| 本文已被:浏览 4301次 下载 9591次 |
 码上扫一扫! |
|
|
| 异构平台数学库MAGMA性能测试与分析 |
|
肖玄基1,2, 张云泉1,3, 李玉成1,3, 袁良1,2
|
|
1.中国科学院 软件研究所 并行计算实验室, 北京 100190;2.中国科学院大学, 北京 100190;3.中国科学院 软件研究所 计算机科学国家重点实验室, 北京 100190
|
|
| 摘要: |
| MAGMA是第一个面向下一代体系架构(多核CPU和GPU)开源的线性代数软件包,它采用了诸多针对异构平台的优化方法,包括混合同步、通信避免和动态任务调度.它在功能、数据存储、接口上与LAPACK相似,可以发挥GPU的巨大计算能力进行数值计算.对MAGMA进行了测试分析.首先对矩阵分解算法进行分析;然后通过测试结果,分析MAGMA有效的优化和并行方法,为MAGMA使用、优化提供有益的建议;最后提出了一种对于矩阵分块算法的自适应调优的方法,经过测试,对于方阵的SGEQRF函数加速比达到1.09,对于高瘦矩阵的CGEQRF函数加速比达到1.8. |
| 关键词: MAGMA GPU 矩阵分解 优化 线性代数软件包 |
| DOI: |
| 分类号: |
| 基金项目:国家自然科学基金(61133005);国家高技术研究发展计划(863)(2012AA010902, 2012AA010903);国家重大专项核高基项目(2009ZX01036-001-002) |
|
| Performance Testing and Analysis of MAGMA Library on Hybrid Architecture (CPU+GPU) |
|
XIAO Xuan-Ji1,2, ZHANG Yun-Quan1,3, LI Yu-Cheng1,3, YUAN Liang1,2
|
|
1.Laboratory of Parallel Computing, Institute of Software, The Chinese Academy of Sciences, Beijing 100190, China;2.University of Chinese Academy of Sciences, Beijing 100190, China;3.State Key Laboratory of Computer Science, Institute of Software, The Chinese Academy of Sciences, Beijing 100190, China
|
| Abstract: |
| MAGMA is an open source high performance linear algebra package first developed for next-generation of heterogeneous/ hybrid architectures (CPUs+GPUs) with a dense linear algebra library similar to LAPACK in functionality, data storage, and interface. This paper presents performance testing and analysis of MAGMA. It first studies the matrix decomposition algorithm in MAGMA, then provides some useful suggestions of MAGMA usage and optimization through massive testing and source code analysis, and finally proposes a method for auto-tuning matrix decomposition block algorithms. In this test, the speedup is 1.09 for SGEQRF of square matrix and 1.8 for CGEQRF in terms of tall and skin matrix. |
| Key words: MAGMA GPU matrix decomposition optimization linear algebra package |