Skip to content

Latest commit

 

History

History
64 lines (43 loc) · 4.2 KB

README.md

File metadata and controls

64 lines (43 loc) · 4.2 KB
Raw

Keep in Balance: Runtime-reconfigurable Intermittent Deep Inference

Overview

This project develops a middleware plugin (referred to as DynBal) that reconfigures the inference engine at runtime. During intermittent DNN inference, DynBal dynamically optimizes the inference configuration for each layer in order to balance the data reuse and data refetch costs, taking into account the varying level of intermittency at runtime.

We implemented our DynBal design on the Texas Instruments device MSP-EXP432P401R. It is an ARM-based 32-bit MCU with 64KB SRAM and single instruction multiple data (SIMD) instructions for accelerated computation. An external NVM module (Cypress CY15B104Q serial FRAM) was integrated to the platform.

DynBal was integrated with the Stateful intermittent inference engine for evalution purposes, although it can be easily integrated into most existing intermittent inference engines.

DynBal contains two key design components which interacts with the inference engine at runtime:

  • Performance estimation: implements an indirect metric, referred to as the Usage Span, which is used to evaluate an inference configuration under a specific level of intermittency
  • Runtime reconfiguration: dynamically updates the inference configuration parameters using feedback from the performance estimation component and heuristics derived at design time.

Demo video: https://www.youtube.com/watch?v=U45VH-SFxKs

Experimental data (including parameter selections): link

Table of Contents

Directory/File Structure

Below is an explanation of the directories/files found in this repo.

  • common/conv.cpp, common/fc.cpp, common/pooling.cpp, common/op_handlers.cpp, common/op_utils.*: functions implementing various neural network layers and auxiliary functions shared among different layers.
  • common/cnn_common.*, common/intermittent-cnn.*: main components of the Stateful intermittent inference engine.
  • common/dynbal*: functions that implement the aforementioned key design components of DynBal.
  • common/platform.*, common/plat-mcu.* and common/plat-pc.*: high-level wrappers for handling platform-specific peripherals.
  • common/my_dsplib.*: high-level wrappers for accessing different vendor-specific library calls performing accelerated computations.
  • common/counters.* : helper functions for measuring runtime overhead.
  • dnn-models/: pre-trained models and python scripts for model training, converting different model formats to ONNX and converting a model into a custom format recognized by the lightweight inference engine.
  • dnn-models/dynbal.py: functions for offline analysis of the Usage Span metric.
  • msp432/: platform-speicific hardware initialization functions.
  • tools/: helper functions for various system peripherals (e.g., UART, system clocks and external FRAM).

Getting Started

Prerequisites

Here are basic software and hardware requirements to build DynBal along with the Stateful intermittent inference engine:

Setup and Build

  1. Prepare vendor-supplied libraries for hardware-accelerated computation. git submodule update --init --recursive will download them all.
  2. Convert the provided pre-trained models with the command python3 dnn-models/transform.py --target msp432 --stateful (cifar10|har|kws) to specify the model to deploy from one of cifar10, har or kws.
  3. Download and extract MSP432 driverlib, and copy driverlib/MSP432P4xx folder into the msp432/ folder.
  4. Import the folder msp432/ as a project in CCStudio.