gpuOwL: an OpenCL program for Mersenne primality testing

[Runtime Error]

Originally Posted by kriesel

Thanks! Windows build posted at https://mersenneforum.org/showpost.p...5&postcount=35. Assorted warnings during build for Windows are a well established feature. Let's see how it runs.

It is entirely possible that this is user error, but I have not been successful getting v7.2-129-g2a16a0b to run the merged PRP and P-1. For example, a freshly unzipped download of that build along with only a worktodo containing the following line does not begin with a P-1 for me.

PRP=XXXXXXXXXXXXXXX,1,2,115xxxxxx,-1,77,2

In contrast, the above line for merged PRP/P-1 works for me on a freshly unzipped v7.2-93-ga5402c5-dirty. Also, when trying to run P-1 standalone, I see the "line ignored" error. Is the new build functioning as expected on your side? Thank you.

[kriesel]

Originally Posted by preda

With these changes gpuowl also does not do second-stage anymore, and does not do the merged first-stage + PRP anymore. In exchange, the first-stage is error-checked with a GEC equivalent, so that's buletproof (bar implementation errors :)

All this is still work-in-progress to a large degree. It's not very documented, and it's not polished, etc. But the basics are there. I do use it myself this way:
- I run P-1 first-stage on the GPU (on Radeon VII), with B1=2M at the wavefront,
- I continue with second-stage on the CPU with mprime, with either 256GB or 128GB of RAM

Originally Posted by Runtime Error

It is entirely possible that this is user error, but I have not been successful getting v7.2-129-g2a16a0b to run the merged PRP and P-1.

Error is entirely possible.
V7.2-129 is like V6.11-380 in that it requires separate PFactor= and PRP= worktodo lines.

Usual V7.x style merged PRP & P-1 syntax including bounds results in PRP:

>type worktodo.txt
B1=1000000,B2=300000000;PRP=<aid>,1,2,65778397,-1,75,1
>gpuowl-win.exe
20221212 18:33:39  GpuOwl VERSION v7.2-129-g2a16a0b
20221212 18:33:39  GpuOwl VERSION v7.2-129-g2a16a0b
20221212 18:33:39  config: -user kriesel -cpu test-radeonvii -block 2000 -proof 9 -autoverify 8 -noclean -d 0 -maxAlloc 15G -B1 1000000 -mprimeDir C:\Users\user\Documents\prime95 -use NO_ASM
20221212 18:33:39  device 0, unique id ''
20221212 18:33:39 test-radeonvii 65778397 FFT: 3.50M 1K:7:256 (17.92 bpw)
20221212 18:33:39 test-radeonvii 65778397 OpenCL args "-DEXP=65778397u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=7u -DAMDGPU=1 -DWEIGHT_STEP=0.054682824827785866 -DIWEIGHT_STEP=-0.051847648923945229 -DIWEIGHTS={0,-0.10100711914894978,-0.19181180017912944,-0.27344456197321132,-0.34683183367030057,-0.41280646847106556,-0.47211719547370123,-0.050874233614341763,} -DFWEIGHTS={0,0.11235586098671806,0.23733556147150281,0.3763574638101177,0.53099929168199911,0.7030160352689856,0.89435986818581947,0.053601151097261476,} -DNO_ASM=1  -cl-std=CL2.0 -cl-finite-math-only "
20221212 18:33:45 test-radeonvii 65778397 OpenCL compilation in 5.30 s
20221212 18:33:45 test-radeonvii 65778397 PRP starting from beginning
20221212 18:33:47 test-radeonvii 65778397 OK         0 on-load: blockSize 2000, 0000000000000003
20221212 18:33:47 test-radeonvii 65778397 validating proof residues for power 9
20221212 18:33:47 test-radeonvii 65778397 Proof using power 9
20221212 18:33:53 test-radeonvii 65778397     10000 a1a698ad083370e1  606
20221212 18:33:59 test-radeonvii 65778397     20000 f6f81bcc7846da93  607
20221212 18:34:05 test-radeonvii 65778397 Stopping, please wait..
20221212 18:34:06 test-radeonvii 65778397 OK     30000   0.05% a2cb16e3a47f877d  609 us/it + check 1.27s + save 0.25s; ETA 11:07
20221212 18:34:06 test-radeonvii  Exiting because "stop requested"
20221212 18:34:06 test-radeonvii  Bye

Separate PFactor only style worktodo line runs P-1:

>type worktodo.txt
B1=1000000,B2=300000000;PFactor=<aid>,1,2,65778397,-1,75,1
>gpuowl-win.exe
20221212 18:34:38  GpuOwl VERSION v7.2-129-g2a16a0b
20221212 18:34:38  GpuOwl VERSION v7.2-129-g2a16a0b
20221212 18:34:38  config: -user kriesel -cpu test-radeonvii -block 2000 -proof 9 -autoverify 8 -noclean -d 0 -maxAlloc 15G -B1 1000000 -mprimeDir C:\Users\user\Documents\prime95 -use NO_ASM
20221212 18:34:38  device 0, unique id ''
20221212 18:34:38 test-radeonvii 65778397 FFT: 3.50M 1K:7:256 (17.92 bpw)
20221212 18:34:38 test-radeonvii 65778397 OpenCL args "-DEXP=65778397u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=7u -DAMDGPU=1 -DWEIGHT_STEP=0.054682824827785866 -DIWEIGHT_STEP=-0.051847648923945229 -DIWEIGHTS={0,-0.10100711914894978,-0.19181180017912944,-0.27344456197321132,-0.34683183367030057,-0.41280646847106556,-0.47211719547370123,-0.050874233614341763,} -DFWEIGHTS={0,0.11235586098671806,0.23733556147150281,0.3763574638101177,0.53099929168199911,0.7030160352689856,0.89435986818581947,0.053601151097261476,} -DNO_ASM=1  -cl-std=CL2.0 -cl-finite-math-only "
20221212 18:34:43 test-radeonvii 65778397 OpenCL compilation in 5.27 s
20221212 18:34:43 test-radeonvii 65778397P1 P1: no savefile found, starting from the beginning
20221212 18:34:44 test-radeonvii 65778397P1  0.00% @0/1442200 B1(1000000) 0000000000000001
20221212 18:34:44 test-radeonvii 65778397P1  0.03% K d456f693b8a92d40  554
20221212 18:34:49 test-radeonvii 65778397P1  0.69%   8baf77ff18501526  537
20221212 18:34:55 test-radeonvii 65778397P1  1.39%   8bb893cb8c0dc4f5  528
20221212 18:35:00 test-radeonvii 65778397P1  2.08%   17900932f4f57702  536
20221212 18:35:06 test-radeonvii 65778397P1  2.77% K 46ad773345c5f2d2  540
20221212 18:35:11 test-radeonvii 65778397P1  3.47%   29e8dfa8f512bf61  564
20221212 18:35:17 test-radeonvii 65778397P1  4.16%   ebf6440ca5de84da  558
20221212 18:35:23 test-radeonvii 65778397P1  4.85%   d07a6a6df5a20faf  565
20221212 18:35:29 test-radeonvii 65778397P1  5.55% K 0cdab0bb00a2ee22  570

And, one must use the -B1 config or command line options, not this style of PFactor line in gpuowl v7.2-129, which will pass the illegal-for-mprime Pfactor line to mprime unchanged, as in this prime95 communication window snippet:

[Comm thread Dec 12 16:40] Updating computer information on the server
[Comm thread Dec 12 16:40] Sending expected completion date for M74218489: Dec 21 2022
[Comm thread Dec 12 16:40] Sending expected completion date for M222257821: Dec 21 2022
[Comm thread Dec 12 16:40] Done communicating with server.
[Main thread Dec 12 18:57] Restarting all worker windows to process .add file.
[Main thread Dec 12 18:57] Restarting all worker windows using new settings.
[Main thread Dec 12 18:57] Illegal line in worktodo.txt file: B1=1000000,B2=300000000;PFactor=<aid redacted>,1,2,65778397,-1,75,1

[Runtime Error]

Originally Posted by kriesel

Error is entirely possible.

Yes! Thank you so much for the detailed response, it enabled me to sort it out & get it working. B1 successfully finished on gpuowl, and now B2 is successfully running on prime95.

Originally Posted by kriesel

And, one must use the -B1 config or command line options, not this style of PFactor line in gpuowl v7.2-129, which will pass the illegal-for-mprime Pfactor line to mprime unchanged, as in this prime95 communication window snippet:

Ah this makes sense, thank you!

I appreciate the help

[DrobinsonPE]

Originally Posted by kriesel

Thanks! Windows build posted at https://mersenneforum.org/showpost.p...5&postcount=35

Thanks! I tried this with a RX6500XT and it seemed to work. The B2 is queued up in mprime on a 64GB machine. Not sure taking over 6 hours to complete the B1 works well with feeding a CPU. I probably need to graduate to a bigger GPU.

C:\Users\drobinson\gpuowl\v72129>gpuowl-win
20221213 19:29:38  GpuOwl VERSION v7.2-129-g2a16a0b
20221213 19:29:38  GpuOwl VERSION v7.2-129-g2a16a0b
20221213 19:29:38  config: -log 20000 -device 0 -maxAlloc 3G -mprimeDir mprime
20221213 19:29:38  device 0, unique id ''
20221213 19:29:38 gfx1034-0 118423787 FFT: 6.50M 1K:13:256 (17.38 bpw)
20221213 19:29:38 gfx1034-0 118423787 OpenCL args "-DEXP=118423787u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=13u -DAMDGPU=1 -DWEIGHT_STEP=0.54217396852433863 -DIWEIGHT_STEP=-0.35156472589349252 -DIWEIGHTS={0,-0.15906339058883701,-0.29282561895205711,-0.40531117373910508,-0.49990439478945886,-0.15890259474564797,-0.2926903998408783,-0.40519746303822279,} -DFWEIGHTS={0,0.18915027459705397,0.41407837557424887,0.68155168861587412,0.99961765226670651,0.18892294014103519,0.41380804074344879,0.68123021987759502,}  -cl-std=CL2.0 -cl-finite-math-only "
20221213 19:29:40 gfx1034-0 118423787 OpenCL compilation in 2.32 s
20221213 19:29:40 gfx1034-0 118423787P1 P1: no savefile found, starting from the beginning
20221213 19:29:40 gfx1034-0 118423787P1  0.00% @0/2885600 B1(2000000) 0000000000000001
20221213 19:29:45 gfx1034-0 118423787P1  0.01% K 3ce7e1f9366a695f 7123
20221213 19:30:53 gfx1034-0 118423787P1  0.35%   a61b92431c995a06 7114
20221214 01:16:04 gfx1034-0 118423787P1 100.00% K 346efd2d4588d52e 7120
20221214 01:16:04 gfx1034-0 118423787P1 completed
20221214 01:16:04 gfx1034-0  Bye

[preda]

Originally Posted by DrobinsonPE

Not sure taking over 6 hours to complete the B1 works well with feeding a CPU. I probably need to graduate to a bigger GPU.

Either that, or you may lower the B1, e.g. to B1=1M, which would reduce the time roughly in half.

[DrobinsonPE]

Originally Posted by preda

Either that, or you may lower the B1, e.g. to B1=1M, which would reduce the time roughly in half.

You're killing me, Smalls. Here I am trying to justify....who am I kidding. I have a 25 year anniversary next month so I am not buying any more computer toys for a while. I could give that a try. I ended up feeding the second half to a computer with 32GB memory . It took about 3 hours to complete so possibly it would be balanced at B1=1M. I should try again this weekend.

[TheJudger]

Hi!

H100 PCIe 80 GB limited to 310 Watt, quick&dirty GPUowl build:

./gpuowl -log 10000 -iters 200000 -prp 77936867
20221217 21:17:24  GpuOwl VERSION v7.2-130-g967853c-dirty
20221217 21:17:24  GpuOwl VERSION v7.2-130-g967853c-dirty
20221217 21:17:24  Note: not found 'config.txt'
20221217 21:17:24  config: -log 10000 -iters 200000 -prp 77936867
20221217 21:17:24  device 0, unique id ''
20221217 21:17:24 NVIDIA H100 PCIe-0 77936867 FFT: 4M 1K:8:256 (18.58 bpw)
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 OpenCL args "-DEXP=77936867u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=8u -DMM_CHAIN=1u -DMM2_CHAIN=2u -DWEIGHT_STEP=0.33644726404543274 -DIWEIGHT_STEP=-0.25174750481886216 -DIWEIGHTS={0,-0.44011820345520131,-0.37306474779553728,-0.29798072935699788,-0.21390437908665341,-0.11975874301407295,-0.014337887291734644,-0.44814572555075455,} -DFWEIGHTS={0,0.78609128957452257,0.5950610473469905,0.42446232150303748,0.2721098723818392,0.1360521812214803,0.014546452690911484,0.81207258201996746,}  -cl-std=CL2.0 -cl-finite-math-only "
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867

20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 OpenCL compilation in 0.01 s
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 PRP starting from beginning
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 OK         0 on-load: blockSize 400, 0000000000000003
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 validating proof residues for power 8
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 Proof using power 8
20221217 21:17:27 NVIDIA H100 PCIe-0 77936867 OK     10000   0.01% fc4f135f7cf4ad29  243 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:17:30 NVIDIA H100 PCIe-0 77936867 OK     20000   0.03% 3cd1bd9d5e09cbc5  243 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:17:33 NVIDIA H100 PCIe-0 77936867 OK     30000   0.04% c4e0ff35e3290d98  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:17:35 NVIDIA H100 PCIe-0 77936867 OK     40000   0.05% dffe1b1b0d748128  244 us/it + check 0.13s + save 0.10s; ETA 05:17
20221217 21:17:38 NVIDIA H100 PCIe-0 77936867 OK     50000   0.06% 52e286945371ed29  244 us/it + check 0.13s + save 0.10s; ETA 05:17
20221217 21:17:41 NVIDIA H100 PCIe-0 77936867 OK     60000   0.08% 0945da4dc08bdd95  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:17:43 NVIDIA H100 PCIe-0 77936867 OK     70000   0.09% 7131fa4eb77f4bb2  243 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:17:46 NVIDIA H100 PCIe-0 77936867 OK     80000   0.10% 8d76071d27ee4221  243 us/it + check 0.12s + save 0.10s; ETA 05:16
20221217 21:17:49 NVIDIA H100 PCIe-0 77936867 OK     90000   0.12% 0bacff453b2f470e  243 us/it + check 0.12s + save 0.10s; ETA 05:16
20221217 21:17:51 NVIDIA H100 PCIe-0 77936867 OK    100000   0.13% 6d7296b9e2830f50  244 us/it + check 0.12s + save 0.10s; ETA 05:16
20221217 21:17:54 NVIDIA H100 PCIe-0 77936867 OK    110000   0.14% 8cbfd4435622bda7  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:17:57 NVIDIA H100 PCIe-0 77936867 OK    120000   0.15% 79ae5dad855057ad  244 us/it + check 0.12s + save 0.10s; ETA 05:16
20221217 21:17:59 NVIDIA H100 PCIe-0 77936867 OK    130000   0.17% 50c97bcbf876231f  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:02 NVIDIA H100 PCIe-0 77936867 OK    140000   0.18% e1db15f897271496  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:05 NVIDIA H100 PCIe-0 77936867 OK    150000   0.19% 127631386c6a9b17  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:07 NVIDIA H100 PCIe-0 77936867 OK    160000   0.21% 25b7b6206fc6f085  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:10 NVIDIA H100 PCIe-0 77936867 OK    170000   0.22% 416816b0d9f4bba8  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:13 NVIDIA H100 PCIe-0 77936867 OK    180000   0.23% 6bee5d054f770861  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:15 NVIDIA H100 PCIe-0 77936867 OK    190000   0.24% f37f068f014b18a0  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:18 NVIDIA H100 PCIe-0 77936867 Stopping, please wait..
20221217 21:18:18 NVIDIA H100 PCIe-0 77936867 OK    200000   0.26% f0b04b45b0855bd2  244 us/it + check 0.13s + save 0.10s; ETA 05:16
20221217 21:18:18 NVIDIA H100 PCIe-0  Exiting because "stop requested"
20221217 21:18:18 NVIDIA H100 PCIe-0  Bye

Oliver

[rebirther]

Can gpuowl running TF? If yes how the speed is compared to mfakto / mfaktc and GPU support? Cant find some useful informations.

[kriesel]

Originally Posted by rebirther

Can gpuowl running TF? If yes how the speed is compared to mfakto / mfaktc and GPU support? Cant find some useful informations.

Not for years. Mihai briefly tried on TF in gpuowl and quit it since its performance was not advantageous, and it only worked with rocm for the few old versions in which it appeared. Run mfaktx. See also old thread posts around https://www.mersenneforum.org/showpo...&postcount=540

[moebius]

Originally Posted by TheJudger

Hi!

H100 PCIe 80 GB limited to 310 Watt, quick&dirty GPUowl build:[CODE]

./gpuowl -log 10000 -iters 200000 -prp 77936867
20221217 21:17:24  GpuOwl VERSION v7.2-130-g967853c-dirty
20221217 21:17:24  GpuOwl VERSION v7.2-130-g967853c-dirty
20221217 21:17:24  Note: not found 'config.txt'
20221217 21:17:24  config: -log 10000 -iters 200000 -prp 77936867
20221217 21:17:24  device 0, unique id ''
20221217 21:17:24 NVIDIA H100 PCIe-0 77936867 FFT: 4M 1K:8:256 (18.58 bpw)
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 OpenCL args "-DEXP=77936867u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=8u -DMM_CHAIN=1u -DMM2_CHAIN=2u -DWEIGHT_STEP=0.33644726404543274 -DIWEIGHT_STEP=-0.25174750481886216 -DIWEIGHTS={0,-0.44011820345520131,-0.37306474779553728,-0.29798072935699788,-0.21390437908665341,-0.11975874301407295,-0.014337887291734644,-0.44814572555075455,} -DFWEIGHTS={0,0.78609128957452257,0.5950610473469905,0.42446232150303748,0.2721098723818392,0.1360521812214803,0.014546452690911484,0.81207258201996746,}  -cl-std=CL2.0 -cl-finite-math-only "
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 OpenCL compilation in 0.01 s
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 PRP starting from beginning
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 OK         0 on-load: blockSize 400, 0000000000000003
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 validating proof residues for power 8
20221217 21:17:25 NVIDIA H100 PCIe-0 77936867 Proof using power 8
20221217 21:17:27 NVIDIA H100 PCIe-0 77936867 OK     10000   0.01% fc4f135f7cf4ad29  243 us/it + check 0.13s + save 0.10s; ETA 05:16

With this graphic accelerator I was better at guessing again, I had 235 us/it.

gpuOwl benchmarks online

[kracker]

Out of the loop, but is stage 2 of P1 still doable by gpuowl or is Prime95 required to do that stage?