One-Dimensional Systolic Arrays for Multidimensional Convolution and Resampling

We present one-dimensional systolic arrays for performing two- or higher-dimensional convolution and resampling. These one-dimensional arrays are characterized by the fact that their I/0 bandwidth requirement is independent of the size of the convolution kernel. This contrasts with alternate two-dimensional array solutions, for which the I/0 bandwidth must increase as the kernel size increases. The proposed architecture is ideal for VLSI implementation—an arbitrarily large kernel can be handled by simply extending the linear systolic array with simple processors of the same type, so that one processor corresponds to each kernel element.