qconv is a performant C++ library using AVX2 intrinsics to implement fully connected layers and quantitized convolutional layers with kernel 3x3. All of the implemented layers are tested by google tests, and benchmarked by google benchmark.

Supported os: Linux

Known dependencies:
cmake
ninja
clang

sudo snap install cmake --classic
sudo apt-get -y install ninja-build
sudo apt install clang
sudo apt install libstdc++-12-dev
sudo apt-get install libc++-dev
sudo apt install libc++abi-dev

A Docker image can be builded and runned with the following commands (casted from the root of the repo):
sudo ./source/scripts/misc/docker/build.sh
sudo ./source/scripts/misc/docker/run.sh

After the setup is done, the repo can be configured, builded, tested and benchmarked with the following commands (casted from the root of the repo):
./source/scripts/configure.sh release
./source/scripts/build.sh release
./source/scripts/test.sh release
./source/scripts/bench.sh release

The default compiler of qconv is clang. gcc is also perfectly fine to compile qconv, the only advantage of clang is that experience shows that the google benchmark results are more consistent with the clang compiled binaries than with the gcc compiled binaries. Compilers can be switched by changing in ./source/scripts/build/init.sh CMAKE_CXX_COMPILER=clang++
to
CMAKE_CXX_COMPILER=g++