摘要
BLAS是当前科学计算领域重要的底层支持数学库之一,其中的3级BLAS函数应用最为广泛.本文基于国产申威1600平台,提出了一种基础线性代数库BLAS的三级函数通用矩阵乘GEMM的高性能实现方法.在单核上,使用乘加指令、循环展开、软件流水线指令重排、SIMD向量化运算、寄存器分块技术等与平台架构相关的技术手段,实现汇编级手工优化;在多核上,提出了适用于该平台的多线程加速方案.实验结果显示,在单核串行性能测试中,与知名开源数学库Goto BLAS相比,我们实现了平均4.72倍的加速效果;在多核并行扩展测试中,4线程版的性能则平均达到了单线程版性能的3.02倍.
BLAS is one of the most important basic underlying math library for scientific computing, in which the level 3 BLAS functions are most widely used. In this paper, we provide a high-performance method to implement Level 3 BLAS functions based on domestic Sunway 1600 platform. To make it clear, we take GEMM as an example. For the implementation on single-core, we apply many tuning techniques related to the specific platform, such as multiply-add instructions, loop unrolling, software pipelining and instruction rearrangement, SIMD operations, and register blocking to push up the performance. For the multi-core implementation, we propose an efficient multi-threaded method. Compared with Goto BLAS, one of the famous open-source BLAS, the experiments show that our serial single-threaded method achieves a speedup of 4.72. What’s more, the average speedup of 4-threaded execution towards the single-threaded one can also reach 3.02.
出处
《计算机系统应用》
2016年第12期234-239,共6页
Computer Systems & Applications
基金
国家自然科学基金(91530103
91530323)
