基于地址信息表的共享内存访问方法
集成电路应用
陶中玉
格兰菲智能科技股份有限公司
摘要: 在通用图形处理器(GPGPU)的并行计算中, 同一个工作组的工作项可对同一共享内存空间进行访问。在传统的共享内存访问方法中,每个工作项独立对共享内存存储库进行访问。当多个工作项访问的数据格式小于存储库宽度时,访问效率极低,并且易出现存储库冲突。该文提出一种基于地址信息表的共享内存访问方法,将相同地址的共享内存请求进行合并,以合并后的请求,对共享内存进行读写操作,以此减少共享内存的访问次数和存储库冲突,提高共享内存访问效率。
中图分类号:TN402, TP391文献标识码:ADOI:10.19339/j.issn.1674-2583.2026.010.001
中文引用格式:陶中玉. 基于地址信息表的共享内存访问方法[J].集成电路应用,2026,43(10):1-9.
英文引用格式:Tao Zhongyu. Shared memory access method based on address information table[J].Application of IC,2026,43(10):1-9.
中文引用格式:陶中玉. 基于地址信息表的共享内存访问方法[J].集成电路应用,2026,43(10):1-9.
英文引用格式:Tao Zhongyu. Shared memory access method based on address information table[J].Application of IC,2026,43(10):1-9.
Shared memory access method based on address information table
Tao Zhongyu
Glenfly Tech Co., Ltd.
Abstract: General-Purpose Graphics Processing Units (GPGPU) parallel computing, work items within the same work group can access a common shared-memory address space. In conventional shared-memory access schemes, each work item independently issues requests to shared-memory banks. When multiple work items access data elements whose size is smaller than the bank width, bandwidth utilization can be severely degraded, and bank conflicts may occur frequently. To address this issue, this paper proposes an address-information-table-based shared-memory access method that coalesces requests targeting the same address. The coalesced requests are then used to perform shared-memory read and write operations. By reducing the number of shared-memory transactions, the proposed method mitigated bank conflicts and improves shared-memory access efficiency.
Key words : General-Purpose Graphics Processing Unit (GPGPU); shared memory; bank conflict; address information table
引言
通用图形处理器(GPGPU)的并行计算通常由多个工作组协作完成,每个工作组由多个工作项构成,每个工作项可对同一共享内存(Shared Memory, SM)进行访问[1-2]。SM可以根据(m,n,k)三个均为大于1的正整数参数进行配置[3]:m为存储库(bank)的个数[4],n为存储库宽度的字节指数,即2n为存储库字节宽度,k为SM包含的存储行,整个共享内存的大小则为: m×2n×k字节。共享内存存储布局如图1所示。

本文详细内容请下载:
http://www.chinaaet.com/resource/share/2000007331
投稿邮箱:ICAPP@belling.com.cn
作者信息:
陶中玉
(格兰菲智能科技股份有限公司,上海201203)

此内容为AET网站原创,未经授权禁止转载。
