The sliding window acts as a cache for the partial sums of the
convolution operation.  For image magnification, this results in
a significant performance boost, especially for interpolation
with large kernels.