Skip to content

tensil-ai/tensil

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Tensil

Build status

Tensil toolchain flow

Flow

Tutorials

For in-depth end-to-end instructions check our tutorials.

Documentation

For reference documentation see our website.

Setup

  1. Pull and run Tensil docker container (see below);
  2. Download and install Xilinx Vitis or Vivado;
  3. Download and install Xilinx PYNQ for your FPGA development platform;
  4. Copy Tensil PYNQ driver drivers/tcu_pynq to /home/xilinx/tcu_pynq on your FPGA development platform.

Pull and run docker container

docker pull tensilai/tensil
docker run -u $(id -u ${USER}):$(id -g ${USER}) -v $(pwd):/work -w /work -it tensilai/tensil bash

Compile AI/ML model

Compile AI/ML model (ResNet20 v2 CIFAR) for specific TCU architecture and FPGA development platform, PYNQ Z1 in this example.

From ONNX

tensil compile -a /demo/arch/pynqz1.tarch -m /demo/models/resnet20v2_cifar.onnx -o "Identity:0" -s true

From frozen TensorFlow graph

tensil compile -a /demo/arch/pynqz1.tarch -m /demo/models/resnet20v2_cifar.pb -o "Identity" -s true

Other ML frameworks are supported by converting to ONNX

Run bit accurate Tensil emulator

tensil emulate -m resnet20v2_cifar_onnx_pynqz1.tmodel -i /demo/models/data/resnet_input_1x32x32x8.csv

Make Verilog RTL

Make Verilog RTL for specific TCU architecture and FPGA development platform, PYNQ Z1 in this example.

tensil rtl -a /demo/arch/pynqz1.tarch -s true

Create Vivado design

Create Vivado design for specific FPGA development platform. We include detailed steps in our PYNQ Z1 tutorial. If you get stuck, we can help! Please reach out to us at [email protected] or in Discord.

PYNQ Z1 design

Run AI/ML model on FPGA

Use PYNQ and Jupyter notebooks to run AI/ML model on FPGA. (See in notebooks.)

Resnet on PYNQ

For maintainers

Additional setup steps

  1. Download and install OpenJDK 11 from Azul;
  2. Download and install Verilator;
  3. Download test models:
wget https://github.com/tensil-ai/tensil-models/archive/main.tar.gz
tar xf main.tar.gz
mv tensil-models-main models
rm main.tar.gz

Run RTL tool from source code

./mill rtl.run -a ./arch/pynqz1.tarch -s true

Run compiler from source code

./mill compiler.run -a ./arch/pynqz1.tarch -m ./models/resnet20v2_cifar.onnx -o "Identity:0" -s true

Run emulator from source code

./mill emulator.run -m resnet20v2_cifar_onnx_pynqz1.tmodel -i ./models/data/resnet_input_1x32x32x8.csv

Run full test suite

./mill __.test -l org.scalatest.tags.Slow

To run a single RTL test in, for example, the Accumulator module, and also output a VCD file, do:

./mill rtl.test.testOnly tensil.tcu.AccumulatorSpec -- -DwriteVcd=true -z "should accumulate values"

View VCD files

To view the latest VCD file generated:

./scripts/gtkwave/display-latest-vcd.py

To view a specific VCD file:

./scripts/gtkwave/display-vcd.sh <vcd_file>

Build and push AWS ECS image for web compiler (Tensil Explorer)

docker build -f docker/web/Dockerfile -t tensil-web-compiler .
aws ecr get-login-password | docker login --username AWS --password-stdin <ACCOUNT ID>.dkr.ecr.<REGION>.amazonaws.com
docker tag tensil-web-compiler <ACCOUNT ID>.dkr.ecr.<REGION>.amazonaws.com/tf2rtl-web-compiler
docker push <ACCOUNT ID>.dkr.ecr.<REGION>.amazonaws.com/tf2rtl-web-compiler

Get help