The source code of a variant of Genefer for OpenCL is available on assembla:
[...]\branches\yves\2013\OclGenefer and [...]\branches\yves\2013\Common

I tested it with NVidia SDK on Fermi and Kepler GPUs and with Intel SDK for OpenCL, running on the CPU because Intel HD Graphics have no FP64. I have no ATI card, but the code doesn't depend on any external library then it should run on ATI cards.

It is about as fast as GeneferCUDA on my computers but the algorithm is different then speed ratio depends on hardware and exponent.

It is not finished: error is not computed and many improvements are still possible.

The source code is "standard" C++, then you can compile it with Visual Studio 2012 (Express) or gcc with NVidia or ATI or Intel SDK for OpenCL.

Yves

Here's some benches on NVidia: this version doesn't like Fermi architecture :-(

You can compile it with cl.h and OpenCL.lib from any OpenCL SDK (Intel, Nvidia or ATI).
The executable will mount the OpenCL.dll installed by graphics driver(s).
On my laptop, with a HD 4000 and a GeForce GT 740M, any OpenCL binary can run on Intel GPU and on NVidia GPU. I don't have to generate one with Intel SDK and another one with NVidia SDK.

App crashed with "cannot create program".

Added current status to my dropbox. If someone is interested tell me.

What are command lines for the app?

What are command lines for the app?

Test results on Asus HD7950: [...]

I had relatively little difficulty building Yves' application with VS2012 Express as an x64 app.

What are command lines for the app?

Hello Yves, it has been a while. You might recall me from helping get the original genefer running on PowerMac. I wrote the checkpointing code that the various flavors of genefer use today.

I'm excited to see you back in action on this. I've been waiting for someone to port this to OpenCL so that I can run it on my Mac Pro as it has an ATI card.

Now if you could only port a command line version of Proth to OpenCL... :-)

To get this to build on Mac, I had to make this change:

#ifdef __APPLE__
#include <OpenCL/cl.h>
#else
#include <CL/cl.h>
#endif

in OclGenefer.cpp.

Also, to use clang (Apple's newer incarnation of gcc) I had to make this change:

virtual ~ITransform() {};

although llvm (Apple's older incarnation of gcc) allowed the original code:

virtual ~ITransform() = 0 {};

and the updated version.

Unfortunately, I am running into this problem with clang:

OclGenefer.cpp:176:29: error: no matching conversion for functional-style cast from 'std::ifstream' (aka 'basic_ifstream<char>') to 'std::istreambuf_iterator<char>'
const std::string source((std::istreambuf_iterator<char>(std::ifstream("Genefer.cl", std::ifstream::in))), std::istreambuf_iterator<char>());
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/include/c++/4.2.1/bits/streambuf_iterator.h:48:11: note: candidate constructor (the implicit copy constructor) not viable: no known conversion from 'std::ifstream' (aka 'basic_ifstream<char>') to 'const std::istreambuf_iterator<char, std::char_traits<char> >' for 1st argument
class istreambuf_iterator
^
/usr/include/c++/4.2.1/bits/streambuf_iterator.h:98:7: note: candidate constructor not viable: no known conversion from 'std::ifstream' (aka 'basic_ifstream<char>') to 'istream_type &' (aka 'basic_istream<char, std::char_traits<char> > &') for 1st argument
istreambuf_iterator(istream_type& __s) throw()
^
/usr/include/c++/4.2.1/bits/streambuf_iterator.h:102:7: note: candidate constructor not viable: no known conversion from 'std::ifstream' (aka 'basic_ifstream<char>') to 'streambuf_type *' (aka 'basic_streambuf<char, std::char_traits<char> > *') for 1st argument
istreambuf_iterator(streambuf_type* __s) throw()
^
/usr/include/c++/4.2.1/bits/streambuf_iterator.h:94:7: note: candidate constructor not viable: requires 0 arguments, but 1 was provided
istreambuf_iterator() throw()


llvm doesn't give as useful a message:

OclGenefer.cpp: In constructor ‘Program::Program(bool)’:
OclGenefer.cpp:176: error: invalid conversion from ‘void*’ to ‘std::basic_streambuf<char, std::char_traits<char> >*’
OclGenefer.cpp:176: error: initializing argument 1 of ‘std::istreambuf_iterator<_CharT, _Traits>::istreambuf_iterator(std::basic_streambuf<_CharT, _Traits>*) [with _CharT = char, _Traits = std::char_traits<char>]’

I'm not good enough with C++ streams to fix this.

Test result with HD7970Ghz using Rebirther's Dropbox genefer.exe:



Was using 18% of CPU (X6 1100T) and up to 99% of GPU. The high use of GPU is too aggressive and program does not finish long runs.
Not shown in screen shot: 572186^131072+1 took 3424.5 seconds and 2418^262144+1 took 8000+ seconds.

Hi Mark,

Also, to use clang (Apple's newer incarnation of gcc) I had to make this change:
virtual ~ITransform() {};
although llvm (Apple's older incarnation of gcc) allowed the original code:
virtual ~ITransform() = 0 {};
and the updated version.

Test result with HD7970Ghz using Rebirther's Dropbox genefer.exe:

Was using 18% of CPU (X6 1100T) and up to 99% of GPU.

The latest app with result, Yves you need a version number...

Test result with HD7970Ghz using Rebirther's Dropbox genefer.exe [...]

Ugh! The Radeon HD 5870 only supports cl_amd_fp64, not cl_khr_fp64. I can't run the code on my MacPro. When it tries to use the CPU, I get this error:

inline Complex _MulWeightOut(const Complex z, const Complex aw, const double nInv, const double g)
^

Running on platform 'Apple' and device 'Intel(R) Xeon(R) CPU W3530 @ 2.80GHz'.


Error detected on GPU device.

I don't expect it to be fast on the CPU though, but I wouldn't expect this error either.

Running Catalyst 12.8. That might explain the extra pragma warning and the test result.

HD7970Ghz test result with new version:



Same %18 CPU and 99% GPU usage. Catalyst 12.8. I'll update to 13.4 and run again. Not shown is 572186^131072+1 that took 5985.4 sec.

Ugh! The Radeon HD 5870 only supports cl_amd_fp64, not cl_khr_fp64. I can't run the code on my MacPro.

Tested HD7970Ghz with Catalyst 13.4. Behavior is different. Now uses around 6% of CPU but still 99% of GPU.
No #pragma warning, faster times, but maybe a memory leak when testing 676754^262144+1?
In this screen shot you can see memory use maxed out at 6GB, is not using any GPU, so has gone zombie:



I tested -q "676754^262144+1", -q "75898^524288+1" and -q "475856^524288+1" and all had memory usage that quickly went crazy after only seconds, as shown by Task Manager.
Note this test was straight after updating to Catalyst 13.4.
I reran the test after a reboot and got some slightly better times for high N and slightly worse times for low N.

Ugh! The Radeon HD 5870 only supports cl_amd_fp64, not cl_khr_fp64. I can't run the code on my MacPro.

It should run with cl_amd_fp64 (I think: I can't find cl_amd_fp64 instruction set).
But 5 instructions only are called : +, - , *, fma, rint.

There was an error in the logic of the defines : cl_amd_fp64 was not set with OpenCL 1.2. I solved this. You can download the latest release.
I removed "_MulWeightOut", it was an unreachable code.

Tested HD7970Ghz with Catalyst 13.4. Behavior is different. Now uses around 6% of CPU but still 99% of GPU.
No #pragma warning, faster times, but maybe a memory leak when testing 676754^262144+1?
In this screen shot you can see memory use maxed out at 6GB, is not using any GPU, so has gone zombie:
I tested -q "676754^262144+1", -q "75898^524288+1" and -q "475856^524288+1" and all had memory usage that quickly went crazy after only seconds, as shown by Task Manager.

Tested HD7970Ghz [...] had memory usage that quickly went crazy after only seconds.

Tested and confirmed. The program took all my 16GB memory and cpu was at full usage (memory leak).

Bug is fixed, 676754^262144+1 test uses 39.0MB memory, only 2% of CPU and 97% of GPU. This is a much better combination, 99% GPU makes the computer less responsive.
^524288 tests use 1.0% of CPU, 43.9MB and 98% GPU. Usage depends on exponent being tested.

Full test run on HD7970Ghz with Catalyst 13.4:



Is all sweetness and light. What's the next step for this algorithm?
Great work Yves!

Full test run on HD7970Ghz with Catalyst 13.4.
[...] What's the next step for this algorithm?

My benchmark was incorrect.
That's good news, because OclGenefer runs faster than it was displayed!
During the benchmark (which is similar to Genefer code), 24 loops are computed (to initialize memory and caches) and then the true benchmark starts. But I forget to add a synchronization point at the end of the 24 loops, then the computation of these transforms biased estimates (especially for large n).
The error is corrected in OclGenefer 2013-08-10.
I'm on holiday and tested the program on a GeForce GT 740M (GK107, 2 SMX, 384 shaders). OclGenefer is faster than GeneferCUDA on it.
Then today, GeneferCUDA is (still :-) ) faster on Fermi, but OclGenefer is faster on Kepler with few SMX (it should be faster on GK107, i.e. GT 640 and GTX 650). I cannot extrapolate to GK104 (GTX 680) or GK110 (GTX 780 & Titan)... a benchmark on the Titan would be very interesting.

New assembla OclGenefer.cpp revision 381 code run on HD7970Ghz:



Benchmarks for large N are definitely faster. New device initialisation information.
I guess the full probable prime tests are unaffected by the latest code change?

GFN-OCL 2^20 1.84 ms/mul
2^20 Goal=1.5 ms/mul => another 18% drop required
GFN-OCL 2^22 7.67 ms/mul
2^22 Goal=5.6 ms/mul => another 27% drop required

New device initialisation information.

I guess the full probable prime tests are unaffected by the latest code change?

GFN-OCL 2^20 1.84 ms/mul
2^20 Goal=1.5 ms/mul => another 18% drop required
GFN-OCL 2^22 7.67 ms/mul
2^22 Goal=5.6 ms/mul => another 27% drop required

A new version is available on assembla (2013-08-12).
It is really faster than the previous one on my laptop, I hope that it is fast on high and mid range graphics cards.
Max error is computed.

Any benchmark is welcome.

If tests are ok, the next version will be a beta release of geneferOpenCL.

Latest assembla Genefer.cl and OclGenefer.cpp revision 382 code run on HD7970Ghz:



Of course I will run full test after the challenge.

GFN-OCL 2^20 1.46 ms/mul
2^20 Goal=1.5 ms/mul => Goal achieved!
GFN-OCL 2^22 5.61 ms/mul
2^22 Goal=5.6 ms/mul => Goal achieved!
Congrats!

And then it will run faster on a HD 7970 GHz than GeneferCUDA on a GeForce GTX Titan :o)

The error was not computed in the previous release.
HD 7950 is about as fast as the HD 7970 GHz... is it an overclocked HD7950 ?

The HD 7950 is faster than a GeForce GTX Titan with full DP enabled!
The relative GPU value (genefer mark / price) is 4 / 1.

Great, I'm working on the OpenCL version of the real genefer app.

Yves, should I be able to use my 5870 or not? If not, I'm okay with that (albeit disappointed).

Yves, should I be able to use my 5870 or not? If not, I'm okay with that (albeit disappointed).

On my stock GTX 460:

GeneferCUDA:

I think that timings are incorrect. If they are correct, then something appears to be impacting the overall throughput. I commented out the check for CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE.

Running on platform 'Apple', device 'ATI Radeon HD 5870', version 'OpenCL 1.1 ' and driver '1.0'.
Clock frequency = 850 MHz, compute units = 20.
Global mem size = 1024 MB, cache size = 0 kB (None), cache line size = 0 Bytes.
Local mem size = 32 kB (dedicated), Constant mem size = 64 kB.
Max workgroup size = 1024.

2199064^8192+1 Time: 169 us/mul. Err: 0.2344 51956 digits
1798620^16384+1 Time: 173 us/mul. Err: 0.2344 102481 digits
1471094^32768+1 Time: 171 us/mul. Err: 0.2734 202102 digits
1203210^65536+1 Time: 177 us/mul. Err: 0.2344 398482 digits
984108^131072+1 Time: 178 us/mul. Err: 0.2584 785521 digits
804904^262144+1 Time: 249 us/mul. Err: 0.2486 1548156 digits
658332^524288+1 Time: 252 us/mul. Err: 0.2378 3050541 digits
538452^1048576+1 Time: 266 us/mul. Err: 0.2188 6009544 digits
440400^2097152+1 Time: 304 us/mul. Err: 0.2275 11836006 digits
360204^4194304+1 Time: 352 us/mul. Err: 0.2168 23305854 digits
294612^8388608+1 Time: 337 us/mul. Err: 0.2344 45879398 digits

102^64+1 is a probable prime. (0.1 sec., err = 1.82e-11)
15000250^64+1 is a probable prime. (0.3 sec., err = 0.406)
120^128+1 is a probable prime. (0.2 sec., err = 5.09e-11)
10000038^128+1 is a probable prime. (0.5 sec., err = 0.406)
278^256+1 is a probable prime. (0.3 sec., err = 3.53e-10)
5684328^256+1 is a probable prime. (0.9 sec., err = 0.172)
46^512+1 is a probable prime. (0.5 sec., err = 1.64e-11)
4619000^512+1 is a probable prime. (1.8 sec., err = 0.195)
824^1024+1 is a probable prime. (1.6 sec., err = 9.08e-09)
3752220^1024+1 is a probable prime. (3.7 sec., err = 0.195)
150^2048+1 is a probable prime. (2.4 sec., err = 5.24e-10)
3066672^2048+1 is a probable prime. (7.1 sec., err = 0.219)
1534^4096+1 is a probable prime. (7.3 sec., err = 8.2e-08)
2485064^4096+1 is a probable prime. (14.5 sec., err = 0.215)
30406^8192+1 is a probable prime. (20.7 sec., err = 5.25e-05)
2030234^8192+1 is a probable prime. (28.9 sec., err = 0.234)
67234^16384+1 is a probable prime. (44.9 sec., err = 0.00037)
1651902^16384+1 is a probable prime. (57.9 sec., err = 0.22)
70906^32768+1 is a probable prime. (92.1 sec., err = 0.000671)
1277444^32768+1 is a probable prime. (113.7 sec., err = 0.227)
48594^65536+1 is a probable prime. (174.8 sec., err = 0.000458)
857678^65536+1 is a probable prime. (222.8 sec., err = 0.148)
62722^131072+1 is a probable prime. (365.5 sec., err = 0.00122)
572186^131072+1 is a probable prime. (429.6 sec., err = 0.105)
24518^262144+1 is a probable prime. (904.9 sec., err = 0.000275)

I think that timings are incorrect. If they are correct, then something appears to be impacting the overall throughput. I commented out the check for CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE. [...]

clock() is Elapsed Time on Windows.
clock() is Cpu Time on Unix.

A new version is available on assembla (2013-08-14).

It supports "cl_khr_fp64" and "cl_amd_fp64".

A bug was corrected, a possible final round-off error. It occurred with composite numbers but not with primes!

This may be the last release of this program: the next one is a beta of "GeneferOCL" (the real genefer app).

I expect I'll be releasing GeneferOCL as a beta/app_info app later today. I'm playing with it right now.

It appears to be happy running on ANYTHING. On my computer, specifying "--device 0" runs on my GTX 460 and "--device 1" runs on all 4 cores of my Core2Quad.

Although I don't have this hardware on this computer, it should also run on an ATI or Intel GPU just as easily.

So caution is warranted when running this on computers with multiple OpenCL capable devices.

This raises the possibility of officially supporting not just AMD/AT GPUs but also Intel GPUs, and possibly CPU-based multicore apps. (Well, that last part might be wishful thinking. It looks like GeneferOCL running on all 4 cores of my computer is about 10 times SLOWER than GenefX64 running on a single core.)

OpenCL on CPU:

[...], it should also run on an ATI or Intel GPU just as easily.

Intel HD Graphics doesn't support double FP.

Intel CPU does but that's incredibly slow!
On my laptop:

geneferocl -q 212346^32768+1 -d 1 Running on platform 'Intel(R) OpenCL', device 'Intel(R) Core(TM) i3-3217U CPU @ 1.80GHz' Estimated total run time for 212346^32768+1 is 4:44:56

geneferavx -q "212346^32768+1" Estimated total run time for 212346^32768+1 is 0:04:04

And according to Intel's OpenCL compiler: "all kernel functions were successfully vectorized".
Does Intel's SDK emulate double FP ?

Hi,

testing with my 7950 the estimated runtime (after 3% work done) is 27050 seconds. GPU load is 99%, CPU is about 1%. I am using this app_info.xml:

<app_info>
<app>
<name>genefer</name>
<user_friendly_name>Genefer OCL</user_friendly_name>
</app>
<file_info>
<name>geneferocl-windows.exe</name>
<executable/>
</file_info>
<app_version>
<app_name>genefer</app_name>
<version_num>206</version_num>
<api_version>7.0.64</api_version>
<avg_ncpus>1.000000</avg_ncpus>
<max_ncpus>1.000000</max_ncpus>
<file_ref>
<file_name>geneferocl-windows.exe</file_name>
<main_program/>
</file_ref>
<platform>windows_intelx86</platform>
<coproc>
<type>ATI</type>
<count>1.000000</count>
</coproc>
</app_version>
</app_info>

____________
DeleteNull

Latest assembla OclGenefer.cpp revision 386 code with full run on HD7970Ghz:



Are we expecting the same "B" limits as GeneferCUDA?

Are we expecting the same "B" limits as GeneferCUDA?

Generalized Fermat Number b Limits
The upper bound m = 8192, b = 2650000, Err = 0.2813
The upper bound m = 16384, b = 2280000, Err = 0.2969
The upper bound m = 32768, b = 1840000, Err = 0.2969
The upper bound m = 65536, b = 1525000, Err = 0.2969
The upper bound m = 131072, b = 1270000, Err = 0.2969
The upper bound m = 262144, b = 995000, Err = 0.2813
The upper bound m = 524288, b = 815000, Err = 0.2813
The upper bound m = 1048576, b = 695000, Err = 0.3047
The upper bound m = 2097152, b = 580000, Err = 0.2969
The upper bound m = 4194304, b = 475000, Err = 0.3125
The upper bound m = 8388608, b = 400000, Err = 0.3125

Generalized Fermat Number b Limits
The upper bound m = 8192, b = 2670000, Err = 0.2910
The upper bound m = 16384, b = 2210000, Err = 0.2969
The upper bound m = 32768, b = 1780000, Err = 0.2969
The upper bound m = 65536, b = 1505000, Err = 0.2969
The upper bound m = 131072, b = 1240000, Err = 0.2969
The upper bound m = 262144, b = 1015000, Err = 0.3047
The upper bound m = 524288, b = 825000, Err = 0.3057
The upper bound m = 1048576, b = 680000, Err = 0.3047
The upper bound m = 2097152, b = 555000, Err = 0.2969
The upper bound m = 4194304, b = 455000, Err = 0.2813
The upper bound m = 8388608, b = 385000, Err = 0.3125

From the database:

Running on platform 'AMD Accelerated Parallel Processing', device 'Tahiti', version 'OpenCL 1.2 AMD-APP (1124.2)' and driver '1124.2 (VM)'.


Starting initialization...
Initialization complete (2.464 seconds).
Testing 216560^1048576+1...
Estimated total run time for 216560^1048576+1 is 7:39:03
216560^1048576+1 is complete. (5594760 digits) (err = 0.0469) (time = 9:48:38) 06:53:04


Running on platform 'AMD Accelerated Parallel Processing', device 'Tahiti', version 'OpenCL 1.2 AMD-APP (1124.2)' and driver '1124.2 (VM)'.


Starting initialization...
Initialization complete (3.205 seconds).
Testing 216878^1048576+1...
Estimated total run time for 216878^1048576+1 is 14:56:13
216878^1048576+1 is complete. (5595428 digits) (err = 0.0488) (time = 7:51:30) 14:44:38

Running on platform 'AMD Accelerated Parallel Processing', device 'Tahiti', version 'OpenCL 1.2 AMD-APP (1124.2)' and driver '1124.2 (VM)'.


Starting initialization...
Initialization complete (3.078 seconds).
Testing 216862^1048576+1...
Estimated total run time for 216862^1048576+1 is 7:52:07
216862^1048576+1 is complete. (5595394 digits) (err = 0.0469) (time = 8:20:41) 16:17:09

Running the full blown geneferocl app using the -l option, [...]As you can see, the limits are similar but not identical.

A few interesting things. The first and most obvious is that it works.

The second is that it doesn't tell you the exact device model. That's a shame.

The third is that the estimates vary by more than I'm accustomed to seeing on my system. These three results are from the same computer, but it's not mine.

My computer has two devices:
device 0 = HD7970
device 1 = HD7950

I downloaded DistrRTgen source code and compared OpenCL initialisation.
I found a difference and modified genefer code (now, initialisations are similar). Mike, could you update GeneferOCL?

I noticed that DistrRTgen uses boinc_get_opencl_ids() to get platform and device ids. Mike, do you know if we should use this function?

The latest OpenCL beta build is available for download. Its version number is 3.1.2-1, and can be downloaded via this thread.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

On my system (Core2Quad and GTX 460) GeneferCUDA uses almost no CPU at all. GeneferOCL seems to use an entire CPU core as well as the entire GPU.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

On my system (Core2Quad and GTX 460) GeneferCUDA uses almost no CPU at all. GeneferOCL seems to use an entire CPU core as well as the entire GPU.

Hi Yves,

the output with both versions of geneferocl-windows.exe is:

Device List:
0: GPU device 'Tahiti' on 'AMD Accelerated Parallel Processing'.
1: GPU device 'Tahiti' on 'AMD Accelerated Parallel Processing'.
2: CPU device ' Intel(R) Core(TM) i7-3930K CPU @ 3.20GHz' on 'AMD Accelerated Parallel Processing'.
____________
DeleteNull

Are we expecting the same "B" limits as GeneferCUDA?

Running the full blown geneferocl app using the -l option, the b limits mostly look to be similar. Let's see if I can cut and paste something useful...

GeneferOCL:

Generalized Fermat Number b Limits
The upper bound m = 8192, b = 2670000, Err = 0.2910
The upper bound m = 16384, b = 2210000, Err = 0.2969
The upper bound m = 32768, b = 1780000, Err = 0.2969
The upper bound m = 65536, b = 1505000, Err = 0.2969
The upper bound m = 131072, b = 1240000, Err = 0.2969
The upper bound m = 262144, b = 1015000, Err = 0.3047
The upper bound m = 524288, b = 825000, Err = 0.3057
The upper bound m = 1048576, b = 680000, Err = 0.3047
The upper bound m = 2097152, b = 555000, Err = 0.2969
The upper bound m = 4194304, b = 455000, Err = 0.2813
The upper bound m = 8388608, b = 385000, Err = 0.3125

The higher limits are green and the lower limits are red. As you can see, the limits are similar but not identical.

Some of DeleteNull's GeneferOCL tasks have now been validated (one with a CPU, the other with a Tesla GPU), so we now officially have our first credit awarded to GeneferOCL and an ATI/AMD GPU.

This is a milestone many people have been waiting for.

Many thanks and congratulation to Yves for the great achievement! (This is, of course, just the latest advance for which Yves deserves our thanks.)
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Running on platform 'AMD Accelerated Parallel Processing', device 'AMD Athlon(tm) 64 X2 Dual Core Processor 4200+', version 'OpenCL 1.2 AMD-APP (1124.2)' and driver '1124.2 (sse2)'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 8.73 ms/mul. Err: 0.2344 51956 digits



ATI 3550 or somehting like that.....

____________
wbr, Me. Dead J. Dona

The second is that it doesn't tell you the exact device model. That's a shame.

Running on platform 'AMD Accelerated Parallel Processing', device 'AMD Athlon(tm) 64 X2 Dual Core Processor 4200+', version 'OpenCL 1.2 AMD-APP (1124.2)' and driver '1124.2 (sse2)'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 8.73 ms/mul. Err: 0.2344 51956 digits



ATI 3550 or somehting like that.....

If anyone builds a Linux or Mac build of GeneferOCL before we do, the source has been changed to 3.1.2-2, which has a few minor tweaks to the utility functions. The only significant change is that under Linux and Mac, the benchmarks should produce the correct results.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Wow, OpenCL is almost twice as fast on your Titan! That is totally unexpected.

I tuned geneferocl. It's a bit faster on my computer for large exponents.

I committed the new parameters on assembla, in the branch [...]\branches\yves\2013\OclGenefer.

I think that it is faster on a 'Tahiti', but only a real experiment can answer to the question. Please, could someone compile it and run the bench on a HD79x0?

Mike, could you also test it on your GTX 460? Because, I don't know why, but your card doesn't like my algorithm :o) If it is the card that found 75898^524288+1 with GeneferCUDA, I understand it!

If new tuning is faster on Tahiti and Fermi, I will update the real genefer.

Thanks, Yves

The GeneferOCL we've been running is a 64 bit version. Until today, I had not tried running a 32 bit version. I would expect the 32 bit and 64 bit builds to run at the same speed, and in fact their benchmark speeds are the same. This behavior is consistent with GeneferCUDA.

However, the B limits are NOT the same.

64 bit GeneferOCL:

Command line: geneferocl-windows-x64.exe -l


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 460', version 'OpenCL 1.1
CUDA' and driver '320.57'.

Generalized Fermat Number b Limits
The upper bound m = 8192, b = 2670000, Err = 0.2910
The upper bound m = 16384, b = 2210000, Err = 0.2969
The upper bound m = 32768, b = 1780000, Err = 0.2969
The upper bound m = 65536, b = 1505000, Err = 0.2969
The upper bound m = 131072, b = 1240000, Err = 0.2969
The upper bound m = 262144, b = 1015000, Err = 0.3047
The upper bound m = 524288, b = 825000, Err = 0.3057
The upper bound m = 1048576, b = 680000, Err = 0.3047
The upper bound m = 2097152, b = 555000, Err = 0.2969
The upper bound m = 4194304, b = 455000, Err = 0.2813
The upper bound m = 8388608, b = 385000, Err = 0.3125

Command line: geneferocl-windows-x86.exe -l


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 460', version 'OpenCL 1.1
CUDA' and driver '320.57'.

Generalized Fermat Number b Limits
The upper bound m = 8192, b = 2720000, Err = 0.2969
The upper bound m = 16384, b = 2210000, Err = 0.2969
The upper bound m = 32768, b = 1830000, Err = 0.3008
The upper bound m = 65536, b = 1490000, Err = 0.2969
The upper bound m = 131072, b = 1235000, Err = 0.3008
The upper bound m = 262144, b = 1015000, Err = 0.2891
The upper bound m = 524288, b = 840000, Err = 0.3047
The upper bound m = 1048576, b = 690000, Err = 0.3008
The upper bound m = 2097152, b = 565000, Err = 0.3066
The upper bound m = 4194304, b = 470000, Err = 0.3125
The upper bound m = 8388608, b = 385000, Err = 0.3125

While I wasn't surprised that there were small differences between GeneferCUDA and GeneferOCL, this does surprise me a bit because the math done on the GPU shouldn't be affected by the integer size on the CPU, especially since that doesn't affect the double precision floating point on either the CPU or the GPU.

While I wasn't surprised that there were small differences between GeneferCUDA and GeneferOCL, this does surprise me a bit because the math done on the GPU shouldn't be affected by the integer size on the CPU, especially since that doesn't affect the double precision floating point on either the CPU or the GPU.

The cos/sin tables are computed during initialisation by the CPU.
I think that it is more accurate than the GPU sin/cos functions.

Win32 binaries still use FP80 for internal computation. x64 uses SSE2 and then FP64. If you compile the win32 app with "SSE2 instruction set", the results may be identical...?

I tuned geneferocl. It's a bit faster on my computer for large exponents.

I committed the new parameters on assembla, in the branch [...]\branches\yves\2013\OclGenefer.

I think that it is faster on a 'Tahiti', but only a real experiment can answer to the question. Please, could someone compile it and run the bench on a HD79x0?

Mike, could you also test it on your GTX 460? Because, I don't know why, but your card doesn't like my algorithm :o) If it is the card that found 75898^524288+1 with GeneferCUDA, I understand it!

If new tuning is faster on Tahiti and Fermi, I will update the real genefer.

Thanks, Yves

If anyone has BOTH an Nvidia and an AMD GPU in the same system, I'd like to know which one GeneferOCL chooses.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Here's some benchmarks for some Fermis.

These are on AMD PII 1100Ts systems @ 3.8GHZ accept the 570 which is on a 2600K @ 4.3GHZ
All GPUs at stock speed

GeForce GTX 470

C:\>genefercuda-windows.exe -b
genefercuda 3.1.2-2 (Windows 32-bit CUDA)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: genefercuda-windows.exe -b

Generalized Fermat Number Bench
2199064^8192+1 Time: 111 us/mul. Err: 0.2188 51956 digits
1798620^16384+1 Time: 119 us/mul. Err: 0.2188 102481 digits
1471094^32768+1 Time: 177 us/mul. Err: 0.2031 202102 digits
1203210^65536+1 Time: 271 us/mul. Err: 0.2031 398482 digits
984108^131072+1 Time: 527 us/mul. Err: 0.2070 785521 digits
804904^262144+1 Time: 742 us/mul. Err: 0.2031 1548156 digits
658332^524288+1 Time: 1.35 ms/mul. Err: 0.2266 3050541 digits
538452^1048576+1 Time: 2.5 ms/mul. Err: 0.2031 6009544 digits
440400^2097152+1 Time: 4.77 ms/mul. Err: 0.1953 11836006 digits
360204^4194304+1 Time: 10 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 20.9 ms/mul. Err: 0.1797 45879398 digitsght


C:\>geneferocl-windows.exe -b
geneferocl 3.1.2-2 (Windows 64-bit OpenGL)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: geneferocl-windows.exe -b


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 470', version 'OpenCL 1.1
CUDA' and driver '320.18'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 90.3 us/mul. Err: 0.2344 51956 digits
1798620^16384+1 Time: 98.9 us/mul. Err: 0.2266 102481 digits
1471094^32768+1 Time: 134 us/mul. Err: 0.2344 202102 digits
1203210^65536+1 Time: 195 us/mul. Err: 0.2656 398482 digits
984108^131072+1 Time: 322 us/mul. Err: 0.2188 785521 digits
804904^262144+1 Time: 654 us/mul. Err: 0.2188 1548156 digits
658332^524288+1 Time: 1.19 ms/mul. Err: 0.2188 3050541 digits
538452^1048576+1 Time: 2.46 ms/mul. Err: 0.2266 6009544 digits
440400^2097152+1 Time: 5 ms/mul. Err: 0.2266 11836006 digits
360204^4194304+1 Time: 10.5 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 22.2 ms/mul. Err: 0.1895 45879398 digits
Genefer Mark = 483.


GeForce GTX 560ti

C:\>genefercuda-windows.exe -d 1 -b
genefercuda 3.1.2-2 (Windows 32-bit CUDA)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: genefercuda-windows.exe -d 1 -b

Generalized Fermat Number Bench
2199064^8192+1 Time: 97.7 us/mul. Err: 0.2188 51956 digits
1798620^16384+1 Time: 111 us/mul. Err: 0.2188 102481 digits
1471094^32768+1 Time: 150 us/mul. Err: 0.2031 202102 digits
1203210^65536+1 Time: 239 us/mul. Err: 0.2031 398482 digits
984108^131072+1 Time: 444 us/mul. Err: 0.2070 785521 digits
804904^262144+1 Time: 752 us/mul. Err: 0.2031 1548156 digits
658332^524288+1 Time: 1.48 ms/mul. Err: 0.2266 3050541 digits
538452^1048576+1 Time: 2.97 ms/mul. Err: 0.2031 6009544 digits
440400^2097152+1 Time: 6.09 ms/mul. Err: 0.1953 11836006 digits
360204^4194304+1 Time: 12.3 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 25.6 ms/mul. Err: 0.1797 45879398 digits

C:\>geneferocl-windows.exe -d 1 -b
geneferocl 3.1.2-2 (Windows 64-bit OpenGL)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: geneferocl-windows.exe -d 1 -b


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 560 Ti', version 'OpenCL
1.1 CUDA' and driver '320.18'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 68.4 us/mul. Err: 0.2344 51956 digits
1798620^16384+1 Time: 73.2 us/mul. Err: 0.2266 102481 digits
1471094^32768+1 Time: 106 us/mul. Err: 0.2344 202102 digits
1203210^65536+1 Time: 190 us/mul. Err: 0.2656 398482 digits
984108^131072+1 Time: 366 us/mul. Err: 0.2188 785521 digits
804904^262144+1 Time: 791 us/mul. Err: 0.2188 1548156 digits
658332^524288+1 Time: 1.58 ms/mul. Err: 0.2188 3050541 digits
538452^1048576+1 Time: 3.28 ms/mul. Err: 0.2266 6009544 digits
440400^2097152+1 Time: 6.95 ms/mul. Err: 0.2266 11836006 digits
360204^4194304+1 Time: 14.5 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 30.3 ms/mul. Err: 0.1895 45879398 digits
Genefer Mark = 353.


GeForce GTX 570

C:\>genefercuda-windows.exe -b
genefercuda 3.1.2-2 (Windows 32-bit CUDA)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: genefercuda-windows.exe -b

Generalized Fermat Number Bench
2199064^8192+1 Time: 91.6 us/mul. Err: 0.2188 51956 digits
1798620^16384+1 Time: 98 us/mul. Err: 0.2188 102481 digits
1471094^32768+1 Time: 145 us/mul. Err: 0.2031 202102 digits
1203210^65536+1 Time: 226 us/mul. Err: 0.2031 398482 digits
984108^131072+1 Time: 410 us/mul. Err: 0.2070 785521 digits
804904^262144+1 Time: 596 us/mul. Err: 0.2031 1548156 digits
658332^524288+1 Time: 1.09 ms/mul. Err: 0.2266 3050541 digits
538452^1048576+1 Time: 2.03 ms/mul. Err: 0.2031 6009544 digits
440400^2097152+1 Time: 3.91 ms/mul. Err: 0.1953 11836006 digits
360204^4194304+1 Time: 8.13 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 16.7 ms/mul. Err: 0.1797 45879398 digits

C:\>geneferocl-windows.exe -b
geneferocl 3.1.2-2 (Windows 64-bit OpenGL)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: geneferocl-windows.exe -b


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 570', version 'OpenCL 1.1
CUDA' and driver '310.70'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 81.8 us/mul. Err: 0.2344 51956 digits
1798620^16384+1 Time: 88.8 us/mul. Err: 0.2266 102481 digits
1471094^32768+1 Time: 118 us/mul. Err: 0.2344 202102 digits
1203210^65536+1 Time: 170 us/mul. Err: 0.2656 398482 digits
984108^131072+1 Time: 273 us/mul. Err: 0.2188 785521 digits
804904^262144+1 Time: 527 us/mul. Err: 0.2188 1548156 digits
658332^524288+1 Time: 957 us/mul. Err: 0.2188 3050541 digits
538452^1048576+1 Time: 1.91 ms/mul. Err: 0.2266 6009544 digits
440400^2097152+1 Time: 3.91 ms/mul. Err: 0.2266 11836006 digits
360204^4194304+1 Time: 8.2 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 17.3 ms/mul. Err: 0.1895 45879398 digits
Genefer Mark = 618.


GeForce GTX 580

C:\>genefercuda-windows.exe -b
genefercuda 3.1.2-2 (Windows 32-bit CUDA)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: genefercuda-windows.exe -b

Generalized Fermat Number Bench
2199064^8192+1 Time: 98.4 us/mul. Err: 0.2188 51956 digits
1798620^16384+1 Time: 94.8 us/mul. Err: 0.2188 102481 digits
1471094^32768+1 Time: 114 us/mul. Err: 0.2031 202102 digits
1203210^65536+1 Time: 159 us/mul. Err: 0.2031 398482 digits
984108^131072+1 Time: 278 us/mul. Err: 0.2070 785521 digits
804904^262144+1 Time: 489 us/mul. Err: 0.2031 1548156 digits
658332^524288+1 Time: 867 us/mul. Err: 0.2266 3050541 digits
538452^1048576+1 Time: 1.61 ms/mul. Err: 0.2031 6009544 digits
440400^2097152+1 Time: 3.09 ms/mul. Err: 0.1953 11836006 digits
360204^4194304+1 Time: 6.42 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 13.3 ms/mul. Err: 0.1797 45879398 digits

C:\>geneferocl-windows.exe -b
geneferocl 3.1.2-2 (Windows 64-bit OpenGL)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: geneferocl-windows.exe -b


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 580', version 'OpenCL 1.1
CUDA' and driver '314.22'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 67.3 us/mul. Err: 0.2344 51956 digits
1798620^16384+1 Time: 71.8 us/mul. Err: 0.2266 102481 digits
1471094^32768+1 Time: 74.5 us/mul. Err: 0.2344 202102 digits
1203210^65536+1 Time: 112 us/mul. Err: 0.2656 398482 digits
984108^131072+1 Time: 186 us/mul. Err: 0.2188 785521 digits
804904^262144+1 Time: 396 us/mul. Err: 0.2188 1548156 digits
658332^524288+1 Time: 752 us/mul. Err: 0.2188 3050541 digits
538452^1048576+1 Time: 1.52 ms/mul. Err: 0.2266 6009544 digits
440400^2097152+1 Time: 3.13 ms/mul. Err: 0.2266 11836006 digits
360204^4194304+1 Time: 6.64 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 14.2 ms/mul. Err: 0.1895 45879398 digits
Genefer Mark = 767.


Any specific testing needed let me know.
____________
Largest Primes to Date:
As Double Checker: SR5 109208*5^1816285+1 Dgts-1,269,534
As Initial Finder: SR5 243944*5^1258576-1 Dgts-879,713

Here's some benchmarks for some Fermis.

These are on AMD PII 1100Ts systems @ 3.8GHZ accept the 570 which is on a 2600K @ 4.3GHZ
All GPUs at stock speed

I do have a second 560ti, exact same Zotac model as first bench but on my 2600K SB. The benchmarks are slightly slower. Here's the bench from the 2600K with the PII 1100T below it to compare.

GTX 560ti on 2600K

C:\>genefercuda-windows.exe -d 1 -b
genefercuda 3.1.2-2 (Windows 32-bit CUDA)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: genefercuda-windows.exe -d 1 -b

Generalized Fermat Number Bench
2199064^8192+1 Time: 97.7 us/mul. Err: 0.2188 51956 digits
1798620^16384+1 Time: 112 us/mul. Err: 0.2188 102481 digits
1471094^32768+1 Time: 151 us/mul. Err: 0.2031 202102 digits
1203210^65536+1 Time: 243 us/mul. Err: 0.2031 398482 digits
984108^131072+1 Time: 449 us/mul. Err: 0.2070 785521 digits
804904^262144+1 Time: 771 us/mul. Err: 0.2031 1548156 digits
658332^524288+1 Time: 1.5 ms/mul. Err: 0.2266 3050541 digits
538452^1048576+1 Time: 3.01 ms/mul. Err: 0.2031 6009544 digits
440400^2097152+1 Time: 6.25 ms/mul. Err: 0.1953 11836006 digits
360204^4194304+1 Time: 12.5 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 26.1 ms/mul. Err: 0.1797 45879398 digits

C:\>geneferocl-windows.exe -d 1 -b
geneferocl 3.1.2-2 (Windows 64-bit OpenGL)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: geneferocl-windows.exe -d 1 -b


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 560 Ti', version 'OpenCL
1.1 CUDA' and driver '310.70'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 69.6 us/mul. Err: 0.2344 51956 digits
1798620^16384+1 Time: 74.2 us/mul. Err: 0.2266 102481 digits
1471094^32768+1 Time: 107 us/mul. Err: 0.2344 202102 digits
1203210^65536+1 Time: 195 us/mul. Err: 0.2656 398482 digits
984108^131072+1 Time: 374 us/mul. Err: 0.2188 785521 digits
804904^262144+1 Time: 806 us/mul. Err: 0.2188 1548156 digits
658332^524288+1 Time: 1.59 ms/mul. Err: 0.2188 3050541 digits
538452^1048576+1 Time: 3.32 ms/mul. Err: 0.2266 6009544 digits
440400^2097152+1 Time: 6.95 ms/mul. Err: 0.2266 11836006 digits
360204^4194304+1 Time: 14.6 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 30.6 ms/mul. Err: 0.1895 45879398 digits
Genefer Mark = 350.


GTX 560ti on 1100T

C:\>genefercuda-windows.exe -d 1 -b
genefercuda 3.1.2-2 (Windows 32-bit CUDA)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: genefercuda-windows.exe -d 1 -b

Generalized Fermat Number Bench
2199064^8192+1 Time: 97.7 us/mul. Err: 0.2188 51956 digits
1798620^16384+1 Time: 111 us/mul. Err: 0.2188 102481 digits
1471094^32768+1 Time: 150 us/mul. Err: 0.2031 202102 digits
1203210^65536+1 Time: 239 us/mul. Err: 0.2031 398482 digits
984108^131072+1 Time: 444 us/mul. Err: 0.2070 785521 digits
804904^262144+1 Time: 752 us/mul. Err: 0.2031 1548156 digits
658332^524288+1 Time: 1.48 ms/mul. Err: 0.2266 3050541 digits
538452^1048576+1 Time: 2.97 ms/mul. Err: 0.2031 6009544 digits
440400^2097152+1 Time: 6.09 ms/mul. Err: 0.1953 11836006 digits
360204^4194304+1 Time: 12.3 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 25.6 ms/mul. Err: 0.1797 45879398 digits

C:\>geneferocl-windows.exe -d 1 -b
geneferocl 3.1.2-2 (Windows 64-bit OpenGL)
Copyright 2001-2013, Yves Gallot
Copyright 2009, Mark Rodenkirch, David Underbakke
Copyright 2010-2012, Shoichiro Yamada, Ken Brazier
Copyright 2011-2013, Iain Bethune, Michael Goetz, Ronald Schneider

Command line: geneferocl-windows.exe -d 1 -b


Running on platform 'NVIDIA CUDA', device 'GeForce GTX 560 Ti', version 'OpenCL
1.1 CUDA' and driver '320.18'.

Generalized Fermat Number Bench
2199064^8192+1 Time: 68.4 us/mul. Err: 0.2344 51956 digits
1798620^16384+1 Time: 73.2 us/mul. Err: 0.2266 102481 digits
1471094^32768+1 Time: 106 us/mul. Err: 0.2344 202102 digits
1203210^65536+1 Time: 190 us/mul. Err: 0.2656 398482 digits
984108^131072+1 Time: 366 us/mul. Err: 0.2188 785521 digits
804904^262144+1 Time: 791 us/mul. Err: 0.2188 1548156 digits
658332^524288+1 Time: 1.58 ms/mul. Err: 0.2188 3050541 digits
538452^1048576+1 Time: 3.28 ms/mul. Err: 0.2266 6009544 digits
440400^2097152+1 Time: 6.95 ms/mul. Err: 0.2266 11836006 digits
360204^4194304+1 Time: 14.5 ms/mul. Err: 0.2031 23305854 digits
294612^8388608+1 Time: 30.3 ms/mul. Err: 0.1895 45879398 digits
Genefer Mark = 353.

I wonder if it's the CPU, driver version, GPU chip quality or something else. I'll try different driver version a little later.
____________
Largest Primes to Date:
As Double Checker: SR5 109208*5^1816285+1 Dgts-1,269,534
As Initial Finder: SR5 243944*5^1258576-1 Dgts-879,713

I tuned geneferocl. It's a bit faster on my computer for large exponents.

The latest OpenCL beta build is available for download. Its version number is 3.1.2-1, and can be downloaded via this thread.