I have been working on an OpenCL version of wwww for a few months (on and off). It's been slow due to a change in architecture and a lack of free time to work on it.

The original version started with the architecture of wwww. I quickly realized that the sieving code in wwww couldn't feed the GPU fast enough to keep it busy. I then turned to ppsieve. Although ppsieve has a faster sieve, I found that it had the same problem. I then found primesieve (at code.google.com). It is the fastest implementation of a sieve I found.

The end product is a framework called GPUSieve. It is a framework for problems of this nature. By 'problem' I mean a program that sieve for primes, then for each prime, do something with it on the GPU. It is written in C++ and is multi-threaded, something which the other programs cannot do.

The new code is screaming fast compared to the current CPU only code. I estimate that my ATI Radeon 5770 could complete a Wieferich range for 3.5e16 in a year, that is more than twice the total amount of work done on the search since it started in the 80's. The WallSunSun code should be able to do at least 4e15 in a year, but I haven't done enough time testing of that code.

I hope to release the source code tomorrow along with Windows and Mac builds. I only have the ATI Radeon 5770 to test on (and a really old ATI Radeon Mobility), thus I am looking for some people to help beta test on their hardware. Once I post the code, I will provide details on how I want you to test it.

I can test with several different NVidia GPUs. These projects require a 64-bit OS for the current PRPnet CPU app. For the OpenCL app, is this still true, or will 32-bit be possible?



____________
141941*2^4299438-1 is prime!

Nvidia GPU can also work on OPENCL, so if you have app I am ready to test it.
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

I can test with several different NVidia GPUs. These projects require a 64-bit OS for the current PRPnet CPU app. For the OpenCL app, is this still true, or will 32-bit be possible?

I have a couple of machines with NVidia GPU in 32-bit Windows systems, so I can test it out if you want to build them.


____________
141941*2^4299438-1 is prime!

I have a 4650 in a w7 x64, can it be used for testing? It has no DP support.

Do I need DP? If not I might be test on a couple W7-64 machines I have one of each ATI Juniper and Redwood series.

I have a HD 6470 and an HD 6310 available to test.
____________

I have posted the code (along with a Mac build) here: http://home.roadrunner.com/~mrodenkirch/wwwwcl.zip.

My estimate for 3.5e16 in a year was on a ATI Radeon HD 5870. I have a Windows box with an ATI Radeon 6450, but the program hangs on that. I suspect it is due to mixing a 32-bit exe with the 64-bit OpenCL library. I'll try to fix that in the coming days.

Since I can't provide a Windows build yet, there is a VS2008 solution included. If you have VS2008 (or newer) and want to build it, you just need to install the AMD APP OpenCL development toolkit.

I don't have access to any Linux systems, so I cannot build the code for Linux. I have also been lazy in that I didn't write a makefile. I built on the Mac with this command line:

gcc *.c *.cpp opencl/*.cpp wwww/*.cpp sieve/*.cpp -m64 -framework OpenCL -lstdc++ -O2 -o wwwwcl

I assume that on Linux you just need to replace the "-framework OpenCL" with the appropriate -I and -l options to include the OpenCL headers and libraries on your system.

Execution is fairly simple:

./wwwwcl -p1334568e11 -P1334569e11 -TWieferich -t5 -b2000

where -t tells the program how many threads to execute and -b is the number of blocks per thread. Here is what to expect for output:

wwwwcl v2.0
Compiled Apr 29 2012 with GCC 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.1.00)
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456892569421833, 9.560M p/sec, 1.07 CPU cores, 92.6% done. ETA 29 Apr 18:52
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 265.35 sec. (1.85 init + 263.37 sieve) at 9557118 p/sec.
Processor time: 282.66 sec. (1.86 init + 280.80 sieve).
Seconds spent in CPU and GPU: 1066.99 (cpu), 124.43 (gpu)
Percent of time spent in CPU vs. GPU: 0.90 (cpu), 0.10 (gpu)
CPU/GPU utilization: 0.20 (cores), 0.02 (devices)

You will see more detail with the -v switch. I recommend starting with a smaller number of threads and blocks as you will use those to find the optimal settings for you GPU.

As part of the testing you should run ranges that have known Wieferich/WallSunSun special instances and ranges without them and thus verify the output.

I think it very likely that this program will either not run or will crash for some of you and might hang the system of a few others. (I experienced a few hangs during my initial development, so you aren't alone.) In all seriousness I strongly suggest starting with small ranges, 1e8 or so, and small numbers of threads and blocks until you get more comfortable running the program and feel that it will run safely on your hardware.

What I'm looking for is problems compiling and linking, problems at runtime, i.e. OpenCL errors output by the program, and any other serious issues.

This code will NOT work with PRPNet, at least not yet. I have to make some changes to both the client and server to support this program and I really don't want to support it until the major kinks are worked out.

here is the output from my iMac using an ATI Radeon HD 5750:

Download-32bit
Download-64bit

This is a win build, I dont know if its working because I have no ATI card.

If you use the -l option, wwwwcl will list the available platforms and devices on your system. The default is to use platform 0, device 0. You can override with the -p and -d options. You will need this if you have multiple GPUs on your system.

I ask you to test other ranges, not just the example I gave. I suggest that you look at special instances on the PRPNet server to get some ranges to test. You will need to use the -s option to change the threshold for special instances. It defaults to 100.

You should also vary the values for -t and -b so that you can determine the optimal settings for your GPU.

On some systems, the GPU is so fast that the GPU is starved for work. You can run multiple instances of wwwwcl concurrently to keep the GPU fed. To reach the 3.5e16 yearly throughput, I could run 3 instances of wwwwcl.

I can speed up the GPU code more, but it was more important for me to get this beta out first. I don't know when I'll get around to further speed-ups.

When I try to run the 64 bit Win build with

wwwwcl -p1334568e11 -P1334569e11 -TWieferich -t5 -b2000

I get an error message saying something like

The application could not be started (0xc000007b). Click "OK" to close the application.

Tried to run under Windows using a ATI Mobility Radeon HD 5470.

Neither windows build will run for me. I get an immediate program "stopped working" message without any output or a log file. This happens when I run with just the -l option.

i7-860 (12GB RAM)
GPU #1 - GT 530
GPU #2 - 9600 GSO

Win7 Enterprise x64

I tested the compiled win versions.

Test machine was an i7-2720QM with HD6470M running Win 7 x64 Enterprise, I tested the same range as rogue:

x86:

Hmm, when checking wwwwcl.exe with Dependency Walker it says there is a file missing called sysntfy.dll

Neither windows build will run for me. I get an immediate program "stopped working" message without any output or a log file. This happens when I run with just the -l option.

i7-860 (12GB RAM)
GPU #1 - GT 530
GPU #2 - 9600 GSO

Win7 Enterprise x64

Rebirther, your 64-bit app works fine on my NVIDIA card:
Win 7 Ultimate (64-bit)
i7 875K @2.93GHz (HT disabled, no other CPU tasks running)
9800 GT
Visual C++ 2010 Redistributable Package (x64) installed.


Examples:

d:\man\wwww>wwwwcl.exe -p5423711e9 -P5423715e9 -TWieferich -t2 -b3000 -s1000
wwwwcl v2.0
Sieve started: 5423711000000000 <= p < 5423715000000000
5423711968552697 is a special instance (+1 -488 p)

Sieve complete: 5423711000000001 <= p < 5423715000000000 110405720 primes tested
Elapsed time: 17.61 sec. (0.22 init + 17.00 sieve) at 6268625 p/sec.
Processor time: 24.60 sec. (0.22 init + 24.38 sieve).
Seconds spent in CPU and GPU: 17.60 (cpu), 25.24 (gpu)
Percent of time spent in CPU vs. GPU: 0.41 (cpu), 0.59 (gpu)
CPU/GPU utilization: 0.51 (cores), 0.73 (devices)

d:\man\wwww>wwwwcl.exe -p14490200e9 -P14490204e9 -TWieferich -t2 -b3000 -s1000
wwwwcl v2.0
Sieve started: 14490200000000000 <= p < 14490204000000000

Sieve complete: 14490200000000001 <= p < 14490204000000000 107488122 primes tested
Elapsed time: 17.51 sec. (0.36 init + 16.82 sieve) at 6139807 p/sec.
Processor time: 24.76 sec. (0.36 init + 24.40 sieve).
Seconds spent in CPU and GPU: 17.53 (cpu), 25.68 (gpu)
Percent of time spent in CPU vs. GPU: 0.41 (cpu), 0.59 (gpu)
CPU/GPU utilization: 0.51 (cores), 0.75 (devices)

So it works it seems. Top marks! Though I am slightly worried about the cpu usage. It almost used two full cores. Is that worth switching over to GPU for?

Edit: System: Win 7 64bit, i5 2500k, gtx 570, nothing else eating away at the cpu.

@all: The win app is only for ATI cards, you will get an error with a missing ati driver file and the app is crashing on nvidia cards.

Through reading here, some got it running on nvidia cards. I dont know how and why???

I tested the compiled win versions.

Test machine was an AMD E-350 with HD6310 running Win 7 x64 Ultimate, I tested the same range as rogue:

x86:

@all: The win app is only for ATI cards, you will get an error with a missing ati driver file and the app is crashing on nvidia cards.

Through reading here, some got it running on nvidia cards. I dont know how and why???

I needed only to make sure the libraries in the system matched the app (32 vs 64-bit).

Can you make a list of libs and your driver version pls?

@all: The win app is only for ATI cards, you will get an error with a missing ati driver file and the app is crashing on nvidia cards.

Through reading here, some got it running on nvidia cards. I dont know how and why???

I needed only to make sure the libraries in the system matched the app (32 vs 64-bit).

Can you make a list of libs and your driver version pls?

Well, I installed the package from MS and it worked immediately. Links: x64 and x86

You may need to copy the files msvcp100.dll and msvcr100.dll from 'Windows\system32' to the folder of the executable. That Dependency Walker utility (available via the link Toshio posted) could be useful.

NVIDIA driver is 290.53

I have already both files and its not working. How did you run it? I have a batch file + exe, thats all.

Do you got a popup with ati driver?

Intel i5-2500k
Nvidia 560Ti ( beta 301.xx drivers)
windows 7 64 bit
64 bit app wwwwcl 2.0
GPU usage is about 31 %
and uses over over one core ( 35% ) on quad processor.


wwwwcl -p1334568e11 -P1334569e11 -TW
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456886350927309, 12.00M p/sec, 1.39 CPU cores, 86.4%
Sieve complete: 133456800000000001 <= p < 133456900000000000
Elapsed time: 214.60 sec. (2.41 init + 212.16 sieve) at 1181
Processor time: 297.20 sec. (2.87 init + 294.33 sieve).
Seconds spent in CPU and GPU: 441.44 (cpu), 83.89 (gpu)
Percent of time spent in CPU vs. GPU: 0.84 (cpu), 0.16 (gpu)
CPU/GPU utilization: 0.20 (cores), 0.04 (devices)
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

If your elapsed time is much greater than processor time, then you should reduce the number of threads.

Win 7 Pro 64bit / i7-2700K stock speed / GTX570 stock speed (more details at the end if you're interested)

GT440 (OEM version - 144 shaders)

i7920 @ stock (HT on)

Vista Home Premium x64

I've made some code changes to increase gpu load. Changed single thread sieve to multi-thread and what I've got.

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t6 -b2000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456849813392709, 41.95M p/sec, 5.65 CPU cores, 49.8% done. ETA 02 May 01:
21
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes
tested
Elapsed time: 65.63 sec. (4.77 init + 60.76 sieve) at 38640587 p/sec.
Processor time: 364.20 sec. (23.12 init + 341.08 sieve).
Seconds spent in CPU and GPU: 270.71 (cpu), 49.33 (gpu)
Percent of time spent in CPU vs. GPU: 0.85 (cpu), 0.15 (gpu)
CPU/GPU utilization: 0.17 (cores), 0.03 (devices)

If someone want to test, binary is here (32bit win): http://dl.dropbox.com/u/1452459/wwwwcl.exe

I think I know what you changed. I had removed ParallelPrimeSieve from the sieve folder since I wasn't using it. I can re-add it then have App.cpp instantiate it instead of PrimeSieve. Nevertheless you can send me your code changes just in case you did something differently. The biggest PITB might be dependencies within ParallelPrimeSieve for other libraries.

If your elapsed time is much greater than processor time, then you should reduce the number of threads.

If your elapsed time is much greater than processor time, then you should reduce the number of threads.

This might not hold true for the Fermi NVidia cards. Using -t4 or -t2 slows things down slightly on my GT440 OEM card compared to -t5. Increasing to -t8 (and using -b3000), however, almost doubles performance (and -t16 crashes the system).

I've made some code changes to increase gpu load. Changed single thread sieve to multi-thread and what I've got.

Please do not post exes without code. I maintain the code base and the official builds (with help from rebirther). If you don't mind, could you send me the code change you made. I prefer e-mail rather than the forum or PM. My e-mail is rogue @ wi.rr.com.

Rebirther's 64bit app on 64bit Win7, GTX580 with driver version 266.58 and CUDAToolkit 3.2

Sesef version

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t6 -b2000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456892074820551, 13.19M p/sec, 1.41 CPU cores, 92.1% done. ETA 02 May 12:42
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 240.01 sec. (22.85 init + 217.09 sieve) at 10566198 p/sec.
Processor time: 311.66 sec. (24.01 init + 287.65 sieve).
Seconds spent in CPU and GPU: 384.62 (cpu), 658.20 (gpu)
Percent of time spent in CPU vs. GPU: 0.37 (cpu), 0.63 (gpu)
CPU/GPU utilization: 0.18 (cores), 0.30 (devices)

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t8 -b3000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456882799494283, 31.45M p/sec, 2.83 CPU cores, 82.8% done. ETA 02 May 12:44
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 93.85 sec. (6.44 init + 87.38 sieve) at 27021435 p/sec.
Processor time: 246.70 sec. (8.99 init + 237.71 sieve).
Seconds spent in CPU and GPU: 257.82 (cpu), 321.73 (gpu)
Percent of time spent in CPU vs. GPU: 0.44 (cpu), 0.56 (gpu)
CPU/GPU utilization: 0.13 (cores), 0.16 (devices)

Rogue 64 bit

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t6 -b2000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456895638206683, 19.89M p/sec, 1.50 CPU cores, 95.6% done. ETA 02 May 12:54
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 128.09 sec. (0.59 init + 127.48 sieve) at 19797948 p/sec.
Processor time: 191.58 sec. (0.59 init + 190.99 sieve).
Seconds spent in CPU and GPU: 575.66 (cpu), 0.05 (gpu)
Percent of time spent in CPU vs. GPU: 1.00 (cpu), 0.00 (gpu)
CPU/GPU utilization: 0.17 (cores), 0.00 (devices)

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t8 -b3000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456894820528821, 19.82M p/sec, 1.40 CPU cores, 94.8% done. ETA 02 May 12:56
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 128.47 sec. (0.58 init + 127.86 sieve) at 19740249 p/sec.
Processor time: 179.96 sec. (0.58 init + 179.39 sieve).
Seconds spent in CPU and GPU: 785.99 (cpu), 51.53 (gpu)
Percent of time spent in CPU vs. GPU: 0.94 (cpu), 0.06 (gpu)
CPU/GPU utilization: 0.13 (cores), 0.01 (devices)
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

If you too many threads I think that some of the additional GPU time is spent waiting to execute in the GPU. It would be interesting if you vary -t and -b more. For example, try -t3 -b10000.

The error with clGetKernelWorkGroupInfo is due to the requirement for OpenCL 1.1. The code in wwwwcl is not supported by OpenCL 1.0. Can anyone tell me the minimum driver needed to support OpenCL 1.1? I'll add that information to the readme.

Can anyone tell me the minimum driver needed to support OpenCL 1.1? I'll add that information to the readme.

As you requested

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456895194058081, 19.96M p/sec, 1.58 CPU cores, 95.2% done. ETA 02 May 13:52
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 127.83 sec. (0.61 init + 127.14 sieve) at 19838761 p/sec.
Processor time: 201.29 sec. (0.61 init + 200.68 sieve).
Seconds spent in CPU and GPU: 260.22 (cpu), 69.95 (gpu)
Percent of time spent in CPU vs. GPU: 0.79 (cpu), 0.21 (gpu)
CPU/GPU utilization: 0.33 (cores), 0.09 (devices)

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t2 -b15000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456895395957667, 19.96M p/sec, 1.56 CPU cores, 95.4% done. ETA 02 May 13:55
Sieve complete: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 127.77 sec. (0.61 init + 127.04 sieve) at 19848544 p/sec.
Processor time: 198.70 sec. (0.61 init + 198.09 sieve).
Seconds spent in CPU and GPU: 132.52 (cpu), 66.87 (gpu)
Percent of time spent in CPU vs. GPU: 0.66 (cpu), 0.34 (gpu)
CPU/GPU utilization: 0.50 (cores), 0.25 (devices)

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t1 -b25000
wwwwcl v2.0
Sieve started: 133456800000000000 <= p < 133456900000000000
p=133456889338901581, 12.41M p/sec, 0.98 CPU cores, 89.3% done. ETA 02 May 13:59
Sieve interrupted: 133456800000000001 <= p < 133456900000000000 2535955284 primes tested
Elapsed time: 205.08 sec. (0.61 init + 204.28 sieve) at 12365704 p/sec.
Processor time: 200.82 sec. (0.61 init + 200.21 sieve).
Seconds spent in CPU and GPU: 209.72 (cpu), 69.48 (gpu)
Percent of time spent in CPU vs. GPU: 0.75 (cpu), 0.25 (gpu)
CPU/GPU utilization: 1.00 (cores), 0.33 (devices)
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

For NVidia, it looks like you need something above the 280.13 (Linux) or 280.19(Windows) drivers.

@rogue: Is it possible to add a function which do some short tests with settings for -t and -b and give an output to determine the optimal setting/throughput?

Regards Odi

9600 GSO

AMD Phenom 9100e

Win7 enterprise x64

I got the same error with toolkit 3.2 and 4.2 (compiled app). Driver version 285.62. Iam not sure why. The app should run with 3.2 but 4.2 need a new driver.

@Honza: Do you create the app and run or do you only installed toolkit?

One of the slowest possible NVidia cards...


8400M GS

Core2 T8100

Vista Home Premium 32-bit

I use 301.xx series on my 560Ti works perfect
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

Nvidia 560TI beta 301.xx drivers Win 7 64 bit
wwwwcl64.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel

wwwwcl64.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel


wwwwcl64.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel
____________
92*10^^1585996-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^⁴⁹⁸¹⁰¹-1 CRUS PRIME
2022202116^¹³¹⁰⁷²+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

Nvidia 560TI beta 301.xx drivers Win 7 64 bit
wwwwcl64.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel

wwwwcl64.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel


wwwwcl64.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel

wwwwcl.exe -p1334568e11 -P1334569e11 -TWieferich -t3 -b10000
wwwwcl v2.0
OpenCL Error: Invalid value
in call to clGetKernelWorkGroupInfo
kernelName: wieferich_kernel

[quote]Nvidia 560TI beta 301.xx drivers Win 7 64 bit

This is most likely due to the driver not supporting OpenCL 1.1. I could change the code to remove this dependency, but it could hurt performance.

I've updated the source (and Mac build) to include the changes provided by sesef. No other changes were made. I'll let rebirther build the Win32 and Win64 exes as I still can get them to build and run correctly.

Make sure that you test ranges with known special instances.

Also, please test some WallSunSun ranges (-TWallSunSun). They are a little slower, but the software can verify that the code on the GPU was executed correctly whereas it can't do that for Wieferich.

I hope to test the PRPNet changes this weekend so that we can start using wwwwcl in production next week.

Win 2008R2 SP2 x64, ATI 7950

Win 7 Pro 64bit / i7-2700K stock speed / GTX570 stock speed

Win 2008R2 SP2 x64, ATI 7950

wwwwcl64.exe -p1334568e11 -P1334569e11 -TWieferich -t4 -b2000

Give around 14,25Mp/sec with GPU usage ~70%.
32-bit version does 10,6Mp/sec with GPU usage ~50%.
I wasn't able to get more GPU usage using differrent -t and -b switches. According to task manager, process still uses only around 1 CPU core.

-TWallSunSun crashes the app.

c:\temp>wwwwcl64.exe -p1334568e11 -P1334569e11 -TWallSunSun -t2 -b1000 -v
wwwwcl v2.0
Defaulting to platform 0
Defaulting to device 0

Win7 x64 on i5-2500, GTX 580 with 301.xx driver.
GPU usage ~35%.

I have tested with larger -b value, one error, smaller -b values no errors.

[code]wwwwcl64.exe -p1334568e11 -P1334569e11 -TWallSunSun -t4 -b10000 -v
wwwwcl v2.0
Defaulting to platform 0
Defaulting to device 0
workGroupSize = 2240000 = 10000 * 32 * 7 (blocks * workGroupSizeMultiple * deviceComputeUnits)
Sieve started: 133456800000000000 <= p < 133456900000000000
OpenCL Error: Invalid command queue
in call to clEnqueueReadBuffer
argument: prime