基于OpenMP的Winograd并行矩阵乘算法应用研究

为了提高半经典分子动力学模拟中矩阵乘法效率,通过一种稀疏矩阵分解方法化简矩阵乘法,基于OpenMP实现矩阵相乘的Winograd并行算法。该算法将Winograd算法中各部分依次采用OpenMP并行计算,降低了数据通信。在16核服务器上测试表明,该方法能够显著提高半经典分子动力学模拟中矩阵乘法效率,并行加速比能够达到9.47,并具有良好的可扩展性,为大分子体系的模拟提供了可能。

Application and research on Winograd parallel algorithm of matrix multiplication based on OpenMP

In order to improve the efficiency of the matrices multiplication in the semi classical molecular dynamics simulation, simplified through a the matrix multiplication sparse matrix decomposition method. And then, achieved the matrix multiplication by Winograd parallel algorithm based on OpenMP. This algorithm calculated parallelly each part of Winograd algorithm in order by OpenMP to reduce the data communication. Tested on the server with 16 cores, this method could significantly improve the efficiency of the matrix multiplication in the semi classical molecular dynamics simulation. Its parallel speedup ratio can reach 9. 47 and has good scalability. It provided probability for the simulation in large molecular system.