Runtime for Gati CNN accelerator¶

Build¶

Install Dependencies¶

Arch:

sudo pacman -S python3 python-pip pkg-config cmake

Fedora:

sudo dnf install python3-devel python-pip cmake

Ubuntu/Debian:

sudo apt install python3-dev python3 python3-pip pkg-config cmake

MacOs

brew install python pkg-config cmake

On board¶

On the board, you would need additional dependencies that allow CPU-FPGA communication:

First, check if “Vaaman FPGA communication” is checked in the overlay config, find a detailed how-to here. Consider, rebooting after this.

Next, add Vicharak’s apt repo to your board’s apt config. Follow this guide to do this.

After this is setup, run:

sudo apt update && sudo apt upgrade
sudo apt install rah-service

Create and activate a Python virtual environment:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Compile¶

cd /path/to/gaticc
./scripts/install_deps.sh
mkdir build
cmake -B build
cmake --build build && sudo cmake --install build
pip install -e .

Compile with debug flags:

cmake -B build -DCMAKE_BUILD_TYPE=Debug

Usage¶

See,

gaticc -h

for usage instructions.

[!NOTE] If you encounter:
gaticc: error while loading shared libraries: libprotobuf-lite.so.32: cannot open shared object file: No such file or directory
Refresh the system library cache:
sudo ldconfig
Then try running gaticc -h again.

Python Interface¶

Here’s an example script to run simulation of a model (install model files from the model zoo):

import gati

onnx_path = "tests/models/mnist_6_28_int8.onnx"
print(np.argmax(np.squeeze(gati.sim(onnx_path, np.load("mnist_2.py")), axis=1), axis=-1))

For more examples, checkout examples/ directory. It contains, scripts to compile, run, summarize etc.

For full api doc, read python/gati.py (or feed it to your favorite LLM).