76.5-Gb/s Viterbi Decoder for Convolutional Codes on GPU

This letter presents an optimized Viterbi decoder of convolutional codes on graphics processing unit (GPU) for software defined radio (SDR) platforms. Before the forward process, channel messages are interleaved with coalesced global memory access and the interleaved messages are represented with 4 bits to improve shared memory efficiency. Moreover, we optimize on-chip memory allocations of the forward process to accelerate instruction execution. Excluding the data transfer latency between host and device, the proposed Viterbi decoder achieves 22.2 and 76.5-Gb/s throughput on Tesla V100 and RTX4090, respectively. Compared with related works, the throughput speedups achieved by the proposed decoder are from 2.06× to 2.93×.