[CI] Pack R package tarball with pre-built xgboost.so (with GPU support) (#6827)

* Add scripts for packaging R package with GPU-enabled libxgboost.so

* [CI] Automatically build R package tarball

* Add comments

* Don't build tarball for pull requests

* Update the installation doc

Jenkinsfile

@@ -65,6 +65,7 @@ pipeline {
             'build-gpu-cuda10.1': { BuildCUDA(cuda_version: '10.1') },
             'build-gpu-cuda10.2': { BuildCUDA(cuda_version: '10.2', build_rmm: true) },
             'build-gpu-cuda11.0': { BuildCUDA(cuda_version: '11.0') },
+            'build-gpu-rpkg': { BuildRPackageWithCUDA(cuda_version: '10.0') },
             'build-jvm-packages-gpu-cuda10.0': { BuildJVMPackagesWithCUDA(spark_version: '3.0.0', cuda_version: '10.0') },
             'build-jvm-packages': { BuildJVMPackages(spark_version: '3.0.0') },
             'build-jvm-doc': { BuildJVMDoc() }
@@ -264,6 +265,24 @@ def BuildCUDA(args) {
   }
 }
 
+def BuildRPackageWithCUDA(args) {
+  node('linux && cpu_build') {
+    unstash name: 'srcs'
+    def container_type = 'gpu_build_r_centos6'
+    def docker_binary = "docker"
+    def docker_args = "--build-arg CUDA_VERSION_ARG=10.0"
+    if (env.BRANCH_NAME == 'master' || env.BRANCH_NAME.startsWith('release')) {
+      sh """
+      ${dockerRun} ${container_type} ${docker_binary} ${docker_args} tests/ci_build/build_r_pkg_with_cuda.sh ${commit_id}
+      """
+      echo 'Uploading R tarball...'
+      path = ("${BRANCH_NAME}" == 'master') ? '' : "${BRANCH_NAME}/"
+      s3Upload bucket: 'xgboost-nightly-builds', path: path, acl: 'PublicRead', includePathPattern:'xgboost_r_gpu_linux_*.tar.gz'
+    }
+    deleteDir()
+  }
+}
+
 def BuildJVMPackagesWithCUDA(args) {
   node('linux && mgpu') {
     unstash name: 'srcs'

doc/build.rst

@@ -2,18 +2,15 @@
 Installation Guide
 ##################
 
-.. note:: Pre-built binary wheel for Python
+.. note:: Pre-built binary wheel for Python: now with GPU support
 
-  If you are planning to use Python, consider installing XGBoost from a pre-built binary wheel, available from Python Package Index (PyPI). You may download and install it by running
+  If you are planning to use Python, consider installing XGBoost from a pre-built binary wheel, to avoid the trouble of building XGBoost from the source. You may download and install it by running
 
   .. code-block:: bash
 
-    # Ensure that you are downloading one of the following:
-    #   * xgboost-{version}-py2.py3-none-manylinux1_x86_64.whl
-    #   * xgboost-{version}-py2.py3-none-win_amd64.whl
     pip3 install xgboost
 
-  * The binary wheel will support GPU algorithms (`gpu_hist`) on machines with NVIDIA GPUs. Please note that **training with multiple GPUs is only supported for Linux platform**. See :doc:`gpu/index`.
+  * The binary wheel will support the GPU algorithm (``gpu_hist``) on machines with NVIDIA GPUs. Please note that **training with multiple GPUs is only supported for Linux platform**. See :doc:`gpu/index`.
   * Currently, we provide binary wheels for 64-bit Linux, macOS and Windows.
   * Nightly builds are available. You can go to `this page
     <https://s3-us-west-2.amazonaws.com/xgboost-nightly-builds/list.html>`_, find the
@@ -23,6 +20,21 @@ Installation Guide
 
       pip install <url to the wheel>
 
+.. note:: (EXPERIMENTAL) Pre-built binary package for R: now with GPU support
+
+  If you are planning to use R, consider installing ``{xgboost}`` from a pre-built binary package, to avoid the trouble of building XGBoost from the source. The binary package will let you use the GPU algorithm (``gpu_hist``) out of the box, as long as your machine has NVIDIA GPUs.
+
+  Download the binary package from the Releases page. The file name will be of the form ``xgboost_r_gpu_linux_[version].tar.gz``. Then install XGBoost by running:
+
+  .. code-block:: bash
+
+    # Install dependencies
+    R -q -e "install.packages(c('data.table', 'magrittr', 'jsonlite', 'remotes'))"
+    # Install XGBoost
+    R CMD INSTALL ./xgboost_r_gpu_linux.tar.gz
+
+  Currently, we provide the binary package for 64-bit Linux.
+
 
 ****************************
 Building XGBoost from source

tests/ci_build/Dockerfile.gpu_build_r_centos6

@@ -0,0 +1,112 @@
+ARG CUDA_VERSION_ARG
+FROM nvidia/cuda:$CUDA_VERSION_ARG-devel-centos6
+ARG CUDA_VERSION_ARG
+
+# Environment
+ENV DEBIAN_FRONTEND noninteractive
+ENV DEVTOOLSET_URL_ROOT http://vault.centos.org/6.9/sclo/x86_64/rh/devtoolset-4/
+
+COPY CentOS-Base.repo /etc/yum.repos.d/
+
+# Install all basic requirements
+RUN \
+    yum install -y epel-release && \
+    yum -y update && \
+    yum install -y tar unzip wget xz git patchelf readline-devel libX11-devel libXt-devel \
+                   xorg-x11-server-devel openssl-devel texlive-* && \
+    yum install -y $DEVTOOLSET_URL_ROOT/devtoolset-4-gcc-5.3.1-6.1.el6.x86_64.rpm \
+                   $DEVTOOLSET_URL_ROOT/devtoolset-4-gcc-gfortran-5.3.1-6.1.el6.x86_64.rpm \
+                   $DEVTOOLSET_URL_ROOT/devtoolset-4-libquadmath-devel-5.3.1-6.1.el6.x86_64.rpm \
+                   $DEVTOOLSET_URL_ROOT/devtoolset-4-gcc-c++-5.3.1-6.1.el6.x86_64.rpm \
+                   $DEVTOOLSET_URL_ROOT/devtoolset-4-binutils-2.25.1-8.el6.x86_64.rpm \
+                   $DEVTOOLSET_URL_ROOT/devtoolset-4-runtime-4.1-3.sc1.el6.x86_64.rpm \
+                   $DEVTOOLSET_URL_ROOT/devtoolset-4-libstdc++-devel-5.3.1-6.1.el6.x86_64.rpm
+
+ENV PATH=/opt/python/bin:/usr/local/ninja:/opt/software/packages/bin:/opt/R/3.3.0/bin:$PATH
+ENV LD_LIBRARY_PATH=/opt/software/packages/lib:/opt/R/3.3.0/lib64:$LD_LIBRARY_PATH
+ENV CC=/opt/rh/devtoolset-4/root/usr/bin/gcc
+ENV CXX=/opt/rh/devtoolset-4/root/usr/bin/c++
+ENV CPP=/opt/rh/devtoolset-4/root/usr/bin/cpp
+ENV F77=/opt/rh/devtoolset-4/root/usr/bin/gfortran
+
+# A few packages have to be built from the source because CentOS 6 is a very old distribution and
+# the system packages are not sufficiently up-to-date to build R 3.3.0. We'll want to update to
+# CentOS 7 after the 1.4.0 release. Tracking issue: dmlc/xgboost#6791.
+#
+# Why choose an old Linux distro? This is so that the resulting xgboost.so is compatible with a
+# wide range of Linux OSes currently in operation. See https://www.python.org/dev/peps/pep-0571/
+RUN \
+    wget https://zlib.net/fossils/zlib-1.2.5.tar.gz  && \
+    wget https://sourceware.org/pub/bzip2/bzip2-1.0.6.tar.gz  && \
+    wget http://tukaani.org/xz/xz-5.2.2.tar.gz  && \
+    wget https://ftp.pcre.org/pub/pcre/pcre-8.40.tar.gz  && \
+    wget https://www.openssl.org/source/old/1.0.0/openssl-1.0.0k.tar.gz  && \
+    wget --no-check-certificate https://curl.se/download/curl-7.47.1.tar.gz  && \
+    tar xf zlib-1.2.5.tar.gz  && \
+    tar xf bzip2-1.0.6.tar.gz  && \
+    tar xf xz-5.2.2.tar.gz  && \
+    tar xf pcre-8.40.tar.gz  && \
+    tar xf openssl-1.0.0k.tar.gz  && \
+    tar xf curl-7.47.1.tar.gz  && \
+    cd zlib-1.2.5  && \
+    ./configure --prefix=/opt/software/packages  && \
+    make -j$(nproc)  && \
+    make install  && \
+    cd ../bzip2-1.0.6  && \
+    sed -i 's/CFLAGS=-Wall/CFLAGS=-fPIC -Wall/g' Makefile  && \
+    make -f Makefile-libbz2_so  && \
+    make clean  && \
+    make -j$(nproc)  && \
+    make -n install PREFIX=/opt/software/packages  && \
+    make install PREFIX=/opt/software/packages  && \
+    cd ../xz-5.2.2  && \
+    ./configure --prefix=/opt/software/packages  && \
+    make -j$(nproc)  && \
+    make install  && \
+    cd ../pcre-8.40  && \
+    ./configure --enable-utf8 --prefix=/opt/software/packages  && \
+    make -j$(nproc)  && \
+    make install  && \
+    cd ../curl-7.47.1  && \
+    ./configure --prefix=/opt/software/packages --with-ssl && \
+    make -j$(nproc)  && \
+    make install  && \
+    export CFLAGS="-I/opt/software/packages/include"   && \
+    export LDFLAGS="-L/opt/software/packages/lib"  && \
+    cd .. && \
+    # R 3.3.0
+    wget -nv -nc https://cran.r-project.org/src/base/R-3/R-3.3.0.tar.gz  && \
+    tar xf R-3.3.0.tar.gz  && \
+    cd R-3.3.0  && \
+    ./configure --prefix=/opt/R/3.3.0 --enable-R-shlib  && \
+    make -j$(nproc)  && \
+    make install  && \
+    # Python
+    wget -nv -nc -O Miniconda3.sh https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
+    bash Miniconda3.sh -b -p /opt/python && \
+    /opt/python/bin/python -m pip install auditwheel && \
+    # CMake
+    wget -nv -nc https://cmake.org/files/v3.13/cmake-3.13.0-Linux-x86_64.sh --no-check-certificate && \
+    bash cmake-3.13.0-Linux-x86_64.sh --skip-license --prefix=/usr && \
+    # Ninja
+    mkdir -p /usr/local && \
+    cd /usr/local/ && \
+    wget -nv -nc https://github.com/ninja-build/ninja/archive/v1.10.0.tar.gz --no-check-certificate && \
+    tar xf v1.10.0.tar.gz && mv ninja-1.10.0 ninja && rm -v v1.10.0.tar.gz && \
+    cd ninja && \
+    /opt/python/bin/python ./configure.py --bootstrap
+
+ENV GOSU_VERSION 1.10
+
+# Install lightweight sudo (not bound to TTY)
+RUN set -ex; \
+    wget -O /usr/local/bin/gosu "https://github.com/tianon/gosu/releases/download/$GOSU_VERSION/gosu-amd64" && \
+    chmod +x /usr/local/bin/gosu && \
+    gosu nobody true
+
+# Default entry-point to use if running locally
+# It will preserve attributes of created files
+COPY entrypoint.sh /scripts/
+
+WORKDIR /workspace
+ENTRYPOINT ["/scripts/entrypoint.sh"]

tests/ci_build/build_r_pkg_with_cuda.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+set -e
+set -x
+
+if [ "$#" -ne 1 ]
+then
+  echo "Build the R package tarball with CUDA code. Usage: $0 [commit hash]"
+  exit 1
+fi
+
+commit_hash="$1"
+
+make Rpack
+mv xgboost/ xgboost_rpack/
+
+mkdir build
+cd build
+cmake .. -GNinja -DUSE_CUDA=ON -DR_LIB=ON
+ninja
+cd ..
+
+rm xgboost
+# This super wacky hack is found in cmake/RPackageInstall.cmake.in and
+# cmake/RPackageInstallTargetSetup.cmake. This hack lets us bypass the normal build process of R
+# and have R use xgboost.so that we've already built.
+rm -v xgboost_rpack/configure
+rm -rfv xgboost_rpack/src
+mkdir -p xgboost_rpack/src
+cp -v lib/xgboost.so xgboost_rpack/src/
+echo 'all:' > xgboost_rpack/src/Makefile
+echo 'all:' > xgboost_rpack/src/Makefile.win
+mv xgboost_rpack/ xgboost/
+tar cvzf xgboost_r_gpu_linux_${commit_hash}.tar.gz xgboost/