Abstract: This paper analyses the implementation mechanism of GOTOBLAS library, especially the high performance implementation of its General Matrix-matrix Multiplication(GEMM) routines, and shows how to achieve high performance. It can be concluded that the effects of hierarchical memory on application performance are brought into the design of new computational models. Experimental results verifies the high performance of GOTOBLAS compared with public domain available BLAS, and testifies the important effect of hierarchical memory on algorithm performance.

Key words: GOTOBLAS, general matrix-matrix multiplication, hierarchical memory, blocking algorithm, computational models

摘要： 对GOTOBLAS库(GOTO)的实现机制，尤其是其中的一般矩阵乘法部分的实现进行了分析。结合近年来的一些研究成果，讨论了如何高效地实现矩阵相乘操作，把存储层次对程序性能的影响提高到计算模型的高度。对比实验表明，GOTO库的性能远远高于没有考虑存储层次的一般BLAS库。证明了GOTO库性能上的优越性和将存储层次引入计算模型的必要性。

关键词: GOTOBLAS库, 一般矩阵乘法, 存储层次, 分块算法, 计算模型

CLC Number: 

• TP311