SHF Small: A Compiler-Based Auto-Tuning Framework for Many-Core Code Generation

该项目为数据并行优化和代码生成提供了基于编译器的基础，可以在一系列不同的多核和众核架构中产生非常高的性能，同时提高应用程序员的生产力。 该项目针对两种不同的架构，一个传统的四核微处理器和一个240核的图形处理单元（GPU），以及一个结合两者的异构系统。 该技术代表了与当代编译器组织的重大偏离，尽可能多地自动化优化，但开放映射过程，为精明的程序员提供他们经常渴望的控制。 该系统的独特功能包括：（1）转换配方作为程序员或高级编译器的优化和代码生成接口;（2）多面体框架，以数学方式表示代码结构和鲁棒代码生成的优化;以及（3）自动调整技术，以复杂地描述和系统地评估一系列可能的计算实现。 其中最新颖的是转换配方，它排除了OpenMP，CUDA和OpenCL代码生成的共性，从而支持异构平台并促进后期绑定决定使用哪个实现。 以前工作的软件系统正在发布，并加速了图书馆和科学应用。 有了正确的转换方法供自动调优系统评估，自动生成的代码会产生高性能，在某些情况下具有竞争力，甚至比手动编码的实现更好。 PI计划将此功能扩展到更广泛的应用程序和架构特性。

This project provides a compiler-based foundation for data-parallel optimization and code generation that can yield very high performance across a range of different multi-core and many-core architectures while increasing the productivity of application programmers. The project is targeting two distinct architectures, a conventional quad-core microprocessor and a graphics processing unit (GPU) with 240 cores, as well as a heterogeneous system that combines both. The technology represents a significant departure from the organization of contemporary compilers, automating as much of optimization as possible, but opening up the mapping process to provide savvy programmers the control they so often desire. The unique features of the system include: (1) transformation recipes as a programmer's or high-level compiler's interface to optimization and code generation; (2) a polyhedral framework to mathematically represent code structures and optimizations for robust code generation; and (3) auto-tuning technology to compactly describe and systematically evaluate a range of possibleimplementations of a computation. The most novel of these is the is the transformation recipe, which factors out the commonality of OpenMP, CUDA and OpenCL code generation, thus supporting the heterogeneous platform and facilitating late-binding decisions of which implementation to use. The software system from previous work is being released, and has accelerated both libraries and scientific applications. With the right transformation recipe for the auto-tuning system to evaluate, the resulting automatically-generated code yields high performance, in some cases competitive or even better than manually-coded implementations. The PIs plan to expand this capability to a broader set of applications and architectural features.