gpuOwL: an OpenCL program for Mersenne primality testing

[ewmayer]

Originally Posted by kriesel

Nonzero pseudorandomly selected shift for gpuowl PRP would be useful. It would make life easier for uncwilly et al in the double, triple, quad checking effort, and gpuowl results could be checked with gpuowl.

Question for George: Is the Primenet server set up to allow non-Prime95/mprime clients which support shift to do both initial-test and DC, or does it only allow such same-client-for-both-tests for Prime95/mprime?

[Prime95]

Originally Posted by ewmayer

Question for George: Is the Primenet server set up to allow non-Prime95/mprime clients which support shift to do both initial-test and DC, or does it only allow such same-client-for-both-tests for Prime95/mprime?

No way to know without looking at the PHP code. If the server does not consider that a valid double-check, then I'll need to fix the server's PHP code.

[ewmayer]

Originally Posted by Prime95

No way to know without looking at the PHP code. If the server does not consider that a valid double-check, then I'll need to fix the server's PHP code.

Actually, it might make sense to keep things that way - only officially "trusted" clients (which I believe is just yours ATM) are allowed to run both tests on a given exponent, as a way to prevent a malicious user from submitting matching pairs of nonzero residues in order to accumulate project credit. But whichever way you & Aaron decide on, probably a good idea to check the server code to see what it is doing in regards to this.

[LaurV]

Originally Posted by ewmayer

Question for George: Is the Primenet server set up to allow non-Prime95/mprime clients which support shift to do both initial-test and DC, or does it only allow such same-client-for-both-tests for Prime95/mprime?

First case. We DC our own old work (done with P95) and also 100M-digits new work (done with cudaLucas) using cudaLucas, all mentioned programs use shifts, and the server never got angry with us, but happily accepting our results as DC. We would be grateful if this behavior is kept. Edit: at least for "presumed honest" users.

[kriesel]

Is it possible to save the opencl compile result at gpuowl launch for reuse, on NVIDIA K80 on Google Colaboratory, which is ubuntu 18.04.1?

[preda]

Originally Posted by kriesel

Is it possible to save the opencl compile result at gpuowl launch for reuse, on NVIDIA K80 on Google Colaboratory, which is ubuntu 18.04.1?

Are you asking because the opencl compilation is slow, and the launch frequent? (how slow is it -- how large would be the benefit?)

It might be possible, OpenCL does offer some binary kernel support (that could be used to save the compilation initially and reload it on subsequent runs of the same exponent on the same driver&hardware)

[kriesel]

Originally Posted by preda

Are you asking because the opencl compilation is slow, and the launch frequent? (how slow is it -- how large would be the benefit?)

It might be possible, OpenCL does offer some binary kernel support (that could be used to save the compilation initially and reload it on subsequent runs of the same exponent on the same driver&hardware)

The launch is frequent; every twelve hours of gpu run time or less on Colab. It's a small optimization. I asked because I recalled it as a capability present in the past on linux (perhaps only with rocm, and I don't know what driver's present on Colab or how to ask it). Since I'm offering Colab scripts for reuse in my reference threads, it could have larger impact than just my own use. Here's a recent resume timing. Time stamps in UTC as usual.

2019-11-05 18:39:52 gpuowl 
2019-11-05 18:39:52 Note: no config.txt file found
2019-11-05 18:39:52 config: -use ORIG_X2 -block 200 -log 120000 -maxAlloc 10240 -user kriesel -cpu colab/K80 
2019-11-05 18:39:52 355000033 FFT 20480K: Width 256x4, Height 256x4, Middle 10; 16.93 bits/word
2019-11-05 18:39:52 OpenCL args "-DEXP=355000033u -DWIDTH=1024u -DSMALL_HEIGHT=1024u -DMIDDLE=10u -DWEIGHT_STEP=0x1.0d27019dccb6fp+0 -DIWEIGHT_STEP=0x1.e6fb0d049fbefp-1 -DWEIGHT_BIGSTEP=0x1.ae89f995ad3adp+0 -DIWEIGHT_BIGSTEP=0x1.306fe0a31b715p-1 -DORIG_X2=1  -I. -cl-fast-relaxed-math -cl-std=CL2.0"
2019-11-05 18:39:54 

2019-11-05 18:39:54 OpenCL compilation in 1686 ms
2019-11-05 18:40:00 355000033 P1 B1=2760000, B2=66240000; 3981446 bits; starting at 3981445

At full 12 hour session duration, it's only about 40 ppm. Some are experiencing much earlier termination; 1 hour is not unusual; ~480 ppm in that case.

[kriesel]

Originally Posted by kriesel

At full 12 hour session duration, it's only about 40 ppm. Some are experiencing much earlier termination; 1 hour is not unusual; ~480 ppm in that case.

That's at the fast end of compile time, 1.6 seconds. I have seen up to 3.4 seconds. It's every Colab session or every fft length change, whichever comes first. I finally got a P100 I think, during a 411M P-1 stage 1, and it's way faster than a K80; ~6ms/iter instead of ~23.

[philbo0042]

I apologize if this has been asked several times. I am a complete noob to almost all of what GPU computing requires.

I have searched and perused this thread looking for information on how to set up gpuOwL with my Nvidia GTX 860M. I have attached gpuOwL.log that shows error codes I do not understand. I have installed mysys2 and minggw but may have selected the wrong settings during installation. I have also installed Visual Studio.

Can someone help, please?

Any information would be greatly appreciated. If additional information is needed from me, please let me know and I will be more than glad to provide it.

Thank you in advance!

[kriesel]

Originally Posted by philbo0042

Can someone help, please?

Try adding -use ORIG_X2
as in

gpuowl-win -device 0 -use ORIG_X2 -user kriesel -cpu condorella/rx480

in a command line or batch file, or put the option in config.txt
It's been mentioned before, but there are a LOT of posts to search through.

[kriesel]

In the gpuowl v6.11-9-g9ae3189 built-in help, it says in part

-B1                : P-1 B1 bound, default 500000
-B2                : P-1 B2 bound, default B1 * 30

What it doesn't say is minimum B1 is 15015.

2019-11-10 17:04:34 B1=10000 too small, adjusted to 15015
2019-11-10 17:04:34 102001127 FFT 5632K: Width 256x4, Height 64x4, Middle 11; 17.69 bits/word
2019-11-10 17:04:37 OpenCL args "-DEXP=102001127u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=11u -DWEIGHT_STEP=0x9.f10dfae44e32p-3 -DIWEIGHT_STEP=0xc.e00add36b0
688p-4 -DWEIGHT_BIGSTEP=0x9.837f0518db8a8p-3 -DIWEIGHT_BIGSTEP=0xd.744fccad69d68p-4 -DORIG_X2=1  -I. -cl-fast-relaxed-math -cl-std=CL2.0"
2019-11-10 17:04:41 OpenCL compilation in 3321 ms
2019-11-10 17:04:42 102001127 P1 B1=15015, B2=300000; 21677 bits; starting at 0

There seems to be no way to run only stage 1. Minimum B2 is > B1.

2019-11-10 17:19:11 B2=15015 too small, adjusted to 30030

Also, a later run with larger bounds fails as follows, apparently by design. (And yes it has a " on a line all by itself.)

2019-11-10 17:48:33 B2=850000 too small, adjusted to 1700000
2019-11-10 17:48:33 102001127 FFT 5632K: Width 256x4, Height 64x4, Middle 11; 17.69 bits/word
2019-11-10 17:48:33 OpenCL args "-DEXP=102001127u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=11u -DWEIGHT_STEP=0x9.f10dfae44e32p-3 -DIWEIGHT_STEP=0xc.e00add36b0
688p-4 -DWEIGHT_BIGSTEP=0x9.837f0518db8a8p-3 -DIWEIGHT_BIGSTEP=0xd.744fccad69d68p-4 -DORIG_X2=1  -I. -cl-fast-relaxed-math -cl-std=CL2.0"
2019-11-10 17:48:35 OpenCL compilation in 2736 ms
2019-11-10 17:48:36 102001127 P1 wants B1=850000 but savefile has B1=100000. Fix B1 or move savefile
2019-11-10 17:48:36 'C:\msys64\home\ken\gpuowl-compile\v6.11-9-g9ae3189\102001127\102001127.p1.owl' invalid
2019-11-10 17:48:36 102001127 P1 wants B1=850000 but savefile has B1=100000. Fix B1 or move savefile
2019-11-10 17:48:36 'C:\msys64\home\ken\gpuowl-compile\v6.11-9-g9ae3189\102001127\102001127-old.p1.owl' invalid
2019-11-10 17:48:36 Exiting because "invalid savefiles found, investigate why
"

There is apparently no ability to extend a stage 1 run or presumably a stage 2 run on the same exponent.