Superscalar RISC-V Processor with SIMD Vector Extension

With the increasing number of digital products in the market, the need for robust and highly configurable processors rises. The demand is convened by the stable and extensible open-sourced RISC-V instruction set architecture. RISC-V processors are becoming popular in many fields of applications and research. This thesis presents a dual-issue superscalar RISC-V processor design with dynamic execution. The proposed design employs the global sharing scheme for branch prediction and Tomasulo algorithm for out-of-order execution. The processor is capable of speculative execution with five checkpoints. Data flow in the instruction dispatch and commit stages is optimized to achieve higher instruction throughput. The superscalar processor is extended with a customized vector instruction set of single-instruction-multiple-data computations to specifically improve the performance on machine learning tasks. According to the definition of the proposed vector instruction set, the scratchpad memory and element-wise arithmetic units are implemented in the vector co-processor. Different test programs are evaluated on the fully-tested superscalar processor. Compared to the reference work, the proposed design improves 18.9% on average instruction throughput and 4.92% on average prediction hit rate, with 16.9% higher operating clock frequency synthesized on the Intel Arria 10 FPGA board. The forward propagation of a convolution neural network model is evaluated by the standalone superscalar processor and the integration of the vector co-processor. The vector program with software-level optimizations achieves 9.53× improvement on instruction throughput and 10.18× improvement on real-time throughput. Moreover, the integration also provides 2.22× energy efficiency compared with the superscalar processor along