Hnswlib with JNA (Java Native Access)

This project contains a JNA (Java Native Access) implementation built on top of the native Hnswlib (Hierarchical Navigable Small World Graph) which offers a fast approximate nearest neighbor search. It includes some modifications and simplifications in order to provide Hnswlib features with native like performance to applications written in Java. Differently from the original Python implementation, the multi-thread support is not included in the bindings itself but it can be easily implemented on the Java side. Hnswlib-jna works in collaboration with a shared library which contains the native code. For more information, please check the sections below.

Dependencies

Pre-Generated Shared Library

The jar file includes some pre-generated libraries for Windows, Debian Linux and MacOS (x86-64) which should allow an easy integration and abstract all complexity related to compilation. An extra library for Debian Linux (aarch64) is also available for tests with AWS Graviton 2. In the case of operating system issues, a runtime exception will be thrown and the manual compilation will be advised.

On Windows, the Build Tools for Visual Studio 2019 (C++ build tools) is required.

Using in Your Project

Add the following dependency in your pom.xml:

<dependency>
    <groupId>com.stepstone.search.hnswlib.jna</groupId>
    <artifactId>hnswlib-jna</artifactId>
    <version>1.4.0</version>
</dependency>

For more information and implementation details, please check hnswlib-jna-example.

Manual Compilation (Whenever it is advised)

This section includes more information about how to compile the shared libraries on Windows, Linux and Mac for different architectures (e.g., x86-64, aarch64). If you were able to run the example project on your PC, this section can be ignored.

Compiling the Shared Library

To generate the shared library required by this project, binding.cpp needs to be compiled using a C compiler (e.g., clang or gcc) with C++11 support, at least. The library can be generated with clang via:

clang++ -O3 -shared bindings.cpp -I hnswlib -o <project_folder>/lib/libhnswlib-jna-x86-64.dylib

Note: The shared library's name must be: libhnswlib-jna-ARCH.EXT where ARCH is the canonical architecture name (e.g., x86-64 for AMD64, or aarch64 for ARM64) and EXT is dylib for MacOS, for windows use dll, and linux so.

Instructions for Windows

Using Visual Studio Build Tools

Download and install LLVM;
Download and install Build Tools for Visual Studio 2019 (or higher);
Compile the bindings using clang:

clang++ -O3 -shared bindings.cpp -I hnswlib -o <project_folder>/lib/libhnswlib-jna-x86-64.dll

This procedure will generate the 3 necessary files: libhnswlib-jna-x86-64.dll, libhnswlib-jna-x86-64.exp and libhnswlib-jna-x86-64.lib.

Using MinGW64

Download and install LLVM;
Make sure that LLVM's bin folder is in your PATH;
Download MinGW-w64 with Headers for Clang;
Unpack the archive and include MinGW64's bin folder into your PATH as well;
Compile the bindings using clang:

clang++ -O3 -target x86_64-pc-windows-gnu -shared bindings.cpp -I hnswlib -o <project_folder>lib/libhnswlib-jna-x86-64.dll -lpthread

This procedure will generate libhnswlib-jna-x86-64.dll.

Instructions for Linux

Download and install clang (older versions might trigger compilation issues, so it is better use a recent version);
Compile the bindings using clang:

clang++ -O3 -fPIC -shared -std=c++11 bindings.cpp -I hnswlib -o <project_folder>/lib/libhnswlib-jna-x86-64.so

This procedure will generate libhnswlib-jna-x86-64.so.

Reading the Shared Library in Your Project

Once the shared library is available, it is necessary to tell the JVM and JNA where it is located. This can be done by setting the property jna.library.path via JVM parameters or system properties.

Via JVM parameters

-Djna.library.path=<project_folder>/lib

Programmatically via System Class

System.setProperty("jna.library.path", "<project_folder>/lib");