Coalesced memory access to global memory; CUDA Vs Opencl

Question

Are there different set of requirements for coalesced access for CUDA and OpenCL? I think it has to be hardware feature,so the API should really not matter. Any pointers??

Answer 1

That's correct. CUDA iteself doesn't have any requirements for coalesced access; the hardware does, and the hardware has the same requirements regardless of whether you're programming it in CUDA, OpenCL, DirectCompute, OpenGL, or something else. On CPU hardware the same could be said about cache behaviour; cache line sizes and cache behaviour don't care if you're using C, Java, or Python.