It seems intuitive that when the kernel depends on image content, runtimes will be nearly doubled since the algorithm must make a pass over the data to compute the kernel and then another pass to apply the kernel, but what's interesting is that through parallel processing optimizations, kernel portions can be computed in advance and applied in tiles to speed up computation.
It seems intuitive that when the kernel depends on image content, runtimes will be nearly doubled since the algorithm must make a pass over the data to compute the kernel and then another pass to apply the kernel, but what's interesting is that through parallel processing optimizations, kernel portions can be computed in advance and applied in tiles to speed up computation.