Archives Discussions

[erman_amd]

Hi,

I want to measure my kernel performance (kernel execution time) vs. number of compute unit. I have 5870 card (20 cores). How to use only 10 of these compute units to execute the kernel?. 

I have read it months ago in this forum, but I can not find it. It's like I have to set an environment variable or some setting in Visual Studio, Windows?

Thanks

[nou]

well eac workgroup is executing on one compute unit. so you can execute only so many worgroups as much you want use a compute units.

in theory you should see a linear speedup. question is what will happen if kernel is memory bound.

[himanshu_gautam]

yeah sure.

I think you can also use device fission extension. It should be easier this way.

EDIT: Thanks for the correction.

[LeeHowes]

Device fission doesn't work on the GPU. The hardware doesn't maintain separate work queues for different cores. Nou's suggestion is the only practical way at the moment, I think.