standalone_vta

A maintained, unified, and extended Versatile Tensor Accelerator (VTA) ecosystem.

Overview

This repository addresses the limitations of the original VTA project by providing:

• Unified Simulation: A consistent input format (raw binary files) for both functional (C++) and cycle-accurate (CHISEL) simulators.
• Extended Cycle-Accurate Simulation: Enriched cycle-accurate simulation with multiple test cases for different submodules.
• Standalone Compiler: An open-source, TVM-independent compiler for generating VTA binaries.

This project aims to improve the VTA's usability and applicability, particularly in safety-critical systems like aeronautics. The VTA is an open-source hardware accelerator designed to efficiently execute matrix multiplications, a core operation in Convolutional Neural Networks (CNNs).

Repository Structure

• tutorials/ : Contains a Jupyter tutorial in two parts explaining and illustrating the use of the VTA compiler.
• src/: Contains the source code of the project.
  • compiler/: Contains the VTA compiler that generates the binaries from JSON file.
  • simulators/: Both functional (C++) and cycle-accurate (CHISEL) simulators.
• examples/: The examples to run.
  • Makefile: Use make helpto get the different examples to run.
• config/vta_config.json: The JSON file that defines the VTA hardware parameters.
• environment_setup\standalone-vta.yml: The file to setup the conda environment for executing the project.

Getting Started

To get started with this repository, follow these steps:

1. Clone the repository:

   git clone https://github.com/onera/standalone-vta.git
   cd standalone_vta
   

2. Run the examples:

   cd examples
   make help
   make matrix_16x16
   

   It results in two folders: compiler_output/ and simulators_output.