[Hardware] OGR-27 OpenCL core write attempt

Konrad Rzepecki hannibal at astral.lodz.pl
Wed Aug 15 15:49:52 EDT 2012


Dnia środa, 15 sierpnia 2012 10:15:01 Jeff Lawson pisze:

> Unfortunately, I don't think it will be possible for you to eliminate
> the choose data tables since they are intrinsic to the OGR algorithm
> and the parameters that we have selected for the project.

Too bad. So I must look deeply how it is generated and make fast way to make 
it accessible to OpenCL kernel.

> What performance are you currently seeing on your OpenCL core relative
> to the CPU core for OGR?

I'm far far away from see any benchmarks. For now I have only this kernel. But 
I don't expect high computing rate since algorithm need lot of memory and is 
hard to scale it to multicore processing.

I estimate for my AMD HD6670 - it has 480 800MHz stream processors which give 
it  6 compute units. It have 32kb of local memory, my kernel use ~1kb of it. 
So at max each compute unit can run up to 32 stubs on same time, but safe 
estimate will be 16 on same time. Assuming computing speed of single work unit 
as "real 800MHz CPU" it will have ~ 3 Mnodes/s.

All of it give 6 computing units x 16 work units x 3 Mnodes/s = 288 Mnodes/s. 
Nothing special but still it is two times more than my AMD FX-4100 CPU. 

-- 
  Konrad Rzepecki (Hannibal)


More information about the Hardware mailing list