Lib.rs

Rookie attempt to rewriting BLAS FORTRAN 77 and Intel Math Library Kernels in modern native rust. ONLY x86_64

refs I took:

• https://www.netlib.org/blas/ ← good for overview
• https://www.netlib.org/lapack/explore-html/ ←/
• https://icl.utk.edu/~mgates3/docs/lapack.html ←/
• https://suif.stanford.edu/papers/lam-asplos91.pdf <- good for understanding cache blocking techniques & optimizations
• https://www.intel.com/content/www/us/en/docs/onemkl/developer-reference-dpcpp/2025-2/blas-routines.html ← good details, very clear
• https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html ← good man for intrinsics
• https://doc.rust-lang.org/core/arch/x86_64/index.html#functions ←/
• https://public.dhe.ibm.com/software/dw/cell/BLAS_Prog_Guide_API_v3.1.pdf <- api ref, mostly similar everywhere

TODO:

• lvl1: rotmg, rotm
• lvl2: gemv - all
• lvl3: gemm - all
• handle NaN, over/underflow, return vs panic and many more edge cases :(
• test code ref from this repo, currently ai-generated
• multithreading, GPU maybe (NOT CUDA)?
• LESS BAD BRANCHING, more SIMD, LESS FN CALLS, better DATA_READ, better CACHE
• MAKE GEMV & GEMM FAST FOR ALL PATHS!!!
• bench is somewhat UNFAIR AND NOT UNIFORM

To bench, run harness using cargo bench --bench bencher <kernel> or all ref and build using RUSTFLAGS target-cpu=x86-64-v3, otherwise rustc will throw error.

Install Intel oneAPI Toolkit then copy the .dll or .so into the project root:

• windows > Copy-Item "C:\Program Files (x86)\Intel\oneAPI\compiler\2026.0\bin\libiomp5md.dll" .
• linux > cp /home/ronakgh97/intel/oneapi/compiler/latest/lib/libiomp5.so .
• mac > sell that shit

All are single threaded, ran on i7 14650hx, rust 1.97.1, last updated on 8/7/2026

Axpy

Dot

Gemv

Gemv_T

Gemm_F_F

Wanna improve the bench and added more kernels perf test? Send me a PR