Dataflow Accelerator Examples

for PYNQ on Zynq and Alveo

This repository contains a variety of customized FPGA neural network accelerator examples built using the FINN compiler, which targets few-bit quantized neural networks with emphasis on generating dataflow-style architectures customized for each network.

The examples here come with pre-built bitfiles, PYNQ Python drivers and Jupyter notebooks to get started, and you can rebuild them from source. Both PYNQ on Zynq and Alveo are supported.

Need help with a problem in this repo, or got a question? Feel free to ask for help in the GitHub discussions. In the past, we also had a Gitter channel. Please be aware that this is no longer maintained by us but can still be used to search for questions previous users had.

Quickstart

We recommend PYNQ version 3.0.1, but older installations of PYNQ should also work. For PYNQ v2.6.1, please refer for set-up instructions to FINN-examples v0.0.5.

Zynq

For ZYNQ boards, all commands below must be prefixed with sudo or by first going into sudo su.

First, source the PYNQ and XRT virtual environment:

source /etc/profile.d/pynq_venv.sh
source /etc/profile.d/xrt_setup.sh

Next, ensure that your pip and setuptools installations are up-to-date on your PYNQ board:

python3 -m pip install pip==23.0 setuptools==67.1.0

Since we are going to install finn-examples without build-isolation, we need to ensure all dependencies are installed. For that, install setuptools_scm as well:

python3 -m pip install setuptools_scm==7.1.0

Install the finn-examples package using pip:

# remove previous versions with: pip3 uninstall finn-examples
pip3 install finn-examples --no-build-isolation
# to install particular git branch:
# pip3 install git+https://github.com/Xilinx/finn-examples.git@dev --no-build-isolation

Retrieve the example Jupyter notebooks using the PYNQ get-notebooks command. An example of how to run the Jupyter notebook server, assuming we are forwarding port 8888 from the target to some port on our local machine, is also shown below:

# on PYNQ boards, first cd /home/xilinx/jupyter_notebooks
pynq get-notebooks --from-package finn-examples -p . --force
jupyter-notebook --no-browser --allow-root --port=8888

Alveo

For Alveo we recommend setting up everything inside a virtualenv as described here.

First, create & source a virtual environment:

conda create -n <virtual-env> python=3.8
conda activate <virtual-env>

Next, ensure that your pip and setuptools installations are up-to-date:

python3 -m pip install --upgrade pip==23.0 setuptools==67.2.0

Finally, we can now install Pynq, FINN-examples and Jupyter (please note to source the XRT environment before):

pip3 install pynq==3.0.1
python3 -m pip install setuptools_scm==7.1.0 ipython==8.9.0
pip3 install finn-examples --no-build-isolation
# to install particular git branch:
# pip3 install git+https://github.com/Xilinx/finn-examples.git@dev --no-build-isolation
python3 -m pip install jupyter==1.0.0

Retrieve the example Jupyter notebooks using the PYNQ get-notebooks command. An example of how to run the Jupyter notebook server is also shown below:

pynq get-notebooks --from-package finn-examples -p . --force
jupyter-notebook --no-browser --port=8888

---

You can now navigate the provided Jupyter notebook examples, or just use the provided accelerators as part of your own Python program:

from finn_examples import models
import numpy as np

# instantiate the accelerator
accel = models.cnv_w2a2_cifar10()
# generate an empty numpy array to use as input
dummy_in = np.empty(accel.ishape_normal(), dtype=np.uint8)
# perform inference and get output
dummy_out = accel.execute(dummy_in)

Example Neural Network Accelerators

Dataset	Topology	Quantization	Supported boards	Supported build flows
CIFAR-10	CNV (VGG-11-like)	several variants: 1/2-bit weights/activations	Pynq-Z1 ZCU104 Ultra96 U250	Pynq-Z1 ZCU104 Ultra96 U250
MNIST	3-layer fully-connected	several variants: 1/2-bit weights/activations	Pynq-Z1 ZCU104 Ultra96 U250	Pynq-Z1 ZCU104 Ultra96 U250
ImageNet	MobileNet-v1	4-bit weights & activations 8-bit first layer weights	Alveo U250	Alveo U250
ImageNet	ResNet-50	1-bit weights 2-bit activations 4-bit residuals 8-bit first/last layer weights	Alveo U250	Alveo U250
RadioML 2018	1D CNN (VGG10)	4-bit weights & activations	ZCU104	ZCU104
MaskedFace-Net	BinaryCoP Contributed by TU Munich+BMW	1-bit weights & activations	Pynq-Z1	Pynq-Z1
Google Speech Commands v2	3-layer fully-connected	3-bit weights & activations	Pynq-Z1	Pynq-Z1
UNSW-NB15	4-layer fully-connected	2-bit weights & activations	Pynq-Z1 ZCU104 Ultra96	Pynq-Z1 ZCU104 Ultra96
GTSRB	CNV (VGG-11-like)	1-bit weights & activations	Pynq-Z1	Pynq-Z1

Please note that you can target other boards (such as the Pynq-Z2 or ZCU102) by changing the build script manually, but these accelerators have not been tested.

We welcome community contributions to add more examples to this repo!

Supported Boards

Note that the larger NNs are only available on Alveo or selected Zynq boards.

finn-examples provides pre-built FPGA bitfiles for the following boards:

• Edge: Pynq-Z1, Ultra96 and ZCU104
• Datacenter: Alveo U250

It's possible to generate Vivado IP for the provided examples to target any modern Xilinx FPGA of sufficient size. In this case you'll have to manually integrate the generated IP into your design using Vivado IPI. You can read more about this here.

Rebuilding the bitfiles

All of the examples here are built using the FINN compiler, and can be re-built or customized. See the build/README.md for more details.