Optimizing the Semantic Segmentation CNN SalsaNext on a Vertical Vector Processor

The vertical vector architecture V<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>PRO is a potential candidate for deployment in an advanced driver assistance system. A typical use case in this field is the semantic segmentation of lidar point clouds. The convolutional neural network SalsaNext is used to investigate the performance of semantic segmentation on a vertical vector architecture and assess whether the applied optimizations layer fusion and kernel tuning help to leverage the architectural features for performance improvements.It is shown that the optimizations reduce the inference time by 23.1% compared to the baseline, resulting in a final runtime of 468 ms. The optimized efficiency, measured in frames per second normalized to frequency and number of processing elements, is 11.9% above the Nvidia references. The final MAC utilization of 65% outperforms the references by 6.7 percentage points. The results show that semantic segmentation can be executed efficiently on the V<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>PRO and that the optimizations are effective on a vertical vector architecture.