mRNA: Efficienct mapping space exploration for Reconfigurable Neural Accelerator

Introduction

mRNA is a mapping space exploration tool for MAERI. MAERI is a DNN Spatial Accelerator which is able to support flexible Dataflows. It aims at efficient mapping of datalfows emanating from the diverse deep learning landscape. For details of MAERI, please refer to the ASPLOS 2018 paper below and for its code base, you can go to here

Hyoukjun Kwon, Ananda Samajdar, and Tushar Krishna, MAERI: Enabling Flexible Dataflow Mapping over DNN Accelerators via Reconfigurable Interconnects, ASPLOS 2018

mRNA generates the DNN Model which is implemented by Tensorflow, extracts the necessary information from each layer, and generate the optimal mapping strategy which map a layer onto MAERI. It has three basic functionalities:

Efficienctly search the optimal DNN mapping for MAERI
Mapping space exploration tool for developers to understand the impact of mapping over MAERI
Generate the Network configuration(s) for MERI.

For details on mRNA, please refer to the ISPASS 2019 paper below.

Zhongyuan Zhao, Hyoukjun Kwon, Sachit Kuhar, Weiguang Sheng, Zhigang Mao, and Tushar Krishna, mRNA: Enabling Efficient Mapping Space Exploration for a Reconfigurable Neural Accelerator, ISPASS 2019

Tool Setup

If you want to use Front-end Paser, you should have python3, Tensorflow environment and lark-parser package. The install information of Tensorflow goes to here. The install information of lark-parser package goes to here. And the for information of installing python3 please go to here.
For using mRNA, you should install Gglags, the download and install information of Gflags please go to here.
Install instruction:

1. Enter into mRNA folder
2. cd build
3. cmake ../
4. make

We also provide an executable file inside build folder so that you do not need to install gflags. However, if this does not work, you should install mRNA according to the afore mentioned steps.

Command Guide

Front-end Parser

The Front-end Parser get the DNN Model as input and output the information of each layer include DNN Model name, layer type and name. And four dimensional representation of input, weight, output and hidden unit (for LSTM). The command example is:

python3 network_parser.py #dnn_model#.py

And all the output parameters of each layer for a DNN mdoel are located inside param folder. The input folder contains all the layer representations for four DNN Model (alexnet, resnet, vggnet and rnn). We also provide the Tensorflow implementation so that you can try to generate the parameters of each layer for a specific model.

Analyzer

You can use --help command to check all the descriptions of each command

build/MAERI_Mapper --help

You can see the architecture of MAERI (32 MSes in this example) under different scales by typing command

build/MAERI_Mapper -show_maei num_ms=32

If you want to understand the impact of mapping over MAERI, example below is the command of mapping a convolutional layer of alexnet, and generating the profile result for this layer

1. cd mNRA/input/alexnet
2. ../../build/MAERI_Mapper -model_para="Model_parameter_4.txt" -show_energy 
                            -config_file="Config_file.txt" -num_ms=256 -dn_bw=128 -rn_bw=256 
                            -performance

If you want to generate the configuration, you should add -genconfig option in the commmand.

../../build/MAERI_Mapper -model_para="Model_parameter_2.txt" -show_energy 
                         -config_file="Config_file.txt" -dn_bw=128 -rn_bw=256 
                         –performance -genconfig

Output file

The output file of mRNA is generated int the same path as the input file. The name scheme of the output profile file is as follows

DNN-#DNNModelName#-#Layer Type##Layer Number#_#MS Num#-#DSW#-#RSW#-#Opt Goal#.txt

The DSW and RSW represents the bandwidth of DS and RS Network. Example of outputfile is:

DNN-GoogleNet-CONV1_MAERI-MS-256-DSW-128-RSW-256_Opt-performance.txt

abhinavhimanshu / mRNA