This Go package implements Manhattan Distance and Cosine Similarity for 32-bit floats. Under the hood it will use SIMD instructions to accelerate the computations when running on ARM64. Alternatively if you build for a non-ARM64 platform or with CGO disabled it will use pure Go implementations of the functions. I hope to add support for AVX (AMD64 SIMD instructions) at a later date.

The NEON intrinsic versions are significantly faster:

$ go test -bench=.
...
cpu: Apple M1
BenchmarkCosineSimilarity-8               895324              1330 ns/op
BenchmarkCosineSimilarityPureGo-8         241170              4969 ns/op
BenchmarkManhattanDistance-8             1223034               980.6 ns/op
BenchmarkManhattanDistancePureGo-8        318074              3742 ns/op

BenchmarkCosineSimilarityPureGo and BenchmarkManhattanDistancePureGo benchmark the pure Go fallback implementations.

Use CosineSimilarity to compute the cosine similarity between two vectors, and ManhattanDistance to compute the 'Manhattan distance' between them. The vectors are represented as []float32, and the routines will panic if the input slices are different lengths.

None of the functions handle NaN's, so garbage in will give you garbage output.

The optimized versions that use SIMD are implemented in C files via intrinsics. They come with their own unit tests which can run via the Makefile. The unit tests use doctest. Currently the Makefile assumes you have a recent version of clang++ with support for NEON intrinsics.

• amd64 AVX intrinsics acceleration of Manhattan Distance and Cosine Similarity.
• float64 support