Skip to content

Latest commit

 

History

History
88 lines (63 loc) · 3.95 KB

README.md

File metadata and controls

88 lines (63 loc) · 3.95 KB

ONNX to HLS Generator Toolkit

This project is a generator toolkit designed to take an ONNX model quantized using Brevitas, and generate a High-Level Synthesis (HLS) representation of it. The generated HLS code is optimized for hardware implementation and uses a library of HLS tools. The toolkit transforms the input ONNX model into an HLS representation in the form of C++ (.cpp) and header (.hpp) files that are ready for synthesis on FPGA platforms.

Features

  • ONNX Model Conversion: Supports conversion of ONNX models, including models quantized using Brevitas and floating-point models.
  • Node Optimizations: Removes unsupported nodes, merges certain nodes, and adds split nodes for compatibility with HLS tools.
  • Hardware Optimization: Sets FIFO depths, parallelism, and divides the model into multiple top levels.
  • Generated Output: The tool generates .hpp and .cpp files, which include top-level wrappers, profilers, configuration, and parameters for HLS.

Table of Contents

Usage

To use this toolkit, ensure that you have the following dependencies:

Requirements:

  • Python 3.x
  • ONNX
  • Brevitas (for quantized ONNX models)
  • Additional libraries as needed for ONNX manipulation and visualization

You can install the required Python libraries by running:

pip install onnx brevitas networkx matplotlib

To use the toolkit, run the following command in the terminal:

python onnx_to_hls.py --model <path_to_onnx_model> [additional options]

This will take the ONNX model and generate HLS-compatible C++ and header files in the specified output directories.

Command-Line Arguments

  • --model: (Required) Path to the ONNX model file.
  • --result_hpp_path: (Optional) Path to save the generated .hpp files (default: ./include_hw/).
  • --result_cpp_path: (Optional) Path to save the generated .cpp files (default: ./src_hw/).
  • --input_mem_width: (Optional) Input memory width (default: 64).
  • --output_mem_width: (Optional) Output memory width (default: 64).
  • --max_subgraph_weight: (Optional) Maximum subgraph weight (default: 250,000).
  • --subgraph_weight_margin: (Optional) Allowed subgraph weight margin (default: 1.1).
  • --brevitas: (Optional) Flag to indicate if the ONNX model is quantized using Brevitas (default: False).
  • --float: (Optional) Flag to treat the model as a floating-point model (default: False).
  • --profiler_json: (Optional) Path to the profiler JSON file (default: ./profiled_activations.json).
  • --stop_node: (Optional) Last node in the graph (default: '').

Example

Here’s an example of how to use the toolkit:

python onnx_to_hls.py --model ./models/my_model.onnx \
    --result_hpp_path ./include_hw/ \
    --result_cpp_path ./src_hw/ \
    --input_mem_width 64 \
    --output_mem_width 64 \
    --brevitas

This command will take the ONNX model located at ./models/my_model.onnx, and generate .hpp files in ./include_hw/ and .cpp files in ./src_hw/ based on the model's architecture. The model is assumed to be quantized using Brevitas.

Output Files

  • Top-Level Wrapper Files: dnn_top_level_wrapper.hpp, dnn_top_level_wrapper.cpp
  • Main Logic Files: dnn_top_level.cpp
  • Profilers: dnn_profilers.hpp, dnn_profilers.cpp
  • Configuration Files: dnn_config.hpp
  • Parameter Files: dnn_params.hpp
  • Visualization Files: my_model_after_onnx.png, my_model_after_creating_subgraphs.png this allows the user to visualize the internal structure of the model after the toolkit has processed it.

Further Development

  • Support for more complex ONNX layers and edge cases will be included in upcoming updates.
  • Support for other quantization libraries and formats will be added.