high performance computing on graphics processing units: hgpu.org

In real-time graphics, the strict separation of programming languages and environments for host (CPU) code and GPU code results in code duplication, subtle compatibility bugs, and additional development and maintenance costs. In contrast, popular general-purpose GPU (GPGPU) programming models like CUDA and C++ AMP avoid many of these issues by presenting unified programming environments where both host and GPU code are written in the same language, can be in the same file, and share lexical scopes. To bring the benefits of unified programming to real-time graphics, this paper examines graphics-specific challenges that complicate the development of such a unified model and explores how to overcome them in a widely used programming language. We observe that GPU code specialization, a key optimization in real-time graphics, requires coordination between parameters that are compile-time-constant in GPU code but are assigned values at runtime in host code based on dynamic data. Current methods used to implement specialization do not translate to a unified environment where host and GPU code share declarations of these parameters. Furthermore, this compile-time vs. runtime coordination is not innately expressible in the popular languages used in this domain. In this paper, we create a unified environment for real-time graphics programming in C++ by co-opting existing features of the language and implementing them with alternate semantics to express the services required. Specifically, we co-opt C++ attributes and virtual functions, which enables us to provide first-class support for specialization in our unified system. By co-opting existing features, we enable programmers to use familiar C++ programming techniques to write host and GPU code together, while still achieving efficient generated C++ and HLSL code via our source-to-source translator.