[REVIEW] Added IDE changes for JNI files to have intelliSense and changes for rapids-compose [skip ci]

[razajafri]

This PR adds the ability to use intellisense with JNI code while using rapids-compose. Thank you @trxcllnt for all your help in setting this up

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[razajafri]

not exactly sure if we should be modifying this variable. @tgravescs @revans2 @jlowe please comment

[trxcllnt]

Just a note, this isn't required for rapids-compose. The script in compose only cares about the C++ build dir, and is setting -Dnative.build.path= as part of the mvn package command.

[razajafri]

Yes, but this will help manage the java bindings in a similar way cpp build is managed.
I still need to add the branch name to this property

[revans2]

I am not a fan of us messing with the project.build.directory.

[razajafri]

Is this necessary? If others are OK I would like to remove this. I don't think we need this to build on the host directly.

[revans2]

It follows the conventions of the other libraries. and is where we would find libcudf if it were installed instead of just being built.

[razajafri]

@revans2 @jlowe thanks for the review, is there anything else you want me to change?

[trxcllnt]

tl;dr: these two commits are the minimal set of changes I've tested that work in compose:

1. pom.xml
2. CMakeLists.txt

@jlowe The changes to the CMakeLists.txt are to support source-building the RAPIDS projects against each other (without installing). We don't have an "install" location in the traditional sense, because we only run make and not make install.

The loose pattern in our CI and build.sh scripts is:

• <PROJ>_HOME env vars point at the repository root
• <PROJ>_ROOT env vars point at the install location, or occasionally the C++ build dir

I've replicated this pattern in rapids-compose for consistency:

RMM_HOME=rapids/rmm
RMM_ROOT=rapids/rmm/build/release
CUDF_HOME=rapids/cudf
CUDF_ROOT=rapids/cudf/cpp/build/release

With CMake I'd override the paths on the command line like this:

cmake -D RMM_INCLUDE=$RMM_HOME/include \
      -D RMM_LIBRARY=$RMM_ROOT/librmm.so \
      -D CUDF_INCLUDE=$CUDF_HOME/cpp/include \
      -D CUDF_LIBRARY=$CUDF_ROOT/libcudf.so \
    ..

But @razajafri and I couldn't find a good way to use mvn package and pass those arguments onto CMake. We can modify pom.xml to forward on each CLI argument (like -DARROW_STATIC_LIB), but there isn't a good default for each of those vars if you invoke mvn package standalone. Defaulting to nothing causes CMake to think the empty string should be the location.

So the alternative is to have the CMakeLists.txt look in additional locations via environment variables. Neither RMM nor cuDF have lib or include directories in their C++ build dirs, which is why we set RMM_INCLUDE and CUDF_INCLUDE to their source locations.

[jlowe]

Thanks for the clarification, @trxcllnt. That helps explain why $RMM_HOME was showing up in the JNI CMakeLists.txt but not the cpp version.

[razajafri]

@jlowe do you have any other comments?

[jlowe]

$CUDF_ROOT/lib should remain in case the user set this environment variable to be the install prefix.