So this is more of a general question. I'm doing some image processing on a webcam feed, so speed is crucial. I've been using Images, as pixel-wise access in kernels is faster than for buffers, but Images have the added overhead of needing to execute two copy commands to transfer the data to/from memory. Another way to do this would be to use mapped buffers - this would take a speed hit on the kernel side, but removes the need for the two memory transfers. Has anybody experienced/played with this tradeoff? Any advice?

Thanks in advance,
Spencer