Understanding output extend of convolution kernels when blending