LL with OpenCL

[msft]

Originally Posted by pinhodecarlos

Can I help testing the windows version?

It would be helpful if you do so.

[pinhodecarlos]

My Windows 7 laptop graphic card is an ATI RADEON HD7670M but I don't think it has DP capability.

[kracker]

finally, getting somewhere. My problem was trying to compile 64 bit code(the samples) on a 32 bit compiler... Stupid me.

[msft]

Originally Posted by kracker

http://devgurus.amd.com/thread/166951

Well, cygwin it is.

Yes, This code abducted from MatrixMulDouble.

[msft]

Originally Posted by pinhodecarlos

My Windows 7 laptop graphic card is an ATI RADEON HD7670M but I don't think it has DP capability.

I agree.

[msft]

Originally Posted by msft

Sometime clAmdFft-1.10.321 have precision problem.

posted.

[msft]

Originally Posted by msft

posted.

Additional information.
ooura fft.

#include <math.h>
#include <stdio.h>
void fft(int n, double theta, double ar[], double ai[])
{
    int m, mh, i, j, k;
    double wr, wi, xr, xi;

    for (m = n; (mh = m >> 1) >= 1; m = mh) {
        for (i = 0; i < mh; i++) {
            wr = cos(theta * i);
            wi = sin(theta * i);
            for (j = i; j < n; j += m) {
                k = j + mh;
                xr = ar[j] - ar[k];
                xi = ai[j] - ai[k];
                ar[j] += ar[k];
                ai[j] += ai[k];
                ar[k] = wr * xr - wi * xi;
                ai[k] = wr * xi + wi * xr;
            }
        }
        theta *= 2;
    }
    /* ---- unscrambler ---- */
    i = 0;
    for (j = 1; j < n - 1; j++) {
        for (k = n >> 1; k > (i ^= k); k >>= 1);
        if (j < i) {
            xr = ar[j];
            xi = ai[j];
            ar[j] = ar[i];
            ai[j] = ai[i];
            ar[i] = xr;
            ai[i] = xi;
        }
    }
}
int main()
{
        double ar[1024],ai[1024];
        double pi,theta;
        int i,j,k,n;
        n=32;
        pi = 3.141592653589793;
        theta=-2.0*pi/n;
        for(i=0;i<n;i++)
        {
                ar[i]=i;
                ai[i]=i;
        }
        fft(n,theta,ar,ai);
        printf("ooura %.15e %.15e \n",ar[1],ai[1]);
}

result:

ooura -1.784507262017418e+02 1.464507262017418e+02

[msft]

New Version.
Speed UP.
Fix Compile warning error.
HD7750:

Iteration 1380000 M( 1398269 )C, 0x6650b95f00196d18, n = 73728, CUDALucas v1.66 err = 0.1108 (0:18 real, 1.7757 ms/iter, ETA 0:17)
Iteration 1390000 M( 1398269 )C, 0x554ae339bfea8fae, n = 73728, CUDALucas v1.66 err = 0.1108 (0:17 real, 1.7712 ms/iter, ETA 0:00)
M( 1398269 )P, n = 73728, CUDALucas v1.66
476.88user 2099.99system 41:21.70elapsed 103%CPU (0avgtext+0avgdata 220496maxresident)k
0inputs+163464outputs (0major+2656901minor)pagefaults 0swaps

[Bdot]

Hi msft,

I just wanted to check if you are aware that the AMD support guy is waiting for some data from you?

What value N does it fail for? Is N=40 or N=32 or does it fail irrespective of N? What about N=16?What is your desired tolerance in terms of RMSE? Or can you report the maximum relative error between corresponding numerical values that you are observing?

With the test program you delivered, he could have tested this in his own, but this is how support works ...

When I see your results here, it really seems as if only specific FFTs are failing, and sometimes you get valid results?

(I'm just very interested in this project - please don't feel pushed by me.)

[kracker]

In PM, he said he is on a trip and will be back this weekend.

@Bdot: So I finally got Ubuntu x64 working.. do you think it is easier to compile for windows on windows or on linux(gcc-mingw)?

[msft]

Hi Bdot,
Thank you for your support.