基于多面体模型的深度学习张量编译器

深度学习应用对人工智能领域特定软硬件架构提出了很高的要求，现有深度学习框架和人工智能芯片的发展，仍无法满足编程易用性和计算力的需求。当前，在人工智能芯片上部署深度学习模型的人力物力成本仍然居高不下，面向深度学习的编译技术是阻碍深度学习应用发展的关键挑战。为了解决上述挑战，本项目开展基于多面体模型的深度学习编译器研究，具体包括：1）设计多面体模型友好的领域特定编程语言；2）兼顾计算空间和数据空间变换的多面体模型循环变换技术；3）面向人工智能芯片领域特定体系结构的轻量级代价模型；4）基于多面体模型的稀疏张量运算自动代码生成技术。本项目旨在实现一套面向人工智能芯片的自动高效编译器，显著降低当前手动生成优化代码的大量人力物力成本。本项目的研究有望解决人工智能芯片上自动编译的关键问题，为领域特定体系结构的代价模型以及基于稀疏运算的深度学习应用提供有效的解决方案。

Deep learning applications pose a severe demanding expectance to the domain-specific software and hardware of artificial intelligence. The development of existing deep learning frameworks and artificial intelligence chips, however, are still far from the expected programmability and computing power. Currently, the human and material costs required by the deployment of deep learning models on artificial intelligence chips are still very expensive. The deep learning compiler is becoming the crucial challenge inhibiting the development of deep learning applications.. .To solve this challenge, we study the deep learning compiler based on the polyhedral model in this project by (1) designing a domain-specific programming language that can benefit the power of the polyhedral model; (2) integrating the affine loop transformations by transforming both iteration and data spaces; (3) implementing a light-weight cost model for domain-specific architectures of artificial intelligence; and (4) automating code generation for sparse tensor computations in polyhedral compilation. The purpose of this project is to implement an effective tensor compiler for artificial intelligence chips, reducing the human and material costs required by existing manual approaches of code generation techniques for deep learning applications. More importantly, this project is promising to address the essential issues of compilation techniques for artificial intelligence chips, and also provide reasonable solutions to customizable cost models for domain-specific architectures and deep learning applications of sparse tensor computations.