OxiBLAS 0.2.1 Released — Pure Rust BLAS/LAPACK That Outperforms OpenBLAS

The mathematical foundation of the COOLJAPAN scientific computing ecosystem just leveled up.

We released OxiBLAS 0.2.1 — a complete, production-grade reimplementation of BLAS (Basic Linear Algebra Subprograms) and LAPACK written entirely in Rust.

No C. No Fortran. No external shared libraries.
No FFI overhead. No build hell.
Just clean, memory-safe, blazing-fast linear algebra that compiles to a single static binary and runs everywhere.

Why OxiBLAS 0.2.1 is a game changer

For decades, high-performance numerical computing meant depending on battle-tested but heavy C/Fortran libraries like OpenBLAS, Intel MKL, or Apple Accelerate.

These tools are powerful but suffer from:

• Complex build systems and dependency hell
• Security risks from large native codebases
• Platform-specific binaries and vendor lock-in
• Difficulty in WASM, embedded, or no_std environments

OxiBLAS 0.2.1 ends all of that.

It delivers competitive or superior performance while being 100% memory-safe and portable.
Notable results:

• Apple M3 (AArch64): f32 GEMM reaches 172% of OpenBLAS
• x86_64: DGEMM matches or exceeds OpenBLAS on large matrices
• Factorization speedups: LU 14.5–23× faster, Cholesky 4.5–10.2× faster vs. naive implementations

Technical Deep Dive: How We Rebuilt BLAS/LAPACK in Pure Rust

The architecture mirrors industry best practices but is radically optimized for modern Rust:

1. SIMD Microkernel Layer (oxiblas-core)
   Custom intrinsics with runtime feature detection (AVX-512 → AVX2+FMA → NEON → scalar).

2. Runtime Auto-Tuner + Cache-Aware Blocking
   RuntimeAutoTuner selects optimal block sizes at runtime. BLIS-style macro/micro kernels with prefetching.

3. BLAS (oxiblas-blas)
   Full Level 1–3 + batched/strided operations, Einstein summation (24 patterns), packed/banded variants.

4. LAPACK (oxiblas-lapack)
   Blocked + recursive + parallel variants of LU, Cholesky, QR (WY representation fixed), SVD, EVD, Schur, mixed-precision refinement.

5. Sparse (oxiblas-sparse)
   9 formats (CSR, CSC, COO, …) with multifrontal factorizations and iterative solvers (GMRES, CG, BiCGStab, etc.).

Key Rust advantages:

• Zero-cost abstractions and static dispatch
• Safe abstractions over raw SIMD
• Full no_std + alloc support in core crates
• Arena allocation + index-based structures

What’s inside 0.2.1 (released March 16)

• Stability improvements and CI hardening
• Sparse test import fixes and BLAS autotune type-cast corrections
• Dependency updates
• Production readiness confirmed through 2,922 passing library tests + 287 doctests
• Enhanced documentation and regression benchmark framework

This is the foundation

OxiBLAS is now the official linear algebra backend for the entire COOLJAPAN scientific stack:

• SciRS2 / NumRS2 — all core numerical operations
• SkleaRS — ML primitives and algorithms
• OxiMedia — computer vision and tensor pipelines
• ToRSh / OxiRAG — high-performance tensor and vector math
• Future integration with OxiLean for formally verified numerics

Repository: https://github.com/cool-japan/oxiblas

Star the repo if you want high-performance scientific computing without the traditional toolchain headaches.

The era of “just link OpenBLAS” is coming to an end.

Pure Rust numerical linear algebra is here — and it’s fast, safe, and sovereign.

— KitaSan at COOLJAPAN OÜ March 16, 2026