FPGA Codec System of Learned Image Compression with Algorithm-Architecture Co-Optimization

Learned Image Compression (LIC) has shown a coding ability competitive to traditional standards. To address the complexity issue of LIC, various hardware accelerators are required. As one category of accelerators, FPGA has been used because of its good reconfigurability and high power efficiency. However, the prior work developed the algorithm of LIC neural network at first, and then proposed an associated FPGA hardware. This separate manner of algorithm and architecture development can easily cause a layout problem such as routing congestion when the hardware utilization is high. To mitigate this problem, this paper gives an algorithm-architecture co-optimization of LIC. We first restrict the input and output channel parallelism with some constraints to ease the routing issue with more DSP usage. After that, we adjust the numbers of channels to increase the DSP efficiency. As a result, compared with one recent work with a fine-grained pipelined architecture, we can reach up to 1.5x faster throughput with almost the same coding performance on the Kodak dataset. Compared with another recent work accelerated by Xilinx DPU, we can reach faster throughput with better coding performance.