Single Path NAS with MAESTRO

Built on a stable version of MAESTRO. (H. Kwon et al., Understanding Reuse, Performance, and Hardware Cost of DNN Dataflows: A Data-Centric Approach, MICRO 2019). Given cache sizes and hardware accelerator, you can estimate the inference latency & energy consumption of your input CNN. Now the package supports conv2d blocks, inverted-residual blocks, and squeeze-and-excitation blocks.

Package Dependences

C++ compiler (g++)

SCONS build system (scons)

Boost libarary (libboost-all-dev)

Python 2.7 or later

Pytorch

How to compile the code

scons --clean scons

How to run the program

./run.sh

How to change the parameters

Change the contents of "run.sh" For parameters other than listed below, please ignore it; active development is going on them so correct functionailty is not guaranteed.

--print_res=true/false : If set true, MAESTRO prints out detailed cost information to the screen

--print_res_csv_file=true/false : If set true, MAESTRO prints out a csv file that contains various statistics

--print_log_file=true/false : If set true, MAESTRO prints out a log file that contains various information of detailed computation patterns to "log.txt"

--DFSL_file='data/DFSL_description/MnasNet-A1_rs.m' : Specify the target dataflow and layer description file

--noc_bw=64 : NoC bandwidth

--noc_hop_latency=1 : NoC latency per hops

--noc_mc_support=true : NoC multicast support (In current dev version it's always on)

--num_pes=256 : Number of PEs

--num_pe_alus=1 : PE ALU vector width

--l1_size=32 : l1 buffer size

--l2_size=512 : l2 buffer size

How to change the DNN model and dataflow

Create a DFSL file under "data/DFSL_description" and point the file using --DFSL_file parameter in "run.sh"

For syntax of the DFSL file, please refer to other DFSL files in data/DFSL_description.

NB: you must put the dataflow under the "data/DFSL_description" directory

How to profile default models:

cd data/pytorch_example

python main.py --model "MnasNet-A1"

Supports "MnasNet-A1", "MobileNet-V2", "MobileNet-V3(large)", "MobileNet-V3(small)", "ProxylessNet(mobile)" and "SinglepathNAS".

How to profile self-defined models:

Use main.py, specify the block arguments by yourself as well as stem/head architecture.

Experiments:

Hardware configuration:

L1 cache: 128

L2 Cache: 5408

Frequency: 2.2G Hz

Model	Params	Multi&Add	Cycles	Estimated Runtime
MnasNet-A1	3887038(3.9M)	330M	25.3M	11.5s
MobileNet-V2	3504872(3.5M)	320M	24.0M	10.9s
MobileNet-V3(large)	5476416(5.5M)	233M	19.8M	9.0s
MobileNet-V3(small)	2534656(2.5M)	65M	6.39M	2.9s
ProxylessNet(mobile)	4080512(4.1M)	336M	26.2M	11.9s
SinglepathNAS	4414216(4.4M)	360M	29.8M	13.5s

Contributors

Ruitao Yi ([email protected])

Hyoukjun Kwon ([email protected])

Prasanth Chatarasi ([email protected])

Felix (Sheng-Chun) Kao ([email protected])