Variable Bit-Precision Vector Extension for RISC-V Based Processors

Neural Network model execution is becoming an increasingly compute intensive task. With advances in optimisation techniques such as using lower-bit width precision, need for quantization and model compression, we need to find efficient ways of implementing these techniques. Most Instruction Set Architectures(ISA) do not support low bit-width vector instructions. In this work, we present an extension for the vector specification of the RISC-V ISA, which is targeted towards supporting the lower bit-widths or variable precision (1 to 16 bits) Multiply and Accumulate (MAC) operations. We demonstrate our proposed ISA extension by integrating it with a RISC-V processor named PicoRV32, which is considered as the baseline processor in the proposed work. We introduce the feature of bit-serial multiplication along with variable bit precision support to demonstrate the advantage over a 16 bit baseline processor model. We also build an assembler for the proposed instructions for easier integration into the testbench of the RTL model. We implement the processor on to a Xilinx Zynq based FPGA. We observe that, compared to the baseline RISC-V Vector processor which only supports 8, 16 and 32-bit vector instructions, our processor with variable precision support (1 to16 bits) performs 1.14x faster on an average on a matrix multiplication test program. The proposed processor architecture reduces the memory footprint by up to 1.88x as compared with a baseline 16-bit vector processor.