The JNI bindings are built with clang++ for Mac (arm64) and for Linux (x86). The build runs on a Mac and uses Docker for cross compilation. The docker image used for building (and possible at runtime) is defined in docker-runtime/Dockerfile. This image is pushed to the Docker Hub (https://hub.docker.com/r/pityka/base-ubuntu-libtorch/tags?page=1&ordering=last_updated), and by default the build will pull it.

• needs a macs with clang++ installed and set up for cross compiling.
• needs docker on the mac
• on the mac needs the libtorch shared libraries in /usr/local/lib/. Read on.
• needs bloop (https://scalacenter.github.io/bloop/) and sbt (https://www.scala-sbt.org/).

1. make prepare - this creates bloop build definitions from sbt the sbt build definition
2. Copy from one to the other
3. make test
4. make test-linux - builds for linux and runs tests in a linux container (needs docker for mac, will pull pityka/base-ubuntu-libtorch from docker hub)
5. make test-cuda - runs tests in a cuda enabled remote docker context

• make publishLocal - publishes artifacts locally
• make publish - publishes artifacts to github packages.
• bash publish.sh - publishes the artifacts to Maven Central

The Makefile will generate Java and C++ sources and create a JNI native library.

parser/ holds a parser for parseable.h which is a subset of the libtorch/include/ATen/Functions.h. This parser generates cpp (./wrapper.cpp) and java JNI code (./aten-scala/core/src/main/java/ATen.java). The cpp code is compiled to ./aten-scala/jni-osx/src/main/resources/libatenscalajni.dylib.

aten-scala/ holds an sbt project with two subprojects. aten-scala/core is the Java counterpart of the binding. It depends on Scala because std::tuple is translated into Scala tuples. aten-scala/jni-osx contains no source code but its Maven artifact contains the native library.

Libtorch is linked dynamically during runtime.

On OSX the jni shared library always looks for libtorch in /usr/local/lib/. You can get the libtorch libraries with pip3 install torch then copy the necessary files to /usr/local/lib/. E.g. cp /usr/local/lib/python3.9/site-packages/torch/lib/* /usr/local/lib/. However one needs to copy /usr/local/lib/python3.9/site-packages/torch/.dylibs/libomp.dylib as well to /usr/local/.dylibs/ (depends on the libtorch version).

On Linux there is no @rpath baked in the jni shared library. Use LD_LIBRARY_PATH to make sure the linker finds libtorch. See the docker-runtime/Dockerfiles on how to do this.

See docker-runtime/Dockerfile. This image is pushed to the Docker Hub (https://hub.docker.com/r/pityka/base-ubuntu-libtorch/tags?page=1&ordering=last_updated).

The image also has Java 8 and sbt. It is ready to make use of this library on Linux.

See LICENSE file.