摘要
为了提高嵌入式图形处理器的纹理单元效率，提出了一种多端口纹理高速缓存（Texture Cache）结构。该结构采用了基于块的光栅化和块交错的纹理内存组织，能够充分发掘数据间相关性，提高了Cache命中率；此外该结构采用Cache预取技术，有效隐藏了访存延迟；为了进一步提高数据吞吐率，设计了4个读端口，可支持并行读取4个纹素。仿真结果表明，设计的Cache可达到92％左右命中率，访存性能可达到零延迟内存系统的90％，数据吞吐率是单端口Cache的3～4倍。
In order to improve GPU＇s texture unit efficiency ,a multiported texture cache architecture is proposed . The architecture employs a tiling rasterization order and a block interleaving memory organization ,which can fully exploit the data locality and improve cache hit rate . In addition , the architecture employs cache prefetching technology ,which can hide memory latency . In order to further improve data throughput , four read ports supporting 4 parallel reading are designed .Simulation results show that the hit rate of the proposed cache is about 92% and can attain 90% of the performance of a zero latency memory system .The data throughput is about 3~4 times of the single ported cache .
出处
《微电子学与计算机》
CSCD
北大核心
2013年第11期27-30,34,共5页
Microelectronics & Computer
基金
国家自然科学基金(61234003)
关键词
图形处理器
光栅化
纹理高速缓存
分层细化
预取
多端口
Graphics Processing Unit
rasterization
texture cache
mipmap
prefetching
multiported
