A Performance Comparison of Sort and Scan Libraries for GPUs

TL;DR

This paper benchmarks seven GPU libraries for sorting and scanning, revealing significant performance variation and highlighting the lack of a single library that is both fast and portable.

Contribution

It provides a comprehensive performance comparison of multiple GPU libraries for fundamental primitives, highlighting their strengths and weaknesses.

Findings

Large performance variation among libraries

No single library offers both optimal performance and portability

Benchmark results guide library selection for GPU applications

Abstract

Sorting and scanning are two fundamental primitives for constructing highly parallel algorithms. A number of libraries now provide implementations of these primitives for GPUs, but there is relatively little information about the performance of these implementations. We benchmark seven libraries for 32-bit integer scan and sort, and sorting 32-bit values by 32-bit integer keys. We show that there is a large variation in performance between the libraries, and that no one library has both optimal performance and portability.