探索基于 HLS 的自动化硬件生成的代码语言模型：基准、基础设施和分析.pdf

1、1Exploring Code Language Models for Automated HLS-based Hardware Generation:Benchmark,Infrastructure and Analysis ASP-DAC 2025Jiahao Gai,Hao(Mark)Chen,Zhican Wang,Hongyu Zhou,Wanru Zhao,Nicholas Lane,Hongxiang FanOutlineI.IntroductionII.DatasetIII.ModelIV.FrameworkV.EvaluationVI.Discussion23The Era

2、of Generative AILLM-assisted code generation:Github Copilot1,Deepminds AlphaCode2Over 50 pre-trained models and more than 170 programming language datasets releasedAutomated Hardware Design Generation:Verilog,SystemVerilog1 Chen,Mark,et al.Evaluating large language models trained on code.arXiv prepr

3、int arXiv:2107.03374(2021).2 Li,Yujia,et al.Competition-level code generation with alphacode.Science 378.6624(2022):1092-1097.Challenge 1:Data Availability of HDLC+=40.52 times HDLPython=2.26*104 times HDL4Challenge 2:Difficulty in Transferring Pretrained KnowledgeMost code LLMs pre-trained on softw

4、are programming languageDifferent from HDL5Challenge 3:Cost of GenerationHDL implementations require 34 times more tokens than HLS6A Code LLM for HLS GenerationChallenge 1&2:HLS shares main semantic/syntax with C/C+,which makes knowledge transfer possible and reduces dataset requirementsChallenge 3:

5、HLS generation is more cost-efficient at inference timeDataset+Model+Generation Framework7Research QuestionsWhether the existing public data is enough for the training HLS-Gen LLM?What performance can be achieved using existing public data?Can advanced techniques,such as CoT,help HLS-Gen?8OutlineI.I

6、ntroductionII.DatasetIII.ModelIV.FrameworkV.EvaluationVI.Discussion9Format of DatasetInput:Natural language description from developerOutput:HLS design10Dataset Collection52 designs,42000 HLS programs from HLSyn3 and ML4Accel4113 https:/ https:/ Thakur,Shailja,et al.Verigen:A large language model fo