支持原位计算的高效三角矩阵乘法向量化方法

向量化算法映射是向量处理器的难点问题。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: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数据搬移完全重叠,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。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%。

Efficient vectorization method of triangular matrix multiplication supporting in-place calculation

The vectorization of algorithm mapping for vector processors is a critical issue. An efficient vectorization method of triangular matrix multiplication which supports the in-place calculation was presented. L1D was configured as SRAM and the ping pong pattern with double buffering was designed to smooth the data transfers of multilevel storage structure, which made the kernel computation overlap the DMA data transfer fully and run with peak speed throughout, so then the optimal computation efficiency was achieved. Irregular triangular matrix multiplication computation was evenly distributed to all vector processing elements to fully exploit multiple levels of parallelism for vector processor. Result matrix was stored in multiplier matrix, thus, the in-place calculation was achieved and the memory space was saved. Experimental results show that the performance of triangular matrix multiplication attained from the presented vectorization method achieves 1053.7 GFLOPS and the efficiency of that reaches to 91.47%.