面向飞腾处理器平台的快速卷积算法优化  

Fast convolution algorithm optimization for Phytium processor

作　　者：赵亚飞[1] 杨耀功 王永刚 魏继增[1] ZHAO Yafei;YANG Yaogong;WANG Yonggang;WEI Jizeng(College of Intelligence and Computing,Tianjin University,Tianjin 300354,China;Phytium Technology Co.,Ltd.,Tianjin 300459,China)

机构地区：[1]天津大学智能与计算学部,天津300354 [2]飞腾信息技术有限公司,天津300459

出　　处：《上海理工大学学报》2024年第6期610-619,共10页Journal of University of Shanghai For Science and Technology

基　　金：国家自然科学基金资助项目(61402321);天津市自然科学基金资助项目(23JCYBJC01770);2024年第一批天津市制造业高质量发展专项资金资助项目(24ZGNGX00020)。

摘　　要：为解决卷积神经网络难以在计算资源受限设备上部署的问题,面向国产FT-2000/4多核处理器提出一种高性能的快速卷积算法FastInfer。采用分块策略优化通用矩阵乘法,将处理器访问频率高的数据存入更靠近处理器的缓存中,从而提高计算过程中的访存效率。配合分块方案设计实现高性能的矩阵乘法微内核,使用向量外积运算更新数据,提高计算访存比,实现最大程度掩盖访存指令的延迟。最终实验结果表明,FastInfer在FT-2000/4处理器上的峰值计算性能达到99.56 GFLOPS。在不同输入规模的通用矩阵乘法测试中,FastInfer性能是OpenBLAS算法的1.07倍和1.52倍。在卷积测试中,FastInfer性能是ARM Compute Library算法的1.32倍,实现了在FT-2000/4多核处理器上的高性能卷积计算。To address the issue of deploying convolutional neural networks on resource-constrained devices,a high-performance fast convolution algorithm(FastInfer)was proposed for the FT-2000/4 multi-core processor.The algorithm optimized general matrix multiplication using a block-based strategy,storing frequently accessed data closer to the processor's cache to improve memory access efficiency during computation.In addition,a high-performance matrix multiplication microkernel was designed and implemented,utilizing vector outer product operations to update data and enhance the memory-to-computation ratio.This design maximized the masking of memory instruction latency.Experimental results demonstrated that FastInfer achieved a peak computational performance of 99.56 GFLOPS on the FT-2000/4 processor.In tests with general matrix multiplication at various input scales,FastInfer outperformed OpenBLAS by 1.07 and 1.52 times.In convolution tests,FastInfer performed 1.32 times better than the ARM Compute Library,achieving high-performance convolution computation on the FT-2000/4 multi-core processor.

关 键 词：深度学习 快速卷积算法 并行计算 通用矩阵乘法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]