You read that right! I've managed to port PPSieve-CUDA to OpenCL and compile it with the ATI compiler!

The code base is very similar to PPSieve-CUDA, and will eventually be merged with it. But for now, if you have Linux and an ATI GPU, please download:

PPSieve-OpenCL (source, on the redcl branch)

And give it a shot with the usual test procedure:

32 bit: ./ppsieve-cl-boinc-x86-linux -p42070e9 -P42070010e6 -k 1201 -K 9999 -N 2000000 -c 60
64 bit: ./ppsieve-cl-boinc-x86_64-linux -p42070e9 -P42070010e6 -k 1201 -K 9999 -N 2000000 -c 60

It should output the following factors:

42070000070587 | 9475*2^197534+1
42070000198537 | 3373*2^1046686+1
42070003101727 | 4207*2^1054290+1
42070003511309 | 6057*2^1043547+1
42070006307657 | 1513*2^1771812+1
42070006388603 | 2059*2^1816098+1
42070007177519 | 5437*2^1121592+1
42070007396759 | 7339*2^1803518+1
42070008823897 | 4639*2^952018+1
42070008858187 | 2893*2^317690+1

Please provide the output from stderr.txt, especially in case of error. Please also provide as much detail about your system as possible, including your ATI GPU model number, driver version, and Stream Processing Unit clock speed if possible.

Also note that this first version doesn't use vectorized arithmetic, so it may be possible to make it ~1.5-2x faster. The first goal is to prove that this can run on an ATI GPU. :)
____________

sorry, my HD3200 isnt able to run an OpenCL-Kernel :(
____________

I entered this into the command line:

./ppsieve-cl-boinc-x86_64-linux -p42070e9 -P42070010e6 -k 1201 -K 9999 -N 2000000 -c 60

./ppsieve-cl-boinc-x86_64-linux: error while loading shared libraries: libOpenCL.so: cannot open shared object file: No such file or directory

I'm not sure how much work is required to set up for running an OpenCL app. That's part of what I'm testing here.

I downloaded the ATI Stream SDK from here and went through this installation procedure (PDF). I'm hoping the full installation isn't necessary, and that you only need to grab the lib/x86/libOpenCL.so or lib/x86_64/libOpenCL.so file and put it in the same directory with my app. But I really don't know.

Edit: Also, the Stream SDK version was 2.1 when I downloaded it, and that's what I used to compile this, but they're serving 2.2 now. Hopefully, that won't make a difference.
____________

OS: Debian GNU Linux Squeeze
GPU: ATI HD 5750 1024MB
Driver: 10.7
SDK: ati-stream-sdk-v2.2-lnx64

./ppsieve-cl-boinc-x86_64-linux -p42070e9 -P42070010e6 -k 1201 -K 9999 -N 2000000 -c 60
ppsieve version cl-0.0.1-alpha (testing)
Compiled Aug 28 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000

# more stderr.txt
shmget in attach_shmem: Invalid argument
Can't set up shared mem: -1
Will run in standalone mode.
Sieve started: 42070000000000 <= p < 42070010000000
Thread 0 starting
Error: Getting Platforms. (clGetPlatformsIDs)
called boinc_finish

EDIT: I don't know if this is important. At the the beginning of the test i use the wrong terminal and test the app on a host with a nvidia-gpu (GTX260). The app runs without problems.
____________

O.K., I tested it again, with the libOpenCl.so from SDK V2.2. I get the same message:

./ppsieve-cl-boinc-x86_64-linux: error while loading shared libraries: libOpenCL.so: cannot open shared object file: No such file or directory

O.K., I tested it again, with the libOpenCl.so from SDK V2.2. I get the same message:

./ppsieve-cl-boinc-x86_64-linux: error while loading shared libraries: libOpenCL.so: cannot open shared object file: No such file or directory

Is it expecting the file to be in a location other than where it launches from?

Looks like you'll (both/all) have to follow this entire installation procedure.

Don't forget to get this ICD registration file and extract it in the root directory, as root.
____________

Don't forget to get this ICD registration file and extract it in the root directory, as root.

Looks like it was using your CPU, not your GPU. Probably because you're using an AMD processor. Try adding a "--device 1" to the end of the command line.
____________

So, out of curiosity, why is Linux the first platform to test an application on?

I would like to help test this ATI app, but right now I simply do not know enough Linux to actively do so. I am afraid that I must wait for a windows testing app, or learn a lot about Linux.

By the way, thanks Ken (and anyone else involved) for starting ATI development!
____________
May the Force be with you always.

So, out of curiosity, why is Linux the first platform to test an application on?

./ppsieve-cl-boinc-x86_64-linux -p42070e9 -P42070010e6 -k 1201 -K 9999 -N 2000000 -c 60 --device 1
ppsieve version cl-0.0.1-alpha (testing)
Compiled Aug 28 2010 with GCC 4.3.3
nstart=76, nstep=32, gpu_nstep=32
ppsieve initialized: 1201 <= k <= 9999, 76 <= n <= 2000000

# more stderr.txt
shmget in attach_shmem: Invalid argument
Can't set up shared mem: -1
Will run in standalone mode.
Sieve started: 42070000000000 <= p < 42070010000000
Thread 0 starting
Invalid MIT-MAGIC-COOKIE-1 keySIGSEGV: segmentation violation
Stack trace (10 frames):
./ppsieve-cl-boinc-x86_64-linux[0x414d8d]
/lib/libpthread.so.0(+0xef60)[0x7f0558473f60]
/home/tocx/download/ati-stream-sdk-v2.2-lnx64/lib/x86_64/libatiocl64.so(+0xca3b0)[0x7f055654f3b0]
/home/tocx/download/ati-stream-sdk-v2.2-lnx64/lib/x86_64/libatiocl64.so(+0x10b640)[0x7f0556590640]
/home/tocx/download/ati-stream-sdk-v2.2-lnx64/lib/x86_64/libatiocl64.so(clCreateCommandQueue+0x69)[0x7f0556533879]
./ppsieve-cl-boinc-x86_64-linux[0x410e15]
./ppsieve-cl-boinc-x86_64-linux[0x40e83e]
./ppsieve-cl-boinc-x86_64-linux[0x40ad5e]
/lib/libc.so.6(__libc_start_main+0xfd)[0x7f0557976c4d]
./ppsieve-cl-boinc-x86_64-linux[0x409fd9]

Exiting...

suspending the wus on gpu gives the same result
---
trying with other devices

shmget in attach_shmem: Invalid argument
Can't set up shared mem: -1
Will run in standalone mode.
Sieve started: 42070000000000 <= p < 42070010000000
Thread 0 starting
Invalid MIT-MAGIC-COOKIE-1 keyCreating Command Queue. (clCreateCommandQueue)
called boinc_finish
____________

Well, I'm stumped. A segfault is hard enough to debug when you do have access to the hardware it happens on. Since it doesn't happen on a CPU, I don't.

I was hoping that someone would come along with an Intel processor and AMD graphics card, run the test, and find that it worked fine there. Since that hasn't happened, I've uploaded more-or-less the same code compiled with debugging on. If that doesn't produce a useful error message, I'm afraid ATI/OpenCL is at a standstill for now.
____________

Would love to help out, but my ATI cards are all in WIN boxes...


____________
141941*2^4299438-1 is prime!

Would love to help out, but my ATI cards are all in WIN boxes...

Would love to help out, but my ATI cards are all in WIN boxes...

Yeah, OK, it'll probably be easier to build a Windows version than to teach all of you how to use Linux. I'm working on it.
____________

Alright, folks, I stuck a Windows executable in that zipfile. It works on my CPU, but I kind of expect it will also segfault on a real GPU.

But test away; maybe we'll get lucky somewhere.
____________

Alright, folks, I stuck a Windows executable in that zipfile. It works on my CPU, but I kind of expect it will also segfault on a real GPU.

But test away; maybe we'll get lucky somewhere.

I get a "Null Platform found Exiting Application" message.


AMD Athlon x2 5600+
Win7 Enterprise 64-bit
HD 4650
Driver 10.7 (also tested with 10.5 with same error).



____________
141941*2^4299438-1 is prime!

No; nVIDIA already has a version specifically made for it. You can try compiling it for nVIDIA if you want, but it's not likely to be faster than the CUDA version.
____________

Scott, did you install the ATI Stream SDK? I think you have to do that. I haven't yet seen a way to run this app without the full installation.
____________

Scott, did you install the ATI Stream SDK? I think you have to do that. I haven't yet seen a way to run this app without the full installation.

HD 4650

It's working for me... :)

i7 980X + HD5850 (GPU clock: 725 MHz, Memory clock: 1000 MHz) + GTX460
Win7 Prof. x64
Catalyst 10.8, SDK 2.2

First try:

It's working for me... :)
...
Second try:

>ppsieve-cl-x86-windows.exe -p42070e9 -P42070030e6 -k 1201 -K 9999 -N 2000000 -c 60 --device 1 ppsieve version cuda-0.1.5a (testing) nstart=76, nstep=32, gpu_nstep=32 ppsieve initialized: 1201 <= k <= 9999, 76 <= n < 2000000 Sieve started: 42070000000000 <= p < 42070030000000 Thread 0 starting Detected 18 multiprocessors (90 SPUs) on device 1. [...] Thread 0 completed Waiting for threads to exit Sieve complete: 42070000000000 <= p < 42070030000000 Found 27 factors count=955289,sum=0x2dbc17167afb6a8d Elapsed time: 25.47 sec. (0.03 init + 25.44 sieve) at 1184985 p/sec. Processor time: 1.62 sec. (0.03 init + 1.59 sieve) at 18945683 p/sec. Average processor utilization: 1.16 (init), 0.06 (sieve)

About 96% load on the ATI GPU, correct results. :)

hd 4850, catalyst 10.8, sdk 2.2
windows 7 x64

Thread 0 starting Device 1 looks like an 10-core CPU, not a GPU. Adjusting. Detected 10 multiprocessors (50 SPUs) on device 1. Error: Building Program (clBuildProgram)

Catalyst 10.8, SDK 2.2, Windows 2008 R2 x64, HD5850.

>ppsieve-cl-x86-windows.exe -p42070e9 -P42070010e6 -k 1201 -K 9999 -N 2000000 -c 60 --device 2

The application has failed to start because its side-by-side configuration is incorrect. Please see the application event log or use the command-line sxstrace.exe tool for more detail.

(getting same error when running ppsieve-cl-x86-windows.exe without any parameters)

using sxstrace I extracted following

Try installing the 32-bit version of the Microsoft Visual C++ 2008 redistributable package.

P.S. Please also use code tags judiciously. When text inside them is too long, they make the forum overflow my screen to the right.
____________

Try installing the 32-bit version of the Microsoft Visual C++ 2008 redistributable package.

Alas my ATI 5770 bought in april is defunct. Put it in yesterday after it having idled in the package for two months. I will coordinate a replacement on monday with the vendor. Then i am free to look into this.

Alright, I have a new test version ready for everyone, v0.0.2 alpha. This version is vectorized, so it could be up to twice as fast as the last version; though I'm suspecting it might be slower than the last version. Should that be the case, I have a version with no vectorization ready to go as well.

Vectorization is controlled with the -v parameter, so please try -v 2, -v 4, and maybe -v 3 (though that's likely to be really slow!). The default is 4.

Also, this code should now run on 4xxx series cards, though I will need comparisons to previous runs on 5xxx cards as well.

Good luck!

P.S. Gerrit, I hope your card gets well soon.
____________

I've neither linux or a decent windows version *g* nor an amd grafics card, but i've tried the opencl version

Hm, yeah, I probably should take the -g out at this point...OK, done!

Edit: In any case, you'd probably need to compile with nVIDIA's compiler to get this to work on nVIDIA.
____________

okay, got another error

hd 4850, catalyst 10.8, sdk 2.2
windows 7 x64

with -v 2

Elapsed time: 15.69 sec. (0.03 init + 15.65 sieve) at 653138 p/sec. Processor time: 1.29 sec. (0.05 init + 1.25 sieve) at 8191947 p/sec.

with -v 3

Error: Building Program (clBuildProgram)

HD 4650, Catalyst 10.7, SDK 2.2
Win7 64-bit

with -v 2

Average processor utilization: 0.99 (init), 3.96 (sieve)

Scott, I'm wondering about that too, but I doubt the CPU is the problem. Try --device 2 and --device 3 just to make sure, though.

HD5850 again:

-v 2:
Elapsed time: 22.72 sec. (0.04 init + 22.69 sieve) at 1328786 p/sec.
Processor time: 2.73 sec. (0.03 init + 2.70 sieve) at 11170287 p/sec.

-v 3:
Elapsed time: 26.23 sec. (0.04 init + 26.19 sieve) at 1151050 p/sec.
Processor time: 3.06 sec. (0.05 init + 3.01 sieve) at 10012744 p/sec.

-v 4:
Elapsed time: 25.09 sec. (0.04 init + 25.05 sieve) at 1203483 p/sec.
Processor time: 2.96 sec. (0.06 init + 2.90 sieve) at 10389565 p/sec.
____________

OK, a new version, 0.1.0-beta, is out. Major changes you should notice:

- It only runs on GPUs. So you shouldn't need to pass --device anymore unless you have two or more.
- It defaults to -v 2. So you shouldn't need to pass that anymore either.

So this should make it possible to run the app on BOINC PPSE WUs, with an appropriate app_info.xml file. (No, I don't know how to create one.)

Other things you might find interesting:
- -m works now. Default is 7, but you can fiddle to your heart's content. Not that I expect any significant improvements, but I've been wrong before.
- Better error messages. Not that you should notice.

Based on the number of crashes I saw from the previous version, I'm not sure any OpenCL app can come out of beta yet. But hopefully this is as stable as it can be. If you see any major issues, post them here.

Oh, and Scott, I have no answers for you about the 100% CPU usage on the 4650. I can only guess that either some driver is too old or it's a quirky card. I will say that I'm glad I didn't buy one. (I was thinking about it!)
____________

Win7 x64
SDK 2.2
Catalyst 10.7
HD4870

-v2

00:53:22 (3460): Can't set up shared mem: -1. Will run in standalone mode.
Sieve started: 42070000000000 <= p < 42070010000000
Thread 0 starting
Detected 160 multiprocessors (800 SPUs) on device 1.

Thread 0 completed
Sieve complete: 42070000000000 <= p < 42070010000000
count=318533,sum=0xb9f8cbeb13d00db3
Elapsed time: 14.10 sec. (0.05 init + 14.06 sieve) at 727317 p/sec.
Processor time: 1.64 sec. (0.06 init + 1.58 sieve) at 6488672 p/sec.
Average processor utilization: 1.33 (init), 0.11 (sieve)
00:53:36 (3460): called boinc_finish

00:53:57 (3572): Can't set up shared mem: -1. Will run in standalone mode.
Sieve started: 42070000000000 <= p < 42070010000000
Thread 0 starting
Detected 160 multiprocessors (800 SPUs) on device 1.
Error: Building Program (clBuildProgram)
00:53:58 (3572): called boinc_finish

00:54:20 (1656): Can't set up shared mem: -1. Will run in standalone mode.
Sieve started: 42070000000000 <= p < 42070010000000
Thread 0 starting
Detected 160 multiprocessors (800 SPUs) on device 1.

Thread 0 completed
Sieve complete: 42070000000000 <= p < 42070010000000
count=318533,sum=0xb9f8cbeb13d00db3
Elapsed time: 15.62 sec. (0.05 init + 15.57 sieve) at 656588 p/sec.
Processor time: 2.14 sec. (0.06 init + 2.07 sieve) at 4927488 p/sec.
Average processor utilization: 1.33 (init), 0.13 (sieve)
00:54:36 (1656): called boinc_finish

Oh, and Scott, I have no answers for you about the 100% CPU usage on the 4650. I can only guess that either some driver is too old or it's a quirky card. I will say that I'm glad I didn't buy one. (I was thinking about it!)

Thanks, Scott! I now see that makes sense. I now have no idea where I got the 7, but searching for it I ran across this PDF. My "BLOCKSIZE" (which doesn't mean much in OpenCL except that it's a constant) is 128. That happens to be the right number of threads to fill a pair of wavefronts. Apparently four wavefronts avoid latency issues, so it makes sense that even m's would be better.

Now, no offense, Scott, but I'm not going to optimize for your low-end GPU specifically. But if I can get some tests of higher-end GPUs verifying this (and I expect I will), I'll release a version where m defaults to 2 tomorrow.

P.S. Anyone want to try -v 4 with -m 2? I doubt it will help, but like I said, I've been wrong before.
____________

Some test results for 0.1.0-beta on the HD5850:

default:
Elapsed time: 22.83 sec. (0.04 init + 22.80 sieve) at 1322316 p/sec.
Processor time: 2.90 sec. (0.05 init + 2.85 sieve) at 10559889 p/sec.

-m1:
Elapsed time: 68.73 sec. (0.04 init + 68.69 sieve) at 438898 p/sec.
Processor time: 3.20 sec. (0.06 init + 3.14 sieve) at 9614226 p/sec.

-m2:
Elapsed time: 40.08 sec. (0.03 init + 40.05 sieve) at 752812 p/sec.
Processor time: 2.90 sec. (0.05 init + 2.85 sieve) at 10559889 p/sec.

-m2, -v 4:
Elapsed time: 36.81 sec. (0.04 init + 36.78 sieve) at 819732 p/sec.
Processor time: 2.98 sec. (0.05 init + 2.93 sieve) at 10279039 p/sec.

-m3:
Elapsed time: 27.54 sec. (0.03 init + 27.50 sieve) at 1096096 p/sec.
Processor time: 2.68 sec. (0.05 init + 2.64 sieve) at 11434671 p/sec.

-m4:
Elapsed time: 21.34 sec. (0.04 init + 21.29 sieve) at 1415782 p/sec.
Processor time: 2.78 sec. (0.06 init + 2.71 sieve) at 11106090 p/sec.

-m8:
Elapsed time: 20.24 sec. (0.05 init + 20.19 sieve) at 1493280 p/sec.
Processor time: 2.59 sec. (0.05 init + 2.54 sieve) at 11855581 p/sec.

-m12:
Elapsed time: 20.22 sec. (0.04 init + 20.18 sieve) at 1493724 p/sec.
Processor time: 2.95 sec. (0.06 init + 2.89 sieve) at 10445728 p/sec.

-m16:
Elapsed time: 19.83 sec. (0.04 init + 19.80 sieve) at 1522851 p/sec.
Processor time: 2.64 sec. (0.03 init + 2.61 sieve) at 11571616 p/sec.

-m20:
Elapsed time: 19.94 sec. (0.05 init + 19.89 sieve) at 1515501 p/sec.
Processor time: 2.62 sec. (0.03 init + 2.59 sieve) at 11641324 p/sec.

-m24:
Elapsed time: 20.48 sec. (0.04 init + 20.44 sieve) at 1474724 p/sec.
Processor time: 3.04 sec. (0.05 init + 3.00 sieve) at 10064893 p/sec.

-m32:
Elapsed time: 19.71 sec. (0.04 init + 19.67 sieve) at 1532762 p/sec.
Processor time: 3.26 sec. (0.03 init + 3.23 sieve) at 9335555 p/sec.

Also tried m 5, 6, 9, 10 and some numbers above 32, but they all were slower that the default.

@tepek: Your app_info works, but you don't need to include the file_info for pps_sr2sieve_20090322.sieveinput, as ppsieve doesn't need the sievefile.
____________

Now, no offense, Scott, but I'm not going to optimize for your low-end GPU specifically. But if I can get some tests of higher-end GPUs verifying this (and I expect I will), I'll release a version where m defaults to 2 tomorrow.

P.S. Anyone want to try -v 4 with -m 2? I doubt it will help, but like I said, I've been wrong before.

Thanks, Scott! I now see that makes sense. I now have no idea where I got the 7, but searching for it I ran across this PDF. My "BLOCKSIZE" (which doesn't mean much in OpenCL except that it's a constant) is 128. That happens to be the right number of threads to fill a pair of wavefronts. Apparently four wavefronts avoid latency issues, so it makes sense that even m's would be better.

@Elgrande71: A computation error means just that: your GPU found a factor that your CPU couldn't verify. Usually, this means you've overclocked too much!

By the way, for PPSieve, there's not much reason to overclock the memory, if such things are separate in ATI like they are in nVIDIA. You can under-clock the memory, overclock the shaders, and get better performance (as long as you don't overclock too much).

@tepek

<cmdline>-m 16</cmdline>

I wonder if the different -m performance across the cards so far is related to ATI employing different wavefront sizes in different cards? The 58xx/57xx and 48xx use 64, but a 5450 card has wavefront size of 32 (I haven't found the size for the 4650 yet...some cards are also have wavefronts of size 16, but I don't think any of these are OpenCL capable???).

Alright, I've pushed a new version that just changes the default -m to match what CL_KERNEL_WORK_GROUP_SIZE returns. It's so similar that I didn't even bother changing the version number. (But the build date is today.) It also prevents using a -m higher than the default, as I think it returns a maximum.

Since the default is now the maximum possible size, I'd like to see what various cards set -m to on the test run. (It's printed to stdout.) It might make sense to make the default -m some fraction of what it is now. So let me know what your cards produce.

Thanks!
____________

HD 4650 results...Something is not quite right with this one...

OK, that didn't work. I've reverted to the previous build. I'll have to figure out something else later.
____________

finally got time to try this...

HD4670, Ubuntu 10.04 x86_64 (collatz works fine on this)

i have done the ICD registration and ldd looks good
i use this wrapper

Jip, did you complete all of the installation procedure (PDF), including setting up all necessary environment variables?

Vato: I'm afraid I don't have any personal experience with cards being found. (Since I don't have one. :P)
____________

check echo $ATISTREAMSDKROOT
____________

I think you need the complete SDK install because the OpenCL code is distributed as source code. The OpenCL SDK has to compile it, every time, before it can run!

I am aware that the current SDK allows compiling in binary OpenCL code; however, the app is currently set up to "bake in" certain constants.
____________

OK, I've pushed a new version that will report the compile error. Let's see what's wrong.
____________

Have now put 32bit and 64bit SDK libs in /lib32 and /lib64 respectively, re-run ldconfig, downloaded latest ppsieve-cl.zip, tried boinc and non-boinc and 32bit and 64bit executables - and still haven't got very far. Running under strace just shows lots of mmap() calls succeeding prior to calling clone() and futex(). Still HD4670 Ubuntu 10.04 x86_64 - any further hints gratefully received.