DocHero AI - Best paraphrasing and translation tool for academic and professional writing | DocHero AI - Best paraphrasing and translation tool for academic and professional writing

Cross-Modality Program Representation Learning for Electronic Design Automation with High-Level Synthesis

Zongyue Qin, Yunsheng Bai, Atefeh Sohrabizadeh et al. (7 total)

2024-06-13

2024 ACM/IEEE 6th Symposium on Machine Learning for CAD (MLCAD)

10.1145/3670474.3685952

11 citations

摘要

In recent years, domain-specific accelerators (DSAs) have gained popularity for applications such as deep learning and autonomous driving. To facilitate DSA designs, programmers use high-level synthesis (HLS) to compile a high-level description written in C/C++ into a design with low-level hardware description languages that eventually synthesize DSAs on circuits. However, creating a high-quality HLS design still demands significant domain knowledge, particularly in microarchitecture decisions e...

查看文献

问题

The research addresses the challenge of automating microarchitecture decisions in high-level synthesis (HLS) for domain-specific accelerators (DSAs). Existing methods fail to fully leverage both source code and control data flow graph (CDFG) modalities for HLS design quality prediction.

方法

The study introduces ProgSG, a model that combines source code sequence and graph modalities in a deep, fine-grained manner. It uses an attention-summary architecture and node-to-token message passing. A pre-training method based on compiler data flow analysis is also proposed.

关键发现

ProgSG reduces the RMSE of design performance predictions by up to 22%. It identifies designs with 1.10x and 1.26x (up to 8.17x and 13.31x) performance improvement in design space exploration compared to HARP and AutoDSE, respectively.

3个要点

ProgSG effectively integrates source code and CDFG modalities for improved HLS design quality prediction.
The attention-summary architecture and node-to-token message passing enable deep, fine-grained interaction between modalities.
Pre-training GNNs with compiler analysis tasks alleviates the scarcity of labeled designs and enhances performance.

学术详情点击展开

研究对象:Medium-complexity kernels functioning as fundamental building blocks for larger applications, selected from MachSuite benchmark and Polyhedral benchmark suite (Polybench).

干预措施:ProgSG model incorporating graph-summary-augmented sequence representation and fine-grained node token interaction.

结果指标:Latency in cycle counts (perf), utilization rate of block RAM (util-BRAM), digital signal processors (util-DSP), flip-flop (util-FF), and lookup-tables (util-LUT).

样本量:42 kernels with 70:15:15 training, validation, and testing split.

统计方法:Mean squared error (MSE) between predicted and ground-truth targets, Rooted Mean Square Error (RMSE).

未来研究方向:Exploration using more advanced modeling for long sequences.

关键发现:Cross-modality model outperforms single-modality models in terms of rooted mean square error (RMSE). ProgSG surpasses ProgSG-ca, ProgSG-si, and GreaseLM, demonstrating the superiority of fine-grained node token interaction.

临床意义:The core idea of using both modalities together with their alignment is general and can be adapted for other tasks.

生成于 3/26/2026