Compiling Streaming Programs onto FPGA Hardware

我们建议通过在硬件中自动实现程序的计算密集型部分来加速软件程序。 我们选择现场可编程门阵列（FPGA）作为我们的目标硬件平台。 FPGA是可编程计算机芯片，可以配置为实现任何数字电路。 FPGA擅长并行实现计算，人们可以将FPGA视为可配置的计算机，可以为任何应用程序定制。 FPGA已经是一个50亿美元的市场，广泛应用于数字硬件行业。 然而，作为硬件平台，今天的FPGA只能由经验丰富的硬件设计人员使用，因此，它们的范围和使用受到一定的限制。 最近有一个强大的推动力，以提高FPGA的易用性和扩大其用户群。 我们提出了一个流程，其中某一类软件程序，称为流程序，可以自动映射到FPGA上，加快程序运行时间，开发工作量低，成本低。 流程序代表了一种新的计算范例，它允许程序员有效地表达程序中的并行性--非常适合FPGA实现的并行性。 本质上，“流”是可以并行操作的独立数据元素的集合。 流计算对于描述当今数字社会中普遍存在的多媒体应用（例如，数字数据、视频和音频应用）特别有用。我们计划使用现有的流编程语言，称为布鲁克，这是流行的C编程语言的扩展。 我们建议研究，设计和实现一个框架，接受布鲁克程序作为输入，并自动将其转换为FPGA硬件。 将在一组示例应用程序上进行实验性评估。

We propose to accelerate software programs by automatically implementing the computationally-intensive portions of programs in hardware. We choose field-programmable gate arrays (FPGAs) as our target hardware platform. FPGAs are programmable computer chips that can be configured to implement any digital circuit. FPGAs excel at implementing computations in parallel and one can think of FPGAs as configurable computers that can be customized for any application. FPGAs are already a 5 billion dollar market, widely used in the digital hardware industry. However, being hardware platforms, today's FPGAs can only be used by experienced hardware designers, and as such, their scope and use has been somewhat restricted. There has recently been a strong push to improve the ease-of-use of FPGAs and expand their user base. We propose a flow wherein a certain class of software programs, called streaming programs, can be automatically mapped onto an FPGA, speeding program run-time with low development effort and cost. Streaming programs represent a new paradigm in computation that allow the programmer to efficiently express parallelism within the program -- parallelism that is well-suited for FPGA implementation. In essence, "streams" are collections of independent data elements that can be operated-on in parallel. Stream computing is particularly useful for describing multimedia applications pervasive in today's digital society, e.g. digital data, video and audio applications. We plan to use an existing streaming programming language, called Brook, which is an extension of the popular C programming language. We propose to research, design and implement a framework that accepts Brook programs as input and automatically translates them into FPGA hardware. Evaluation will be done experimentally on a set of sample applications.