一种高效的支持原位计算的三角矩阵乘法向量化方法

向量处理器的向量化算法映射是难点问题。提出一种高效的支持原位计算的三角矩阵乘法向量化方法,采用将L1D配置为SRAM模式,用双缓冲的乒乓方式平滑多级存储结构的数据传输,使得内核的计算与DMA数据搬移完全重迭,让内核始终以峰值速度运行,从而取得最佳的计算效率；将不规则的三角矩阵乘法计算均衡分布到各个向量处理单元,充分开发向量处理器的多级并行性；将结果矩阵保存在乘数矩阵中,实现原位计算,节省了存储空间。在Matrix上的实验结果表明,提出的向量化方法使三角矩阵乘法性能达到1053.7GFLOPS,效率为91.47%。

An Efficient Vectorization of Triangular Matrix Multiplication Supporting In-Place Calculation

The Vectorization of algorithm mapping for vector processors is a critical issue. An efficient vectorization of triangular matrix multiplication supporting in-place calculation was presented. L1 data cache was configured as SRAM and a double buffering scheme was designed to smooth the data transfers between SRAM and external memory, which makes kernel computation and DMA data transfer to be fully overlapped, and kernel computation achieves optimization computation efficiency with peak speed. Irregular triangular matrix multiplication computation ware evenly distributed to all vector processing elements to fully exploit multiple levels of parallelism for vector processor. In-place calculation makes the result matrix to be saved to multiplier matrix and saves memory space. Experimental results on Matrix show that the performance of presented vectorization of triangular matrix multiplication achieves 1053.7 GFLOPS, and an efficiency of 91.47%.