FPGA-based systolic deconvolution architecture for upsampling

Abstract

A deconvolution accelerator is proposed to upsample n × n input to 2n × 2n output by convolving with a k × k kernel. Its architecture avoids the need for insertion and padding of zeros and thus eliminates the redundant computations to achieve high resource efficiency with reduced number of multipliers and adders. The architecture is systolic and governed by a reference clock, enabling the sequential placement of the module to represent a pipelined decoder framework. The proposed accelerator is implemented on a Xilinx XC7Z020 platform, and achieves a performance of 3.641 giga operations per second (GOPS) with resource efficiency of 0.135 GOPS/DSP for upsampling 32 × 32 input to 256 × 256 output using a 3 × 3 kernel at 200 MHz. Furthermore, its high peak signal to noise ratio of almost 80 dB illustrates that the upsampled outputs of the bit truncated accelerator are comparable to IEEE double precision results.

Keywords: Deep learning; FPGA; Transposed convolution; Upsample.