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

[Dubslow]

Originally Posted by fivemack

You're measuring the size of the memory required to write down the candidate number N; what's relevant for CUDALucas is the amount of memory required to perform an FFT-based multiply of numbers the size of N, which is a good deal bigger (firstly because you can only fit about eighteen bits of N into each 64-bit floating-point number, and secondly because the CUDA FFT uses a lot of scratch space to allow it to use hundreds of threads).

Ah, okay. Makes much more sense, especially the threadedness. Thanks.

[aaronhaviland]

Originally Posted by Lorenzo

All done works!

Iteration 10000 M( 290000033 )C, 0x9e2181b8405a2511, n = 16777216, CUDALucas v1.2

FFT size 16M!

Err, actually not: the cufft size is initialiased as only half of n:

cufftPlan1d(&plan, n/2, CUFFT_Z2Z, 1)

[Lorenzo]

Originally Posted by aaronhaviland

Err, actually not: the cufft size is initialiased as only half of n:

cufftPlan1d(&plan, n/2, CUFFT_Z2Z, 1)

Too bad.

Oops, i find in FAQ:

Exponents: 2<=Exponent< 151,150,000
Exponent< 39,800,00  2MB FFT size
Exponent< 79,600,00  4MB FFT size
Exponent<159,200,00  8MB FFT size

No more question

[Brain]

Originally Posted by Lorenzo

No more question

No, there are still basic questions (for me):
1. what is the max expo possible? how much GPU memory is needed for each FFT size?
2. what are the fft size borders beyond 4m?
i only tried up to 4m on my own.

I will try later but anybody must already have tested it...

[msft]

Hi ,aaronhaviland

Originally Posted by aaronhaviland

Err, actually not: the cufft size is initialiased as only half of n:

cufftPlan1d(&plan, n/2, CUFFT_Z2Z, 1)

GIMPS(Prime95,GIMPS Web,MLucas,GLucas,...) use real data size.

[aaronhaviland]

Originally Posted by msft

Hi ,aaronhaviland

GIMPS(Prime95,GIMPS Web,MLucas,GLucas,...) use real data size.

Right, but I thought the question was about the limitations of CUFFT's transform sizes.

[msft]

Originally Posted by aaronhaviland

Right, but I thought the question was about the limitations of CUFFT's transform sizes.

NVIDIA not announce CUFFT's work area size and limitations ,we can only do the experiments.

[Wizzard]

Sorry people, want try CUDALucas too, I downloaded CUDALucas.1.2b.zip, extracted, installed CUDA Toolkit 3.0 from nvidia website, wanted to compile CUDALucas using make and got an error: CUDALucas.cu:87:26: fatal error: cutil_inline.h: Directory or file does not exist. What is wrong? Do I need some CUDA patch? Which one? I am lost... I am using Ubuntu64 on Core2 Duo and GTS250.

[Dubslow]

Check your Toolkit version. I'm pretty sure the current release is either 4.1 or 4.2. There's also some instructions you have to follow during installation regarding various *PATH variables.

[Wizzard]

Okay, I replaced Toolkit with version 4.0, downloaded also GPU Computing SDK. Now I moved a little farther, and stopped at the error:

gcc: setup.cu: Directory or file does not exist
gcc: warning: '-x c++' after last input file has no effect
gcc: no input files
make: *** [setup.o] Error 1

[Dubslow]

Unless CUDA threw an error about the gcc version being >4.5, I can't help. Sorry.