Program performance optimization is essential for achieving high execution efficiency, yet it remains a challenging task that requires expertise in both software and hardware. Large Language Models (LLMs), trained on high-quality code from platforms like GitHub and other open-source sources, have shown promise in generating optimized code for simple snippets. However, current LLM-based solutions often fall short when tackling project-level programs due to the complexity of call graphs and the intricate interactions among functions. In this paper, we emulate the process a human expert might follow when optimizing project-level programs and introduce a three-phase framework POLO (PrOject-Level Optimizer) to address this limitation. First, we profile the program to identify performance bottlenecks using an iterative weighting algorithm. Next, we conduct structural analysis by scanning the project and generating a graph that represents the program's structure. Finally, two LLM agents collaborate in iterative cycles to rewrite and optimize the code at these hotspots, gradually improving performance. We conduct experiments on open-source and proprietary projects. The results demonstrate that POLO accurately identifies performance bottlenecks and successfully applies optimizations. Under the O3 compilation flag, the optimized programs achieved speedups ranging from 1.34x to 21.5x.

Self-attention mechanisms have recently caused many concerns on Natural Language Processing (NLP) tasks. Relative positional information is important to self-attention mechanisms. We propose Faraway Mask focusing on the (2m + 1)-gram words and Scaled-Distance Mask putting the logarithmic distance punishment to avoid and weaken the self-attention of distant words respectively. To exploit different masks, we present Positional Self-Attention Layer for generating different Masked-Self-Attentions and a following Position-Fusion Layer in which fused positional information multiplies the Masked-Self-Attentions for generating sentence embeddings. To evaluate our sentence embeddings approach Multiple Positional Self-Attention Network (MPSAN), we perform the comparison experiments on sentiment analysis, semantic relatedness and sentence classification tasks. The result shows that our MPSAN outperforms state-of-the-art methods on five datasets and the test accuracy is improved by 0.81%, 0.6% on SST, CR datasets, respectively. In addition, we reduce training parameters and improve the time efficiency of MPSAN by lowering the dimension number of self-attention and simplifying fusion mechanism.