In our research work, we address the optimization issues with TVM support on RISC-V with P extension for subword SIMD. Our work supports QNN flow and fixed-point flow for RISC-V P extension, and provides the tuner for binary point position selection for fixed-point type. As our prior work is more focused on the support with TVM compute OP, this work focuses on the TVM support and optimization issues with hybrid OP on RISC-V P extensions.
We use the post processing operator, NMS(non-maximum suppression), as a case study to illustrate the flow to optimize hybrid OP in the TVM on RISC-V with P extension. It includes the re-writing guideline with hybrid script programs, customized instructions, the better utilization of hybrid-script flow, and llvm compiler optimizations. According to our preliminary experimental result, the instruction counts of NMS on RISC-V P extension is reduced by 4 times with the support of our optimization flow.