A Novel GNN-based Framework with Differentiable Pooling for High-Level Synthesis QoR Prediction

High-Level Synthesis (HLS) enables rapid prototyping of application-specific hardware, allowing designers to develop hardware using C/C++ instead of hardware description languages (HDL). HLS directives (pragmas) provide optimization mechanisms for balancing performance and resource utilization. However, as the number of applied optimizations increases, the number of possible design configurations grows exponentially. Evaluating each design with HLS tools incurs significant computational and time costs, leading to slow and inefficient design space exploration. To accelerate this process, machine learning models, particularly graph neural networks (GNN), have been used to predict the quality of results (QoR) from pre-synthesis representations. However, existing methods still require improvements due to issues such as significant prediction errors caused by information loss during graph convolution and pooling. To address these challenges, we propose a novel GNN-based framework incorporating differentiable pooling (DiffPool) to learn hierarchical representations of HLS designs. By capturing multi-level structural information, our method effectively reduces information loss while improving the accuracy of QoR predictions. Experimental results demonstrate that our model significantly reduces prediction errors in FPGA resource utilization compared to both conventional HLS estimation methods and existing learning-based models. Additionally, we conducted ablation studies to assess the impact of different components on model performance. The results indicate that, in addition to DiffPool, GATv2 and Global Attention further enhance the model’s performance.