English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Sparse convolution plays a pivotal role in emerging workloads, includingpoint cloud processing in AR/VR, autonomous driving, and graph understanding inrecommendation systems. Since the computation pattern is sparse and irregular,specialized high-performance kernels are required. Existing GPU libraries offertwo dataflow types for sparse convolution. The gather-GEMM-scatter dataflow iseasy to implement but not optimal in performance, while the dataflows withoverlapped computation and memory access (e.g.implicit GEMM) are highlyperformant but have very high engineering costs. In this paper, we introduceTorchSparse++, a new GPU library that achieves the best of both worlds. Wecreate a highly efficient Sparse Kernel Generator that generates performantsparse convolution kernels at less than one-tenth of the engineering cost ofthe current state-of-the-art system. On top of this, we design the SparseAutotuner, which extends the design space of existing sparse convolutionlibraries and searches for the best dataflow configurations for training andinference workloads. Consequently, TorchSparse++ achieves 2.9x, 3.3x, 2.2x and1.7x measured end-to-end speedup on an NVIDIA A100 GPU over state-of-the-artMinkowskiEngine, SpConv 1.2, TorchSparse and SpConv v2 in inference; and is1.2-1.3x faster than SpConv v2 in mixed precision training across sevenrepresentative autonomous driving benchmarks. It also seamlessly supports graphconvolutions, achieving 2.6-7.6x faster inference speed compared withstate-of-the-art graph deep learning libraries.

TorchSparse++: Efficient Training and Inference Framework for Sparse Convolution on GPUs