Hardware acceleration for RLNC: A case study based on the xtensa processor with the tensilica instruction-set extension

Random linear network coding (RLNC) can greatly aid data transmission in lossy wireless networks. However, RLNC requires computationally complex matrix multiplications and inversions in finite fields (Galois fields). These computations are highly demanding for energy-constrained mobile devices. The presented case study evaluates hardware acceleration strategies for RLNC in the context of the Tensilica Xtensa LX5 processor with the tensilica instruction set extension (TIE). More specifically, we develop TIEs for multiply-accumulate (MAC) operations for accelerating matrix multiplications in Galois fields, single instruction multiple data (SIMD) instructions operating on consecutive memory locations, as well as the flexible-length instruction extension (FLIX). We evaluate the number of clock cycles required for RLNC encoding and decoding without and with the MAC, SIMD, and FLIX acceleration strategies. We also evaluate the RLNC encoding and decoding throughput and energy consumption for a range of RLNC generation and code word sizes. We find that for GF(2⁸ ) and GF(2¹⁶ ) RLNC encoding, the SIMD and FLIX acceleration strategies achieve speedups of approximately four hundred fold compared to a benchmark C code implementation without TIE. We also find that the unicore Xtensa LX5 with SIMD has seven to thirty times higher RLNC encoding and decoding throughput than the state-of-the-art ODROID XU3 system-on-a-chip (SoC) operating with a single core; the Xtensa LX5 with FLIX, in turn, increases the throughput by roughly 25% compared to utilizing only SIMD. Furthermore, the Xtensa LX5 with FLIX consumes roughly four orders of magnitude less energy than the ODROID XU3 SoC.