ECM for CUDA GPUs in latest GMP-ECM ?

[EdH]

Originally Posted by SethTro

I'm happy to look at this as a bug. I vaguely remember that I wasn't sure if I should always stop or only if the cofactor is composite.

Thanks, but I've moved to using a separate machine with ecm.py and it is working as needed.

I'd leave as is unless others are interested. For me it is no longer an issue.

Again, thank you for considering it.

[storm5510]

There seems to be an oddity about using a GPU in GMP-ECM.

The --help switch shows the maximum size of a B1 value is 2^1018. Using Windows Calculator to calculate log(2) * 1018 says this is a 308 digit value.

If I run this:

echo "2^14447-1" | gpuecm -gpu -maxmem 2048 5e6 10e6

I get this message:

GMP-ECM 7.0.4-dev [configured with MPIR 2.7.2, --enable-gpu, --enable-openmp] [ECM]
Input number is 2^14447-1 (4349 digits)
GPU: Error, input number should be strictly lower than 2^1018
please report internal errors at <ecm-discuss@lists.gforge.inria.fr>.

B1 in my example is 5e6 or 5,000,000. This is obviously smaller than a 308 digit number.

I am missing something here and I cannot determine what it is.

[paulunderwood]

Originally Posted by storm5510

There seems to be an oddity about using a GPU in GMP-ECM.

The --help switch shows the maximum size of a B1 value is 2^1018. Using Windows Calculator to calculate log(2) * 1018 says this is a 308 digit value.

If I run this:

echo "2^14447-1" | gpuecm -gpu -maxmem 2048 5e6 10e6

I get this message:

B1 in my example is 5e6 or 5,000,000. This is obviously smaller than a 308 digit number.

I am missing something here and I cannot determine what it is.

You are missing something. The inputted number is too big. You specified 2^14447-1. The max is 2^1018.

[SethTro]

If you are willing to compile gmp-ecm from source you can test larger input numbers (it's also faster).

See https://www.mersenneforum.org/showthread.php?t=27103

[storm5510]

Originally Posted by paulunderwood

You are missing something. The inputted number is too big. You specified 2^14447-1. The max is 2^1018.

Understood. IMHO, it is not good for much of anything at this size.

[kruoli]

Is there any known progress in implementing stage 1 in (at least somewhat optimised) OpenCL up to 512 or maybe 1024 bits?

[SethTro]

Originally Posted by kruoli

Is there any known progress in implementing stage 1 in (at least somewhat optimised) OpenCL up to 512 or maybe 1024 bits?

I don't know of anyone working on this. I did some quick google searches and didn't find any OpenCl arbitrary precision libraries. The old ECM gpu code was fairly straightforward grade school multiplication algorithm (IIRC) that could be adapted quickly for OpenCl, but if you/someone wanted competitive speeds from OpenCl more code would be needed for montgomery multiplication, various size kernels, etc.

[storm5510]

Originally Posted by SethTro

I don't know of anyone working on this. I did some quick google searches and didn't find any OpenCl arbitrary precision libraries. The old ECM gpu code was fairly straightforward grade school multiplication algorithm (IIRC) that could be adapted quickly for OpenCl, but if you/someone wanted competitive speeds from OpenCl more code would be needed for montgomery multiplication, various size kernels, etc.

AFAIK, nothing meaningful has been done with GMP-ECM for a long time. I stopped using it several years ago. It was not worth the time required to run anything.

OpenCL appears not to work a GPU very hard. I have used gpuOwl a lot, so I know. It does a good job, but there is a limit to its GPU utilization. Around 60% seems to be its limit.

[kriesel]

Originally Posted by storm5510

OpenCL appears not to work a GPU very hard. I have used gpuOwl a lot, so I know. It does a good job, but there is a limit to its GPU utilization. Around 60% seems to be its limit.

That statement is not consistent with my experience. I routinely see 99-100% GPU load indicated in GPU-Z for Gpuowl running on Radeon VII, RX480, or RX550.

[storm5510]

Originally Posted by kriesel

That statement is not consistent with my experience. I routinely see 99-100% GPU load indicated in GPU-Z for Gpuowl running on Radeon VII, RX480, or RX550.

I have no experience with AMD GPU's. All of mine, past and present, are Nvidia based. Something tells me that AMD variations are much better suited to run OpenCL than Nvidia models are. CUDA is much better for Nvidia, it seems. mfaktc cranks out over 3,000 Ghz-d/day on my ever-older RTX-2080. I am satisfied with that.

[kriesel]

NVIDIA quadro k620, gpuowl v6.11-382, gpu-Z v2.52.0, indicated gpu load 100%.
Most recent NVIDIA consumer GPUs have poor DP performance (PRP, LL, P-1) relative to SP performance (TF), while Teslas and AMD typically have decent DP performance ratios. I almost never use RTX2080 or GTX 1650 for anything other than TF because of the 32:1 SP/DP ratio.