TorchSparse: Efficient Point Cloud Inference Engine

• DOI: 10.48550/arxiv.2204.10319
• Publication Date: 2022-01-01
• Authors: Haotian Tang; Zhijian Liu; Xiuyu Li; Yujun Lin; Song Han
• Keywords: Speedup; Benchmark (surveying); Convolution (computer science); FLOPS

Deep learning on point clouds has received increased attention thanks to its wide applications in AR/VR and autonomous driving. These require low latency high accuracy provide real-time user experience ensure safety. Unlike conventional dense workloads, the sparse irregular nature of poses severe challenges running CNNs efficiently general-purpose hardware. Furthermore, existing acceleration techniques for 2D images do not translate 3D clouds. In this paper, we introduce TorchSparse, a high-performance cloud inference engine that accelerates convolution computation GPUs. TorchSparse directly optimizes two bottlenecks convolution: data movement. It applies adaptive matrix multiplication grouping trade better regularity, achieving 1.4-1.5x speedup multiplication. also movement by adopting vectorized, quantized fused locality-aware memory access, reducing cost 2.7x. Evaluated seven representative models across three benchmark datasets, achieves 1.6x 1.5x measured end-to-end over state-of-the-art MinkowskiEngine SpConv, respectively.