CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW)

[ZFR]

Where can I get CUDALucas for Linux? The sourceforge link only has Windows execs.

https://sourceforge.net/projects/cudalucas/

[James Heinrich]

Originally Posted by ZFR

Where can I get CUDALucas for Linux?

https://download.mersenne.ca/CUDALucas

[ZFR]

Originally Posted by James Heinrich

https://download.mersenne.ca/CUDALucas

Thank you.
And the libs? I got only got 6.5 of libcudart.so and libcufft.so. Can I use them?

[James Heinrich]

I have no idea. If someone wants to point me to the libs I can host them too.

[kriesel]

Originally Posted by ZFR

Thank you.
And the libs? I got only got 6.5 of libcudart.so and libcufft.so. Can I use them?

CUDALucas and libs need to match on CUDA version. Driver version needs to support that CUDA version. And gpu compute capability needs to support that CUDA version. A sufficiently old and sufficiently new combination of gpus may be incompatible. (Driver new enough to support the new gpu may not support the old gpu.) So "can I use them" depends on your gpu model. See https://www.mersenneforum.org/showpost.php?p=490615&postcount=4

[ZFR]

My GPU uses CUDA 10.2

Is there a link where I can get the CUDA 10.2 libs? Aside from having to install the whole cuda development toolkit?

[kriesel]

Originally Posted by ZFR

My GPU uses CUDA 10.2

Is there a link where I can get the CUDA 10.2 libs? Aside from having to install the whole cuda development toolkit?

Depending on what linux you are running, you could try something like what's done to get google colab ready for cudalucas:

apt-get install -y cuda-cudart-10-2
apt-get install -y cuda-cufft-dev-10-2

(and there may also be work needed on path)

https://www.mersenneforum.org/showpo...29&postcount=6

[ZFR]

Originally Posted by kriesel

Depending on what linux you are running, you could try something like what's done to get google colab ready for cudalucas:

apt-get install -y cuda-cudart-10-2
apt-get install -y cuda-cufft-dev-10-2

(and there may also be work needed on path)

https://www.mersenneforum.org/showpo...29&postcount=6

Thank you, I'm using Linux Mint, so your instructions worked.

Thanks again both for your help.

[kriesel]

Originally Posted by ZFR

Thank you, I'm using Linux Mint, so your instructions worked.

Thanks again both for your help.

You're welcome.
I had forgotten about the linux setup script linked from https://www.mersenneforum.org/showpo...12&postcount=1
See also https://www.mersenneforum.org/showpo...20&postcount=5
Tune cudalucas for your gpu and test for reliability.
-memtest
fft benchmarking
threads benchmarking
built in self test (which only covers up to 8M fft length)
I recommend at least retest memory annually.
Be aware that CUDALucas does not include a Jacobi check, so some errors may go silently undetected.
Gpuowl now includes LL DC with Jacobi check (but not shift) and may be slightly faster on a modern NVIDIA gpu.

Have fun!

[ZFR]

Originally Posted by kriesel

fft benchmarking
threads benchmarking

After I run the benchmarks, do I just leave the files in the program's folder?

[kriesel]

Originally Posted by ZFR

After I run the benchmarks, do I just leave the files in the program's folder?

Yes. The program uses those files to decide what fft length and threads choices to try first.
Have a look inside them if you haven't yet.
During my early testing, I hit some cases where anomalies showed up in the files, particularly for older CUDA versions, 4.x I think. If the benchmark runs anomalously fast at some lengths or thread combinations, the calculations using those lengths will probably be wrong.
See https://mersenneforum.org/showpost.p...postcount=2591
and https://mersenneforum.org/showpost.p...postcount=2608