Msieve with GPU support - Page 3

mdettweiler · 2009-10-12, 19:10

Just out of curiosity, can the GPU version of msieve do everything that the CPU version can? I see a lot in this thread about polynomial selection, but nothing for QS, NFS postprocessing, etc. If it can do those, how do those compare speed-wise with the CPU version? (All of my systems have integrated graphics, otherwise I'd just check this myself.

)

Also, does the GPU application do *everything* on the GPU, or are there still some tasks that are done on the CPU? I'm having a hard time figuring out from some comments here where stage 2 of NFS poly selection is done.

Thanks.

jasonp · 2009-10-12, 19:10

The overhead from launching kernels should be very minor; doing so takes microseconds, whereas each kernel runs for a good-size fraction of a second.

mdettweiler: The GPU binary can do everything that the ordinary binary can (except degree 4 poly selection). Stage 2 still takes place on the CPU; stage 1 has very regular data access patterns that make it easy to map to a GPU, and this is the only code that runs on a GPU currently.

Boy, forum posts are a poor substitute for multi-way chat :)

henryzz · 2009-10-12, 19:21

Quote:

Originally Posted by jasonp

The overhead from launching kernels should be very minor; doing so takes microseconds, whereas each kernel runs for a good-size fraction of a second.

mdettweiler: The GPU binary can do everything that the ordinary binary can (except degree 4 poly selection). Stage 2 still takes place on the CPU; stage 1 has very regular data access patterns that make it easy to map to a GPU, and this is the only code that runs on a GPU currently.

Boy, forum posts are a poor substitute for multi-way chat :)

what might make things simpler would be running the stage 1 and stage 2 separately like ggnfs does
it would be nice to test things like the gpu version plus the ggnfs second stage
are both of your stages better/faster than ggnfs's?

xilman · 2009-10-12, 19:28

Quote:

Originally Posted by jrk

You need the stage1_core.ptx file in your msieve directory. If you are on Linux you'll need to compile this file per frmky's instructions in the second post here.

Got it. It now works for the C160 posted earlier in the thread!

Now that it's functional I'll be able to start development in earnest.

Thanks for your help.

Paul

henryzz · 2009-10-12, 19:35

10 people viewing this thread currently including 3 supermods and 3 mods
how often does that happen?

jasonp · 2009-10-12, 19:38

Paul: Good to hear.

Henry, I don't know how even the GPU version of stage 1 compares to pol5 (the algorithm used in pol51m0b would be very difficult to port to a GPU). Stage 2 is much more powerful on msieve; pol51opt really cannot produce a polynomial with a skew larger than 1 million, whereas msieve's stage 2 can easily handle 100x that much.

I thought about separating and batching stage 1 and stage 2, and they're well-separated in the source tree, but doing stage 2 immediately once stage 1 finds a hit relieves the user of balancing the load between the two stages. I'm not convinced it's better this way, though.

Also, you mention graphics cards nowadays and everyone's ears perk up :)

jrk · 2009-10-12, 20:32

What has happened when msieve prints "crap"?

jasonp · 2009-10-12, 20:57

If the rational polynomial is P * x - M, the code works by choosing P and then determining M so that the rational polynomial satisfies a specific modular condition. If the resulting M is close enough to a reference number M0, then the algebraic polynomial has a smaller coefficients than usual, and is forwarded to stage 2 for optimization. There's a check on P and M before that happens, and sometimes that check fails. Whether that's because P and M are bad, or the check is erroneous, I don't know yet. There are other internal checks on P and they currently pass.

ltd · 2009-10-12, 21:19

Just wanted to confirm that with the latest release the CPU is idle during stage 1 for me also.

Batalov · 2009-10-13, 04:40

Quote:

Originally Posted by ltd

Just wanted to confirm that with the latest release the CPU is idle during stage 1 for me also.

Tru dat.

I happened to have the old binary (because I was waiting for something to finish before a mandatory reboot after installing the 2.3 drivers). Then I rebooted, installed the toolkit and tried the yesterday's version. Arrgh. Huge kernel activity (red in monitor, if you set it to visible).

Then, the new version (10/12 2:52pm). Great! No kernel activity (almost).
The graphics is quite sluggish, so the card must be doing somethin'.

My card is an ancient 8600 GT, but I couldn't resist trying.

Batalov · 2009-10-13, 05:19

Quote:

Originally Posted by jrk

What has happened when msieve prints "crap"?

Tourette's syndrom?
Actually I was very interested, too. It does say that. A lot.

2009-10-12, 19:10	#24
jasonp Tribal Bullet Oct 2004 110111010101₂ Posts	The overhead from launching kernels should be very minor; doing so takes microseconds, whereas each kernel runs for a good-size fraction of a second. mdettweiler: The GPU binary can do everything that the ordinary binary can (except degree 4 poly selection). Stage 2 still takes place on the CPU; stage 1 has very regular data access patterns that make it easy to map to a GPU, and this is the only code that runs on a GPU currently. Boy, forum posts are a poor substitute for multi-way chat :) Last fiddled with by jasonp on 2009-10-12 at 19:18

2009-10-12, 19:38	#28
jasonp Tribal Bullet Oct 2004 3,541 Posts	Paul: Good to hear. Henry, I don't know how even the GPU version of stage 1 compares to pol5 (the algorithm used in pol51m0b would be very difficult to port to a GPU). Stage 2 is much more powerful on msieve; pol51opt really cannot produce a polynomial with a skew larger than 1 million, whereas msieve's stage 2 can easily handle 100x that much. I thought about separating and batching stage 1 and stage 2, and they're well-separated in the source tree, but doing stage 2 immediately once stage 1 finds a hit relieves the user of balancing the load between the two stages. I'm not convinced it's better this way, though. Also, you mention graphics cards nowadays and everyone's ears perk up :) Last fiddled with by jasonp on 2009-10-12 at 19:38

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Compiling Msieve with GPU support	LegionMammal978	Msieve	6	2017-02-09 04:28
5+ GPU support	TheMawn	GPU Computing	3	2014-07-13 02:31
Support AVX	Unregistered	Information & Answers	5	2011-07-05 17:12
Msieve with GNFS support	R.D. Silverman	Msieve	465	2010-01-11 20:59
Athlon64 support?	JuanTutors	Software	1	2004-06-04 02:46

2009-10-12, 19:10	#23
mdettweiler A Sunny Moo Aug 2007 USA (GMT-5) 14151₈ Posts	Just out of curiosity, can the GPU version of msieve do everything that the CPU version can? I see a lot in this thread about polynomial selection, but nothing for QS, NFS postprocessing, etc. If it can do those, how do those compare speed-wise with the CPU version? (All of my systems have integrated graphics, otherwise I'd just check this myself. ) Also, does the GPU application do everything on the GPU, or are there still some tasks that are done on the CPU? I'm having a hard time figuring out from some comments here where stage 2 of NFS poly selection is done. Thanks.

2009-10-12, 19:35	#27
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 2³·3·5·7² Posts	10 people viewing this thread currently including 3 supermods and 3 mods how often does that happen?

2009-10-12, 20:32	#29
jrk May 2008 3·5·73 Posts	What has happened when msieve prints "crap"?

2009-10-12, 20:57	#30
jasonp Tribal Bullet Oct 2004 3,541 Posts	If the rational polynomial is P * x - M, the code works by choosing P and then determining M so that the rational polynomial satisfies a specific modular condition. If the resulting M is close enough to a reference number M0, then the algebraic polynomial has a smaller coefficients than usual, and is forwarded to stage 2 for optimization. There's a check on P and M before that happens, and sometimes that check fails. Whether that's because P and M are bad, or the check is erroneous, I don't know yet. There are other internal checks on P and they currently pass.

2009-10-12, 21:19	#31
ltd Apr 2003 2²·193 Posts	Just wanted to confirm that with the latest release the CPU is idle during stage 1 for me also.