mfaktc: a CUDA program for Mersenne prefactoring

[Mark Rose]

I've had hundreds of lines in worktodo.txt before. Interesting.

[Gordon]

Originally Posted by Karl M Johnson

Found my first bug for mfaktc by getting overnight TF work.
It would seem that 128 lines in worktodo file cause some kind of overflow, and mfaktc complains about invalid format in such a case (and doesn't start).
Will report further findings later.

I currently have 1723 lines in my worktodo and it's running sweet as a nut.

The new version puts out just over 500 days/day on my 970, previous version was about 470 days/day

[James Heinrich]

Originally Posted by TheJudger

P.P.S. How high is the demand for Wagstaff and/or LessClasses?

I would normally request LessClasses, but my newest GPU is a 670 so I don't have any immediate need for it until a GTX 1080 magically falls out of the sky

[kladner]

CUDA 8.0 made little difference on my currently-running 580. No surprise, for something that old.

[flashjh]

Originally Posted by James Heinrich

I would normally request LessClasses, but my newest GPU is a 670 so I don't have any immediate need for it until a GTX 1080 magically falls out of the sky

I just compiled the 32 and 64 bit versions of lessclass, Wagstaff, and lessclass Wagstaff.

I emailed to Oliver a few minutes ago, I'm sure he'll post them soon ;)

[flashjh]

Originally Posted by kladner

CUDA 8.0 made little difference on my currently-running 580. No surprise, for something that old.

On this note, I'm sad to say that the workhorse 580s are soon obsolete based on CUDA versions -- If they're not already ;)

Current compiler warning:

nvcc warning : The 'compute_20', 'sm_20', and 'sm_21' architectures are deprecated, and may be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).

Won't change much either way, just sad.

[flashjh]

Originally Posted by Karl M Johnson

Found my first bug for mfaktc by getting overnight TF work.
It would seem that 128 lines in worktodo file cause some kind of overflow, and mfaktc complains about invalid format in such a case (and doesn't start).
Will report further findings later.

Back in my HayDay of running mfaktc I had thousands of lines in many instances running with no problems. Have you checked for any non-text characters in the file? That might cause a problem.

[Mark Rose]

Originally Posted by flashjh

Back in my HayDay of running mfaktc I had thousands of lines in many instances running with no problems. Have you checked for any non-text characters in the file? That might cause a problem.

I did have a problem with line endings once, IIRC.

[Karl M Johnson]

Originally Posted by Mark Rose

I did have a problem with line endings once, IIRC.

Nailed it!
Turns out, it was not related to mfaktc, but which EOL worktodo.txt used.
I've realised this only when I've noticed a 'Macintosh' EOL being used n Notepad++ and I'm not sure how this happened.
False alarm

[kladner]

Originally Posted by flashjh

On this note, I'm sad to say that the workhorse 580s are soon obsolete based on CUDA versions -- If they're not already ;)

Current compiler warning:

nvcc warning : The 'compute_20', 'sm_20', and 'sm_21' architectures are deprecated, and may be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).

Won't change much either way, just sad.

One point I am uncertain about: I have always run 32 bit versions of mfaktc, because they have a very slight performance edge over the 64 bit. My experiment was with the 64 bit version of CUDA 8.0. Still, any differences were within normal fluctuations in everyday running.

[James Heinrich]

Originally Posted by Karl M Johnson

Turns out, it was not related to mfaktc, but which EOL worktodo.txt used.

Vaguely related: I discovered while writing a script to fetch new worktodo that mfaktc always rewrites all of worktodo.txt and always uses \r\n even if the file had \n line-endings (which mfaktc can read fine).