一种龙芯平台上多媒体指令优化时地址非对齐问题的解决方案

Solution to Address Misalignment in Multimedia Instructions Optimization on Loongson Platform

在龙芯平台多媒体指令优化过程中,通常用浮点存取指令存取需并行计算的整数.若这些整数存放在非自然对齐的内存地址上,会导致优化函数的性能显著下降.为了保证优化函数在访问非对齐数据时也有同样的性能,本文采用龙芯通用指令中的非对齐存取指令实现多媒体指令对非对齐数据的存取需求.非对齐存取指令是成对使用的,两条非对齐存取指令的处理时长大概是单条浮点存取指令的五倍左右,故需要合理安排非对齐存取指令的使用.基于此,本文先设计了龙芯平台上64位的非对齐访存函数接口,同时保留现有访存接口;然后设计接口自适应择优算法,用以根据程序上下文灵活选取这些访存接口;最后对LibYUV库的优化函数应用接口自适应择优算法进行测试.结果表明,在数据非对齐时,多媒体指令优化函数出现性能提升比例较小甚至普遍下降的情况;而使用接口自适应择优算法后,所有优化函数平均保持近40%的性能提升比例.

During the optimization of the multimedia instructions in the Loongson platform,floating-point access instructions are usually used to access integers that need to be calculated in parallel.If these integers are stored in memory addresses that are not naturally aligned,the performance of the optimization function will be significantly reduced.In order to ensure that the optimized function also has the same performance when accessing non-aligned data,this article uses the non-aligned access instruction in the Loongson universal instruction to complete the multimedia instruction’s access to non-aligned data.Non-aligned access instructions are used in pairs and the processing time of two non-aligned access instructions is about five times larger than a single floating-point access instruction.Therefore,it is required to use access instructions reasonably.Based on that,this article designed the 64-bit non-aligned memory access function interface on the Loongson platform firstly and retaining the existing memory access interface;then designed the interface adaptive optimization algorithm to flexibly select these memory access interfaces according to the program context;and finally the optimization function of the LibYUV library is tested by using the adaptive optimization algorithm of the interface.The results show that w hen the data is not aligned,the performance improvement ratio of the multimedia instruction optimization function is small or even generally decreased;and after the interface adaptive optimization algorithm is used,all optimization functions maintain an average performance improvement ratio of nearly 40%.