Aicraft
Skip to main content

Changelog

All notable changes to Aicraft are documented here.

Format: Semantic Versioning

[1.0.0] — 2026-03-01​

Initial public release 🎉

Core​

  • Tensor engine: N-dimensional tensors with broadcasting, slicing, and reshape
  • Autograd: Reverse-mode automatic differentiation with 22 differentiable ops
  • Arena allocator: Checkpoint/restore memory management with zero per-tensor malloc

Layers​

  • ac_dense — Fully-connected layer with configurable activation
  • ac_conv2d — 2D convolution (coming in v1.1)
  • ac_batchnorm — Batch normalisation
  • ac_dropout — Dropout regularisation

Activations​

  • ReLU, LeakyReLU, Sigmoid, Tanh, Softmax, GELU

Loss Functions​

  • Mean Squared Error (MSE)
  • Cross-Entropy
  • Binary Cross-Entropy
  • Huber Loss

Optimisers​

  • SGD, SGD with Momentum
  • Adam, AdamW

Backends​

  • SIMD: AVX2, AVX-512, ARM NEON hand-tuned kernels
  • Vulkan: 14 GLSL compute shaders for GPU acceleration

Quantisation​

  • Post-training INT8 quantisation with asymmetric per-tensor scaling

Serialisation​

  • Binary format for weights (ac_save_weights / ac_load_weights)

[Unreleased]​

Planned for v1.1​

  • Conv2D layer with im2col optimisation
  • MaxPool2D / AvgPool2D layers
  • RNN / LSTM cells
  • ONNX export

Planned for v1.2​

  • Automatic mixed precision (FP16/BF16)
  • Multi-GPU support (Vulkan)
  • Metal backend for Apple Silicon
  • WebAssembly build target

How to Upgrade​

Aicraft is header-only, so upgrading is simple:

cd Aicraft
git pull origin main

Then recompile your project. No ABI concerns.

Reporting Issues​

Found a bug? Please open an issue at:

https://github.com/TobiasTesauri/Aicraft/issues