TPSieve CUDA Testing Thread

Ken_g6 · 2010-09-04, 21:37

You asked for it, and I've finally made it. Download TPSieve-CUDA here.

I haven't done extensive testing on twin primes, so probably somebody should go over a short range (100G? Perhaps 1T?) with TPSieve-CUDA to make sure it gets the same factors.

I hope it works well for everyone!

Karl M Johnson · 2010-09-04, 22:18

Could you please specify an example to run? I see it needed cudart static linking - that file comes with cuda sdk. Was it compiled with cuda toolkit 3.1 ?

Ken_g6 · 2010-09-04, 22:25

OK. Supposing you downloaded the 480000-484999_30aug2010.txt sieve file, if you run:

./tpsieve-cuda-x86_64-linux -i 480000-484999_30aug2010.txt -p 710005180000000 -P 710005200000000

It should output:
710005185071411 | 5012115*2^481782+1
710005192340203 | 4018161*2^483419-1

very quickly. (I tested this on the emulator, so it runs really slow for me!) Expand the range, and you should get more of Mdettweiler's results.

./tpsieve-cuda-x86_64-linux -i 480000-484999_30aug2010.txt -p 710T -P 715T

would produce all of them, for instance.

Edit: Compiled with the 2.3 toolkit. One place to get the appropriate libcudart.so would be here or here.

Karl M Johnson · 2010-09-05, 06:50

Where do I get that fancy file?
I dont need that libcudart.so file, I'm on windows.

Karl M Johnson · 2010-09-05, 08:47

Ah, I got it working.
Found that fancy file from this thread
Here's the output:

Code:

tpsieve-cuda>tpsieve-cuda-x86-windows.exe -i 480000-484999_30aug2010.txt -p 710T -P 715T
tpsieve version cuda-0.1.5b (testing)
Found K's from 3 to 9999999.
Found N's from 480000 to 484999.
nstart=480000, nstep=27, gpu_nstep=27
Read 18013513 terms from NewPGen format input file `480000-484999_30aug2010.txt'
ppsieve initialized: 3 <= k <= 9999999, 480000 <= n <= 484999
Sieve started: 710000000000000 <= p < 715000000000000
Thread 0 starting
Detected GPU 0: GeForce GTX 285
Detected compute capability: 1.3
Detected 30 multiprocessors.
710001064441429 | 1473435*2^480477+1
710001781836203 | 3090555*2^482969+1
710002017069043 | 1947711*2^484889-1
710002639870109 | 7153191*2^483771+1
710003699276149 | 5489211*2^481645-1
710004831474721 | 9156609*2^482469+1
710005185071411 | 5012115*2^481782+1
710005192340203 | 4018161*2^483419-1
710005390472317 | 3240861*2^484861+1
710005916032213 | 5469669*2^482131+1
710006212449883 | 9438471*2^480253+1
710006478541837 | 942801*2^484681-1
p=710007273971713, 121.2M p/sec, 0.34 CPU cores, 0.1% done. ETA 05 Sep 23:11
710007380861971 | 3067731*2^482247-1
710007392845019 | 7483995*2^483443-1
710007480582299 | 1724049*2^480073-1
710008202353481 | 5813421*2^481371-1
710008811001043 | 9322383*2^480292-1
710008912579171 | 6024705*2^482149-1
710009562402587 | 5037609*2^482129-1
710010162887723 | 6614673*2^481762+1
710010987465557 | 6749691*2^483663+1
710011016356171 | 1349535*2^480408-1
710011368918931 | 7281273*2^482722-1
710011521417881 | 8617299*2^483945+1
710013019046899 | 3562503*2^481238-1
710013536554247 | 2683773*2^482840-1
p=710013762297857, 108.1M p/sec, 0.46 CPU cores, 0.3% done. ETA 05 Sep 23:50
710013880633081 | 4357815*2^480333+1
710013961546411 | 6488649*2^484015+1
710014319798129 | 1676877*2^480670-1
710014611723727 | 3195591*2^483289+1
710015165703751 | 1844445*2^483863+1
710016591664817 | 2155857*2^482100+1
710017445315627 | 9930375*2^480732-1
710017473222427 | 8642289*2^480555+1
710018153777579 | 5008965*2^484938-1
710018465445529 | 9185721*2^480167+1
p=710019807338497, 100.7M p/sec, 0.50 CPU cores, 0.4% done. ETA 06 Sep 00:21
710020260919457 | 1584663*2^483746-1

Now, will compiling x64 win binaries cause trouble?
I've had a "out of memory" error, even though I had like 1GB out of 4 free.

P.S.
It's not using GPU completely.
Peak GPU usage is reported 40%.
But I guess you already know that ?

amphoria · 2010-09-05, 09:16

I tried on a GTX465 on 64-bit linux using a range I had already tested so that I could compare the results. However I didn't get very far before getting an error.

Quote:

./tpsieve-cuda-x86_64-linux -i 480000-484999_19jun2010.txt -p 510T -P 515T
tpsieve version cuda-0.1.5b (testing)
Compiled Sep 4 2010 with GCC 4.3.3
Found K's from 3 to 9999999.
Found N's from 480000 to 484999.
nstart=480000, nstep=26, gpu_nstep=26
Read 18977477 terms from NewPGen format input file `480000-484999_19jun2010.txt'
ppsieve initialized: 3 <= k <= 9999999, 480000 <= n <= 484999
Sieve started: 510000000000000 <= p < 515000000000000
Thread 0 starting
Detected GPU 0: GeForce GTX 465
Detected compute capability: 2.0
Detected 11 multiprocessors.
510000064759291 | 604839*2^481707-1
510000994356869 | 2198475*2^482446+1
510001808585051 | 6049827*2^482948+1
510001965458981 | 9867039*2^480087-1
510002179900517 | 3334131*2^481253+1
510002930897567 | 8814495*2^481041+1
510003018137897 | 7665489*2^480401-1
510003129240001 | 4959981*2^480291+1
510003356427241 | 2391561*2^483615-1
510003644411923 | 7580307*2^484486-1
510003728553343 | 8313309*2^482255-1
510003886955161 | 3607413*2^482256-1
510004210312339 | 5073345*2^483515-1
Cuda error: cudaStreamCreate: out of memory

Karl M Johnson · 2010-09-05, 09:27

Amorphia, that's exactly the same error I had on x86 windows.
Here it pops again:

Code:

tpsieve-cuda-x86-windows.exe -i 480000-484999_30aug2010.txt -p 900T -P 901T
tpsieve version cuda-0.1.5b (testing)
Found K's from 3 to 9999999.
Found N's from 480000 to 484999.
nstart=480000, nstep=27, gpu_nstep=27
Read 18013513 terms from NewPGen format input file `480000-484999_30aug2010.txt'
ppsieve initialized: 3 <= k <= 9999999, 480000 <= n <= 484999
Sieve started: 900000000000000 <= p < 901000000000000
Thread 0 starting
Detected GPU 0: GeForce GTX 285
Detected compute capability: 1.3
Detected 30 multiprocessors.
900000899028509 | 3182751*2^483513-1
900001860603749 | 9998469*2^481563+1
900001934059139 | 1853133*2^482022-1
900002540407273 | 8064075*2^482811+1
900002726446853 | 5749455*2^480565-1
900003355059173 | 3695019*2^484373-1
900003556591063 | 9754467*2^480376-1
900003917464219 | 7522179*2^481393-1
900004723972547 | 5306133*2^484306+1
900005287423111 | 6879159*2^482887-1
900007745466833 | 451935*2^481504+1
900009608245457 | 5425383*2^480786+1
p=900010489954305, 87.42M p/sec, 0.50 CPU cores, 1.0% done. ETA 05 Sep 15:17
900010638873601 | 378417*2^481830-1
900011291258897 | 6507645*2^482813+1
900011626245037 | 1340685*2^481238+1
900012104179271 | 645705*2^484085-1
900016125631741 | 2968161*2^483961+1
900016501038581 | 8124711*2^484951+1
900016817068751 | 75363*2^484216+1
900017662186813 | 525711*2^480789+1
900018252281867 | 6892521*2^484727-1
900020059012663 | 8598285*2^481068+1
900021150322181 | 4615461*2^482939+1
900021336561331 | 8746389*2^484435+1
900021361527311 | 3408945*2^482966-1
p=900021998075905, 95.90M p/sec, 0.54 CPU cores, 2.2% done. ETA 05 Sep 15:08
900022619382521 | 6958245*2^482800-1
900022833913493 | 6580995*2^483721+1
900022917366103 | 4560555*2^482723-1
900023322448907 | 1472211*2^480431-1
900024288211007 | 4808679*2^480371+1
900026935242913 | 3056079*2^482407-1
900028117404131 | 5600343*2^481214-1
900029413721059 | 815793*2^483818-1
900029829750299 | 4354917*2^483802-1
900030812047093 | 5639913*2^483592+1
p=900033017561089, 91.82M p/sec, 0.53 CPU cores, 3.3% done. ETA 05 Sep 15:08
900033789053611 | 7495449*2^481653-1
900034220560883 | 9094419*2^483205-1
900034570657763 | 9890505*2^480175-1
900035606160989 | 5867385*2^481157-1
900037077781057 | 8390829*2^481741+1
900037229605601 | 1863285*2^484553-1
900038990324497 | 3815157*2^482054+1
900040739108881 | 3513243*2^482350+1
900041542191221 | 6049533*2^482774-1
900042730035877 | 9304977*2^481916+1
900043309201403 | 136581*2^482397+1
p=900044056969217, 91.99M p/sec, 0.55 CPU cores, 4.4% done. ETA 05 Sep 15:08
900044321638183 | 8388129*2^484645-1
900044489973593 | 8240649*2^483659+1
900044550938063 | 7226823*2^484696-1
900045358508729 | 7763775*2^483076-1
900047216136989 | 2338305*2^482753-1
900047780897267 | 4008369*2^483695+1
900048470300299 | 2963115*2^481453-1
900048762025013 | 383355*2^480270-1
900049228276043 | 8622855*2^483971-1
900049467999349 | 660627*2^481816-1
900049796295679 | 2937537*2^483980-1
900052042582919 | 385575*2^484714+1
900052572323899 | 7711221*2^484603+1
900053267475361 | 7173609*2^483949+1
900053714040401 | 633879*2^480079+1
900053996550817 | 6894867*2^480856-1
p=900055633248257, 96.46M p/sec, 0.54 CPU cores, 5.6% done. ETA 05 Sep 15:06
900055866972487 | 6849789*2^483481-1
900056741014807 | 2245995*2^482732-1
900056770768759 | 814365*2^482000-1
900057523274303 | 642045*2^480196+1
900057941699027 | 3370071*2^480999+1
900058480102739 | 9883737*2^484374+1
900060511991023 | 8680035*2^484611-1
900060730024969 | 7366341*2^482195+1
900060738679177 | 1099155*2^483395-1
900063136569923 | 6597225*2^483763-1
900063669798383 | 5873829*2^481137-1
900064551341591 | 9219153*2^483872+1
900064734779653 | 7558803*2^483916-1
900065290605601 | 7338225*2^482126-1
900065587257671 | 7356405*2^481242-1
900065728724587 | 8091525*2^484942-1
p=900067529342977, 99.13M p/sec, 0.51 CPU cores, 6.8% done. ETA 05 Sep 15:04
900067553916287 | 8588259*2^483407+1
900068224207921 | 5907333*2^480414-1
900068309721587 | 4858185*2^483053+1
900069742623089 | 7249299*2^483067-1
900071614223911 | 974289*2^484133-1
900072154118867 | 3615069*2^480585-1
900072931824211 | 6749313*2^480668-1
900073013900513 | 1479111*2^482079-1
900073241850151 | 3667035*2^484867-1
900075811775299 | 5091681*2^482559+1
900076383783517 | 6995187*2^481406-1
p=900079152807937, 96.86M p/sec, 0.52 CPU cores, 7.9% done. ETA 05 Sep 15:03
900079180930459 | 8088465*2^482743+1
900080177837117 | 9137745*2^481706+1
900081068828399 | 116547*2^480962+1
900082664855509 | 4331577*2^481696-1
900084606014311 | 7923375*2^480228+1
900084897625079 | 7498953*2^482154+1
900085127281819 | 3059145*2^480229-1
900086877470243 | 2313279*2^483107+1
900087308304337 | 5166585*2^482543-1
p=900090529857537, 94.80M p/sec, 0.52 CPU cores, 9.1% done. ETA 05 Sep 15:03
900090831293629 | 9965295*2^481322-1
900091902233021 | 9990753*2^481446-1
900095462990077 | 9537003*2^480320+1
900096420949717 | 8525847*2^481988-1
900096832377143 | 1048245*2^483598+1
900096929852677 | 2153943*2^481358-1
900098509267721 | 7751367*2^481256+1
900099157340237 | 9244893*2^480360-1
900099669905143 | 9687819*2^484633+1
900101450465951 | 1940013*2^484300+1
p=900101851332609, 94.34M p/sec, 0.52 CPU cores, 10.2% done. ETA 05 Sep 15:03
900102028546621 | 2525739*2^482461+1
900102230642357 | 9699093*2^482344+1
900102319841591 | 8400777*2^481706-1
900102426091157 | 3881955*2^483157-1
900102488675867 | 337989*2^481711+1
900102580103633 | 9216783*2^482100+1
900102741563621 | 2272611*2^480277+1
900103553433571 | 7722345*2^483866-1
900104117029049 | 505821*2^480631-1
900105270926371 | 8850651*2^483739-1
900105302568581 | 7921695*2^482577+1
900106241542903 | 5146383*2^482750-1
900107926468921 | 5710305*2^481576-1
900110050114909 | 8376111*2^480199+1
900110665560263 | 7689909*2^483029-1
p=900112516399105, 88.77M p/sec, 0.52 CPU cores, 11.3% done. ETA 05 Sep 15:04
900113522958017 | 4037511*2^483349+1
900113818670537 | 2881989*2^483381+1
900114293440121 | 9168045*2^484941-1
900114895651987 | 1452225*2^484402+1
900116209588091 | 9696105*2^483814-1
900119156145683 | 1042815*2^481830+1
900120506278387 | 7095909*2^480787+1
900120924004501 | 8100075*2^480374-1
900121363886917 | 7647465*2^481616-1
900122553451477 | 5630019*2^482963+1
900122578082093 | 4954053*2^480218-1
900122941358243 | 6183075*2^482076+1
p=900123264303105, 89.57M p/sec, 0.55 CPU cores, 12.3% done. ETA 05 Sep 15:05
900124366202023 | 5674563*2^484186-1
900126758233421 | 757953*2^481946+1
900127178299367 | 2843865*2^484173-1
900128009292293 | 5166717*2^484434-1
900128886707417 | 4666179*2^484269+1
900129038645509 | 4965093*2^481026-1
900129732119477 | 8266305*2^480961-1
900131300857937 | 9063171*2^481141-1
900131738429219 | 7244415*2^480666+1
900131757667309 | 9087471*2^481013-1
900132376847051 | 8236677*2^480080+1
900133053419231 | 4099683*2^480928-1
p=900134169493505, 90.88M p/sec, 0.57 CPU cores, 13.4% done. ETA 05 Sep 15:05
900136707397699 | 4107933*2^480032+1
900138484885813 | 9160035*2^484905-1
900139177590013 | 8492325*2^481494+1
900141735781361 | 2119167*2^480722+1
900141821615489 | 5879169*2^480907+1
900143923881347 | 8023179*2^481613+1
900144031193809 | 1315365*2^482686+1
p=900144518938625, 86.23M p/sec, 0.60 CPU cores, 14.5% done. ETA 05 Sep 15:06
900146872538153 | 6797847*2^483152+1
900146934657221 | 31053*2^482810-1
900148138109243 | 6732345*2^484116-1
900149419577411 | 7471065*2^483931+1
900150206766011 | 4495635*2^480957-1
900152425932013 | 2111517*2^480778+1
900152581520117 | 8415135*2^481699-1
900153500000561 | 4769493*2^483890-1
900153813027347 | 4079283*2^481164+1
p=900154149060609, 80.25M p/sec, 0.58 CPU cores, 15.4% done. ETA 05 Sep 15:08
900155149794317 | 9979257*2^481892+1
900155755904123 | 2521533*2^483994+1
900158445390817 | 3743577*2^480768-1
900158816255227 | 6612759*2^481673-1
900159262356971 | 7420557*2^482234-1
900159600424079 | 7586655*2^481304+1
900159986271683 | 943605*2^481215+1
Cuda error: cudaStreamCreate: out of memory

tpsieve-cuda>pause
Press any key to continue . . .

Ken_g6 · 2010-09-05, 15:45

I get the feeling I have a severe memory leak on the GPU that I didn't know I had. Someone helped me with the stream synchronization code, and it worked, but I'm starting to suspect that each event and stream that is created also has to be destroyed. I'll fix it in the next release.

Ken_g6 · 2010-09-06, 17:20

v0.1.6, of both PPSieve and TPSieve, is released. Many changes and fixes are included.

- Faster on the GPU than 0.1.5b (though about the same as 0.1.5c)
- Uses less CPU
- A huge memory leak on the GPU should be fixed.
- Input files are more often read correctly.
- Many other bugfixes and tweaks.

Get it at the usual URL, in the first post.

Edit: P.S. I've forgotten to post the source location!

amphoria · 2010-09-07, 17:24

Quote:

Originally Posted by Ken_g6

v0.1.6, of both PPSieve and TPSieve, is released. Many changes and fixes are included.

I have completed sieving 510-515T and the factors match those I previously found. I got 138M p/sec on a GTX465 using 0.41 CPU on a single core of a Core i7@3.6GHz. As the single core was not maxed out I decided to try running 2 instances on a single core (the other 3 cores were running instances of LLR). With 2 instances I got a combined throughput of 210M p/sec with 0.68 CPU used. This would suggest that the GTX465 wasn't maxed out either with a single instance.

Ken_g6 · 2010-09-07, 18:28

Quote:

Originally Posted by amphoria

I have completed sieving 510-515T and the factors match those I previously found.

Good!

Quote:

Originally Posted by amphoria

With 2 instances I got a combined throughput of 210M p/sec with 0.68 CPU used. This would suggest that the GTX465 wasn't maxed out either with a single instance.

Interesting! Try fiddling with the -m option (probably going up from 8 in increments of 1), and see if you can make a single instance do any better.

2010-09-04, 21:37	#1
Ken_g6 Jan 2005 Caught in a sieve 5×79 Posts	TPSieve CUDA Testing Thread You asked for it, and I've finally made it. Download TPSieve-CUDA here. I haven't done extensive testing on twin primes, so probably somebody should go over a short range (100G? Perhaps 1T?) with TPSieve-CUDA to make sure it gets the same factors. I hope it works well for everyone!

2010-09-04, 22:25	#3
Ken_g6 Jan 2005 Caught in a sieve 5·79 Posts	OK. Supposing you downloaded the 480000-484999_30aug2010.txt sieve file, if you run: ./tpsieve-cuda-x86_64-linux -i 480000-484999_30aug2010.txt -p 710005180000000 -P 710005200000000 It should output: 710005185071411 \| 50121152^481782+1 710005192340203 \| 40181612^483419-1 very quickly. (I tested this on the emulator, so it runs really slow for me!) Expand the range, and you should get more of Mdettweiler's results. ./tpsieve-cuda-x86_64-linux -i 480000-484999_30aug2010.txt -p 710T -P 715T would produce all of them, for instance. Edit: Compiled with the 2.3 toolkit. One place to get the appropriate libcudart.so would be here or here. Last fiddled with by Ken_g6 on 2010-09-04 at 22:48

2010-09-06, 17:20	#9
Ken_g6 Jan 2005 Caught in a sieve 613₈ Posts	v0.1.6, of both PPSieve and TPSieve, is released. Many changes and fixes are included. - Faster on the GPU than 0.1.5b (though about the same as 0.1.5c) - Uses less CPU - A huge memory leak on the GPU should be fixed. - Input files are more often read correctly. - Many other bugfixes and tweaks. Get it at the usual URL, in the first post. Edit: P.S. I've forgotten to post the source location! Last fiddled with by Ken_g6 on 2010-09-06 at 17:24

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Fast Mersenne Testing on the GPU using CUDA	Andrew Thall	GPU Computing	109	2014-07-28 22:14
Inconsistent factors with TPSieve	Caldera	Twin Prime Search	7	2013-01-05 18:32
tpsieve-cuda slows down with increasing p	amphoria	Twin Prime Search	0	2011-07-23 10:52
Is TPSieve-0.2.1 faster than Newpgen?	cipher	Twin Prime Search	4	2009-05-18 18:36
Thread for non-PrimeNet LL testing	ThomRuley	Lone Mersenne Hunters	6	2005-10-16 20:11

2010-09-04, 22:18	#2
Karl M Johnson Mar 2010 110011011₂ Posts	Could you please specify an example to run? I see it needed cudart static linking - that file comes with cuda sdk. Was it compiled with cuda toolkit 3.1 ?

2010-09-05, 06:50	#4
Karl M Johnson Mar 2010 3×137 Posts	Where do I get that fancy file? I dont need that libcudart.so file, I'm on windows.

2010-09-05, 15:45	#8
Ken_g6 Jan 2005 Caught in a sieve 5×79 Posts	I get the feeling I have a severe memory leak on the GPU that I didn't know I had. Someone helped me with the stream synchronization code, and it worked, but I'm starting to suspect that each event and stream that is created also has to be destroyed. I'll fix it in the next release.