I have an A10-6800k APU and found that the memory space is partitioned into CPU and GPU space. Although APU has shared memory, if I transfer the openCL data to the GPU space, the kernels are much faster. I'm wondering whether this is still the case for new versions of APUs.
Though same physical memory may be shared but still the memory is separated for CPU and GPU. Size of GPU memory can be set with BIOS. With newer APUs having HSA support (like Kaveri), HUMA and HSA enhancements allows the CPU and GPU cores to dynamically share the same virtual memory address space. I guess, the new SVM features in OpenCL 2.0 will allow the same benefit on these APU.