silky / resnet50-pynq

Quantized ResNet50 Dataflow Acceleration on Alveo, with PYNQ

Quantized ResNet50 Dataflow Acceleration on Alveo

This repository contains an implementations of a binary ResNet50 FINN-style dataflow accelerator targeting Alveo boards. It is intended as a showcase of achievable throughput and latency for ImageNet clasiffication on FPGA, using dataflow execution and on-chip weight storage.

Repo organization

The repository is organized as follows:

• src: contains source code and submodules
  • hls: HLS custom building blocks and submodules to FINN librares (FINN and FINN-HLSLib)
  • w1a2-v1.0: pre-build weights, thresholds, directives and configuration files for Binary ResNet50
• compile: contains scripts for accelerator compilation (Vivado HLS CSynth + Vivado Synthesis)
• link: contains scripts for accelerator linking into the Alveo platform with Vitis
• host: python and Jupyter host code, using PYNQ for Alveo

Building the Accelerator

The Accelerator is built using Vitis 2019.2. We recommend using this version, otherwise changes might be required to source and/or Makefiles for things to work.

To build the accelerator, clone the repository (using --recursive to pull submodules), after which:

cd ResNet50-PYNQ/compile
make NET=w1a2_v1.0
cd ../link
make

See the specific Compile and Link documentation for further info.

Running the Demo

After you have built the accelerator, you can install the required files in the host folder. Move in the cloned repo and do make install

cd ResNet50-PYNQ
make install

You can then run the included Jupyter notebook or the Python multithreaded inference example. If you want to use the distributed PYNQ python package, please read below. If you want to run example Python inference code, please see the host code documentation.

PYNQ quick start

Install the resnet50-pynq package using pip:

pip install resnet50-pynq

After the package is installed, to get your own copy of the available notebooks run:

pynq get-notebooks ResNet50

You can then try things out by doing:

cd pynq-notebooks
jupyter notebook

There are a number of additional options for the pynq get-notebooks command, you can list them by typing

pynq get-notebooks --help

You can also refer to the official PYNQ documentation for more information regarding the PYNQ Command Line Interface and in particular the get-notebooks command.

Supported Boards/Shells

Currently, we distribute the overlay only for the following Alveo boards and shells:

Shell	Board
xilinx_u250_xdma_201830_2	Xilinx Alveo U250

Designs are built using Vitis 2019.2.

Author

Lucian Petrica @ Xilinx Research Labs.