Skip to content

Latest commit

 

History

History
125 lines (95 loc) · 4.16 KB

README.md

File metadata and controls

125 lines (95 loc) · 4.16 KB
Raw

SMaLL: Software for rapidly instantiating Machine Learning Libraries

SMaLL is a framework for rapidly instantiating DNN based libraries for different architecture. In particular, the objective of SMaLL is to provide a quick and portable approach for developing DNN inference libraries for edge devices. At the core of SMaLL is a common set of loop nests across the different supported layers. Instead of writing every layer manually, each layer is instantiated from a small set of kernels.

A detailed discussion of the framework can be found here (preprint), and here (ACM Transaction of Embedded System article).

Citation

Please use the following citation when referencing SMaLL

@article{10.1145/3607870,
author = {Sridhar, Upasana and Tukanov, Nicholai and Binder, Elliott and Low, Tze Meng and McMillan, Scott and Schatz, Martin D.},
title = {SMaLL: Software for Rapidly Instantiating Machine Learning Libraries},
year = {2023},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
issn = {1539-9087},
url = {https://doi.org/10.1145/3607870},
doi = {10.1145/3607870},
note = {Just Accepted},
journal = {ACM Trans. Embed. Comput. Syst.},
month = {jul}
}

Build Instructions

Specify µArch

The µarch is specified when building with CMake using the CMAKE_UARCH parameter.

The reference build (REF) is written in purely in C and should work on any architecture with a C compiler.

Modify Compiler Flags

Depending on the architecture and the compiler you are using, the compiler flags in ./CMakeLists.txt might need to be modified. For instance, if you are cross-compiling for arm you might specify the GNU C++ arm compiler

#If Cross Compiling for Arm
cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
project(ContextDependentFusion)
include(CTest)
set(CMAKE_VERBOSE_MAKEFILE off)
set(CMAKE_CXX_COMPILER g++)  -------> set(CMAKE_CXX_COMPILER aarch64-linux-gnu-g++)
set(CMAKE_C_COMPILER gcc) -------> set(CMAKE_C_COMPILER aarch64-linux-gnu-gcc)

Build using CMake

Create a build folder from the Top Level directory and run CMake with the desired CMAKE_UARCH platform. Currently CMAKE_UARCH can be defined as one of the following:

  • REF (the default, float or quantized uint8))
  • ZEN2 (Epyc, but also for x86 architectures)
  • ARM-A72 (Raspberry pi 4 b)
  • ARM-A55 (Snapdragon 888 Silver)
  • ARM-A78 (Snapdragon 888 Gold)
  • ARM-X1 (Snapdragon 888 Prime)
  • Q-ARM7E (For the Arduino Nano 33 BLE, quantized uint8 only, see below)

Except for the Arduino (Q-ARM7E) platform, use the following instructions to build the executables.

$ mkdir build
$ cd build
$ cmake ../ -DCMAKE_UARCH=<µarch>
$ make

The make generates executables from the demo directory and unit tests from the test directory.

Build for Arduino (Q-ARM7E)

The Arduino platform (Q-ARM7E) has restrictions on the code layout. To create an Arduino-friendly layout of a subset of the SMaLL library perform the following starting in the SMaLLFramework root directory.

$ mkdir build
$ cd build
$ cmake .. -DCMAKE_UARCH=Q-ARM7E
$ cd ../demo/small-MCU/
$ ./setup_env.sh ../../build/demo ../.. quantized_arm7E

Note that the command line arguments for the setup script are:

$ ./setup_env.sh <path to build/demo> <path to SMaLLFramework> <platform dir>

This creates a new layout under the build/demo/small subdirectory. If you want to build c++ executables for each model:

$ cd ../../build/demo/small/
$ make

If running on the arduino, uncomment the appropriate model in small.ino

#include "mbed.h"
#define NANO33BLE 0

// #include "quantized/autoencoder.cpp"
// #include "quantized/resnet.cpp"
#include "quantized/dscnn.cpp"

Supported Features

DNN Layers

  • Convolution Layer (1D and 2D square and rectangular filters)
  • Group Convolution (1D and 2D)
  • Depthwise Convolution Layer (only square filters)
  • MaxPool Layer (1D and 2D square and rectangular filters)
  • AveragePool (1D and 2D)
  • Upsample (1D and 2D)

With valid and same padding

  • Dense Layer
  • ReLU Activation
  • LeakyReLU
  • Softmax
  • Dropout(?)
  • Add/Accum
  • Concat