A Novel CNN Accelerator That Enables Fully-Pipelined Execution of Layers

By Donghyun Kang, Jintaek Kang, Soonhoi Ha.

Introduction

MIDAP, Memory In the Datapath Architecture Processor, features bus-free multi-bank on-chip memory architecture. For more details, please refer to our ICCD Paper.

Citing MIDAP

Please cite MIDAP in your publications if it helps your research:

@inproceedings{kang2019novel,
    title = {A Novel Convolutional Neural Network Accelerator That Enables Fully-Pipelined Execution of Layers},
    author = { D. {Kang} and J. {Kang} and H. {Kwon} and H. {Park} and S. {Ha} },
    booktitle = { 2019 IEEE 37th International Conference on Computer Design (ICCD) },
    year = {2019},
    pages = {698--701},
}

This repository includes MIDAP Compiler & MIDAP Simulator

--Midap Simulator can be excuted with dedicated simulator instruction, please see data_structure/simulator_instruction.py

--Midap Compiler code will be refactored & modulized soon..

How to install?

1. Get the code.

   git clone https://github.com/cap-lab/MIDAPSim.git
   cd MIDAPSim

2. Install requirements.

   pip install -r requirements.txt

3. Run the code at the root directory

   python test.py -n test

How to simulate?

You can define your own network with ModelBuilder please refer test code with models/model_builder.py and models/examples.py

1. you can use TestWrapper class for easy simulation.

2. Please refer test.py for more information

   python test.py -h
   python test.py -n test # Test MidapSim

Tensor Virtualization

1. We proposed a novel virtualization technique [DAC Paper] ( it will be available soon) and tested it via MIDAPSim.

2. Tensor virtualization technique is applied to Concat, Upsample, TransposedConv (UpsampleZero + Conv) Layers.