https://community.element14.com/technologies/wireless/w/documents/7491/pageen-USTelligent Community 12https://community.element14.com/technologies/wireless/w/documents/7491/pageThu, 07 Oct 2021 07:35:38 GMT93d5dcb4-84c2-446f-b2cb-99731719e767:5cea8f78-bb85-4eda-955a-63e98015a724sophie0lovehttps://community.element14.com/technologies/wireless/w/documents/7491/page#commentsCurrent Revision posted to Documents by sophie0love on 10/7/2021 7:35:38 AM<br /> <p style="margin:0;">非规则计算是大规模并行应用中普遍存在和影响效率的关键问题.在基于分布式内存的数据并行范例中,如何针对非规则数组引用,有效地生成本地内存访问序列和通信集,是并行编译生成SPMD结点程序所必须解决的重要问题.文中针对两重嵌套循环中,下一层循环边界是上一层循环变量的线性或非线性函数,数组下标是两层循环变量的非线性函数这样一类包含非规则数组引用的并行应用问题,提出了一种在编译时生成通信集的代数算法.并且针对cyclic(k)数据分布和线性对齐模板,借助整数格概念,给出了编译时全局地址和本地地址之间的转换方法.文中还给出了相应的经过通信优化的SPMD结点程序.最后通过实例验证了算法的正确性.该算法的意义在于避免了传统Inspector/Executor非规则计算模型中的Inspector阶段,从而节省了运行时Inspector阶段通过穷举下标生成通信集的巨大开销. </p> <div class="migration-injected-attachments"><div style="font-weight:bold;margin:15px 0 5px 0;">Attachments:</div><div></div><table style="border:0;"><tbody><tr><td width="160"><span class="_se_attachment" id="attid_https://www.element14.com/community/api/core/v3/attachments/20035"><a href="https://community.element14.com/cfs-file/__key/communityserver-wikis-components-files/00-00-00-00-05/_004E7B7C5E97C4891952765E4C88945E2875EE95989884761A90E14FC6961F751062977BD56C_.pdf">community.element14.com/.../_004E7B7C5E97C4891952765E4C88945E2875EE95989884761A90E14FC6961F751062977BD56C_.pdf</a></span></td><td><span style="font-size:80%;">一类非规则并行应用问题的通信集生成算法.pdf</span></td></tr></tbody></table></div><div style="clear:both;"></div>