Archives Discussions

[Gipsel]

I've just read that there is some interest in ATI support for GPGPU computing on the BOINC distributed computing platform. Currently only nvidias CUDA is supported (it looks they completely sponsor the GPUGrid project, formerly known as PS3Grid), but this may change if they get some help putting the ATI GPU detection for the client together.

As ATI cards have a massive advantage at double precision calculations, it could be an opportunity to conquer that niche of enthusiasts spending lots of money just to build one or even more machines for numbercrunching on some distributed computing projects.

[Gipsel]

but this may change if they get some help putting the ATI GPU detection for the client together.

I guess all one has to do is to hand the CALdeviceattribs for every CAL supported device to the client. That should be all they need.

[udeepta]

It has to be a different (likely OS dependent) call to figure out what GPUs you currently have in a system. You cannot depend on CAL since systems that do not have ATI GPUs will not have CAL libs installed.

Originally posted by: Gipsel

but this may change if they get some help putting the ATI GPU detection for the client together.

I guess all one has to do is to hand the CALdeviceattribs for every CAL supported device to the client. That should be all they need.

[Gipsel]

Originally posted by: udeepta@amd It has to be a different (likely OS dependent) call to figure out what GPUs you currently have in a system. You cannot depend on CAL since systems that do not have ATI GPUs will not have CAL libs installed.

That should be the first check of course. If the CAL libs are not installed, there is no CAL support, it does not matter which cards are installed then. But after that one can use the CAL API to determine the capabilities.

Edit: If you look here you can see how it is done with CUDA devices. If you want ATI support for the BOINC platform, just send them a working piece of code to be included there.

[udeepta]

Got it. Looks like an ATI equivalent of string COPROC_CUDA::get(COPROCS&coprocs) needs to be created.