基于神经网络的时延驱动版图规划算法  

Neural Network Based Timing Driven Floorplanning

作　　者：胡卫明[1,2] 凌海滨[1,2] 

机构地区：[1]北京大学计算机科学技术研究所文字信息处理技术国家重点实验室 [2]中国科学院自动化所模式识别国家重点实验室,北京100080

出　　处：《电子学报》2000年第5期21-24,共4页Acta Electronica Sinica

基　　金：中国博士后科学基金资助课题

摘　　要：本文用离散的网格代替连续的版图规划平面 ,把长宽比可变的软模块对应成多个长度和宽度均确定的硬模块 ,给出了相应的时延驱动版图规划问题的形式化描述 ,并提出了基于均场退火网络的新的求解算法 .算法用一个三维二值换位矩阵将问题映射为神经网络 ,建立包含时延约束、重叠约束和优化目标的能量函数 ,再用均场退火方程迭代求解 .对应于同一软模块的硬模块有且只有一个能且只能放置在版图规划平面一个位置上的约束用神经元归一化的方法解决 .本算法已用VisualC ++编程实现 ,实验结果表明 ,这是一种有效的方法 .In this paper,with the continuous floorplanning space being replaced by discrete space,and soft cell with variable aspect being expressed by some hard cells with fixed lengths and heights,a formal description of the timing driven floorplanning is presented,and a mean field annealing approach for the problem is proposed.In the algorithm,with a three dimensional permute matrix of binary variablesubeing used to map the problem to the neural network,the energy function including object item,overlap constrained item and timing constrained item,is presented,and then iteration procedure is put into practice with the mean annealing equation.Normalization of neurons proves that only one of the hard cells corresponding to the same soft cell must be placed once and only once in the plane.The algorithm is programmed with Visual C++ language,and experimental result shows that it is an effective method.

关 键 词：神经网络 时延驱动 版图规划 VC 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]