Towards Composable GPU Programming

Abstract

In this paper, we advocate a composable approach to programming systems with Graphics Processing Units (GPU): programs are developed as compositions of generic, reusable patterns. Current GPU programming approaches either rely on low-level, monolithic code without patterns (CUDA and OpenCL), which achieves high performance at the cost of cumbersome and error-prone programming, or they improve the programmability by using pattern-based abstractions (e.g., Thrust) but pay a performance penalty due to inefficient implementations of pattern composition.We develop an API for GPUs based programming on C++ with STL-style patterns and its compiler-based implementation. Our API gives the application developers the native C++ means (views and actions) to specify precisely which pattern compositions should be automatically fused during code generation into a single efficient GPU kernel, thereby ensuring a high target performance. We implement our approach by extending the range-v3 library which is currently being developed for the forthcoming C++ standards. The composable programming in our approach is done exclusively in the standard C++14, with STL algorithms used as patterns which we re-implemented in parallel for GPU. Our compiler implementation is based on the LLVM and Clang frameworks, and we use advanced multi-stage programming techniques for aggressive runtime optimizations.We experimentally evaluate our approach using a set of benchmark applications and a real-world case study from the area of image processing. Our codes achieve performance competitive with CUDA monolithic implementations, and we outperform pattern-based codes written using Nvidia's Thrust.[Copyright notice will appear here once 'preprint' option is removed.]