Caten.py

Caten is a Python-based Polyhedral Compiler framework designed for deep learning and high-performance computing education and experimentation. It provides Python bindings for the Integer Set Library (ISL) and high-level abstractions for tensor scheduling and optimization.

Vision

Caten aims to bridge the gap between high-level tensor operations (like in PyTorch/NumPy) and low-level loop optimizations (Tiling, Fusion, Vectorization). By leveraging the Polyhedral Model, Caten allows users to:

• Visualize and manipulate loop structures as mathematical objects (Schedule Trees).
• Perform complex transformations like Conv+Pool Fusion or Symbolic Tiling using a Pythonic API.
• Generate optimized code for various backends (CPU, CUDA, Metal).

Features

• caten.isl: Auto-generated, type-safe Python bindings for ISL. No manual C-types management required.
• caten.polyhedral (Planned): A high-level DSL for schedule construction and manipulation.
• Implicit Context: ISL context is managed automatically; no boilerplate code.
• Operator Overloading: Use standard Python operators (|, &, -, +) for set/map operations.

Documentation

• Project Overview & Architecture
• Polyhedral Loop Transformation Spec
• IR & Class Design
• ISL API Coverage

Tutorial

Check out examples/isl_tutorial.ipynb for a hands-on introduction to using ISL bindings in Python.

Installation

Requirements: libisl must be installed (e.g., brew install isl).

# Clone the repository
git clone https://github.com/hikettei/Caten.py.git
cd Caten.py

# Install dependencies
uv sync

Roadmap

1. ISL Bindings: Complete caten.isl.
2. Polyhedral DSL: Implement caten.polyhedral.
3. IR & Kernel: Implement caten.ops and caten.kernel.
4. Runtime & Renderer: Code generation for CPU/C.
5. Auto-Scheduler: Basic search for optimal schedules.

License

MIT