Program Generation for Parallel Platforms

微处理器的时钟速度终于达到了它的实际极限。未来的性能提升只能通过各种形式的并行来实现，比如在一个芯片上集成多个CPU内核：主流并行领域已经开始。这将给高性能库的开发人员带来巨大的负担。最优代码必须仔细调整到每个特定的平台，包括它的内存层次结构、特殊指令集和它提供的并行形式。这个耗时的过程在每一个新平台上都会重复。是时候问这个问题了：计算机能为我们编写这些库吗？本研究的目标是开发一个程序生成系统，该系统可以完全自动化实现和优化大量性能关键库功能。本课程将至少包括线性变换、一组密集线性代数问题、相关、一组解码器和数值积分。程序生成系统将生成针对计算机内存层次结构进行优化的代码，如果需要，这些代码可以并行化，用于矢量架构，共享或分布式内存并行，甚至图形处理单元（gpu）中的流并行，或者用于这些的组合。生成的代码的性能应该与最好的手写代码相竞争。“程序生成”意味着系统只将问题说明作为输入。换句话说，计算机本身编写高度优化的，如果需要的话，已经并行化的源代码。为了实现这一点，关于替代算法和算法优化的知识必须以一种可以由计算机完成的方式形式化。总之，目标是使计算机能够为易于理解的数值功能和广泛的并行平台编写非常快速的库。

The clock speed of microprocessors has finally reached its practical limits. Future performance gains will only be obtained through various forms of parallelism such as integrating multiple CPU cores on one chip: The area of mainstream parallelism has started. This will pose an enormous burden on the developers of high performance libraries. Optimal code has to be carefully tuned to every specific platform including its memory hierarchy, special instruction sets, and the forms of parallelism it provides. This time-consuming process is repeated for every new platform released. It is time to ask the question: Can computers write these libraries for us?The goal of this research is to develop a program generation system that completelyautomates the implementation and optimization of a large class of performance-critical library functionality. This class will at least include linear transforms, a set of dense linear algebra problems, correlation, a set of decoders, and numerical integration. The program generation system will produce code that is optimized to a computer's memory hierarchy and that is parallelized, if required, for vector architectures, shared or distributed memory parallelism, or even streaming parallelism in graphics processing units (GPUs), or a for a combination of those. The performance of the generated code should be competitive with the best hand-written code available. "Program generation" means that the system takes as input only the problem specification. In other words, the computer itself writes highly optimized and, if desired, already parallelized source code. To achieve this, the knowledge about alternative algorithms and about algorithm optimization has to be formalized in a way that it can be done by the computer. In summary, the goal is to enable computers to write very fast libraries for well-understood numerical functionality and for a wide range of parallel platforms.