tiny-dpcpp-nn

The tiny-dpcpp-nn library is a library for the fast evaluation of inference and training of neural networks in which each layer has the same, small width. This implementation utilizes Intel’s DPC++ and, in particular, the joint_matrix extension to utilize Intel’s AMX and XMX hardware to maximize performance.

License

BSD 3-Clause License

• Building and linking
  • Dependencies
  • Build from source using oneAPI
  • Build options
• Network
  • Multi-Layer Perceptrons
  • Fused MLPs
  • ESIMD Kernels
• Encodings
  • Identity Encoding
  • Spherical Harmonics Encoding
  • Grid Encoding
• Benchmarks
• Examples
  • Gradient Descent Test Script
  • Model Training and Evaluation Script
  • Image Learning with PyTorch Extensions Script
• Contributing
  • Structure of the repository
  • Structure of tiny-dpcpp-nn/include
• tiny_dpcpp_nn_pybind Module Documentation
  • Installation
  • Usage