TorchSparse++: Efficient Point Cloud Engine-猫眼课题宝

DOI：

10.1109/cvprw59228.2023.00025

发表时间：

2023-06

期刊：

2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)

影响因子：

通讯作者：

中科院分区：

其他

文献类型：

作者：

Haotian Tang;Shang Yang;Zhijian Liu;Ke Hong;Zhongming Yu;Xiuyu Li;Guohao Dai;Yu Wang;Song Han

点云计算已成为自动驾驶和其他应用程序中越来越重要的工作负载。与密集的二维计算不同，点云卷积具有稀疏和不规则的计算模式，因此需要具有专门的高性能内核的专用推理系统支持。虽然现有的点云深度学习库已经为点云上的卷积开发了不同的数据流，但它们在整个模型的执行过程中假设单个数据流。在这项工作中，我们系统地分析和改进现有的数据流。我们的最终系统 TorchSparse++ 在 NVIDIA A100 GPU 上的推理端对端加速比最先进的 MinkowskiEngine、SpConv 1.2、TorchSparse 和 SpConv v2 分别实现了 2.9×、3.3×、2.2× 和 1.8×。此外，TorchSparse++ 是迄今为止唯一支持自动驾驶中 3D 分割、检测和重建工作负载所有必需原语的系统。代码公开发布于 https://github.com/mit-han-lab/torchsparse。

Point cloud computation has become an increasingly more important workload for autonomous driving and other applications. Unlike dense 2D computation, point cloud convolution has sparse and irregular computation patterns and thus requires dedicated inference system support with specialized high-performance kernels. While existing point cloud deep learning libraries have developed different dataflows for convolution on point clouds, they assume a single dataflow throughout the execution of the entire model. In this work, we systematically analyze and improve existing dataflows. Our resulting system, TorchSparse++, achieves 2.9×, 3.3×, 2.2× and 1.8× measured end-to-end speedup on an NVIDIA A100 GPU over the state-of-the-art MinkowskiEngine, SpConv 1.2, TorchSparse and SpConv v2 in inference respectively. Furthermore, TorchSparse++ is the only system to date that supports all necessary primitives for 3D segmentation, detection, and reconstruction workloads in autonomous driving. Code is publicly released at https://github.com/mit-han-lab/torchsparse.

TorchSparse++: Efficient Point Cloud Engine

文献摘要

相似文献