[1]欧阳有恒*,严大卫.基于 GRU_LSTM 及 RL 算法的伪随机指令生成器[J].计算机技术与发展,2024,34(02):78-83.[doi:10. 3969 / j. issn. 1673-629X. 2024. 02. 012]
　OUYANG You-heng*,YAN Da-wei.Pseudo Random Instruction Generator Based on GRU_LSTM and Reinforcement Learning Algorithms[J].,2024,34(02):78-83.[doi:10. 3969 / j. issn. 1673-629X. 2024. 02. 012]

点击复制

基于 GRU_LSTM 及 RL 算法的伪随机指令生成器()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年02期

页码:

78-83

栏目:

软件技术与工程

出版日期:

2024-02-10

文章信息/Info

Title:

Pseudo Random Instruction Generator Based on GRU_LSTM and Reinforcement Learning Algorithms

文章编号:

1673-629X(2024)02-0078-06

作者:

欧阳有恒^1* ; 严大卫²

1. 南京信息工程大学,江苏 南京 210044;

2. 无锡先进技术研究院,江苏 无锡 214125

Author(s):

OUYANG You-heng1* ; YAN Da-wei2

1. Nanjing University of Information Science and Technology,Nanjing 210044,China;

2. Wuxi Institute of Advanced Technology,Wuxi 214125,China

关键词:

门控循环单元; 长短记忆; 强化学习; 伪随机指令生成; 通用验证方法学

Keywords:

gate recurrent unit ( GRU ) ;  long short - term memory ( LSTM ) ;  reinforcement learning;  pesudo random instructiongeneration; universal verification methodology ( UVM)

分类号:

TP301. 6;TN492

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 02. 012

摘要:

在 CPU 验证过程中,传统伪随机指令生成器通过生成大量合法无序的指令序列,从而实现功能覆盖率或代码覆盖率的验证目标。 然而,没有趋向针对性的指令生成,为达到指标需要耗费大量

的人力及时间成本。 该文以一款基于精简指令集( RISC-V)自研核心为例,在基于通用验证方法学( Universal Verification Methodology,UVM) 的验证平台上设计出一种伪随机指令生成器,并针对普通伪随机指令生成器覆盖率低、收敛速度慢的问题,建立 GRU_LSTM 算法模型,并结合强化学习( Reinforcement Learning,RL) 算法构建新算法模型 RLGRU_LSTM 应用于伪随机指令生成过程,并且针对 RL 方向决策,提出了基于霍夫曼编码的 CPU 指令包编码方式训练 opcode 分布,同时融合了 CPU 指令类型和指令间执行顺序因素,快速捕获人工定向验证预料不到的

验证盲点,有效加快了代码覆盖率达到预期的进程。 该文着重描述伪随机指令生成器及 RLGRU_LSTM 算法对模型训练过程的指导。 实验结果表明,与直接使用伪随机指令生成技术相比,该方法在约定伪随机指令条目下,相比传统伪随机方法能提高约 19% 的覆盖率,收敛至目标覆盖率消耗时长减少 22% 。

Abstract:

In the CPU verification process, the traditional pseudo random instruction generator generates a large number of legallyunordered instruction sequences to achieve the verification goal of function coverage or code coverage. However, there is no trendtowards targeted instruction generation,and it takes a lot of manpower and time to achieve the target. Taking a RISC-V based self -developed core as an example, we design a pesudo random instruction generator on the verification platform based on UniversalVerification Methodology ( UVM) , and establish a GRU _ LSTM algorithm model to solve the problems of low coverage and slowconvergence of ordinary random instruction. A new algorithm model RLGRU_ LSTM combined with Reinforcement Learning ( RL) isapplied to the pesudo random instruction generation process. Aiming at the RL direction decision, LSTM proposes a CPU instructionpackage coding method based on Huffman coding proposed to train opcode distribution,which integrates the CPU instruction type and theexecution order factors between instructions,quickly captures the unexpected verification blind spots of manual directional verification,and effectively speeds up the process of code coverage reaching the expected rate. We focus on the description of pesudo randominstruction generator and RLGRU_ LSTM algorithm guides the model training process. The experimental results show that compared withthe direct use of pesudo random instruction generation technology,the proposed method can improve the coverage rate by about 19%under the agreed pesudo random instruction entry compared with the traditional pseudo random method, and converge to the targetcoverage rate about 22% earlier.

常用功能