A complete deep-learning framework written entirely in pure C. SIMD-optimised, Vulkan-accelerated, header-only. From training to edge inference in a single #include.
$ git clone https://github.com/miaototi/Aicraft.git && cd Aicraft#include "aicraft/aicraft.h" and compile. That's it.Official Teaser
Dead simple
No CMake, no vcpkg, no conan. Drop the header folder into your project, pass -I./include, and build. One translation unit, zero friction.
#include "aicraft/aicraft.h"
int main(void) { ac_init();
// Build a feedforward network AcLayer *net[] = { ac_dense(784, 128, AC_RELU), ac_dense(128, 10, AC_SOFTMAX) };
// Forward + backprop in one line AcTensor *x = ac_tensor_rand((int[]){1,784}, 2); AcTensor *y = ac_forward_seq(net, 2, x); ac_backward(y);
ac_cleanup(); return 0;}Architecture
Aicraft is a vertically integrated stack. No external libraries sit between your code and the hardware.
Capabilities
AVX2, AVX-512, ARM NEON. Every hot path hand-tuned with platform intrinsics and BLIS-style GEMM micro-kernels.
14 GLSL compute shaders for GEMM, activations, and reductions. Cross-vendor GPU acceleration.
22 differentiable ops. Dynamic computational graph with reverse-mode autodiff and O(1) cycle detection.
Post-training quantisation with asymmetric per-tensor scaling. ~4x model compression for edge.
Checkpoint/restore memory management. Zero per-tensor malloc. Constant memory in training.
SGD, Adam, AdamW optimisers. Cross-entropy, MSE, Huber loss. Full training pipeline.
How it works
Add the single header to your C project. No build system changes needed.
Stack layers, pick a loss function and optimiser. Just like Python, but in C.
Forward, backward, step. The autograd engine handles gradient computation.
Quantise to INT8, serialise, and run on anything from x86 to ARM Cortex-M.
How it compares
The best dependency is the one you never add. Aicraft proves you can train a neural network without pulling half the internet into your build.
Explore
Open source
Read the docs, explore the source, or start building.
A project by Tobias Tesauri — T&M Softwares