SHF: Small: Tools for Productive High-performance Computing with GPUs

图形处理单元（GPU）广泛且廉价可用，并且相对于通用CPU已经变得越来越强大。因此，它们是计算科学和数据科学中计算密集型应用的有吸引力的目标。然而，在GPU上运行的软件的开发是耗时的，并且只需要应用程序开发人员社区的一小部分人拥有的专业知识。该项目正在开发一系列工具，以帮助GPU高性能软件的生产性开发，从而降低科学界有效使用GPU的障碍。本研究的中心思想是识别限制GPU内核性能的主要硬件资源瓶颈，以寻求缓解所识别的瓶颈的方式来指导内核的修改。抽象内核仿真沿着与硬件资源延迟/吞吐量参数相关的灵敏度分析用于瓶颈识别。三种使用场景是针对的：（1）OpenMP卸载，（2）特定于域的代码生成器，和（3）CUDA/OpenCL内核。 OpenMP4.0中引入的卸载模型是一种有吸引力的方法，可用于转换现有的遗留代码以及新开发的代码，以提高生产力和可移植性。特定于域的库生成器利用特定于模式的语义来执行超出通用优化编译器范围的优化转换。张量收缩和张量扩张是特别强调的两个领域。对于所有有针对性的使用场景，该工具集旨在帮助开发人员通过模型驱动搜索和自动调优相结合来提高GPU代码的性能。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响评审标准进行评估，被认为值得支持。

Graphical Processing Units (GPUs) are widely and cheaply available and have become increasingly powerful relative to general-purpose CPUs. Therefore, they are attractive targets for compute-intensive applications in computational science and data science. However, development of software to run on GPUs is time-consuming and requires expertise held by only a very small fraction of the application developer community. This project is developing a collection of tools to assist in the productive development of high-performance software for GPUs, so that the barrier to effective use of GPUs by the scientific community can be lowered.A central idea being pursued in this research is the identification of primary hardware resource bottlenecks that limit performance of a GPU kernel, to guide the modification of the kernel in a manner that seeks to alleviate the identified bottleneck. Abstract kernel emulation along with sensitivity analysis with respect to hardware resource latency/throughput parameters are used for bottleneck identification. Three usage scenarios are targeted: (1) OpenMP offload, (2) domain-specific code generators, and (3) CUDA/OpenCL kernels. The offload model introduced in OpenMP 4.0 is an attractive approach for transforming existing legacy codes as well as for newly developed codes, to facilitate productivity and portability. Domain-specific library generators exploit pattern-specific semantics in order to perform optimizing transformations that are beyond the scope of general-purpose optimizing compilers. Tensor contractions and stencils are two domains of particular emphasis. For all targeted usage scenarios, the collection of tools is intended to assist developers improve the performance of GPU code through a combination of model-driven search and auto-tuning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.