[Hardware] OGR-27 OpenCL core write attempt

Konrad Rzepecki hannibal at astral.lodz.pl
Thu Aug 16 11:21:51 EDT 2012


Dnia środa, 15 sierpnia 2012 15:56:44 Frédéric Bastien pisze:
> On Wed, Aug 15, 2012 at 3:49 PM, Konrad Rzepecki
> 
> <hannibal at astral.lodz.pl> wrote:
> > 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.
> 
> In CUDA, there is a constant memory, can you store it in something
> equivalent in OpenCL? What is the size of the table?

They are, but only 64kB - pchoose table have 4MB.

I made 64kB table*, but it use only 10 bits instead of 16 in original one.
This probably increase amount of counting. The question is how big will be 
that increase - know somebody?

I can use global memory, but according some AMD docs, access to global memory 
is 25 times slower than local memory. So, accessing global memory, is rather 
no option, particularly it is accessed very often.


* In each depth, every 64 values, I count max of them, and save. Is this 
correct?

-- 
  Konrad Rzepecki (Hannibal)


More information about the Hardware mailing list