I am happy to report that we successfully tried an NVIDIA OpenCL test version of our AQUA app on volunteer computers.
There is no way to know for sure whether people who ran the work units on their computers had OpenCL SDKs or not, but we didn't see any failures, which means that every user with the appropriate NVIDIA driver version (197.13+) who downloaded the OpenCL app ran it successfully to completion. We have an ATI OpenCL test version of the app too. We can try it ASA the appropriate application plan is functional. -Kamran -----Original Message----- From: David Anderson [mailto:[email protected]] Sent: Wednesday, May 12, 2010 8:09 PM To: Kamran Karimi Cc: [email protected] Subject: Re: [boinc_dev] OpenCL support? I added plan classes cuda_opencl (misnomer, sorry) and ati_opencl. The former requires driver version 197.13+; the latter is a stub that never matches. -- David Kamran Karimi wrote: > Thanks for the clarification. > > I would appreciate it if you could do the app_plan() modifications to > make sure opencl_nvidia is available to everybody for testing. > > -Kamran > > > -----Original Message----- > From: David Anderson [mailto:[email protected]] > Sent: Wednesday, May 12, 2010 3:22 PM > To: Kamran Karimi > Cc: [email protected] > Subject: Re: [boinc_dev] OpenCL support? > > Kamran Karimi wrote: >> We also have two versions of our aqua app for NVIDIA and ATI. They run >> when invoked manually, and accept a "--device" argument to determine >> which GPU to use (as provided to cuda apps before, but I can't find > this >> argument in the sched_customize.cpp file any more). > > It's not supplied by the scheduler. > It's supplied by the client, after it decides which GPU to use. > >> We can try the apps >> under BOINC when it can detect and report OpenCL's presence. > > You can try them right now; > just define an opencl_nvidia plan class, > and add a clause to app_plan() in which you check for > driver version 197.13 or later. > I can add this if you want. > > ATI: we need to wait until ATI includes OpenCL support in their driver. > Asking volunteers to install an SDK, > and having the client check for it, is not viable. > > -- David > _______________________________________________ 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.
