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

[flashjh]

Originally Posted by flashjh

4.11 build is slower than 1.2b for me. I had to install CUDA 4.0 to get the lastest .dll files.

Anyone know why the newer one is slower or something I can do to make it faster? Anyone know how to get the ETA back in 4.11?

Originally Posted by Brain

As CUDALucas does auto-resume from checkpoint files we should recommend not using "-c" any more, do we? I will have to update the GPU guide in the new year...

By the way, the iteration times are so low as I didn't do complete 10000 runs. Kind of a bug.

Last but not least, utilisation for state-of-the-art expos (40M range) is 97% as before. Low utilisation is understandable for small FFT sizes...

I have not tried 1.3, but 1.4.1 is about 2x slower for me than 1.2b. I dropped the -c and -t and only use -D01 for GPU 2. I'm still learning... Is it better to use 1.4.1 to get the non-power-of-2-fft-sizes or use 1.2b to optimuze speed? Thanks.

[Brain]

Originally Posted by Brain

As CUDALucas does auto-resume from checkpoint files we should recommend not using "-c" any more, do we? I will have to update the GPU guide in the new year...

Stupid me: If we don't use -c there are no checkpoints written. So we have to use it. I will have time to test more in the new year but currently I recommend using 1.2(b)!

DLL files download

[Brain]

Hi,
although nobody has yet confirmed that CL >= 1.4 uses more CPU resources I wrote msft this question. He answered that he is investigating and will try to fix it.
I CUDALucas (and all GPU based primality tests)
Thanks to msft!

[msft]

Hi ,

Originally Posted by Brain

although nobody has yet confirmed that CL >= 1.4 uses more CPU resources I wrote msft this question.

Fixed CPU issue.
Thank you,

[msft]

exec file.

[Brain]

Originally Posted by msft

Hi ,

Fixed CPU issue.
Thank you,

Thanks a lot. I noticed you didn't change the version string. For Win64 compilation I will change it to:

const char program_revision[] = "$Revision: 1.4.2 $";

My last test with 1.41 didn't show the high CPU utlilisation any more, see attached. I am a bit confused. Maybe I made a mistake with the "-c" param so that CL wrote it every iteration. Just guessing.

Will now compile again.

[Brain]

Maybe even a bit faster. We'll see.

[Brain]

CUDALucas 1.4.2 for Win64 / CUDA 4.0 / Compute Capability 1.3

[Brain]

CUDALucas 1.4.2 for Win64 / CUDA 4.0 / Compute Capability 2.1

[msft]

CUFFT benchmark with cuda4.0:

CUFFT_D2Z size=512 k time=1.269744 msec
CUFFT_D2Z size=1024 k time=2.609184 msec
CUFFT_D2Z size=1536 k time=4.359898 msec
CUFFT_D2Z size=2048 k time=5.615232 msec
CUFFT_D2Z size=2560 k time=7.277350 msec
CUFFT_D2Z size=3072 k time=8.969321 msec
CUFFT_D2Z size=3584 k time=10.251376 msec
CUFFT_D2Z size=4096 k time=11.749495 msec
CUFFT_D2Z size=4608 k time=13.065844 msec
CUFFT_D2Z size=5120 k time=14.971667 msec
CUFFT_D2Z size=5632 k time=148.589874 msec
CUFFT_D2Z size=6144 k time=19.145430 msec
CUFFT_D2Z size=6656 k time=217.340973 msec
CUFFT_D2Z size=7168 k time=21.095901 msec
CUFFT_D2Z size=7680 k time=24.699974 msec
CUFFT_D2Z size=8192 k time=24.172211 msec

Some fft length was very slow.
Ver 1.42 not avoid this length.

[f11ksx]

V1.41 is running pretty well just like v1.42

9.3 ms/iter for 54M exponent on GTX-580 card with v1.41
8.6 ms/iter with v1.42

Domo