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

[msft]

I understand,lycorn.

Thank you,

[fivemack]

Hi msft.

This is great work: many thanks! I had to copy the .h files from a separate install of MacLucasFFTW into the directory and modify some of the paths in the makefile to get it to work, but it works now.

It's a bit slower than I expected, 3m40s on a GTX275 to test 216091, but that's probably because 131072 is a very large FFT size to use in double precision for so small a number.

Unfortunately my computer crashed the second time I tried testing 216091; I think the graphics card is a bit flaky.

[frmky]

This is getting interesting! I decided to try the exponent 24036583. On a Tesla C1060 GPU, CUDA MacLucasFFTW runs at 0.0153 sec/iter using a 2048K FFT. Using one thread on a 2GHz Opteron K10 CPU, on the same exponent Prime95 runs at 0.055 sec/iter using a 1280K FFT. So, comparing the speed on a top-of-the-line GPU and a notoriously slow for Prime95 CPU, the GPU version runs about 3.5x faster. Also interesting is that after adding

cudaSetDeviceFlags(cudaDeviceBlockingSync);
cudaSetDevice(0);

near the top of the main() function, MacLucasFFTW uses only about 5% of a cpu core.

Assuming the computer doesn't get reset in the next 5 days or that restarting works, I'll let this run to completion.

[msft]

Thank you testing this program,fivemack,

Originally Posted by fivemack

I had to copy the .h files from a separate install of MacLucasFFTW into the directory and modify some of the paths in the makefile to get it to work, but it works now.

Sorry,it is my first Makefile.

Originally Posted by fivemack

It's a bit slower than I expected, 3m40s on a GTX275 to test 216091, but that's probably because 131072 is a very large FFT size to use in double precision for so small a number.

The side effects of parallelization , GPU need more thrhads, tuning target is 2048k or more higher.

Originally Posted by fivemack

Unfortunately my computer crashed the second time I tried testing 216091; I think the graphics card is a bit flaky.

Exactry, What are GTX275 made of ?

[msft]

Hi,frmky.

Originally Posted by frmky

I decided to try the exponent 24036583. On a Tesla C1060 GPU, CUDA MacLucasFFTW runs at 0.0153 sec/iter using a 2048K FFT.

I immediately check 24036583's 2000 iteration checksum, It is correct.

Originally Posted by frmky

cudaSetDeviceFlags(cudaDeviceBlockingSync);
cudaSetDevice(0);
near the top of the main() function, MacLucasFFTW uses only about 5% of a cpu core.

Cpu was 95% spin loop, I add this function, thank you.

Originally Posted by frmky

Assuming the computer doesn't get reset in the next 5 days or that restarting works, I'll let this run to completion.

Thank you for lots of work,

[frmky]

Originally Posted by msft

I immediately check 24036583's 2000 iteration checksum, It is correct.

And so were the next 24 million.

M( 24036583 )P, n = 2097152, MacLucasFFTW v8.1 Ballester

[msft]

Originally Posted by frmky

M( 24036583 )P, n = 2097152, MacLucasFFTW v8.1 Ballester

Prima!!!, Thank you

[lycorn]

Congrats, msft. The code seems to be running fine. 0.0153 sec/iter for a 1280K FFT is better than I can get on a Core2 duo T8300, with BOTH cores crunching the same exponent (best result is ~ 0.017).

[msft]

Thank you, lycorn

New version on GTX260.

$ tar -zxvf MacLucasFFTW.cuda.k.tar.gz
$ make
$ time ./MacLucasFFTW 216091

M( 216091 )P, n = 131072, MacLucasFFTW v8.1 Ballester

real 6m34.691s
user 0m10.025s
sys 0m0.188s

$ time ./MacLucasFFTW 2976221

M( 2976221 )P, n = 262144, MacLucasFFTW v8.1 Ballester

real 129m52.509s
user 19m27.337s
sys 0m1.232s

$ time ./MacLucasFFTW 33333333
10001 2097152

real 2m44.702s
user 0m22.469s
sys 0m1.136s

2048k fft sec/iter = 0.0165

$ time ./MacLucasFFTW 63333333
10001 4194304

real 7m0.095s
user 1m43.026s
sys 0m1.160s

4096k fft sec/iter = 0.042

M131101 to M1548619 1000 iterations check sum compare to Glucas,it is correct.

Thank you,

[nucleon]

Is there any advantage is doing multiple FFTs at the same time on the GPU? i.e. can we get 2x prime checks at the same time is say <50% of the time in doing one check?

-- Craig

[msft]

Hi, nucleon

Originally Posted by nucleon

Is there any advantage is doing multiple FFTs at the same time on the GPU? i.e. can we get 2x prime checks at the same time is say <50% of the time in doing one check?

Unfortunately PCI was too slow for LL-test.

Thank you,