Archives Discussions

[thomasp]

Hi

I plan to code something in OpenCL, using überKernel pattern.

It means that a given kernel would have this structure:

__kernel void my_uber_kernel(void)

{

     while(...)

     {

          if(stage==..)

          {

               device_function_0() ;

          } else

          if(stage==...)

          {

               device_function_1() ;

          }

          // etc...

          stage = stage + 1 ;

     }

}

Each one of device_function_X() potentially contains a substantial amount of code.

I'm wondering if there is known limitations regarding the amount of instructions supported (per thread?) before performances are impacted ?

Does splitting process in small device functions calls help to optimize ?

Or do I have to split process in several kernel calls (so that above-mentioned device_function_X become kernels)