[R-package] Add GPU support for CRAN package

[jameslamb]

Summary

#3188 added an installation strategy for the R package that is compliant with CRAN's restrictions. I originally added support for using a flag to build a GPU-enabled version from that CRAN package, but decided to pull it out of that pull request. I think it deserves more careful thought and testing.

Description

The goal of this issue is to be able to build the CRAN package for the GPU, like this:

sh build-cran-package.sh
R CMD INSTALL \
    --configure-args='--enable-gpu' \
    lightgbm_2.3.2.tar.gz

The following code was removed from that PR, and adding it back would be a good start:

R-package/README.md

#### Custom Installation (Linux, Mac)

You can use `-configure-args to build a GPU-enabled version of the package.

Rscript build_r.R --skip-install
R CMD INSTALL \
    --configure-args='--enable-gpu' \
    lightgbm_2.3.2.tar.gz

#### Custom Installation (Windows)

Since R on Windows does not support the use of `--configure-args`, building a GPU-enabled version of the package on Windows requires the use of an environment variable.

To install from source with GPU on Windows, set environment variable `LGB_USE_GPU` to `true`.

setx LGB_USE_GPU "true"

Restart CMD, then check that it was set:

echo %LGB_USE_GPU%

If it shows `true`, you should see the flag `-DUSE_GPU` in the compiler messages.

R CMD install lightgbm_*.tar.gz

R-package/configure.ac

# Compile for GPU?
AC_ARG_ENABLE(
    [gpu],
    AS_HELP_STRING(
        [--enable-gpu],
        [Pass this flag to install the R package for usage with GPU]
    ),
    [
        LGB_CPPFLAGS+=" -DUSE_GPU"
    ]
)

configure.win

# Compile for GPU?
# Check whether environment variable LGB_USE_GPU has been set to true
if test ${LGB_USE_GPU};
then
    if ${LGB_USE_GPU} = "true";
    then
        LGB_CPPFLAGS="${LGB_CPPFLAGS} -DUSE_GPU"
    fi
fi

[jameslamb]

This has been added to #2302 . Leave a comment here or on that issue if you'd like to contribute this feature!

[StrikerRUS]

I wonder, how will it be useful for users given that they have to install Boost and OpenCL by their own anyway? The only difference will be "CMake vs configure", if I'm not mistaken. Maybe it's better to provide a link for the guide how build from the most recent sources?

[jameslamb]

I don't think it is something we have to do before going to CRAN, but it would be useful some day in the same way that Python users can do this

pip install lightgbm --install-option=--gpu

• no need to have or know about cmake
• no need to have or know about git or GitHub

I agree that it's ok to just have it documented in the README how to build for source, maybe for a long time, but I do think this could be useful one day.

[StrikerRUS]

@jameslamb Would it make sense to add the same feature request for CUDA build? If you are OK, I can create an issue.

[jameslamb]

yes definitely! I think @Laurae2 already did a proof of concept of "R package with CUDA", but dont remember where it was posted.

Either way yes, definitely

[StrikerRUS]

Created #3465.

but dont remember where it was posted.

Here it is

szilard/GBM-perf#12 (comment)

https://gist.github.com/Laurae2/7195cebe65887907a06e9118a3ec7f96

[jameslamb]

oh perfect, thank you!