基于样板的GPU线程快速构造方法

Design of Fast GPU Thread Construction Approach Based on Pattern in General-purpose Computation

阐述通用计算任务可利用GPU的多线程获得并行加速,根据预先设定的样板,对整个线程一次性创建,加快了线程的构造时间。样板分为线性和瓦片两种,分别处理线性和矩形排列的工作组。模拟仿真测试表明,基于样板的线程构造时间一般为8cycles左右,比传统方法获得3倍的提升。该方法易于硬件实现,可以在之前电路上直接扩展,实现功能向前兼容。

This paper describes that general-purpose computing tasks can be accelerated in parallel through the multithreading of GPU.It presents thread construction approach based on patterns to speed up construction time.The pattern is divided into linear and tile,which deal with linear and rectangular working groups respectively.The simulation experiments show that the thread construction time based on pattern is about 8 cycles,which is 3 times faster than the traditional method.The approach is easy to be implemented in hardware,and preserves function compatibility on previous circuit.