异构多核平台下基础数学库寄存器分配方法

Register allocation in base mathematics library for platform of heterogeneous multi-core

针对异构多核处理器协处理器数学函数中由于查表法和寄存器资源不足而导致的性能下降问题,提出一种基于热路径的寄存器分配方法,结合数学函数的相关路径特点和两类寄存器资源使用开销不一致的情况,对热路径和较少使用路径上的寄存器资源进行再分配过程,将较少使用路径上的高效寄存器资源与热路径上的耗时寄存器资源进行交换,最大限度消除热路径上的访存过程,以降低函数较少使用路径上的性能为代价从而提高函数热路径上的性能,从而达到提升函数整体性能的目的。实际数据表明,上述手段能够使协处理器典型数学函数的性能提升18%以上,从而有效发挥协处理器的计算性能。

Aiming at the performance problem caused by look-up table and lacking of registers in the co-processor mathematics library of heterogeneous multi-core, in view of the feature of mathematics path and the condition that the cost of two kinds of registers were different, the paper propounded a register-allocation optimization algorithm based on hot-path which reallocated the register resource on the hot-path and less-used path which exchanged the high-cost registers on the hot-path with the low-cost registers on the less-used path. The access memory process on the hot-paths was eliminated and a obvious improvement was gained on the hot-path with a performance fall of the less-used path. The real data shows that the above- mentioned way can gain a obvious performance improvement of over 18% for typical functions, so the co-proeessor＇s performanee can be improved.