Co-Design for Exascale General Purpose computing on Graphics Processing Units (GPGPU)

随着计算系统向兆兆级发展，它们将被能源效率所主导。 这个项目研究了三个方面，使图形处理单元（GPU）在这种能源占主导地位的制度下进行通用计算。 首先，简单的架构创新允许处理器间通信。 其次，此机制暴露给编程API。 最后，我们解决域特定的密集模板计算编译。 该项目推进了能源约束计算的架构，可编程性和编译，并通过数学上严格的程序转换技术（称为多面体模型）为这些平台的领域特定语言（DSL）的编译奠定了基础。 该研究还对与exascale相关的其他广泛领域做出了次要贡献：通过使用DSL提高最终用户程序员的生产力，影响exascale计算机的未来架构，以及探索专用与通用计算之间的权衡。

As computing systems move to exascale, they will be dominated by energy-efficiency. This project investigates three aspects related to enabling graphics procesing units (GPUs) to do general purpose computing under this energy-dominated regime. First, simple architectural innovations allow inter-processor communication. Second, this mechanism is exposed to the programming API. Finally, we address domain specific compilation of dense stencil computations. The project advances architecture, programmability and compilation for energy constrained computation, and also contributes to the foundations of compiling domain specific languages (DSLs) for these platforms through the mathematically rigorous program transformation techniques known as the polyhedral model. The research also makes secondary contributions to other broad areas relevant to exascale: end-user programmer productivity through the use of DSLs, impacting the future architecture of exascale computers, and exploration of the tradeoff between special-purpose vs general purpose computing.