基于CUDA的GPU条件分支分歧聚合优化策略

分析NVIDIA GPU底层处理SIMD条件分支分歧的方式及其对程序性能产生的影响。在软件层级提出两种利用"聚合"思想的SIMD条件分支分歧优化策略:循环推迟和循环提前。策略将不同SIMD道中选择相同路径的条件分支"聚合"到同一步循环中,减少了SIMD操作的实际次数。使用CUDA对这两种策略进行的试验结果表明,在满足策略使用条件的前提下能够取得预想中的加速比。该策略实现难度较低、可操作性较强。

GPU conditional branch divergence converging optimization strategies based on CUDA

The underlying rules how the NVIDIA GPU deals single instruction multiple data (SIMD) conditional branch divergence and its impact on application performance were analyzed. Based on loop postpone or loop advance, two SIMD conditional branch divergence optimization strategies were proposed, in which conditional branches that choose the same path in different SIMD lanes were merged into one loop step, resulting in reducing the number of SIMD operations. The experimental results of these two strategies by using CUDA show that the application achieves expected speedup when the conditions of applying these strategies are met. These strategies are less difficult to implement and have a strong operability.