Newer X64 build needed - Page 4

ATH · 2016-01-02, 12:59

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

wombatman · 2016-01-02, 20:46

I'm trying to compile MSieve on Ubuntu 14.04 using the following (after doing make clean):

Code:

make all ECM=1 CUDA=1

But it's throwing errors related to omp_get_thread_limit and other "omp" related functions. Is it trying to compile with MPI or am I forgetting something else?

jasonp · 2016-01-02, 23:55

Something in the compile chain wants OpenMP (probably GMP-ECM). Try adding -lomp to the link line (there will probably be more dependencies) or compiling with -mopenmp.

wombatman · 2016-01-03, 03:53

Will do. I did notice it was something with GMP-ECM and compiled it with just CUDA. That worked fine, but when I try and actually run msieve with a C134, I get a CUDA_ERROR_OUT_OF_MEMORY (4x) followed by CUDA_ERROR_DEINITIALIZED immediately as stage 1 of poly selection starts. This is with CUDA 5.5. Any idea what that might be from?

wombatman · 2016-01-03, 07:54

Might have figured it out. I, by way of habit from Windows, ran Msieve with "-t 16" at first, which is when the error popped up. When I took the "-t" parameter out completely, it seems to be fine. I'm going to let it run overnight and confirm that that's the case.

jasonp · 2016-01-03, 13:30

Yes, stage 1 of poly selection is multithreaded and needs a limit on the number of threads that make sense. I would never use more than 4, and beyond about C140 there's basically no difference in stage 1 throughput compared to one thread.

wombatman · 2016-01-03, 15:01

Yeah, it's more that on Windows you can put something like 16, and it will work fine (or at least not throw any errors--maybe performance suffers).

Googulator · 2016-01-04, 10:54

The non-GPU version works fine, thanks.

jasonp · 2016-01-04, 15:24

Does anyone want to test SVN 991 with a CUDA toolkit > 5.5? This replaces the GPU sorting code with CUB and streamlines some of the GPU setup. 64-bit windows builds should compile with 'make all WIN=1 WIN64=1 CUDA=1 NO_ZLIB=1'.

Single-threaded and multithreaded runs work fine for me on a range of input sizes with the v6.5 toolkit. It would be nice if someone can also confirm that it builds and runs okay in linux.

ATH · 2016-01-04, 18:37

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

wombatman · 2016-01-04, 19:29

Working on building it with 7.0, but running into some issues (not with compilation, but immediate crash when run). I'm going to try a few things first.

Edit: Yeah, still an immediate crash when compiled in MSieve. I recall some sort of issue with crossing between 32-bit and 64-bit dlls or something. Also working on getting it compiled in VS2012, but I'm running into issues there with mpir.lib and libecm.lib throwing unresolved external symbol errors at the last step of linking.

Edit 2: Alright, everything compiled fine in VS2012 and seems to work fine (I can start it and watch poly numbers go screaming by on the screen for a few seconds before I stop it). I'll do more testing in a bit, but it looks like you have successfully updated Msieve! Thanks a ton for taking the time to do it.

2016-01-02, 20:46	#35
wombatman I moo ablest echo power! May 2013 29·61 Posts	This is embarrassing... I'm trying to compile MSieve on Ubuntu 14.04 using the following (after doing make clean): Code: make all ECM=1 CUDA=1 But it's throwing errors related to omp_get_thread_limit and other "omp" related functions. Is it trying to compile with MPI or am I forgetting something else?

2016-01-02, 23:55	#36
jasonp Tribal Bullet Oct 2004 3,541 Posts	Something in the compile chain wants OpenMP (probably GMP-ECM). Try adding -lomp to the link line (there will probably be more dependencies) or compiling with -mopenmp. Last fiddled with by jasonp on 2016-01-02 at 23:59

2016-01-04, 19:29	#44
wombatman I moo ablest echo power! May 2013 29·61 Posts	Working on building it with 7.0, but running into some issues (not with compilation, but immediate crash when run). I'm going to try a few things first. Edit: Yeah, still an immediate crash when compiled in MSieve. I recall some sort of issue with crossing between 32-bit and 64-bit dlls or something. Also working on getting it compiled in VS2012, but I'm running into issues there with mpir.lib and libecm.lib throwing unresolved external symbol errors at the last step of linking. Edit 2: Alright, everything compiled fine in VS2012 and seems to work fine (I can start it and watch poly numbers go screaming by on the screen for a few seconds before I stop it). I'll do more testing in a bit, but it looks like you have successfully updated Msieve! Thanks a ton for taking the time to do it. Last fiddled with by wombatman on 2016-01-04 at 20:05

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Newer milestone thread	Uncwilly	Data	3421	2021-07-12 16:10
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

2016-01-02, 12:59	#34
ATH Einyen Dec 2003 Denmark 3⁵·13 Posts	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

2016-01-03, 03:53	#37
wombatman I moo ablest echo power! May 2013 29·61 Posts	Will do. I did notice it was something with GMP-ECM and compiled it with just CUDA. That worked fine, but when I try and actually run msieve with a C134, I get a CUDA_ERROR_OUT_OF_MEMORY (4x) followed by CUDA_ERROR_DEINITIALIZED immediately as stage 1 of poly selection starts. This is with CUDA 5.5. Any idea what that might be from?

2016-01-03, 07:54	#38
wombatman I moo ablest echo power! May 2013 6E9₁₆ Posts	Might have figured it out. I, by way of habit from Windows, ran Msieve with "-t 16" at first, which is when the error popped up. When I took the "-t" parameter out completely, it seems to be fine. I'm going to let it run overnight and confirm that that's the case.

2016-01-03, 13:30	#39
jasonp Tribal Bullet Oct 2004 DD5₁₆ Posts	Yes, stage 1 of poly selection is multithreaded and needs a limit on the number of threads that make sense. I would never use more than 4, and beyond about C140 there's basically no difference in stage 1 throughput compared to one thread.

2016-01-03, 15:01	#40
wombatman I moo ablest echo power! May 2013 3351₈ Posts	Yeah, it's more that on Windows you can put something like 16, and it will work fine (or at least not throw any errors--maybe performance suffers).

2016-01-04, 10:54	#41
Googulator Dec 2015 20₁₀ Posts	The non-GPU version works fine, thanks.

2016-01-04, 15:24	#42
jasonp Tribal Bullet Oct 2004 DD5₁₆ Posts	Does anyone want to test SVN 991 with a CUDA toolkit > 5.5? This replaces the GPU sorting code with CUB and streamlines some of the GPU setup. 64-bit windows builds should compile with 'make all WIN=1 WIN64=1 CUDA=1 NO_ZLIB=1'. Single-threaded and multithreaded runs work fine for me on a range of input sizes with the v6.5 toolkit. It would be nice if someone can also confirm that it builds and runs okay in linux.

2016-01-04, 18:37	#43
ATH Einyen Dec 2003 Denmark 3⁵·13 Posts	What about CUDA 7.0 and 7.5? Are they still not compatible?