Reference material discussion thread

[kriesel]

In reference to building and running gpuowl on Google Colab:

Originally Posted by petrw1

I ran the first part and got several errors....maybe that is as expected:
I created worktodo.txt.....
However I do NOT know what goes into config.txt ... Thanks

THEN when I ran this:

Drive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount("/content/drive", force_remount=True).
  /bin/bash: ./gpuowl: Is a directory

I haven't built it in quite a while, so don't know what the current build behavior or quirks are, but that stream you posted seemed unusual to me. The IP# problems for example.
The build failed, with a fatal compile error, so there is no gpuowl executable as a result, only the directory for it.

For config.txt, options you put on the command line can go in there. In your example,

-use ORIG_X2 -block 200 -log 120000 -maxAlloc 10240 -user petrw1 -cpu colab/K80

If you're not attached to doing the compile of the latest commit, you could try running a build posted by Fan Ming.

[Dylan14]

In the script "conditional gpu runs full" in the Colab reference thread, I'd replace the following lines:

gpu_info = !nvidia-smi # and one for the script to look at
gpu_info = '\n'.join(gpu_info)

to these, which is based on my mlucas compile script:

!nvidia-smi > gpu_info.txt # and one for the script to look at
count_gpu = !grep -c failed gpu_info.txt # this counts the number of instances failed has appeared, should be 0 if we want to run GPU code
if count_gpu != 0:
     #code continues as before

For the parts where you look for specific GPU models, we could run the following

count_T4 = !grep -c T4 gpu_info.txt
count_P100 = !grep -c P100 gpu_info.txt
count_P4 = !grep -c P4 gpu_info.txt
count_K80 = !grep -c K80 gpu_info.txt
# then branch off into the different gpu codes using if/else, like if count_T4 != 0, run T4 code

Not sure if it's more efficient, but it's a good exercise in using grep and its options.

[kriesel]

I'd be interested in knowing of any gpuowl version/ igp model combination that works together. Gpuowl requires some aspects of OpenCL2, atomics as I recall, and DP capability.
Some igps have those available, some don't. UHD630 does. But I don't know of a build that works there. And igps are slow, even for TF, so it's not a priority.

gpuowl-v6.11-318>gpuowl-win
2020-06-12 12:16:38 gpuowl v6.11-318-g3109989
2020-06-12 12:16:38 config: -cpu peregine/uhd630 -user kriesel -device 0 -log 10000 -safeMath -use NO_ASM
2020-06-12 12:16:38 device 0, unique id ''
2020-06-12 12:16:38 peregine/uhd630 worktodo.txt line ignored: ";PRP=N/A,1,2,86243,-1,60,0"
2020-06-12 12:16:38 peregine/uhd630 worktodo.txt line ignored: ";PRP=0,1,2,110503,-1,60,0"
2020-06-12 12:16:38 peregine/uhd630 worktodo.txt line ignored: ";PRP=0,1,2,132049,-1,60,0"
2020-06-12 12:16:38 peregine/uhd630 216091 FFT: 128K 256:1:256 (1.65 bpw)
2020-06-12 12:16:38 peregine/uhd630 Expected maximum carry32: 00000
2020-06-12 12:16:38 peregine/uhd630 using long carry kernels
2020-06-12 12:16:38 peregine/uhd630 OpenCL args "-DEXP=216091u -DWIDTH=256u -DSMALL_HEIGHT=256u -DMIDDLE=1u -DPM1=0 -DWEIGHT_STEP=0xa.34c173fbe77ep-3 -DIWEIGHT_STEP=0xc.8aa2df2079a98p-4 -DNO_ASM=1  -cl-std=CL2.0 -cl-finite-math-only "
2020-06-12 12:16:43 peregine/uhd630 ASM compilation failed, retrying compilation using NO_ASM
2020-06-12 12:16:47 peregine/uhd630 OpenCL compilation error -11 (args -DEXP=216091u -DWIDTH=256u -DSMALL_HEIGHT=256u -DMIDDLE=1u -DPM1=0 -DWEIGHT_STEP=0xa.34c173fbe77ep-3 -DIWEIGHT_STEP=0xc.8aa2df2079a98p-4 -DNO_ASM=1  -cl-std=CL2.0 -cl-finite-math-only  -DNO_ASM=1)
2020-06-12 12:16:47 peregine/uhd630 1:58:26: warning: unsupported OpenCL extension 'cl_khr_int64_base_atomics' - ignoring
#pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable
                         ^
1:59:26: warning: unsupported OpenCL extension 'cl_khr_int64_extended_atomics' - ignoring
#pragma OPENCL EXTENSION cl_khr_int64_extended_atomics : enable
                         ^


error: undefined reference to `_Z8atom_addPU3AS1Vmm()'

error: backend compiler failed build.

2:58:26: warning: unsupported OpenCL extension 'cl_khr_int64_base_atomics' - ignoring
#pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable
                         ^
2:59:26: warning: unsupported OpenCL extension 'cl_khr_int64_extended_atomics' - ignoring
#pragma OPENCL EXTENSION cl_khr_int64_extended_atomics : enable
                         ^

error: undefined reference to `_Z8atom_addPU3AS1Vmm()'

error: backend compiler failed build.

2020-06-12 12:16:47 peregine/uhd630 Exception gpu_error: BUILD_PROGRAM_FAILURE clBuildProgram at clwrap.cpp:246 build
2020-06-12 12:16:47 peregine/uhd630 Bye

It came closer in V6.5-61:

2019-05-30 13:17:51 Note: no config.txt file found
2019-05-30 13:17:51 config: -device 0 
2019-05-30 13:17:51 85469147 FFT 4608K: Width 256x4, Height 64x4, Middle 9; 18.11 bits/word
2019-05-30 13:17:51 using short carry kernels
2019-05-30 13:18:42 OpenCL compilation in 50608 ms, with "-DEXP=85469147u -DWIDTH=1024u -DSMALL_HEIGHT=256u -DMIDDLE=9u  -I. -cl-fast-relaxed-math -cl-std=CL2.0"
2019-05-30 13:18:44 85469147.owl not found, starting from the beginning.
2019-05-30 13:25:50 85469147 EE     2000  0.00%; 95.53 ms/sq; ETA 94d 11:54; 91e7259a0ae0534b (check 96.17s)
2019-05-30 13:25:50 85469147.owl not found, starting from the beginning.
2019-05-30 13:32:39 85469147 EE     2000  0.00%; 156.09 ms/sq; ETA 154d 09:38; 91e7259a0ae0534b (check 96.44s)
2019-05-30 13:32:39 85469147.owl not found, starting from the beginning.
2019-05-30 13:41:12 Note: no config.txt file found
2019-05-30 13:41:12 config: -device 2 
2019-05-30 13:41:12 85469147 FFT 4608K: Width 256x4, Height 64x4, Middle 9; 18.11 bits/word
2019-05-30 13:41:12 using short carry kernels

[M344587487]

iGPU's share RAM with the CPU so unless the CPU can be kept busy largely from cache I don't think there's much point even if it's possible? Regardless I'd be interested in how the iGPU in the Renoir mobile APU's does if it's able at all (it might be, ROCm indicates the earlier APU's have limited unofficial ROCm OpenCL functionality). By rights those APU's should end up quite popular and the iGPU is much more performant than intel's UHD630, although "much more performant" is admittedly still a blip compared to the big boys.

[kriesel]

Originally Posted by M344587487

iGPU's share RAM with the CPU so unless the CPU can be kept busy largely from cache I don't think there's much point even if it's possible? Regardless I'd be interested in how the iGPU in the Renoir mobile APU's does if it's able at all (it might be, ROCm indicates the earlier APU's have limited unofficial ROCm OpenCL functionality). By rights those APU's should end up quite popular and the iGPU is much more performant than intel's UHD630, although "much more performant" is admittedly still a blip compared to the big boys.

Right. I've seen mfakto on an Intel igp reduce prime95 throughput on the cpu by half. And TF is not as memory hungry.
If I had put more effort into making the igp/gpu part of the poll choices, it would probably have split off into a separate poll question. A little more digging revealed this list.

• AMD igpu (https://www.cpu-monkey.com/en/amd_igpus)
• AMD discrete gpu such as RX550 to Radeon VII
• NVIDIA discrete gpu GTX10xx, RTX20XX etc
• Intel igp HDxxx, HD4xxx or UHDxxx etc
• Smartphone gpu (see https://deviceatlas.com/blog/most-used-smartphone-gpu) Many makers

Supposedly there will eventually be Intel discrete gpus (https://www.tomshardware.com/news/in...cs-all-we-know)

Huawei is a possible future PC gpu entrant.

[Dylan14]

I can confirm there is no HWinfo output for the serial number of a Intel GPU. However, using a different utility (GPU-Z) one can get the Device ID. See the attached photograph which shows the output on a laptop with a i7-8750H, GTX 1050 ti and UHD 630.

The downside with this method is that GPU-Z only works in Windows (XP+).

[LaurV]

Originally Posted by kriesel

History and evolution of Mersenne prime software

Antiquity
Precomputer

Say, search, or something else instead of software. What software are we talking about in the pre-computers' era or in antiquity?

A good start could be found in Mr. Caldwell's pages (he talks about antiquity, math, and how LL tests were first implemented long ago by Noll and co).

[DrobinsonPE]

Attached is a spreadsheet with my full testing data (prime95, mfakto, and gpuowl) for the A8-9600. I though about posting it in the GPU Computing forum but there did not seem to be a good spot for it and I did not feel like starting a new thread for an old CPU. I saw that you are keeping a reference thread for integrated graphic so I thought you might be interested in the additional details.

Interesting information:

• The iGPU is faster running gpuowl than the CPU is running prime95.
• With both Prime95 and gpuowl running, gpuowl errored three times in the same spot and ended.
• the iGPU seems to be favored over the CPU. Prime95 CPU utilization dropped to 45% or less when either gpuowl or mfakto were also running.
• The combined output of prime95 and gpuowl together is almost the same as gpuowl alone but with gpuowl failing it does not matter.
• Prime95 does not seem to slow down mfakto at all.
• The most efficient use of this CPU is to only use the iGPU for either PRP or TF.

Feel free to use this information or delete it.

[Xyzzy]

Originally Posted by LaurV

To add to it, if you want to show off to your friends, you buy the best swiss knife, but if you want to get any job done, buy a toolbox (hammer, screwdriver, pliers, etc). Each tool adequate to its purpose is more productive.

https://forum.multitool.org/index.php?topic=5649.0

[Dr Sardonicus]

The worst passenger ship disaster in the US, the burning of the General Slocum on June 15, 1904, killed over a thousand people. The ship had been inspected and found "safe" a month earlier. Contributing to the loss of life were inferior fire hoses which burst, lifeboats which were rusted in place and could not be launched, and life jackets which were stuffed with cork dust rather than solid cork. They soaked up water like sponges, and the rotted fabric split. As a result, many of the fatalities were people who donned life jackets, jumped into the water, were pulled under by the water-laden life jackets, and drowned. Many of them had large amounts of cork dust in their hair when they were pulled from the water. The ship company's records had been falsified to show it had purchased new life jackets.

The ship's owner and the inspector were acquitted at trial, while the ship's captain was sentenced to ten years in prison. He was pardoned in 1911 by President Taft.


edit kriesel: https://en.wikipedia.org/wiki/PS_General_Slocum is an interesting read. Notably, its legal capacity was 2500 passengers, and one of its prior accidents occurred with 4700 passengers aboard. The event Dr. Sardonicus wrote about was not however the worst passenger ship disaster on US waterways. That would be the Sultana, in 1865, returning Civil War POWs, with a badly repaired boiler and carrying far more than its rated capacity, killing an estimated 1500-1900 on a ship rated for 376.

[axn]

Originally Posted by kriesel

Since the cost of P-1 B1 is very small in Gpuowl V7.x, assuming a full PRP will be completed if no factor is found by either stage of P-1, there is good reason for the B1 selection to differ considerably for Gpuowl V7.x, using much larger B1. The basis of the much larger B2 is less clear.

v7 also has faster stage 2 (similar to 30.4), so a doubling of B2 would be understandable. But 5x? Not sure about that.

EDIT:- Since v7 integrates P-1 & PRP, it is probably optimal to do 4x B1 and ~2x B2 of the mersenne.ca bounds, so maybe something like (2.5M, 50M)