As opposed to having an OpenCL program object for all existing kernels and doing the compiling for all of them from the start, evaluate the alternative of creating an OpenCL program for one single kernel every time that kernel is needed.
Evaluate how it impacts the overall application performance considering:
the overhead for creating the OpenCL program
how long clBuildProgram takes (for both AMD and NVIDIA GPUs)
If there is no performance penalty for the host side, implement this model with 1 OpenCL program/kernel.
As opposed to having an OpenCL program object for all existing kernels and doing the compiling for all of them from the start, evaluate the alternative of creating an OpenCL program for one single kernel every time that kernel is needed.
Evaluate how it impacts the overall application performance considering:
If there is no performance penalty for the host side, implement this model with 1 OpenCL program/kernel.