SIMD向量指令的非满载使用方法研究

大规模SIMD体系结构提供了更强的向量并行硬件支持,但是,大量迭代次数不足的循环由于不能提供足够的并行性,难以用等价的向量方式实现。为了更有效地利用SIMD,提出了一种非满载地使用SIMD指令的向量化方法。研究了向量寄存器的使用方式,基于非满载的向量寄存器使用方式实现了非满载的向量操作和短循环的向量化,并将非满载的向量化方法用于一般循环的向量化。提供了收益分析方法来为本向量化方法作精确指导。实验结果表明了该方法的有效性,所选测试用例的目标循环被向量化,平均加速比达到1.2。

Research on Non-full Length Usage of SIMD Vector Instruction

Large-scale SIMD architecture provides stronger vector parallel support on hardware.However,a large number of loops which are short of iterations can not provide sufficient parallelism,and it is difficult to achieve them with the equivalent vector mode.In order to make full use of SIMD,this paper presented a vectorization method which can use non-full length of SIMD vector instruction.This paper studied the vector register usage,achieved a non-full vector operation based on non-full length usage of vector register,which can vectorize short loops.Finally,this method was used to vectorize the common loops.Moreover,This paper provided a benefit analysis method to guide the vectorization method.Experimental results show that the method is available,the target loops of the selected test programs are vectorized and the average speedup is about 1.2.