SIMD-accelerated similarity measures, metrics, distance functions for x86 and Arm. Tuned for Machine Learning applications, mid-size vectors with 100-512 dimensions.

Need something like this in your CMake-based project?

FetchContent_Declare(
    simsimd
    GIT_REPOSITORY https://github.com/ashvardanian/simsimd.git
    GIT_SHALLOW TRUE
)
FetchContent_MakeAvailable(simsimd)
include_directories(${simsimd_SOURCE_DIR}/include)

By default, we use GCC12, -O3, -march=native for benchmarks. Serial versions imply auto-vectorization pragmas.

Cosine distance performance on a single core of a 64-core Arm-based "Graviton 3" CPUs powering AWS c7g.metal instances:

We only use Arm NEON implementation with vectors lengths that are multiples of 128 bits, avoiding any additional head or tail for loops for misaligned data. SVE looses to NEON on very short vectors, but outperforms on longer sequences.

On the x86 AMD Zen2 cores making up the 64-core Threadripper PRO 3995WX the numbers are:

The gap between auto-vectorized code and directly using 128-bit registers is much less pronounced. With AVX2 and 256-bit registers the results should be better, but would be less broadly applicable.

To replicate on your hardware, please run:

cmake -DCMAKE_BUILD_TYPE=Release -DUNSW_BUILD_BENCHMARKS=1 -B ./build && make -C ./build && ./build/simsimd_bench