Contributing

Structure of the repository

The repository consists of several directories: - benchmarks: Contains codes for default performance benchmarks for inference and training on artificial data. - cmake: Contains files required for the cmake build system like common options and .cmake files to find packages. - docs: Contains this documentation. - dpcpp_bindings: Contains the Python - C++ bindings to use tiny-dpcpp-nn in Python. - extern: Contains all extern dependencies either by providing the files directly (like doctest and json) or as submodules. - include: All the header files for this project. Most of the code is contained in the headers since it heavily relies on template classes and functions. - python: Python code. This will be removed in the future. - source: Contains all the .cpp files of the tiny-dpcpp-nn code. - test: Directory which contains all of the doctest unit-test.

Structure of tiny-dpcpp-nn/include

The most important parts of the code as of version 0.1 are contained in include/common, include/encodings and include/network.

The include/common directory contains small helper functions as well as the containers wrapping device memory, namely the classes DeviceMem, DeviceMatrix, and DeviceMatrices, as well as the view classes DeviceMatrixView and DeviceMatricesView. These latter two classes may be used in device kernels, whereas the classes DeviceMem, DeviceMatrix, and DeviceMatrices are only usable on the host.

The include/encodings directory contains all the available encodings, namely Identity, SphericalHarmonics and Grid. These encodings may be used to pad the input data to the required width before running inference or training.

Different loss functions are implemented in include/losses. They are all derived from on abstract loss base class which is defined in include/losses/loss.h

The include/network directory contains the actual network implementation called SwiftNetMLP. The relevant GPU kernels which facilitate the actual training and inference can be found in kernel_* files. As of version 0.1, we only support an esimd verison which can be found in kernel_esimd.h The class network_with_encodings wraps a network and an ecoding in a single object. At some point in the future, this should be consolidated.

Several optimization algorithms for training may be found in include/optmizers. As of version 0.1, they are not yet implemented.