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

[flashjh]

Welcome back!

A few of us have some particular FFTs that won't work and the issue with the program stopping will get addressed. We're using a simple loop to keep CUDALucas going if it stops until the code is fixed.

As for tests:

1) For your cards, you should run the batch file attached for each card. It will take a while and some of the FFTs may fail as you've experienced, but it will create two files that help fine-tune CUDALucas for each card.

2) Run the built-in memtest. CUDALucas -memtest k n. Read from mid Nov threads until now to see more info.

3) Run the built-in test CUDALucas -r. Make sure all residues match.

The results are accepted as long as the exponent(s) don't already have a CUDALucas/mlucas residue. Download the latest version from sourceforge and it will format the results.txt file correctly. Use the format to properly format previous results.

If you have any bugs/suggestions, let us know. Thanks for testing and your contribution.

[LaurV]

Originally Posted by flashjh

The results are accepted as long as the exponent(s) don't already have a CUDALucas/mlucas same residue (i.e. different residues are accepted, the server can't know which one is good, until DC-ed) and as long as you don't use the "user/computer/timestamp" option of cudaLucas. You ca use manual report form to report the results.

underlined text is mine. The rest is a Jerry said.

[flashjh]

Said much better, thanks.

[henryzz]

Originally Posted by owftheevil

The code that generates the security code could be a separate application.

The API to call that program wouldn't be secret though and that could probably be abused.

[chappjc]

Thanks for the advice. I was running r47, so the formatting changes were not included. Once I reformatted the results.txt, PrimeNet recognized it. I left the "AID" part at the end of the line. Should I run the same exponent again just to verify? This result is from the 670. I'll have another result in just 92 hours!

All cards pass all residue tests (CUDALucas -r). I ran a few very short memory tests (i.e. -memtest 6 2), and a longer one is presently running on the 580.

One thing to note about the 580 that always has runtime API errors, is that it is also display card. Often the driver stops responding and recovers (331.82). The other two cards on which I have never seen a runtime error (yet) are on two different machines and are not the display cards.

I ran the batch script, which generated the fft and threads .txt files, but some of the results are surprising to me. At 2592k, the optimal threads drops off:

...
 2048  512  512  256   2.8779
 2240  512  512  256   3.3209
 2304  512  512  128   3.3607
 2352  512  512 1024   3.8242
 2592   64   32   32   3.9552
 2688   64   64   32   4.6925
 2880   64   32   32   4.6117
 3024   64   32   32   5.1544
 3136   64   32   32   4.9940
...

That probably makes sense for a 580 with 3GB, but I just wanted to make sure.

[owftheevil]

Originally Posted by henryzz

The API to call that program wouldn't be secret though and that could probably be abused.

You are right. I thought more about it last night and came to the same conclusion. Personally, I have no problem with changing the license to account for a closed source authenticator.

[owftheevil]

Originally Posted by chappjc

Thanks for the advice. I was running r47, so the formatting changes were not included. Once I reformatted the results.txt, PrimeNet recognized it. I left the "AID" part at the end of the line. Should I run the same exponent again just to verify? This result is from the 670. I'll have another result in just 92 hours!

All cards pass all residue tests (CUDALucas -r). I ran a few very short memory tests (i.e. -memtest 6 2), and a longer one is presently running on the 580.

One thing to note about the 580 that always has runtime API errors, is that it is also display card. Often the driver stops responding and recovers (331.82). The other two cards on which I have never seen a runtime error (yet) are on two different machines and are not the display cards.

I ran the batch script, which generated the fft and threads .txt files, but some of the results are surprising to me. At 2592k, the optimal threads drops off:

...
 2048  512  512  256   2.8779
 2240  512  512  256   3.3209
 2304  512  512  128   3.3607
 2352  512  512 1024   3.8242
 2592   64   32   32   3.9552
 2688   64   64   32   4.6925
 2880   64   32   32   4.6117
 3024   64   32   32   5.1544
 3136   64   32   32   4.9940
...

That probably makes sense for a 580 with 3GB, but I just wanted to make sure.

That looks fishy to me. The third thread parameter flops around a lot, but the first two are usually pretty stable. How are the timings as compared to the fft bench test? I'd like to see the corresponding section of <gpu> fft.txt.

[flashjh]

Originally Posted by chappjc

<>Should I run the same exponent again just to verify? This result is from the 670. I'll have another result in just 92 hours!

You can run it again, but use P95. Otherwise, just let the natural DC process test it (whenever that will happen). Also, if you're in the process of 'verifying' that your cards are stable, I recommend you pull DCs from Primenet or GPU72; that way you will know if your card is producing good results or not. If it mismatches, you can post it here which will tell others not to use CUDALucas to DC/TC the exponent. Sometimes folks will do a quick run on it for you so you can see which one (or both) was wrong. You can always do another run on the GPU, Primenet won't accept the run unless the residue is different.

One thing to note about the 580 that always has runtime API errors, is that it is also display card. Often the driver stops responding and recovers (331.82). The other two cards on which I have never seen a runtime error (yet) are on two different machines and are not the display cards.

I have a 580 with the same issue, and others have this problem with other cards. owftheevil said it's caused by the drivers, but it will get fixed. My 580 is not the display card and it still happens.

[owftheevil]

Originally Posted by flashjh

You can run it again, but use P95. Otherwise, just let the natural DC process test it (whenever that will happen). Also, if you're in the process of 'verifying' that your cards are stable, I recommend you pull DCs from Primenet or GPU72; that way you will know if your card is producing good results or not. If it mismatches, you can post it here which will tell others not to use CUDALucas to DC/TC the exponent. Sometimes folks will do a quick run on it for you so you can see which one (or both) was wrong. You can always do another run on the GPU, Primenet won't accept the run unless the residue is different.

I have a 580 with the same issue, and others have this problem with other cards. owftheevil said it's caused by the drivers, but it will get fixed. My 580 is not the display card and it still happens.

Maybe I misunderstand you, but the problem won't be fixed until Nvidia does something about their drivers. All I'm trying to do is make the batch files unnecessary for restarting CL when the error does occur. It won't take away the fft hangs, resetting drivers etc. By the way I have it working on Linux, but Windows is again another story.

[flashjh]

Originally Posted by owftheevil

Maybe I misunderstand you, but the problem won't be fixed until Nvidia does something about their drivers. All I'm trying to do is make the batch files unnecessary for restarting CL when the error does occur. It won't take away the fft hangs, resetting drivers etc. By the way I have it working on Linux, but Windows is again another story.

Ok, so you can detect and restart, but the 'real' problem is the drivers? I thought it was a good fix. Sorry for the confusion.

If you have the code working for Linux, can you commit/merge it with the changes on SourceForge so I can take a look at it on Windows?

[chappjc]

Originally Posted by owftheevil

That looks fishy to me. The third thread parameter flops around a lot, but the first two are usually pretty stable. How are the timings as compared to the fft bench test? I'd like to see the corresponding section of <gpu> fft.txt.

From "GeForce GTX 580 fft.txt":

 2048   38492887   2.9761
 2160   40551479   3.5742
 2240   42020509   3.6679
 2304   43194913   3.6846
 2592   48471289   3.9861
 2880   53735041   4.6150
 3072   57237889   4.9730
 3136   58404433   4.9740

Do you want to see the full output from -cufftbench 2592 2592 6?