NumKong

Kahan floats are also commonly used in such cases, but I believe there is room for improvement without hitting those extremes. First of all, we should tune the epsilon here: https://github.com/ashvardanian/SimSIMD/blob/f8ff727dcddcd14... As for the 64-bit version, its harder, as the higher-precision `rsqrt` approximations are only available with "AVX512ER". I'm not sure which CPUs support that, but its not available on Sapphire Rapids.

Well now I know I have a bunch of software I want to try https://github.com/unum-cloud/. The author is definitely brilliant. I should benchmark https://github.com/ashvardanian/SimSIMD/blob/main/include/si... and see if it's any faster than what hnswlib currently uses for bruteforce IP search because I'm currently using hnswlib in something and it's become a bottleneck. Or maybe I'll just try usearch directly.

There are several `do {} while ()` loops in the SimSIMD codebase, mostly in SVE [1]. One of the reasons I don't like putting them in many places - mostly in cases with complex conditions, but also to avoid adding one more level of nesting for that scope. [1]: https://github.com/ashvardanian/SimSIMD/blob/f8ff727dcddcd14...

I encourage one to merge into e.g. {NumPy, SciPy, }; are there PRs? Though SymPy.physics only yet supports X,Y,Z vectors and doesn't mention e.g. "jaccard"?, FWIW: https://docs.sympy.org/latest/modules/physics/vector/vectors... https://docs.sympy.org/latest/modules/physics/vector/fields.... #cfd include/simsimd/simsimd.h: https://github.com/ashvardanian/SimSIMD/blob/main/include/si... conda-forge maintainer docs > Switching BLAS implementation: https://conda-forge.org/docs/maintainer/knowledge_base.html#... : conda install "libblas=*=*mkl" conda install "libblas=*=*openblas" conda install "libblas=*=*blis" conda install "libblas=*=*accelerate" conda install "libblas=*=*netlib" numpy-feedstock: https://github.com/conda-forge/numpy-feedstock/blob/main/rec... scipy-feedstock: https://github.com/conda-forge/scipy-feedstock/blob/ma

Over the years, the merits of Intel’s 512-bit Advanced Vector eXtensions (AVX-512) have been extensively debated. Introduced in 2014, CPUs took time to offer robust support. In parallel, Arm Scalable Vector Extensions (SVE), targeting Arm servers, has only gained traction in recent times. Today, the landscape has shifted significantly with Intel’s Sapphire Rapids CPUs on one flank and the AWS Graviton 3 and Ampere Altra chips on the other. Here are compelling reasons to opt for these over the traditional AVX2 and NEON extensions: 1. Masked Loads: Efficient data processing by selectively loading data. 2. Half-Precision Floating Point Math: Accelerated computations with reduced memory footprint. These features proved invaluable for the latest SimSIMD release. The software now processes vector similarities up to 300x faster using NEON, SVE,

It’s a good start, but you can’t generally get even remotely close to hardware potential in Rust, let alone Python. I had to implement a separate C99 library to always trigger the newest SIMD intrinsics, occasionally leveraging SVE on more recent ARM CPUs, that compilers don’t know how to generate. That library is in turn used in USearch, which is designed for Approximate Search, but some users recently reported that they use it for brute force as well… where it performed 20x faster than FAISS. https://github.com/ashvardanian/simsimd

We don't use BLAS. Why? BLAS helps with matrix-matrix multiplications, if you feel lazy and don't want to write the matrix tiling code manually. They bring essentially nothing of value in vector-vector operations, as compilers can properly auto-vectorize simple dot products... Moreover, they generally only target single and double precision, while we often prefer half or quarter precision. All in all, meaningless dependency. What do we use? I wrote a tiny package called SimSIMD. It's idea is to utilize less common SIMD instructions, especially in mixed-typed computations, that are hard for compilers to optimize. It was also a fun exercise to evaluate the performance of new SVE instruction on recent Arm CPUs, like the Graviton 3. You can find the code, the benchmarks, and the results in the repo: https://github.com/ashvardanian/simsimd Sti