A state-of-the-art C++ tensor library optimized for Apple Silicon, providing full NumPy syntax compatibility and einops-style tensor rearrangement for seamless on-device AI inference.

Features

• 🚀 Apple Silicon Optimized - Leverages Metal Performance Shaders and ARM NEON vectorization
• 🐍 NumPy Compatible - Full syntax compatibility for easy Python-to-C++ translation
• 🔄 Einops Support - Intuitive tensor rearrangement with einops-style operations
• 📊 ONNX Ready - Seamless integration with ONNX model inference pipelines
• ⚡ Zero-Copy Operations - Memory-efficient tensor operations where possible
• 🎯 Type Safety - Modern C++20 with compile-time shape checking

Quick Start

Prerequisites

• macOS 11.0+ (Big Sur) with Apple Silicon
• Xcode 13+ or Clang 13+
• CMake 3.20+
• Metal Performance Shaders Framework

Build Instructions

# Clone the repository
git clone https://github.com/your-username/axiom.git
cd axiom

# Create build directory
mkdir build && cd build

# Configure with CMake
cmake .. -DCMAKE_BUILD_TYPE=Release -DAXIOM_BUILD_TESTS=ON

# Build
make -j$(sysctl -n hw.ncpu)

# Run tests
make test

Installation

# Install system-wide
sudo make install

# Or use in your CMake project
find_package(Axiom REQUIRED)
target_link_libraries(your_target Axiom::axiom)

Usage Example

#include <axiom/tensor.hpp>
#include <axiom/ops.hpp>

using namespace axiom;

int main() {
    // Create tensors with NumPy-like syntax
    auto x = tensor::randn({64, 128, 256});
    auto y = tensor::ones({256, 512});
    
    // Matrix operations
    auto result = x.matmul(y);
    
    // Einops-style rearrangement
    auto reshaped = x.rearrange("b h w -> b (h w)");
    
    // Broadcasting and element-wise ops
    auto scaled = (x * 2.0f + 1.0f).relu();
    
    return 0;
}

Development Roadmap

Phase 1: Core Foundation

• Basic tensor class with shape/stride handling
• Memory management and storage abstraction
• CMake build system setup
• Basic arithmetic operations (+, -, *, /)

Phase 2: Operations & Backends

• Metal backend implementation
• Linear algebra operations (matmul, dot, etc.)
• Reduction operations (sum, mean, max, etc.)
• Broadcasting semantics

Phase 3: Advanced Features

• Einops-style rearrangement parser
• NumPy compatibility layer
• Automatic differentiation (optional)
• ONNX integration utilities

Phase 4: Optimization

• Memory pool allocation
• Kernel fusion optimizations
• Multi-threading for CPU operations
• Benchmarking and profiling tools

Performance Goals

• Memory Bandwidth: >90% of theoretical peak on Apple Silicon
• Compute Utilization: >85% GPU utilization for large tensor operations
• Latency: <1ms overhead for tensor creation and basic operations
• Compatibility: 100% pass rate on NumPy compatibility test suite

License

MIT License - see LICENSE file for details.

Citation

@misc{axiom2025,
  title={Axiom: High-Performance Tensor Library for Apple Silicon},
  author={Noah Kay},
  year={2025},
  url={https://github.com/frikallo/axiom}
}