And if tomorrow some new brand will develop? What then? For example, in near future Intel will provide OpenCL support for their CPU+GPU chips. That is, new GPU vendor... IMHO, this support should be generalized to not require such deep hardcoding in BOINC. Abstract "compute device" model should be developed instead. Such device can have some general and maybe some device-specific properties and being described in separate text-based resource file (some XML file maybe). BOINC client could read such resource just as it reads app_info per project now and setup separate scheduling for each different "compute device" it found in list. This would allow to easely extend list of devices w/o big additional efferts from BOINC dev side and project dev side...
----- Original Message ----- From: David Anderson To: Raistmer Cc: [email protected] Sent: Wednesday, December 15, 2010 1:19 AM Subject: Re: [boinc_dev] Plan classes for AMD GPUs BOINC can currently handle only 1 type of GPU per brand per host; i.e. it can handle 1 type of NVIDIA GPU and 1 type of ATI GPU. Removing this restriction if possible but would be a medium-large development project. -- David On 14-Dec-2010 2:15 PM, Raistmer wrote: > And what about anonymous platform? > How can I define that one GPU requires one type of app and another GPU > requires > another app type? Project can have no GPU apps at all in general, but there > can > be third-party GPU apps... > ----- Original Message ----- > *From:* David Anderson <mailto:[email protected]> > *To:* [email protected] <mailto:[email protected]> > *Sent:* Wednesday, December 15, 2010 12:07 AM > *Subject:* Re: [boinc_dev] Plan classes for AMD GPUs > > The plan class mechanism lets projects define their own plan classes, > and they can use these to target app versions > to specific GPU characteristics if they like. > They don't need BOINC to define these plan classes for them. > -- David > > On 11-Dec-2010 2:50 PM, Raistmer wrote: > >> From Lunatics beta thread: > > " Quote from: cenit on Today at 00:35:57 Put for example that I have a HD > > 4870 and a HD 5870 in the same host, with a unified app it would be easy to > > get full performances (with both app, I would have to use the non-HD5 for > > both!) > > > > Devices with different capabilities should have different plan classes > first > > of all. And if there are different plan classes I see no problem to use few > > different app each fine tuned for particular device. > > > > " > > > > Could we expect such different plan classes for OpenCL apps for ATi HD4xxx > > and HD5xxx that drastically differ in capabilities (no scratchpad local > > memory on HD4xxx at all) ? _______________________________________________ > > boinc_dev mailing list [email protected] > <mailto:[email protected]> > > http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, > > visit the above URL and (near bottom of page) enter your email address. > _______________________________________________ > boinc_dev mailing list > [email protected] <mailto:[email protected]> > http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev > To unsubscribe, visit the above URL and > (near bottom of page) enter your email address. _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
