面向高速并行向量-矩阵乘法运算的光学数字信号处理关键技术研究

In this proposal, a widely studied optical vector-matrix multiplier is employed to realize high-speed and parallel MAC operation in an optical digital signal processing architecture. A synchronous driver module is studied for the multi-channel lasers and the photodetector array, the data are loaded into the VMM unit bit by bit to complete the signal encoding and decoding, then the data loading and information extraction can be realized. A plane photodetector is utilized to receive the VMM ouput signals, mutiple VMM units are parallel arranged or the system clock is multiplied, and the digital partition based preprocessing and postprocessing algorithms are optimized, then the channel and accuracy of the optical DSP system can be upgraded. Each laser and each optical modulator element are calibrated along the horizontal axis and the vertical axis respectively, the differences between the statistical measurements of detector responses and the set points are limited to a specific threshold, then the mapping relationship between a VMM core's input node and output node can be established. Signal generation, data buffering, smooth filter, and statistical analysis are performed on a series of pipeline stages, and a low-latency distributed parallel pipeline architecture is investigated to accomplish real-time calibration and regulation of the system, leading to high-speed and stable MAC operation.. Key technology problems, including information loading and extraction, channel and accuracy upgrade, self calibration and correction, for optical digital signal processing are resolved.

本申请拟采用近年来被广泛研究的光学向量-矩阵乘法器实现光学数字信号处理构架的高速并行MAC运算。研究多路同步分布反馈激光器和探测器阵列驱动，将数据分位加载到VMM单元，完成信号编解码，实现数据加载和信息提取；采用面阵探测器表征VMM输出信号，通过VMM单元并联或时域倍频，优化基于数据分割的预处理和后处理算法，实现光学DSP系统路数和精度的升级；分别沿横轴和纵轴方向对各激光器和光调制器元素进行标定，使探测器响应的统计表征与所设参考点之差满足特定的阈值范围，建立VMM核输入节点与输出节点之间的映射关系；将信号发生、数据缓冲、平滑滤波、统计分析等一系列模块置于多级流水线上，研究低延迟的分布并行流水结构，完成系统的实时标定和校正，实现高速稳定的MAC运算。.解决向量-矩阵乘法器的信息加载和提取、光路和计算精度升级、自标定和自校正等光学数字信号处理技术关键问题。

本项目拟采用近年来被广泛研究的光学向量-矩阵乘法器实现光学数字信号处理构架的高速并行MAC运算。研究多路同步分布反馈激光器和探测器阵列驱动，将数据分位加载到VMM单元，完成信号编解码，实现数据加载和信息提取；采用面阵探测器表征VMM输出信号，通过VMM 单元并联或时域倍频，优化基于数据分割的预处理和后处理算法，实现光学DSP系统路数和精度的升级；分别沿横轴和纵轴方向对各激光器和光调制器元素进行标定，使探测器响应的统计表征与所设参考点之差满足特定的阈值范围，建立VMM核输入节点与输出节点之间的映射关系；将信号发生、数据缓冲、平滑滤波、统计分析等一系列模块置于多级流水线上，研究低延迟的分布并行流水结构，完成系统的实时标定和校正，实现高速稳定的MAC运算。通过三年的研究，解决了向量-矩阵乘法器的信息加载和提取、光路和计算精度升级、自标定和自校正等光学数字信号处理技术关键问题。对面向并行高速VMM 运算的光学数字信号处理技术进行了理论和实验研究，研制出了光学数字信号处理实验系统，其主要技术指标达到了如下要求：（1）完成VMM运算的向量阶数为1×16，矩阵阶数为16×16，并且阶数和分辨率可持续升级；（2）每秒进行VMM运算次数为3千万次，计算速率为7.68GMAC/秒；（3）能够实时不丢帧地完成VMM运算，运算结果与输入激励之间的延迟小于6个主频时钟周期。