循环展开是一种常用的编译优化技术,能够有效减少循环开销,提升指令级并行程度和数据局部性,提升循环的执行效能。然而,过度的循环展开会造成指令Cache溢出,增大寄存器压力,循环展开次数太少又会浪费潜在的性能提升机会,因此寻找恰当的...循环展开是一种常用的编译优化技术,能够有效减少循环开销,提升指令级并行程度和数据局部性,提升循环的执行效能。然而,过度的循环展开会造成指令Cache溢出,增大寄存器压力,循环展开次数太少又会浪费潜在的性能提升机会,因此寻找恰当的展开因子是研究循环展开问题的核心。基于GCC开源编译器,面向循环展开问题开展深入的分析与研究,针对指令Cache和寄存器资源对循环展开的影响,提出了一种基于指令Cache和寄存器压力的循环展开因子计算方法,并在GCC编译器中实现了该计算方法。申威和海光平台上的实验结果显示,相较于目前GCC中存在的其它展开因子计算方法,所提出的方法可以获得更为有效的循环展开因子,提升了程序性能。在SPEC CPU 2006测试集上的平均性能分别提升了2.7%和3.1%,在NPB-3.3.1测试集上的分别为5.4%和6.1%。展开更多
Do-loop unrolling is an effective technique for improving performance of applicistion programs. This paper presents a method for unrolling nested Fortran DO-loops using the m4 macro language. m4 is a macro processor w...Do-loop unrolling is an effective technique for improving performance of applicistion programs. This paper presents a method for unrolling nested Fortran DO-loops using the m4 macro language. m4 is a macro processor widely available on UNIX platforms. By using carefully designed m4 macros, Do-loop unrolling becomes much simpler. More over, with this method, code can be written for general ullrolling parameters, allowing the program to be easily tuned bn different computers to reach optimal performance.展开更多
文摘循环展开是一种常用的编译优化技术,能够有效减少循环开销,提升指令级并行程度和数据局部性,提升循环的执行效能。然而,过度的循环展开会造成指令Cache溢出,增大寄存器压力,循环展开次数太少又会浪费潜在的性能提升机会,因此寻找恰当的展开因子是研究循环展开问题的核心。基于GCC开源编译器,面向循环展开问题开展深入的分析与研究,针对指令Cache和寄存器资源对循环展开的影响,提出了一种基于指令Cache和寄存器压力的循环展开因子计算方法,并在GCC编译器中实现了该计算方法。申威和海光平台上的实验结果显示,相较于目前GCC中存在的其它展开因子计算方法,所提出的方法可以获得更为有效的循环展开因子,提升了程序性能。在SPEC CPU 2006测试集上的平均性能分别提升了2.7%和3.1%,在NPB-3.3.1测试集上的分别为5.4%和6.1%。
文摘Do-loop unrolling is an effective technique for improving performance of applicistion programs. This paper presents a method for unrolling nested Fortran DO-loops using the m4 macro language. m4 is a macro processor widely available on UNIX platforms. By using carefully designed m4 macros, Do-loop unrolling becomes much simpler. More over, with this method, code can be written for general ullrolling parameters, allowing the program to be easily tuned bn different computers to reach optimal performance.