CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW) - Page 37

msft · 2011-02-01, 17:44

Hi, Karl M Johnson

Quote:

Originally Posted by Karl M Johnson

By the way, is there a way to restart a certain exponent from a certain iteration ?
Or, at least, from a checkpoint ?

I got a certain exponent to 90k iterations, then got a reboot(I had -c25k flag), and when I started it again, it went from the very beginning.

Code:

msft@ubuntu:~$ ./CUDALucas -c1000 33333333 &
[1] 14609
msft@ubuntu:~$ Iteration 10000 M( 33333333 )C, 0xd717246f501c7d94, n = 2097152, CUDALucas v1.0 

msft@ubuntu:~$ kill 14609
msft@ubuntu:~$ 
[1]+  Done                    ./CUDALucas -c1000 33333333
msft@ubuntu:~$ ./CUDALucas -c1000 c33333333 
caso 2
Iteration 20000 M( 33333333 )C, 0x7f036ff2b230121b, n = 2097152, CUDALucas v1.0

"c33333333" Need "c".

Karl M Johnson · 2011-02-01, 17:51

Works!
Thanks!

kjaget · 2011-02-02, 02:50

Small update. Windows will not rename a file if it is open. This happens in CUDALucas if you run from a checkpoint file - the c24603451 file will remain open. All this means is that CUDALucas won't be able to back up the old checkpoint before updating it. It's not a significant problem - I only noticed it because I needed both the current and backup files to time execution speed.

In any case, I've fixed the problem and included it here. I've also included cudart64_31_9.dll in the archive so it should be everything you need to build and/or run in one shot.

cudalucas.1.0a.winx64.zip

Run times on my factory overclocked GTX 275, along with some rough run times for current work assignments. I know these aren't the most efficient use of the code but it's a good basis for comparison to a CPU.

8.96 msec/iter @ 2M FFT (~ 2.5 days for a 25M LL double check)
18.8 msec/iter @ 4M FFT (~ 11 days for a 47M LL first time run)

Not sure how that compares to Linux versions, but it's definitely fast enough to be useful.

Prime95 · 2011-02-04, 03:05

Quote:

Originally Posted by kjaget

I've also included cudart64_31_9.dll in the archive so it should be everything you need to build and/or run in one shot.

Does Nvidia's license let you include cufft64_31_9.dll? If so, please include that in your zip file so that it truly includes everything you need to run cudalucas.

msft · 2011-02-04, 12:48

Code:

#include <cuda.h>
#include <cuda_runtime.h>
#include <cufft.h>
#include <cutil_inline.h>
int main()
{
        cufftHandle     plan;
          cudaEvent_t start, stop;
        double          *x;
        double          *g_x;
        int i,j,imax;

        imax = 1024*1024*4;
        cutilSafeCall(cudaMalloc((void**)&g_x, sizeof(double)*imax));
        x = ((double *)malloc(sizeof(double)*imax));
        for(i=0;i<imax;i++)x[i]=0;
        cutilSafeCall(cudaMemcpy(g_x, x, sizeof(double)*imax, cudaMemcpyHostToDevice));
        cutilSafeCall( cudaEventCreate(&start) );
        cutilSafeCall( cudaEventCreate(&stop) );

        for(j=1024*1024;j<imax;j+=1024*1024)
        {
                cufftSafeCall(cufftPlan1d(&plan, j, CUFFT_Z2Z, 1));
                cufftSafeCall(cufftExecZ2Z(plan,(cufftDoubleComplex *)g_x,(cufftDoubleComplex *)g_x, CUFFT_INVERSE));
                cutilSafeCall( cudaEventRecord(start, 0) );
                for(i=0;i<10;i++)
                        cufftSafeCall(cufftExecZ2Z(plan,(cufftDoubleComplex *)g_x,(cufftDoubleComplex *)g_x, CUFFT_INVERSE));
                cutilSafeCall( cudaEventRecord(stop, 0) );
                cutilSafeCall( cudaEventSynchronize(stop) );
                float outerTime;
                cutilSafeCall( cudaEventElapsedTime(&outerTime, start, stop) );
                printf("CUFFT_Z2Z size=%d k time=%f msec\n",j/1024,outerTime/10);
                cufftSafeCall(cufftDestroy(plan));
        }

        for(j=1024*1024;j<imax;j+=256*1024)
        {
                cufftSafeCall(cufftPlan1d(&plan, j, CUFFT_D2Z, 1));
                cufftSafeCall(cufftExecD2Z(plan,g_x,(cufftDoubleComplex *)g_x));
                cutilSafeCall( cudaEventRecord(start, 0) );
                for(i=0;i<10;i++)
                        cufftSafeCall(cufftExecD2Z(plan,g_x,(cufftDoubleComplex *)g_x));
                cutilSafeCall( cudaEventRecord(stop, 0) );
                cutilSafeCall( cudaEventSynchronize(stop) );
                float outerTime;
                cutilSafeCall( cudaEventElapsedTime(&outerTime, start, stop) );
                printf("CUFFT_D2Z size=%d k time=%f msec\n",j/1024,outerTime/10);
                cufftSafeCall(cufftDestroy(plan));
        }
        cutilSafeCall(cudaFree((char *)g_x));
        cutilSafeCall( cudaEventDestroy(start) );
        cutilSafeCall( cudaEventDestroy(stop) );
}

CUFFT_Z2Z size=1024 k time=3.043661 msec
CUFFT_Z2Z size=2048 k time=6.288720 msec
CUFFT_Z2Z size=3072 k time=10.626810 msec
CUFFT_D2Z size=1024 k time=1.947040 msec
CUFFT_D2Z size=1280 k time=2.580678 msec
CUFFT_D2Z size=1536 k time=3.186858 msec
CUFFT_D2Z size=1792 k time=3.640893 msec
CUFFT_D2Z size=2048 k time=4.063977 msec
CUFFT_D2Z size=2304 k time=4.664579 msec
CUFFT_D2Z size=2560 k time=5.340890 msec
CUFFT_D2Z size=2816 k time=76.725174 msec
CUFFT_D2Z size=3072 k time=6.547805 msec
CUFFT_D2Z size=3328 k time=98.685196 msec
CUFFT_D2Z size=3584 k time=7.542326 msec
CUFFT_D2Z size=3840 k time=8.636828 msec

Non power of 2 is enhancement.
But not enough.

ET_ · 2011-02-04, 12:56

Quote:

Originally Posted by msft

Code:

#include <cuda.h>
#include <cuda_runtime.h>
#include <cufft.h>
#include <cutil_inline.h>
int main()
{
        cufftHandle     plan;
          cudaEvent_t start, stop;
        double          *x;
        double          *g_x;
        int i,j,imax;

        imax = 1024*1024*4;
        cutilSafeCall(cudaMalloc((void**)&g_x, sizeof(double)*imax));
        x = ((double *)malloc(sizeof(double)*imax));
        for(i=0;i<imax;i++)x[i]=0;
        cutilSafeCall(cudaMemcpy(g_x, x, sizeof(double)*imax, cudaMemcpyHostToDevice));
        cutilSafeCall( cudaEventCreate(&start) );
        cutilSafeCall( cudaEventCreate(&stop) );

        for(j=1024*1024;j<imax;j+=1024*1024)
        {
                cufftSafeCall(cufftPlan1d(&plan, j, CUFFT_Z2Z, 1));
                cufftSafeCall(cufftExecZ2Z(plan,(cufftDoubleComplex *)g_x,(cufftDoubleComplex *)g_x, CUFFT_INVERSE));
                cutilSafeCall( cudaEventRecord(start, 0) );
                for(i=0;i<10;i++)
                        cufftSafeCall(cufftExecZ2Z(plan,(cufftDoubleComplex *)g_x,(cufftDoubleComplex *)g_x, CUFFT_INVERSE));
                cutilSafeCall( cudaEventRecord(stop, 0) );
                cutilSafeCall( cudaEventSynchronize(stop) );
                float outerTime;
                cutilSafeCall( cudaEventElapsedTime(&outerTime, start, stop) );
                printf("CUFFT_Z2Z size=%d k time=%f msec\n",j/1024,outerTime/10);
                cufftSafeCall(cufftDestroy(plan));
        }

        for(j=1024*1024;j<imax;j+=256*1024)
        {
                cufftSafeCall(cufftPlan1d(&plan, j, CUFFT_D2Z, 1));
                cufftSafeCall(cufftExecD2Z(plan,g_x,(cufftDoubleComplex *)g_x));
                cutilSafeCall( cudaEventRecord(start, 0) );
                for(i=0;i<10;i++)
                        cufftSafeCall(cufftExecD2Z(plan,g_x,(cufftDoubleComplex *)g_x));
                cutilSafeCall( cudaEventRecord(stop, 0) );
                cutilSafeCall( cudaEventSynchronize(stop) );
                float outerTime;
                cutilSafeCall( cudaEventElapsedTime(&outerTime, start, stop) );
                printf("CUFFT_D2Z size=%d k time=%f msec\n",j/1024,outerTime/10);
                cufftSafeCall(cufftDestroy(plan));
        }
        cutilSafeCall(cudaFree((char *)g_x));
        cutilSafeCall( cudaEventDestroy(start) );
        cutilSafeCall( cudaEventDestroy(stop) );
}

CUFFT_Z2Z size=1024 k time=3.043661 msec
CUFFT_Z2Z size=2048 k time=6.288720 msec
CUFFT_Z2Z size=3072 k time=10.626810 msec
CUFFT_D2Z size=1024 k time=1.947040 msec
CUFFT_D2Z size=1280 k time=2.580678 msec
CUFFT_D2Z size=1536 k time=3.186858 msec
CUFFT_D2Z size=1792 k time=3.640893 msec
CUFFT_D2Z size=2048 k time=4.063977 msec
CUFFT_D2Z size=2304 k time=4.664579 msec
CUFFT_D2Z size=2560 k time=5.340890 msec
CUFFT_D2Z size=2816 k time=76.725174 msec
CUFFT_D2Z size=3072 k time=6.547805 msec
CUFFT_D2Z size=3328 k time=98.685196 msec
CUFFT_D2Z size=3584 k time=7.542326 msec
CUFFT_D2Z size=3840 k time=8.636828 msec

Non power of 2 is enhancement.
But not enough.

Was the GPU updating the screen? Those values look rather uncommon...

Luigi

msft · 2011-02-04, 13:02

Hi ,ET_

Quote:

Originally Posted by ET_

Was the GPU updating the screen? Those values look rather uncommon...

Weak length.

Andrew Thall · 2011-02-04, 16:27

There are wide variations in the time similar sized transforms based on their factorization: CUFFT (CUDA 3.2) on Fermi supports 2^a * 3^b * 5^c * 7^d transforms, with pure powers of 2 and 3 being pretty fast, but powers of 5 noticeably slower, and products of powers giving orders-of-magnitude differences in runtimes, some quite good, some horrible, depending on which bases and which powers.

I've got tabulations of runtimes based on a complete search of [a, b, c, d] values giving FFTs of length between 2^18 and 2^24. Use it (manually, at present) to pick LL runlengths, and will be correlating it with maximum wordsizes for LL giving acceptable errors. More to understand convolution errors w/ balanced integers than for the Mersenne stuff per se.

Just one more thing I should put online for anyone who's interested.

kjaget · 2011-02-05, 03:12

Another update for the windows version. Looks like the previous version I posted will end up in an infinite loop once a test finishes. The result will be printed to mersarch.txt, but not to the console.

There should be no problem using this version to complete a test started using one of the other windows versions.

I've included a fix in source and executable attached to this post. I didn't add cufft64_31_9.dll because that file is something like 5MB compressed - it's too large to fit as an attachment.

cudalucas.1.0b.winx64.zip

mdettweiler · 2011-02-07, 18:11

After a long hiatus while Gary's GPU was offline, it has now (finally!) finished the last n=35M LL-D assignment it started on back in October:

35000761

aaronhaviland · 2011-02-09, 23:22

There seems to be a couple upper limits to this right now. I tried running higher numbers, and get a couple different errors:

#CUDALucas 151150000
err = 0.353794, increasing n from 8388608
CUDALucas.cu(534) : cufftSafeCall() CUFFT error.

I'm guessing it's because of: "The cuFFT manual states that 1-D ffts are supported for < 8 million elements."

The other is at exponents around 318750000, I hit the memory limit on my 768MB card. At 336000000, it wants over 1Gb.

Combined, these prevent it from being useful for the 100 million digit numbers. (I can't be the only one eyeing this as making that task feasible.)

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Don't DC/LL them with CudaLucas	LaurV	Data	131	2017-05-02 18:41
CUDALucas / cuFFT Performance on CUDA 7 / 7.5 / 8	Brain	GPU Computing	13	2016-02-19 15:53
CUDALucas: which binary to use?	Karl M Johnson	GPU Computing	15	2015-10-13 04:44
settings for cudaLucas	fairsky	GPU Computing	11	2013-11-03 02:08
Trying to run CUDALucas on Windows 8 CP	Rodrigo	GPU Computing	12	2012-03-07 23:20

2011-02-01, 17:51	#398
Karl M Johnson Mar 2010 3·137 Posts	Works! Thanks!

2011-02-02, 02:50	#399
kjaget Jun 2005 3·43 Posts	Small update. Windows will not rename a file if it is open. This happens in CUDALucas if you run from a checkpoint file - the c24603451 file will remain open. All this means is that CUDALucas won't be able to back up the old checkpoint before updating it. It's not a significant problem - I only noticed it because I needed both the current and backup files to time execution speed. In any case, I've fixed the problem and included it here. I've also included cudart64_31_9.dll in the archive so it should be everything you need to build and/or run in one shot. cudalucas.1.0a.winx64.zip Run times on my factory overclocked GTX 275, along with some rough run times for current work assignments. I know these aren't the most efficient use of the code but it's a good basis for comparison to a CPU. 8.96 msec/iter @ 2M FFT (~ 2.5 days for a 25M LL double check) 18.8 msec/iter @ 4M FFT (~ 11 days for a 47M LL first time run) Not sure how that compares to Linux versions, but it's definitely fast enough to be useful.

2011-02-04, 16:27	#404
Andrew Thall Dec 2010 2³ Posts	There are wide variations in the time similar sized transforms based on their factorization: CUFFT (CUDA 3.2) on Fermi supports 2^a * 3^b * 5^c * 7^d transforms, with pure powers of 2 and 3 being pretty fast, but powers of 5 noticeably slower, and products of powers giving orders-of-magnitude differences in runtimes, some quite good, some horrible, depending on which bases and which powers. I've got tabulations of runtimes based on a complete search of [a, b, c, d] values giving FFTs of length between 2^18 and 2^24. Use it (manually, at present) to pick LL runlengths, and will be correlating it with maximum wordsizes for LL giving acceptable errors. More to understand convolution errors w/ balanced integers than for the Mersenne stuff per se. Just one more thing I should put online for anyone who's interested.

2011-02-05, 03:12	#405
kjaget Jun 2005 3·43 Posts	Another update for the windows version. Looks like the previous version I posted will end up in an infinite loop once a test finishes. The result will be printed to mersarch.txt, but not to the console. There should be no problem using this version to complete a test started using one of the other windows versions. I've included a fix in source and executable attached to this post. I didn't add cufft64_31_9.dll because that file is something like 5MB compressed - it's too large to fit as an attachment. cudalucas.1.0b.winx64.zip

2011-02-07, 18:11	#406
mdettweiler A Sunny Moo Aug 2007 USA (GMT-5) 3×2,083 Posts	After a long hiatus while Gary's GPU was offline, it has now (finally!) finished the last n=35M LL-D assignment it started on back in October: 35000761

2011-02-09, 23:22	#407
aaronhaviland Jan 2011 Dudley, MA, USA 73 Posts	There seems to be a couple upper limits to this right now. I tried running higher numbers, and get a couple different errors: #CUDALucas 151150000 err = 0.353794, increasing n from 8388608 CUDALucas.cu(534) : cufftSafeCall() CUFFT error. I'm guessing it's because of: "The cuFFT manual states that 1-D ffts are supported for < 8 million elements." The other is at exponents around 318750000, I hit the memory limit on my 768MB card. At 336000000, it wants over 1Gb. Combined, these prevent it from being useful for the 100 million digit numbers. (I can't be the only one eyeing this as making that task feasible.)