I would like to use texture compression in OpenCL.
Also, to make an effective usage for APUs, you should include a method to get/set pixels from CPU-C++ host side ( so we can skip to duplicate the data in VRAM <-> RAM ).
Let me explain a bit more: Imagine I load a DXT5 image in an APU. That texture could be accessed from OpenCL/GLSL/HLSL ... but also from C/C++. So you should include an extra lib in your SDK to get/set individual pixels on the fly.