.png){kind=icon}
Return
Mapo: Performance model driven GPU memory access code optimization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.future.2026.108606.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• We propose Mapo-PM, a multimodal memory access performance modeling method for GPU programs. It combines static program analysis with dynamic performance analysis to extract multimodal features from source code, compiler intermediate representations, program structure and runtime memory access behavior. These features are unified through a cross- modal fusion (CMF) mechanism to predict memory access performance metrics. During inference, Mapo-PM relies only on static source code, input size and thread block configuration, significantly reducing dependence on runtime profiling. • We design PM2RL, a reinforcement learning (RL)–driven, memory-oriented code optimization mechanism. By deeply coupling a fine-tuned LLM with Mapo-PM and using the performance model’s predictions to construct reward signals, PM2RL guides the LLM to explore and generate code with improved memory access characteristics during iterative optimization, thereby reducing the additional overhead of frequent compilation and runtime performance evaluation. • We construct multimodal GPU memory access performance and code optimization datasets, covering diverse program structures, input scales and optimization strategies. These datasets provide essential support for training and evaluating Mapo-PM as well as fine-tuning the LLM. • We conduct extensive experimental evaluations on multiple GPU platforms, representative benchmark programs and real-world applications. The results demonstrate Mapo’s advantages in memory access performance model accuracy and code optimization effectiveness.
Keywords:
GPU memory access
performance modeling
code optimization
multimodal features
reinforcement learning
Journal
F
IF:
0
Papers:
642
Citations:
0