Axiom

Axiom is a high-performance, open-source C++ tensor library designed to provide the simplicity of NumPy and PyTorch in native code. With cutting-edge features such as SIMD vectorization, BLAS acceleration, and Metal GPU support, Axiom ensures HPC-grade performance along with an intuitive API that aligns closely with what Python developers are familiar with.

Key Features

• Python-like API: Axiom simplifies tensor operations through familiar constructs such as operator overloading and method chaining, making it easy for developers accustomed to NumPy and PyTorch to get started.
• Exceptional Performance: Utilize the power of Accelerate, OpenBLAS, and specially optimized SIMD kernels to achieve high throughput.
• Comprehensive Vectorization: Take advantage of advanced vectorization techniques across various architectures with support for SSE, AVX, AVX-512, ARM NEON, RISC-V, and more.
• GPU Acceleration: Leverage full GPU support via Metal Performance Shaders (MPSGraph) for efficient tensor operations beyond just matrix multiplication.
• Secure Tensor Manipulation: Enjoy a safe programming environment with features like NaN/Inf guards and shape assertions to validate tensor operations.
• Cross-platform Compatibility: Axiom is designed for seamless deployment across platforms with dynamically linked BLAS backends, ensuring consistent performance.

Usage Examples

Get a glimpse of Axiom’s intuitive syntax:

// NumPy: x = np.where(x > 0, x, 0)
auto x = Tensor::where(x > 0, x, 0);

// NumPy: y = x.reshape(2, -1).T
auto y = x.reshape({2, -1}).T();

// PyTorch: z = F.softmax(scores, dim=-1)
auto z = scores.softmax(-1);

For performance insights, Axiom has demonstrated over 3500+ GFLOPS on M4 Pro, substantially outperforming libraries like Eigen and PyTorch in benchmarks.

Advanced Functionalities

Axiom also integrates powerful features for complex tensor manipulations with functionalities such as:

• Einops Integration: Use semantic patterns for tensor rearrangement and reduction, creating readable and expressive code.
• Full LAPACK Coverage: Access advanced linear algebra operations including singular value decomposition, eigenvalue computations, and matrix solvers.
• Efficient I/O: Save and load tensors in optimized formats like FlatBuffers and NumPy, facilitating swift data interchange between C++ and Python.

Axiom is built for developers seeking the performance capabilities of C++ while enjoying the user-friendly experience of high-level tensor libraries. Explore Axiom’s Quick Start guide and Benchmarks section to see the library in action.