gpuOwL: an OpenCL program for Mersenne primality testing

[kracker]

RX 570. Probably can get a little more out of it if I tune for this FFT size...

2020-11-09 06:31:21 GpuOwl VERSION v7.2-16-g1a50f11
2020-11-09 06:31:21 GpuOwl VERSION v7.2-16-g1a50f11
2020-11-09 06:31:21 Note: not found 'config.txt'
2020-11-09 06:31:21 config: -prp 77936867
2020-11-09 06:31:21 device 0, unique id ''
2020-11-09 06:31:22 Ellesmere-0 77936867 FFT: 4M 1K:8:256 (18.58 bpw)
2020-11-09 06:31:22 Ellesmere-0 77936867 OpenCL args "-DEXP=77936867u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=8u -DAMDGPU=1 -DMM_CHAIN=1u -DMM2_CHAIN=2u -DMAX_ACCURACY=1 -DWEIGHT_STEP_MINUS_1=0.33644726404543274 -DIWEIGHT_STEP_MINUS_1=-0.25174750481886216 -DIWEIGHTS={0,-0.25174750481886216,-0.44011820345520131,-0.16213409745771243,-0.37306474779553728,-0.061788266441989627,-0.29798072935699788,-0.47471232907613115,-0.21390437908665341,-0.41180199020062258,-0.11975874301407295,-0.3413572830988989,-0.014337887291734644,-0.26247586476052853,-0.44814572555075455,-0.17414732433395128,}  -cl-std=CL2.0 -cl-finite-math-only "
2020-11-09 06:31:22 Ellesmere-0 77936867 ASM compilation failed, retrying compilation using NO_ASM
2020-11-09 06:31:25 Ellesmere-0 77936867 OpenCL compilation in 3.29 s
2020-11-09 06:31:25 Ellesmere-0 77936867 maxAlloc: 0.0 GB
2020-11-09 06:31:25 Ellesmere-0 77936867 You should use -maxAlloc if your GPU has more than 4GB memory. See help '-h'
2020-11-09 06:31:25 Ellesmere-0 77936867 P1(0) 0 bits
2020-11-09 06:31:27 Ellesmere-0 77936867 OK       800 on-load: blockSize 400, 1579c241dc63eca6
2020-11-09 06:31:27 Ellesmere-0 77936867 validating proof residues for power 8
2020-11-09 06:31:27 Ellesmere-0 77936867 Proof using power 8
2020-11-09 06:31:31 Ellesmere-0 77936867 OK      1600   0.00% 0f62a1fcc1c78fe9 3389 us/it + check 1.44s + save 0.20s; ETA 3d 01:22
2020-11-09 06:32:00 Ellesmere-0 77936867        10000   0.01% fc4f135f7cf4ad29 3384 us/it
2020-11-09 06:32:34 Ellesmere-0 77936867        20000   0.03% 3cd1bd9d5e09cbc5 3385 us/it
2020-11-09 06:33:07 Ellesmere-0 77936867        30000   0.04% c4e0ff35e3290d98 3385 us/it
2020-11-09 06:33:41 Ellesmere-0 77936867        40000   0.05% dffe1b1b0d748128 3386 us/it
2020-11-09 06:34:15 Ellesmere-0 77936867        50000   0.06% 52e286945371ed29 3385 us/it
2020-11-09 06:34:49 Ellesmere-0 77936867        60000   0.08% 0945da4dc08bdd95 3385 us/it

Also on another note, how exactly do I upload proofs? Yeah, yeah I know I'm late to the party... I couldn't find any instructions in the repository(or likely I just did not look hard enough)

[DrobinsonPE]

Originally Posted by M344587487

Nice that you got it working. What is your set up, simply the latest ROCm on Mint 20 like you used elsewhere?

Not quite. That was it running on Windows 10 using the same configuration it was running when I had it in the Deskmini A300.

Here is an interesting comparison. See the attached picture.

Blue Text, Configuration 1 - Deskmini A300, 3200G, 16GB DDR-4 SO-DIMM, SSD
Purple Text, Configuration 2 - B450-HDV V4.0, 3200G, 16GB DDR-4 DIMM, SSD

I am not finished because I still need to get mfakto and gpuowl working on linux to complete the comparison table but I am making progress.

[moebius]

Originally Posted by kracker

RX 570. ..Also on another note, how exactly do I upload proofs?

Thx, read this post
https://mersenneforum.org/showpost.p...3&postcount=19

[kriesel]

Originally Posted by kracker

how exactly do I upload proofs? Yeah, yeah I know I'm late to the party... I couldn't find any instructions in the repository(or likely I just did not look hard enough)

How to upload them? Let me count the ways. https://www.mersenneforum.org/showpo...0&postcount=26
The standalone uploader worked for me in v7.0 and v7.1 which was the transition to proof v2. https://www.mersenneforum.org/showpo...&postcount=110

[thyw]

I am getting this error: "expected maximum carry: ...", then gpuowl terminates

2020-11-09 23:38:06 config: -log 500000 -use NO_ASM -use STATS -safeMath -B1 650000 -B2 22000000
2020-11-09 23:38:06 config: 
2020-11-09 23:38:06 config: 
2020-11-09 23:38:06 config: 
2020-11-09 23:38:06 device 0, unique id ''
2020-11-09 23:38:06 rx460 101023669 FFT: 5.50M 1K:11:256 (17.52 bpw)
2020-11-09 23:38:06 rx460 Expected maximum carry32: 32630000
2020-11-09 23:38:09 rx460 OpenCL args "-DEXP=101023669u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=11u -DPM1=1 -DCARRYM64=1 -DWEIGHT_STEP_MINUS_1=0xc.b9443732e4938p-5 -DIWEIGHT_STEP_MINUS_1=-0x9.1a969d46d6d3p-5 -DNO_ASM=1 -DSTATS=1  -cl-std=CL2.0 -cl-finite-math-only "

but help says

FFT 5.50M [ 8.65M - 106.88M]

Did my gpu go wrong?

[moebius]

Did you try -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only instead of
-safeMath
but that's more of a case for preda to help.

[thyw]

Originally Posted by moebius

Did you try -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only instead of
-safeMath
but that's more of a case for preda to help.

It didn't help, gonna try more parameters.

[moebius]

Originally Posted by thyw

It didn't help, gonna try more parameters.

Which version number of gpuowl do you use, Linux or Windows? I am convinced that someone here can help you.
Please try one of this versions
Fastest Windows Version (with Proofs)
https://mersenneforum.org/showpost.p...94&postcount=8
Linux binary version Ubuntu 18.04
https://mersenneforum.org/showpost.p...1&postcount=40

[kriesel]

gpuowl version Vx.y-n? OS? system ram? I've had issues with gpuowl, cured by adding system ram in low ram systems (bumping them up toward 16GB) especially with multiple gpus.
Try dropping -use STATS, add -maxAlloc with something reasonable (AT MOST total gpu ram minus 1GB or 20% , maybe lower)
Techpowerup says RX460 is a 2GB gpu.

On a 2GB RX550, I get away with the following, usually, as a config.txt;

-device 1 -user kriesel -cpu roa/rx550 -use NO_ASM -maxAlloc 1500 -proof 8
Sometimes not.

2020-11-09 06:07:48 roa/rx550 saved
2020-11-09 06:07:48 roa/rx550 Exception gpu_bad_alloc: GPU size 46137344
2020-11-09 06:07:48 roa/rx550 waiting for background GCDs..
2020-11-09 06:07:48 roa/rx550 Bye

This is on a Windows10 x64 system, running wavefront P-1. It's ok on PRP. Generally a restart gets it through. P-1 stage 2 is a long slog at only 15 buffers.

2020-11-09 18:12:10 gpuowl v6.11-380-g79ea0cc
2020-11-09 18:12:10 config: -device 1 -user kriesel -cpu roa/rx550 -use NO_ASM -maxAlloc 1500 -proof 8
2020-11-09 18:12:10 device 1, unique id ''
2020-11-09 18:12:10 roa/rx550 100906081 FFT: 5.50M 1K:11:256 (17.50 bpw)
2020-11-09 18:12:10 roa/rx550 Expected maximum carry32: 31AE0000
2020-11-09 18:12:13 roa/rx550 OpenCL args "-DEXP=100906081u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=11u -DPM1=1 -DAMDGPU=1 -DCARRYM64=1 -DWEIGHT_STEP_MINUS_1=0xd.5c397eef9d7ep-5 -DIWEIGHT_STEP_MINUS_1=-0x9.6cd7eb4f2e5fp-5 -DNO_ASM=1  -cl-unsafe-math-optimizations -cl-std=CL2.0 -cl-finite-math-only "
2020-11-09 18:12:23 roa/rx550 OpenCL compilation in 10.49 s
2020-11-09 18:12:24 roa/rx550 100906081 P1 B1=1000000, B2=30000000; 1442134 bits; starting at 1442133
2020-11-09 18:12:24 roa/rx550 100906081 P1  1442134 100.00%; 374791 us/it; ETA 0d 00:00; 6b144dfd0aeb22d7
2020-11-09 18:12:25 roa/rx550 P-1 (B1=1000000, B2=30000000, D=30030): primes 1779361, expanded 1899017, doubles 303358 (left 1203152), singles 1172645, total 1476003 (83%)
2020-11-09 18:12:25 roa/rx550 100906081 P2 using blocks [33 - 999] to cover 1476003 primes
2020-11-09 18:12:25 roa/rx550 100906081 P2 using 15 buffers of 44.0 MB each
2020-11-09 18:14:05 roa/rx550 100906081 P1 GCD: no factor
2020-11-09 18:15:32 roa/rx550 100906081 P2   15/2880: 7596 primes; setup  1.26 s,  24.330 ms/prime

Rough estimate 9.5 hours for stage 2.
A 4GB RX550 in the same system, config.txt as follows:
-device 0 -user kriesel -cpu roa/rx550 -proof 8 -maxAlloc 3200 -use NO_ASM

2020-11-09 09:54:05 roa/rx550 100906349 P2   54/2880: 27978 primes; setup  4.88 s,  19.821 ms/prime
2020-11-09 10:03:22 roa/rx550 100906349 P2  108/2880: 27943 primes; setup  5.14 s,  19.737 ms/prime
2020-11-09 10:12:40 roa/rx550 100906349 P2  162/2880: 28015 primes; setup  4.97 s,  19.749 ms/prime

Note 54 buffers, estimated 8.3 hours stage 2, 13% less time.

[thyw]

gpuowl 380, wasn't aware that 364 is faster
Win 10, 3570 8gb ram rx 460
It was working for the longest time, but in the middle of a 100.7M exponent stage 2, it suddenly output this same message about the carry.
Since then, tried using smaller exponent, it outputs the same error. Maxalloc, new instance, every combination of parameters...
It creates a crash file in every time it errors C:\ProgramData\Microsoft\Windows\WER\ReportArchive , it appears to be an 'error - modul version and modul name' It isn't in English, to be fair. Tried to downgrade my driver to the certified stable, same error 20.9.1.


I think i solved it, had to specify device number (-d 1), weird, it never asked for it until now. Thanks for all the help.

[kriesel]

Originally Posted by thyw

I think i solved it, had to specify device number (-d 1), weird, it never asked for it until now. Thanks for all the help.

If 3570 meant this, perhaps the igp suddenly began to show up as an available OpenCL device. (HD2500 listed as a device in gpuowl -h output?)