面向稀疏矩阵向量乘的DMA设计与验证

Design and Verification of DMA for Sparse Matrix-Vector Multiplication

稀疏矩阵向量乘法(SpMV)是迭代法求解大型线性方程组的核心算法,被广泛应用在科研和工程中。高性能共轭梯度算法(HPCG)是评价高性能计算系统性能的测试程序之一,需要多次调用SpMV进行迭代计算。但是,SpMV计算过程中包含大量不规则访存操作,降低了系统计算性能。基于X-DSP项目,在DMA中设计一条面向SpMV的专用数据通道实现不规则访存的功能,提高HPCG算法运算速度。设计代码的验证与综合结果表明预期的功能实现正确,且满足项目对时序、面积和功耗的要求。

Sparse Matrix-Vector Multiplication(SpMV)is the core algorithm for solving large linear equations by iterative method,which is widely used in scientific research and engineering.The High Performance Conjugate Gradient(HPCG)algorithm is one of the test programs for evaluating the performance of high performance computing systems.It requires multiple calls to SpMV for iterative calculations.However,the SpMV calculation process contains a large number of irregular memory access operations,which reduces the system computing performance.Based on the X-DSP project,a dedicated data channel for SpMV is designed in the DMA to realize the irregular memory access function,and the HPCG algorithm operation speed is improved.Verification and synthesis of the design code indicate that the expected functionality is implemented correctly and it meets the project’s timing,area,and power requirements.