深度神经网络(Deep neural network，DNN)是当前许多人工智能应用的基础。现有的DNN开发框架设计的重点是改善开发环境的可用性，在高级编程语言和后端计算设备的支持上做了很多工作，但是在优化方面的工作远远不够。本项目拟在现有DNN开发框架的基础上，探索在各个层次上实现更高效的性能优化，提出一个更高性能的DNN开发框架。拟提出的DNN开发框架包括以下三个层次：在模型层，提出一套自动数据量化机制，简化网络模型；同时提出基于节点间关系的模型切割与分配的方法，为中间层上计算子图的处理提供更好的布局。在中间层，基于将计算描述和调度优化分离的思想，结合软件流水优化、多面体模型优化等优化方法，为硬件层上不同的终端生成高效的代码。在硬件层，提出一套稀疏矩阵计算优化方案，提升矩阵计算的性能；同时，提出一套面向DNN的GPU内存优化机制，指导数据预取，提升底层运算的性能。

Deep neural network (DNN) is the basis of many current artificial intelligence applications. The existing DNN development framework focuses on improving the usability of the development environment, doing a lot of work on the support of high-level programming languages and back-end computing devices, but the work on optimization is not enough. Based on the existing DNN development framework, this project aims to achieve more efficient performance optimization at various levels and propose a higher performance DNN development framework. The proposed DNN development framework includes the following three levels: At the model level, an automatic data quantification mechanism is proposed to simplify the network model; a model cutting and allocation method based on the relationship between nodes is proposed, which provides a better layout for the processing of computational subgraphs on the intermediate layer. At the intermediate layer, based on the idea of separating the computational description and the scheduling optimization, combined with the optimization methods such as software pipeline optimization and polyhedral model optimization, efficient code is generated for different terminals on the hardware layer. At the hardware layer, a set of sparse matrix computing optimization scheme is proposed to improve the performance of matrix computing; a set of GPU memory optimization mechanism for DNN is proposed to guide data prefetching and improve the performance of hardware.