SHF: Small: Software Fundamentals for Manycore Systems

现代计算机有两个处理器，即“中央处理器”(CPU)和“图形处理器”(GPU)。在历史上，CPU一直用于通用计算，GPU用于图形处理，但随着GPU变得更加灵活和可编程，它们越来越多地用于计算密集型通用任务。然而，对GPU编程是困难的，因为GPU程序员必须将他们的程序分成许多并行部分。使这一过程变得更容易的是使用由专家程序员开发的通用并行构建块。研究小组正在与NVIDIA的同事一起构建这些并行原语的库。在这项工作中，PI集中在三个主要任务上：向该库添加新的原语；优化我们包括的原语的技术；以及将该库扩展到支持系统中的许多GPU。

The modern computer has two processors, the "central processing unit"(CPU) and the "graphics processing unit" (GPU). Historically, the CPU has been used for general-purpose computing and the GPU for graphics, but as GPUs have become more flexible and programmable, they are increasingly used for computationally-intense general-purpose tasks. However, programming a GPU is difficult, because GPU programmers must divide their programs into many parallel parts. What makes this process easier is using common parallel building blocks developed by expert programmers.The research group, together with colleagues from NVIDIA, is building a library of these parallel primitives. In this work, the PIs are concentrating on three major tasks: the addition of new primitives to this library; techniques for optimizing the primitives that we are including; and an extension of the library to supporting many GPUs in a system.