Efficient mapping on FPGA of convolution computation based on combined CSA-CPA accumulator

Abstract

In this paper we present some architectures to deal with fast convolution computation based on carry save adders which are intended to be specifically implemented on FPGAs. Carry-save adders are not frequent in FPGA implementations since FPGA has a fast carry propagation path. In this paper we prove that it is possible to use carry-save arithmetic in a efficient way on FPGA for convolution operation. We make use of the specific structure of the FPGA to design an optimized accumulator which is able to deal with carry-save additions as well as carry-propagate additions using the same hardware. This lead to an efficient combined CSA-CPA architecture with fast computation and optimizing the hardware cost. Experimental results for different word lengths are presented to validate our proposal.