A kernel-independent, pipelined architecture for real-time 2-D convolution

A flexible, fault-tolerant architecture to perform 2-D convolution of images for kernels of arbitrary shapes and sizes is proposed. The convolved image is output online as the input image is generated and received, providing optimal turnaround time. The regularity of the computation and the raster-scan input of pixels are used to pipeline computations as well as memory operations. Throughput of (kl/ u) is achieved for a variety of kernel sizes with a k×l array of multiply-accumulate cells, where u is the time for a single fixed-point multiply-accumulate operation. It is shown that by queuing accesses at memory modules the limitation placed by access time on the throughput can be removed. Larger, as well as smaller kernels can be accommodated on a fixed-size array with a single pass of the image and a minor change in the computation flow. The host enjoys absence of any control overhead and only has to supply the pixels and collect the output pixels in simple raster-scan fashion for any kernel size