Trinity Project [ASPLOS 2026]

Three-Dimensional Tensor Program Optimization via Tile-level Equality Saturation

Trinity is the first tensor program optimizer that achieves scalable joint optimization through tile-level equality saturation. Trinity's IR can capture the essence of all three optimization axes (algebraic equivalence, memory I/O, compute orchestration). By leveraging equality saturation, Trinity enables scalable joint optimization across the entire graph.

Project Structure

Trinity-AE/
├── frontend/       # PyTorch model → Trinity IR conversion
├── optimizer/      # Tile-level equality saturation (Rust)
├── backend/        # IR → Triton kernel generation & profiling
├── trinity/        # End-to-end pipeline automation
└── scripts/        # Model definitions & run example pipeline

For detailed documentation, see the README in each directory:

• scripts/ — Setup, usage guide, API reference, and config options
• frontend/ — IR conversion details
• optimizer/ — Equality saturation engine
• backend/ — Triton code generation and profiling

Setup Note

The frontend now expects a vendored CPU TVM wheel under third_party/wheels/. Run ./scripts/setup.sh after placing the wheel there.