LL with OpenCL

[kladner]

Originally Posted by kracker

Ok, well frankly, 3-4 DC's(all cpu) a day is my max at the moment. So not much.

That sounds like a good bit to me. My GPUs take between 20 and 22 hours per DC (~28M).

[TeknoHog]

FWIW:

Adapter 0 - AMD Radeon HD 7900 Series
            New Core Peak   : 1000
            New Memory Peak : 1500
Platform :Advanced Micro Devices, Inc.
Device 0 : Tahiti


start M57885161 fft length = 3145728
err = 0.352051, increasing n from 3145728

start M57885161 fft length = 3276800
Iteration 10000 M( 57885161 )C, 0x76c27556683cd84d, n = 3276800, clLucas v1.00 err = 0.1289 (4:01 real, 24.1701 ms/iter, ETA 388:32:03)
Iteration 20000 M( 57885161 )C, 0xfd8e311d20ffe6ab, n = 3276800, clLucas v1.00 err = 0.1289 (4:00 real, 23.9985 ms/iter, ETA 385:42:34)

I'd expect more with this 7970, but the 1x PCIe slot may be a bottleneck. GPU load is 99%, using the -aggressive option.

I also tested a couple of 5870s, but they keep throwing this error

Error: CommandQueue::enqueueNDRangeKernel() failed. Error code : CL_INVALID_WORK_GROUP_SIZE
Location : Kernels.cpp:425

and tests with known primes fail.

However, I'm quite happy with the overall development so far, nice work msft

[kracker]

See this:

2097152 and 4194304 fft work best.(-f)

Also, you might want to stick with DC's(for now) until we know clLucas can do those ranges.

[TeknoHog]

Originally Posted by kracker

See this:

2097152 and 4194304 fft work best.(-f)

Also, you might want to stick with DC's(for now) until we know clLucas can do those ranges.

Thanks :) Also, that 57885161 was just a speed test (M#48), I'm already working on a DC.

[kracker]

HD 7770

Iteration 10000 M( 22256453 )C, 0x3d9450d492b7e880, n = 1179648, clLucas v1.00 err = 0.2656 (2:39 real, 15.9270 ms/iter, ETA 98:23:35)
Iteration 10000 M( 24732709 )C, 0x81a12a304a754572, n = 1310720, clLucas v1.00 err = 0.2813 (3:04 real, 18.4028 ms/iter, ETA 126:21:56)
Iteration 10000 M( 29412433 )C, 0x27d7d112a73aa203, n = 1572864, clLucas v1.00 err = 0.25 (4:02 real, 24.1517 ms/iter, ETA 197:14:18)
Iteration 10000 M( 30620113 )C, 0x212dca3cec0acde2, n = 1638400, clLucas v1.00 err = 0.25 (5:11 real, 31.0753 ms/iter, ETA 264:13:34)
Iteration 10000 M( 32993419 )C, 0xcf86a69b844e35c0, n = 1769472, clLucas v1.00 err = 0.2813 (7:18 real, 43.7117 ms/iter, ETA 400:26:52)
Iteration 10000 M( 36418493 )C, 0x2f1388379572d5b4, n = 1966080, clLucas v1.00 err = 0.25 (5:30 real, 33.0295 ms/iter, ETA 333:57:52)
Iteration 10000 M( 38955173 )C, 0x8a45e3bbd4e4fc9b, n = 2097152, clLucas v1.00 err = 0.25 (2:03 real, 12.2586 ms/iter, ETA 132:35:49)
Iteration 10000 M( 43792559 )C, 0x7048d84bbfb0f810, n = 2359296, clLucas v1.00 err = 0.2813 (5:54 real, 35.3544 ms/iter, ETA 429:56:54)
Iteration 10000 M( 48375209 )C, 0xf957e240d591a99e, n = 2621440, clLucas v1.00 err = 0.2188 (6:33 real, 39.2538 ms/iter, ETA 527:18:36)
Iteration 10000 M( 57899201 )C, 0xa2ac01bbc76d92ee, n = 3145728, clLucas v1.00 err = 0.25 (9:00 real, 53.9709 ms/iter, ETA 867:43:57)
Iteration 10000 M( 60622229 )C, 0xd81c849f11fd1054, n = 3276800, clLucas v1.00 err = 0.2813 (11:06 real, 66.5953 ms/iter, ETA 1121:12:22)
Iteration 10000 M( 65066623 )C, 0xde7aeb8cc7a2a826, n = 3538944, clLucas v1.00 err = 0.2539 (16:00 real, 96.0663 ms/iter, ETA 1735:51:51)
Iteration 10000 M( 67662869 )C, 0xf854d1dee3fbb5d7, n = 3932160, clLucas v1.00 err = 0.05078 (11:59 real, 71.8933 ms/iter, ETA 1350:59:44)
Iteration 10000 M( 76722161 )C, 0x4b6ba0a6078e4bbb, n = 4194304, clLucas v1.00 err = 0.25 (4:14 real, 24.9663 ms/iter, ETA 540:43:00)

[owftheevil]

The errors are nicely low. Is this using the double double sin and cos data for the ffts?

[kracker]

Originally Posted by owftheevil

The errors are nicely low. Is this using the double double sin and cos data for the ffts?

You're going to have to ask msft about that, I only compiled it... that's all.

EDIT: Just realized I won't be able to submit my DC(yet) when it is doen to Primenet.

[msft]

Originally Posted by LaurV

Ok, played with it, but the result is very odd!

Hi,
Can you check with -threads option?

$ ./clLucas -threads 64 36666666
$ ./clLucas -threads 128 36666666

[msft]

Originally Posted by owftheevil

The errors are nicely low. Is this using the double double sin and cos data for the ffts?

Hi,
Double sin cos,I make fftw like function.

/*
 * Improve accuracy by reducing x to range [0..1/8]
 * before multiplication by 2 * PI.
 */
#define K2PI 6.2831853071795864769252867665590057683943388
#define m2pi(m, n) ((K2PI * (m)) / (n))
static void ap_sincos(int m, int n, double * si, double * co)
{
        double s,c,theta;
        int n14,n24,n34,n44;
        int m14,m44;
        int n18,n28,n38,n48,n58,n68,n78;
        int m88;

        n14 = n;
        n24 = n14 + n14;
        n34 = n14 + n24;
        n44 = n24 + n24;

        m14 = m;
        m44 = m14 + m14;
        m44 = m44 + m44;

        n18 = n;
        n28 = n18 + n18;
        n38 = n18 + n28;
        n48 = n28 + n28;
        n58 = n28 + n38;
        n68 = n38 + n38;
        n78 = n38 + n48;

        m88 = m + m;
        m88 = m88 + m88;
        m88 = m88 + m88;

        if(n18 > m88)
        {
                theta = m2pi(m44,n44);
                s = sin(theta);
                c = cos(theta);
        }
        else if(n28 > m88)
        {
                theta = m2pi(n14-m44,n44);
                s = cos(theta);
                c = sin(theta);
        } else if(n38 > m88)
        {
                theta = m2pi(m44-n14,n44);
                s = cos(theta);
                c = -sin(theta);
        } else if(n48 > m88)
        {
                theta = m2pi(n24-m44,n44);
                s = sin(theta);
                c = -cos(theta);
        } else if(n58 > m88)
        {
                theta = m2pi(m44-n24,n44);
                s = -sin(theta);
                c = -cos(theta);
        } else if(n68 > m88)
        {
                theta = m2pi(n34-m44,n44);
                s = -cos(theta);
                c = -sin(theta);
        } else if(n78 > m88)
        {
                theta = m2pi(m44-n34,n44);
                s = -cos(theta);
                c = sin(theta);
        } else
        {
                theta = m2pi(n44-m44,n44);
                s = -sin(theta);
                c = cos(theta);
        }
        *si = s;
        *co = c;
}

[msft]

Originally Posted by TeknoHog

I also tested a couple of 5870s, but they keep throwing this error

Error: CommandQueue::enqueueNDRangeKernel() failed. Error code : CL_INVALID_WORK_GROUP_SIZE
Location : Kernels.cpp:425

and tests with known primes fail.

Hi,
I found bug with this issue,Fix next version.

[msft]

Originally Posted by LaurV

Trying to squeeze some more juice from the Tahiti, I believe this is what msft is after, isn't he?

Hi,
Error occurs just 30000.