细粒度任务并行GPU通用矩阵乘

发布时间：2018-03-26 22:26

本文选题：通用矩阵乘　切入点：持久化kernel　出处：《计算机工程与科学》2015年05期

【摘要】：稠密线性代数运算对模式识别和生物信息等许多实际应用至关重要,而通用矩阵乘(GEMM)处于稠密线性代数运算的基础地位。在cuBLAS与MAGMA中,GEMM被实现为若干kernel函数,对大型GEMM计算能够达到很高的性能。然而,现有实现对批量的小型GEMM计算性能发挥则较为有限。而且,现有实现也不能在多个具有不同性能的GPU之间自动扩展并达到负载均衡。提出任务并行式GEMM(TPGEMM),用细粒度任务并行的方式实现批量矩阵乘和多GPU矩阵乘。一个或多个GEMM的计算能够被拆分为多个任务,动态地调度到一个或多个GPU上。TPGEMM避免了为批量矩阵乘启动多个kernel函数的开销,对批量矩阵乘能够取得显著高于cuBLAS与MAGMA的性能。在低开销细粒度任务调度的基础上,TPGEMM支持单个GEMM计算在多个GPU间的自动并行,在一台具有四个不同性能GPU的工作站上取得了接近100%的扩展效率。
[Abstract]:Dense linear algebraic operations are very important for many practical applications, such as pattern recognition and biological information, while general-purpose matrix multiplication is the basis of dense linear algebraic operations. In cuBLAS and MAGMA, they are implemented as several kernel functions. High performance can be achieved for large GEMM computing. However, the performance of existing implementations for batch small GEMM computing is limited. The existing implementation can not automatically extend and achieve load balance among multiple GPU with different performance. This paper proposes a task-parallel GPU matrix multiplication, which implements batch matrix multiplication and multiple GPU matrix multiplication by fine-grained task parallelism. The calculation of GEMM can be split into multiple tasks, Dynamically scheduling to one or more GPU. TPGEMM avoids the overhead of starting multiple kernel functions for batch matrix multiplication. The performance of batch matrix multiplication is significantly higher than that of cuBLAS and MAGMA. On the basis of low overhead fine-grained task scheduling, TPGEMM supports automatic parallelism of single GEMM computation between multiple GPU. Nearly 100% expansion efficiency is achieved on a workstation with four different performance GPU.
【作者单位】：南开大学计算机与控制工程学院;
【分类号】：TP338.6

【共引文献】