AMD E-450 APU on Win7 64bit with Boinc 7.0.28 OpenCL problems/questions...
25 Oct 2012 19:41:14 UTC
Topic 196591
(moderation:
Hello,
I am new to boinc/einstein. Before saying anything about how low-end my processor is. I am just trying to test the OpenCL code of einstein, so I can live with it... I have installed the boinc client 7.0.28 to my laptop and I see this on the computer page:
Coprocessors CAL AMD Radeon HD 6200/6300/7300 series (Wrestler) (384MB) driver: 1.4.1741
It is strange since "clinfo" returns
Max memory allocation: 176160768
Global memory size: 402653184
Problem 1:
I am not sure from where that 384MB is coming from?
Problem 2:
It looks like einstein is only running on CPU and not on GPU. I would like to get it to use GPU only and not CPU (not at all!). How can I make it do that?
Problem 3:
I realized that the source code of einstein is using libclfft (from apple?) while AMD has clAmdFFT library freely available. Why? I would imagine clAmdFFT would perform much better?
Thanks,
Evren
AMD E-450 APU on Win7 64bit with Boinc 7.0.28 OpenCL problems/qu
)
Well I created cc_config.xml in C:\ProgramData\BOINC containing:
0
and restarted boinc manager. Einstein is still using CPU :(
RE: Problem 2: It looks
)
Go to Your Account, then on to Einstein@Home preferences, then edit these preferences and:
Uncheck "Use CPU"
Check "Use ATI GPU"
Uncheck "Use Nvidia GPU"
At "Run only the selected applications" check at least "Binary Radio Pulsar Search (Arecibo)" (since these run on the ATI GPU)
Uncheck "If no work for selected applications is available, accept work from other applications?"
Uncheck "Run CPU versions of applications for which GPU versions are available"
Save preferences with the Update preferences button at the bottom of the page.
Next open BOINC Manager.
Advanced view.
Projects tab.
Select (hi-light) Einstein@Home.
Click Update.
That will tell Einstein to only send GPU tasks.
Mind, the OpenCL GPU application here at Einstein works best when it's allowed to use a full CPU core. So you best set BOINC to use all but one core.
Also mind that every GPU application will run on the CPU. So you can't totally disable the use of the CPU for the GPU app.
Rather than aborting all your
)
Rather than aborting all your CPU Wu's, or using the cc_config.xm option, please go the Einstein Project preferences page and untick 'Use CPU', making sure 'Use ATI GPU' is ticked,
Claggy
RE: Hello, I am new to
)
It is 402653184 /1024 / 1024, so here 1MB = 1024*1024 bytes.
To allow us to use a library, it is not enough that it is "free" in the sense of "volunteers need not pay for it", the libraries license also has to be compatible with the license terms of the other source code and libs used in the app, or compatible license terms have to be negotiated specifically for E@H. Plus it should ideally be open source and we need to be able to build the app with it using cross compilation on Linux for Windows because we prefer it that way ;-). Also the FFT lib should have an interface similar to that of FFTW, our reference solution for the CPU source code that all other source code is derived from. All this pretty much narrowed down to using the Apple lib.
Cheers
HB
Yes you are right, I should
)
Yes you are right, I should get glasses :) I read 4GB instead of 400MB... but that shouldnt be a problem.
I tried to change settings in my account preferences, tried to erase project and re-add and it did not help. Boinc was insistently only downloading CPU program only. I had to manually download einsteinbinary_BRP4_1.28_windows_x86_64__opencl-ati.exe from http://albert.phys.uwm.edu/EinsteinAtHome/download/ and make app_info.xml
It has been apparently running good so far...
einsteinbinary_BRP4
Binary Radio Pulsar Search
einsteinbinary_BRP4_1.28_windows_x86_64__opencl-ati.exe
einsteinbinary_BRP4_1.00_graphics_windows_intelx86.exe
einsteinbinary_BRP4
128
opencl_ati_101
0.000000
0.000000
3500000000.000000
ATI
1.000000
einsteinbinary_BRP4_1.28_windows_x86_64__opencl-ati.exe
einsteinbinary_BRP4_1.00_graphics_windows_intelx86.exe
graphics_app
About AMD's FFT libraries. E@H is surely using cuFFT which is similar in licensing. This would be applying double-standard if the reason is AMD's FFT library is not open source.
AMD's FFT library is available for both Linux and Windows and re-distributable (as far as I can tell from the license). At least for windows and linux platforms it should be possible to make it available. If there is some sort of incompatibility in licensing, can you please point out the exact reason?
Since OpenCL version is largely targeted for AMD GPUs, it probably would make sense to use a library which is specifically optimized for it.
Assuming probably the license is not a problem. The question is, if somebody implements possibility of using AMD's FFT library, and if it gives significantly better performance, can E@H accept the code for general usage?
Thanks,
Evren