基于深度图网络的编译器向量化启发式算法

针对现有的深度学习模型将程序代码考虑为一个串行序列而错失较大性能优化空间的问题,提出了一种新的基于深度图网络的程序启发式优化方法.该方法采用图神经网络对程序的数据和依赖图进行建模,自动从源代码中抽取有效程序特征,然后再将抽取的特征输入下游模型进行循环向量化参数预测.在LLVM循环向量测试集上,所提出的方法取得了2.08倍的加速比,与现有方法相比提高了12％的性能.

Using graph neural networks to enhance compiler code vectorization heuristics

To address the problem that the existing deep learning model considered the program code as a serial sequence and missed a larger performance optimization space, this paper proposed a new heuristic optimization method based on the depth graph network. This method used graph neural network to model the program data and dependency graph, automatically extracted effective program features from the source code, and then input the extracted features into the downstream model for cyclic vectorization parameter prediction. On the LLVM cyclic vector test set, the proposed method achieves a speedup of 2.08 times, which improves the performance by 12% compared with the existing method.