Architecture

Aicraft is a vertically integrated stack. No external libraries sit between your code and the hardware.

Stack Overview

┌─────────────────────────────────┐
│       Your Application          │  main.c
├─────────────────────────────────┤
│          aicraft.h              │  Single include
├─────────────────────────────────┤
│   Layers / Loss / Optimizer     │  High-level API
├─────────────────────────────────┤
│       Autograd Engine           │  22 ops, DAG-based
├─────────────────────────────────┤
│        Tensor Core              │  N-dim, broadcasting
├──────────────┬──────────────────┤
│ SIMD Kernels │  Vulkan Compute  │  Backends
│ AVX-512/NEON │  14 GLSL shaders │
├──────────────┴──────────────────┤
│       Arena Allocator           │  Checkpoint/restore
└─────────────────────────────────┘

Design Principles

1. Single header — #include "aicraft/aicraft.h" pulls in everything
2. Zero allocations in the hot path — the arena allocator pre-reserves memory
3. Compile-time dispatch — SIMD backend is selected at compile time via macros
4. DAG-based autograd — reverse-mode autodiff with topological sorting
5. Vulkan compute — optional GPU path with zero CUDA dependency

Memory Model

Aicraft uses a custom arena allocator with checkpoint/restore semantics:

ac_mem_checkpoint();    // Save current memory state
// ... allocate tensors, run forward/backward ...
ac_mem_restore();       // Free everything since checkpoint

This gives constant-memory training with zero per-tensor malloc/free calls.

Tensor Layout

Tensors are stored in row-major (C-contiguous) order with support for:

• Arbitrary dimensions (N-dim)
• Broadcasting
• Strides and views (zero-copy slicing)
• Optional gradient tracking for autograd

Backend Selection

Backend	Flag	Platforms
Scalar (fallback)	default	All
SSE4.2	-msse4.2	x86/x86_64
AVX2	-mavx2 -mfma	x86_64
AVX-512	-mavx512f	x86_64 (Skylake+)
ARM NEON	-mfpu=neon	ARM/AArch64
Vulkan	-lvulkan	Any GPU vendor