Tensorium-MLIR v0.1.0-prealpha

Tensorium-MLIR v0.1.0-prealpha

First pre-alpha release of the Tensorium / Relativity-MLIR toolchain.
This version establishes the foundations of a symbolic → MLIR → LLVM → executable pipeline and demonstrates Kerr–Schild–aware code generation.

Overview

• Initial Relativity dialect with a minimal set of tensor types and ops.
• Specialized rewrite for Kerr–Schild spatial metric (rank-1 update).
• Lowering to vector with fast-math attributes for CPU vectorization (AVX2/FMA).
• Example and tests for Schwarzschild (Kerr–Schild coordinates).

Contents

• Dialect
  • rel.ks_spatial_metric: produces spatial metric γ, its determinant det(γ), and inverse γ⁻¹ from Cartesian coordinates (M=1 by default).
• Passes
  • Kerr–Schild specialization (det/inv via rank-1 formulas).
  • Vector lowering and alignment hints suitable for SLP/AVX2.
• Examples/Tests
  • Schwarzschild KS at sample points: prints γ, det(γ), γ⁻¹ and verifies expected identities.

Status

• Pre-alpha. API and IR are unstable and subject to change.
• Not a full BSSN stack yet.
• Next milestones:
  • BSSN initialization (χ/φ, γ̃, Ã, Γ̃, K).
  • Finite-difference operators (2nd/4th/6th order) and simple RK time integrator.
  • Constraint monitors and convergence tests.

Build

git clone https://github.com/<org>/<repo>.git
cd <repo>
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
cmake --build . -j

Documentation

The README provides a quick start and notes on the pass pipeline. But the entire PDF with all the explanations is coming soon !

Citation

Louis Touzalin. Tensorium-MLIR v0.1.0-prealpha (2025).

License

MIT (see LICENSE).