卷积神经网络广泛应用于图像识别、语音识别、自然语言处理等领域，其训练非常耗时。在美国芯片限售令的封锁下，国防科大自主研制的通用数字信号处理器GPDSP不仅将加速成为传统高性能领域里进口加速器芯片的替代，也有望在深度学习领域作为高性能GPU的一种有效替代来加速深度学习的训练。本项目以国产GPDSP为研究平台，研究卷积神经网络并行算法，提出隐式矩阵乘卷积多级并行算法、基于任务池调度的直接卷积多级并行算法以及基于融合的快速卷积多级并行算法来实现国产GPDSP上的高性能卷积计算；提出基于支持向量机分类模型与试运行相结合的卷积算子自动配置算法，以较低的开销实现最佳性能卷积算法的自动选择；最终构建国产GPDSP上的高性能卷积神经网络库，为基于国产GPDSP的深度学习应用开发降低难度，加快训练速度，促进国产GPDSP的发展和推广应用。

Convolutional neural networks (CNNs) are widely used in various applications such as image recognition, speech recognition and natural language processing, and their training is very time-consuming. Due to Chip Restricted Order from USA, General Purpose Digital Signal Processing (GPDSP) chips independently developed by National University of Defense Technology would replace the import accelerator chips in the fields of both high-performance computing and deep learning. This project takes the domestic GPDSP as the research platform to study parallel algorithms of convolutional neural networks, and proposes an implicit matrix multiplication convolution multi-level parallel algorithm, a direct convolution multi-level parallel algorithm based on task pool scheduling and a fusion-based fast convolution multi-level parallel algorithm for maximizing the performance of convolution on the domestic GPDSP, and then an convolution auto-configuration algorithm based on the combination of the support vector machine classification model and the trial run for getting the best convolution algorithm with low overhead. Finally, a high-performance convolutional neural network library will be created to reduce the difficulty of developing CNNs applications, speed up their training, and promote the development as well as the popularization for the domestic GPDSP.