Thanks for sharing. I don't see why there is an atomic add in the kernel there. It doesn't look like two separate threads should be able to modify the same pixel, based on the block/thread indices?

Each block is for a different image, but each thread within the same block has a small chance of modifying the same pixel.