 2010 Monticello 5×359 Posts	Oliver: Should/is Bdot on the distribution list for pre-release mfaktc? Just to let you, Bdot, know, the plan for mfaktc is to add a few CPU-side calls to the current code for when factors are found or not or work is needed, and to use as much P95 code as possible (preferably, whole source files unmodified) to do the work.

2011-07-03, 21:45	#1051
Christenson Dec 2010 Monticello 5·359 Posts	Hey, what's the preferred exchange rate for spaces to tabs? And are we using the one true style or the other true style for braces? [I like mine to line up, but I'll follow whatever convention]. Add to list: More robust parsing of worktodo.txt....in parse.c.