A GPU implementation of belief propagation decoder for polar codes

We present a Graphics Processing Unit (GPU) implementation of a Belief Propagation (BP) based decoder for polar codes. The BP decoding algorithm is implemented to utilize the parallel computing capability of the GPUs. We show how the algorithm can make use of parallelism both at the thread level and block level, and by utilizing the limited shared memory available on GPUs, a real time decoding performance is achieved. The resulting algorithm is able to achieve a decoding throughput of almost 5Mbps while maintaining a frame error rate below 10^-3 on code blocks of 1024 bits.