Chisel4ml

Chisel4ml is an open-source library for generating highly-parallel dataflow style hardware implementations of Deeply Quantized Neural Networks. These types of networks are trained using frameworks such as Brevitas and QKeras. However, any training framework is supported as long as it export QONNX.

Instalation: from pip

1. pip install chisel4ml.
2. Download a matching jar from github relases.
3. To test first run java -jar chisel4ml.jar (You can change the port and temporary directory using -p and -d (use --help for info)
4. Paste the Python code bellow into a file and run python script.py

import numpy as np
import qkeras
import tensorflow as tf
from chisel4ml import optimize, generate

w1 = np.array([[1, 2, 3, 4], [-4, -3, -2, -1], [2, -1, 1, 1]])
b1 = np.array([1, 2, 0, 1])
w2 = np.array([-1, 4, -3, -1]).reshape(4, 1)
b2 = np.array([2])

x = x_in = tf.keras.layers.Input(shape=3)
x = qkeras.QActivation(
    qkeras.quantized_bits(bits=4, integer=3, keep_negative=True)
)(x)
x = qkeras.QDense(
    4,
    kernel_quantizer=qkeras.quantized_bits(
        bits=4, integer=3, keep_negative=True, alpha=np.array([0.5, 0.25, 1, 0.25])
    ),
)(x)
x = qkeras.QActivation(qkeras.quantized_relu(bits=3, integer=3))(x)
x = qkeras.QDense(
    1,
    kernel_quantizer=qkeras.quantized_bits(
        bits=4, integer=3, keep_negative=True, alpha=np.array([0.125])
    ),
)(x)
x = qkeras.QActivation(qkeras.quantized_relu(bits=3, integer=3))(x)
model = tf.keras.Model(inputs=[x_in], outputs=[x])
model.compile()
model.layers[2].set_weights([w1, b1])
model.layers[4].set_weights([w2, b2])
data = np.array(
    [
        [0.0, 0.0, 0.0],
        [0.0, 1.0, 2.0],
        [2.0, 1.0, 0.0],
        [4.0, 4.0, 4.0],
        [7.0, 7.0, 7.0],
        [6.0, 0.0, 7.0],
        [3.0, 3.0, 3.0],
        [7.0, 0.0, 0.0],
        [0.0, 7.0, 0.0],
        [0.0, 0.0, 7.0],
    ]
)


opt_model = optimize.qkeras_model(model)
accelerators, lbir_model = generate.accelerators(
    model,
    minimize="delay"
)
circuit = generate.circuit(opt_model)
for x in data:
    sw_res = opt_model.predict(np.expand_dims(x, axis=0))
    hw_res = circuit(x)
    assert np.array_equal(sw_res.flatten(), hw_res.flatten())
circuit.delete_from_server()

This will generate a circuit of a simple two layer fully-connected neural network, and store it in /tmp/.chisel4ml/circuit0. If you have verilator installed you can also add the argument: use_verilator=True in the generate.circuit function. In the first case only a firrtl file be generated (this can be converted to verilog using firtool), if you use verilator, however, a SystemVerilog file will also be created.

chisel4ml also supports convolutional layers and maxpool layers. It also has some support for calculating FFTs and log-mel feature energy (audio features) in hardware.

Installation: from source

1. Install mill build tool.
2. Install python 3.8-3.10
3. Create environment python -m venv venv/
4. Activate environment (Linux)source venv/bin/activate
   • Windows .\venv\Scripts\activate
5. Upgrade pip python -m pip install --upgrade pip
6. Install chisel4ml pip install -ve .[dev]
7. Build Python protobuf code make
8. Build Scala code mill chisel4ml.assembly
9. In another terminal run tests pytest --use-verilator -n auto
   • The --use-verilator flag is optional if you have verilator installed, however it is highly recommended, since it is much faster.

ScalaDocs

To create ScalaDocs run mill chisel4ml.docJar and they will be generated to out/chisel4ml/docJar.dest/javadoc.