Newer X64 build needed - Page 5

wombatman · 2016-01-05, 01:15

Here's the exe along with the pthreads dll and sort engine dll and ptx files. I added a CC5.2 ptx compilation. Don't know which are needed or not, but everything is there except for the cudart dlls. They'd be too big for posting here, I think. CPU is an Ivy Bridge.

chris2be8 · 2016-01-05, 17:29

Quote:

Originally Posted by ATH

What about CUDA 7.0 and 7.5? Are they still not compatible?

I made it work with CUDA 7.5 by editing the makefiles to remove CC 1.1 and 1.3:
http://mersenneforum.org/showpost.ph...2&postcount=20

Chris

ATH · 2016-01-06, 00:47

SVN991 compiled fine with CUDA 7.5 without tinkering with the Makefile beyond the usual parameters. It even worked with ZLIB on. What is the benefit of Zlib, it compresses relations? It is better to leave NO_ZLIB=1 ?

It seems to work, it searches for a poly at least. I tried to use parameters from this old RSA896 thread:
http://www.mersenneforum.org/showthread.php?t=17460
but it is not finding the same polynomials or I do not know what I'm doing, which is far more likely.

Anyone have some more recent parameters that should find a polynomial just for a test?

jasonp · 2016-01-06, 01:55

Compiling zlib in allows the binary to read and write compressed relation files.

I don't think there's a controlled test you can run that will find a known polynomial. There's a lot of checking in polyselect stage 2, so if you find any polynomials at all it's probably working fine. Of course with a hot modern GPU you will find stage 1 hits so fast that performing stage 2 will leave the GPU mostly idle.

Sorry to everyone that it took so long to get back to a working state.

LaurV · 2016-01-06, 02:41

Quote:

Originally Posted by ATH

SVN991 compiled fine with CUDA 7.5 without tinkering with the Makefile beyond the usual parameters. It even worked with ZLIB on. What is the benefit of Zlib, it compresses relations? It is better to leave NO_ZLIB=1 ?

If you have enough space, and don't intend to move the files much around, and don't have a terrible slow hdd, then yes. Disadvantages, as shown by the former sentence, are that the uncompressed files are big, taking a lot of space and being difficult to move from folders to folders, or share on the web, the are slow to read when resuming the work, etc. Advantages is that if the zipped file crash - and the probability is not null, because the files are big and not fast to handle/read/write - then you can most probably say bye-bye to all your relations, but when the file is in clear, the crashed lines are just ignored, and most of the relations are still recoverable.

ATH · 2016-01-06, 14:21

Here are the svn 991 files compiled without Zlib:

msieve-svn991-cuda75-haswell.zip
msieve-svn991-cuda75-sandybridge.zip

Gimarel · 2016-01-07, 07:03

On my System (Linux, GeForce GTX 650 Ti) the new version (compiled with Cuda 6.5, using sm30) is about 20% slower than the old version (compiled with Cuda 6.0, using sm20).

For the comparison I disabled the randomization. The card isn't used for the display.

I haven't had the time to try the old version with Cuda 6.5.

debrouxl · 2016-01-07, 07:15

I only use rotating HDDs, and using gzipped relations, as produced by NFS@Home clients for bandwidth reasons, usually saves filtering time compared to dealing with uncompressed relations. Possibly square root time as well, as the bottleneck when reading relations is on I/O.

On modern zlib versions, the CPU cost of dealing with uncompressed output is minimal.

VictordeHolland · 2016-01-07, 12:24

Quote:

Originally Posted by ATH

Here are the Msieve svn988 I compiled a few days ago without CUDA support (before I figured out how to):

msieve-svn988-nogpu-haswell.zip
msieve-svn988-nogpu-sandybridge.zip
msieve-svn988-nogpu-core2.zip

These need libwinpthread-1.dll to work on a machine without minGW and Visual Studio installed.

ATH · 2016-01-07, 15:38

Quote:

Originally Posted by VictordeHolland

These need libwinpthread-1.dll to work on a machine without minGW and Visual Studio installed.

Thanks, I did not get a chance to test them on another computer. I added my "libwinpthread-1.dll" from MSYS2 to all the zip-files.

I also added "nvcuda.dll" to the gpu versions just to be safe, although it might be present on all computers with Nvidia drivers installed?

Googulator · 2016-01-09, 05:29

991-haswell doesn't work unless I delete nvcuda.dll - apparently that DLL is highly driver-specific and must be loaded from the system.

2016-01-06, 00:47	#47
ATH Einyen Dec 2003 Denmark 2·3·5²·23 Posts	SVN991 compiled fine with CUDA 7.5 without tinkering with the Makefile beyond the usual parameters. It even worked with ZLIB on. What is the benefit of Zlib, it compresses relations? It is better to leave NO_ZLIB=1 ? It seems to work, it searches for a poly at least. I tried to use parameters from this old RSA896 thread: http://www.mersenneforum.org/showthread.php?t=17460 but it is not finding the same polynomials or I do not know what I'm doing, which is far more likely. Anyone have some more recent parameters that should find a polynomial just for a test? Last fiddled with by ATH on 2016-01-06 at 00:48

2016-01-06, 01:55	#48
jasonp Tribal Bullet Oct 2004 3565₁₀ Posts	Compiling zlib in allows the binary to read and write compressed relation files. I don't think there's a controlled test you can run that will find a known polynomial. There's a lot of checking in polyselect stage 2, so if you find any polynomials at all it's probably working fine. Of course with a hot modern GPU you will find stage 1 hits so fast that performing stage 2 will leave the GPU mostly idle. Sorry to everyone that it took so long to get back to a working state. Last fiddled with by jasonp on 2016-01-06 at 01:56

2016-01-07, 07:03	#51
Gimarel Apr 2010 2·131 Posts	Slower On my System (Linux, GeForce GTX 650 Ti) the new version (compiled with Cuda 6.5, using sm30) is about 20% slower than the old version (compiled with Cuda 6.0, using sm20). For the comparison I disabled the randomization. The card isn't used for the display. I haven't had the time to try the old version with Cuda 6.5.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Newer milestone thread	Uncwilly	Data	3668	2023-04-20 12:24
Help needed - M77 232 917 celebration build	jbpace	Hardware	19	2018-01-10 16:32
Performance of cuda-ecm on newer hardware?	fivemack	GMP-ECM	14	2015-02-12 20:10
Newer msieves are slow on Core i7	mklasson	Msieve	9	2009-02-18 12:58
Use of large memory pages possible with newer linux kernels	Dresdenboy	Software	3	2003-12-08 14:47

2016-01-06, 14:21	#50
ATH Einyen Dec 2003 Denmark 3450₁₀ Posts	Here are the svn 991 files compiled without Zlib: msieve-svn991-cuda75-haswell.zip msieve-svn991-cuda75-sandybridge.zip

2016-01-07, 07:15	#52
debrouxl Sep 2009 1724₈ Posts	I only use rotating HDDs, and using gzipped relations, as produced by NFS@Home clients for bandwidth reasons, usually saves filtering time compared to dealing with uncompressed relations. Possibly square root time as well, as the bottleneck when reading relations is on I/O. On modern zlib versions, the CPU cost of dealing with uncompressed output is minimal.

2016-01-09, 05:29	#55
Googulator Dec 2015 14₁₆ Posts	991-haswell doesn't work unless I delete nvcuda.dll - apparently that DLL is highly driver-specific and must be loaded from the system.