Towards real-time LiDAR processing on RISC-V-based ASIPs: fast trigonometric approximations via parabolic synthesis

The real-time processing of LiDAR data is a pivotal aspect of numerous autonomous driving and robotics operations. However, the sheer amount of data requiring computationally complex calculations presents a significant challenge for embedded processors operating under tight real-time constraints. So far, embedded processors were mainly used for the retrieval of LiDAR data, offloading computationally demanding tasks to a coprocessor such as a GPU or a specialized vector processor. This work focuses on processing LiDAR data directly on an embedded, low-power, 3-stage RISC-V (RV32IMFCB) core. To study LiDAR performance, the segmentation stage of LeGO-LOAM - a lightweight and ground-optimized LiDAR odometry and mapping framework - was implemented and extended as a reference algorithm. It consists of ground-detection as well as clustering based on raw data from a Velodyne LiDAR. Profiling revealed that most of the runtime is spent computing sine and arc tangent functions. A special function unit based on parabolic synthesis has been developed, capable of computing sine and arc tangent with more than 16-bit accuracy in two clock cycles. Its implementation is described and evaluated against alternative optimizations, such as a fixed-point software approximation and a CORDIC based unit for trigonometric function approximation. The approaches were executed and verified using RTL simulation. Additionally, all variants of the core have been synthesized in a 22nm FD-SOI technology. Best performance is reached using parabolic synthesis, which improves runtime as well as power efficiency by more than 3× with a slight 5% increase in area compared to the baseline model. Finally, the results demonstrate that the real-time goal of 20 FPS for autonomous driving applications is achievable, enabling customized, low-power RISC-V cores to be applied in LiDAR processing pipelines.