Compiler Optimizations for Real-time Graphics Applications

为了满足日益增长的视觉保真度的需求，实时图形渲染需要使用高度并行的图形处理单元（GPU）进行硬件加速。对于图形应用程序，这些GPU通常使用称为“着色器”的小程序来控制，这些小程序在整个渲染管道的各个点上并行运行。这些着色器在主机CPU上的“图形驱动程序”中编译和调度，并且有关将哪些着色器用于哪些对象的决策通常包含在“游戏引擎”的复杂基础设施中。该项目旨在找到自动检测此图形堆栈中的优化机会的方法，并使用算法或机器学习来指导编译和调度决策，以提高各种移动的和桌面GPU的性能。"

"In order to keep up with the demand for ever increasing visual fidelity, real-time graphics rendering requires hardware acceleration using highly parallel graphics processing units (GPUs). For graphics applications, these GPUs are usually controlled using small programs called ""shaders"", which run in parallel at various points throughout a rendering pipeline. These shaders are compiled and scheduled within a "graphics driver" on the host CPU, and decisions about which shaders to use for which objects are typically contained within the complex infrastructure of a "game engine". This project aims to find ways of automatically detect optimization opportunities accross this graphics stack, and use heuristics or machine learning to guide compilation and scheduling decisions to increase performance across a variety of mobile and desktop GPUs."