This repository presents the implementation of SimulatorCoder, an agent powered by large language models (LLMs), designed to generate and optimize deep neural network (DNN) accelerator simulators based on natural language descriptions. By integrating domain-specific prompt engineering including In-Context Learning (ICL), Chain-of-Thought (CoT) reasoning, and a multi-round feedback-verification flow, SimulatorCoder systematically transforms high-level functional requirements into efficient, executable, and architecture-aligned simulator code.

Python 3.12.4

tqdm 4.67.1

absl-py 2.2.1

configparser 7.2.0

numpy 2.1.2

os, subprocess, math

We access LLMs via APIs: GPT-4o, and DeepSeek-R1. All parameters of the LLMs were set to their default values.

Example:

python scale.py -t ./topologies/mlperf/Transformer.csv -c ./configs/scale.cfg >Transformer.txt

The implementation is inspired by https://github.com/ARM-software/SCALE-Sim.git.