论文部分内容阅读
向量优化是指在一台给定机■上,对一个已知的向量计算生成最好的结果代码。本文为在控制政据 CYBER 205上运行的标量和向量源代码提出了性能分析模型。这些模型的精度对于标量代码一般是在30%以内,对于向量代码一般是在10%以内。对于一个已知的并行计算,如果编译程序可以产生一个以上的代码程序,则可以利用来自这些模型的性能估计,以便选择应当执行哪一个代码程序。把含有两个或更多个源代码版本的16个FORTRAN 核作为测试这个方法的标准检查程序。13个核正确地进行了“向量优化”。没有进行常规优化的3个核有17%的平均性能损失。用一批核作为标准检查程序表明向量优化把它的性能改进了四倍多,所有已经正确优化了的核,有98%的在程序上也可改进。
Vector optimization means generating the best result code for a given vector on a given machine. This article presents a performance analysis model for scalar and vector source code running on CYBER 205 control. The accuracy of these models is typically within 30% for scalar codes and within 10% for vector codes. For a known parallel computation, if the compiler can generate more than one code program, the performance estimates from these models can be used to choose which code program to execute. Sixteen FORTRAN cores, containing two or more source code versions, were used as the standard test for testing this method. Thirteen cores performed “vector optimization” correctly. The three cores that did not routinely optimize had an average performance loss of 17%. Using a number of kernels as a standard checking procedure shows that vector optimization improves its performance more than four times, and that all correctly optimized kernels and 98% of them are procedurally improved.