[LightGBM] build v4.3.0 with OpenCL/CUDA support

L/LightGBM/build_tarballs.jl

@@ -1,13 +1,18 @@
 # Note that this script can accept some limited command-line arguments, run
 # `julia build_tarballs.jl --help` to see a usage message.
 using BinaryBuilder, Pkg
+using Base.BinaryPlatforms: arch, tags
 
 name = "LightGBM"
-version = v"3.3.5"
+version = v"4.3.0"
+
+const YGGDRASIL_DIR = "../.."
+include(joinpath(YGGDRASIL_DIR, "fancy_toys.jl"))
+include(joinpath(YGGDRASIL_DIR, "platforms", "cuda.jl"))
 
 # Collection of sources required to complete build
 sources = [
-    GitSource("https://github.com/microsoft/LightGBM.git", "ca035b2ee0c2be85832435917b1e0c8301d2e0e0"),
+    GitSource("https://github.com/microsoft/LightGBM.git", "252828fd86627d7405021c3377534d6a8239dd69"),
     DirectorySource("./bundled"),
 ]
 
@@ -17,27 +22,43 @@ cd $WORKSPACE/srcdir/LightGBM
 git submodule update --init --depth=1
 git submodule update --checkout --depth=1
 
-
-if [[ $target == *"apple-darwin"* ]]; then
-  cmake_extra_args="-DAPPLE=1 -DAPPLE_OUTPUT_DYLIB=1"
-fi
+for p in $WORKSPACE/srcdir/patches/*.patch; do
+  atomic_patch -p1 "${p}"
+done
 
 FLAGS=()
+cmake_extra_args=()
+
+if [[ "${bb_full_target}" == *"apple-darwin"* ]]; then
+  cmake_extra_args+=(-DAPPLE=1 -DAPPLE_OUTPUT_DYLIB=1)
+fi
 
-if [[ "${target}" == *-mingw* ]]; then
-  # Parches windows
-  for p in $WORKSPACE/srcdir/patches/windows/*.patch; do
-    atomic_patch -p1 "${p}"
-  done
-  cmake_extra_args="-DWIN32=1 -DMINGW=1"
+if [[ "${bb_full_target}" == *-mingw* ]]; then
+  cmake_extra_args+=(-DWIN32=1 -DMINGW=1)
   FLAGS+=(LDFLAGS="-no-undefined")
 fi
 
-cmake \
+if [[ "${bb_full_target}" == *-linux* ]]; then
+  cmake_extra_args+=(-DUSE_GPU=1)
+fi
+
+if  [[ ("${bb_full_target}" == *-cuda*) && ("${bb_full_target}" != *-cuda+none*) ]]; then
+  # nvcc writes to /tmp, which is a small tmpfs in our sandbox.
+  # make it use the workspace instead
+  export TMPDIR=${WORKSPACE}/tmpdir
+  mkdir ${TMPDIR}
+  export CUDA_PATH=${WORKSPACE}/destdir/cuda
+  export PATH=$PATH:$CUDA_PATH/bin/
+
+  cmake_extra_args+=(-DUSE_CUDA=1 -DCMAKE_CUDA_FLAGS="-L${CUDA_PATH}/lib -L${CUDA_PATH}/lib/stubs")
+fi
+
+cmake . \
   -DCMAKE_INSTALL_PREFIX=$prefix\
   -DCMAKE_TOOLCHAIN_FILE=${CMAKE_TARGET_TOOLCHAIN} \
   -DCMAKE_BUILD_TYPE=Release \
-  $cmake_extra_args
+  "${cmake_extra_args[@]}"
+
 make -j${nproc} "${FLAGS[@]}"
 make install
 install_license LICENSE
@@ -46,20 +67,43 @@ install_license LICENSE
 # These are the platforms we will build for by default, unless further
 # platforms are passed in on the command line
 platforms = expand_cxxstring_abis(supported_platforms())
+cuda_platforms = expand_cxxstring_abis(CUDA.supported_platforms(; min_version=v"11"))
+
+# CUDA errors for other platforms
+filter!(p -> (Sys.islinux(p) && arch(p) == "x86_64"), cuda_platforms)
 
 # The products that we will ensure are always built
 products = [
     LibraryProduct("lib_lightgbm", :lib_lightgbm),
     ExecutableProduct("lightgbm", :lightgbm),
 ]
 
+all_platforms = [platforms; cuda_platforms]
+
 # Dependencies that must be installed before this package can be built
 dependencies = [
     # For OpenMP we use libomp from `LLVMOpenMP_jll` where we use LLVM as compiler (BSD
     # systems), and libgomp from `CompilerSupportLibraries_jll` everywhere else.
-    Dependency(PackageSpec(name="CompilerSupportLibraries_jll", uuid="e66e0078-7015-5450-92f7-15fbd957f2ae"); platforms=filter(!Sys.isbsd, platforms)),
-    Dependency(PackageSpec(name="LLVMOpenMP_jll", uuid="1d63c593-3942-5779-bab2-d838dc0a180e"); platforms=filter(Sys.isbsd, platforms)),
+    Dependency(PackageSpec(name="CompilerSupportLibraries_jll", uuid="e66e0078-7015-5450-92f7-15fbd957f2ae"); platforms=filter(!Sys.isbsd, all_platforms)),
+    Dependency(PackageSpec(name="LLVMOpenMP_jll", uuid="1d63c593-3942-5779-bab2-d838dc0a180e"); platforms=filter(Sys.isbsd, all_platforms)),
+
+    # GPU support is enabled via OpenCL and is hard to cross compile for other platforms
+    Dependency(PackageSpec(name="OpenCL_jll", uuid="6cb37087-e8b6-5417-8430-1f242f1e46e4"); platforms=filter(Sys.islinux, all_platforms)),
+    BuildDependency(PackageSpec(name="OpenCL_Headers_jll", uuid="a7aa756b-2b7f-562a-9e9d-e94076c5c8ee"); platforms=filter(Sys.islinux, all_platforms)),
+
+    Dependency(PackageSpec(name="boost_jll", uuid="28df3c45-c428-5900-9ff8-a3135698ca75"); compat="=1.76.0", platforms=filter(Sys.islinux, all_platforms)),
 ]
 
-# Build the tarballs, and possibly a `build.jl` as well.
-build_tarballs(ARGS, name, version, sources, script, platforms, products, dependencies; julia_compat="1.6", preferred_gcc_version = v"8.1.0")
+for platform in all_platforms
+    if platform in platforms && CUDA.is_supported(platform)
+        platform["cuda"] = "none"
+    end
+
+    deps = [dependencies; CUDA.required_dependencies(platform; static_sdk=true)]
+
+    should_build_platform(triplet(platform)) || continue
+
+    # Build the tarballs, and possibly a `build.jl` as well.
+    build_tarballs(ARGS, name, version, sources, script, [platform], products, deps; 
+    julia_compat="1.6", preferred_gcc_version = v"8.1", augment_platform_block=CUDA.augment)
+end

L/LightGBM/bundled/patches/lightgbm-cmake-find-cuda.patch

@@ -0,0 +1,72 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 1ff289b9..e71d5b99 100644
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -198,32 +198,28 @@ if(__INTEGRATE_OPENCL)
+ endif()
+
+ if(USE_CUDA)
+-    find_package(CUDA 11.0 REQUIRED)
+-    include_directories(${CUDA_INCLUDE_DIRS})
++    find_package(CUDAToolkit 11.0 REQUIRED)
++    include_directories(${CUDAToolkit_INCLUDE_DIRS})
+     set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler=${OpenMP_CXX_FLAGS} -Xcompiler=-fPIC -Xcompiler=-Wall")
+
+     # reference for mapping of CUDA toolkit component versions to supported architectures ("compute capabilities"):
+     # https://en.wikipedia.org/wiki/CUDA#GPUs_supported
+-    set(CUDA_ARCHS "6.0" "6.1" "6.2" "7.0" "7.5")
+-    if(CUDA_VERSION VERSION_GREATER_EQUAL "11.0")
+-        list(APPEND CUDA_ARCHS "8.0")
++    set(CUDA_ARCHS "60" "61" "62" "70" "75")
++    if(CUDA_VERSION VERSION_GREATER_EQUAL "110")
++        list(APPEND CUDA_ARCHS "80")
+     endif()
+-    if(CUDA_VERSION VERSION_GREATER_EQUAL "11.1")
+-        list(APPEND CUDA_ARCHS "8.6")
++    if(CUDA_VERSION VERSION_GREATER_EQUAL "111")
++        list(APPEND CUDA_ARCHS "86")
+     endif()
+-    if(CUDA_VERSION VERSION_GREATER_EQUAL "11.5")
+-        list(APPEND CUDA_ARCHS "8.7")
++    if(CUDA_VERSION VERSION_GREATER_EQUAL "115")
++        list(APPEND CUDA_ARCHS "87")
+     endif()
+-    if(CUDA_VERSION VERSION_GREATER_EQUAL "11.8")
+-        list(APPEND CUDA_ARCHS "8.9")
+-        list(APPEND CUDA_ARCHS "9.0")
++    if(CUDA_VERSION VERSION_GREATER_EQUAL "118")
++        list(APPEND CUDA_ARCHS "89")
++        list(APPEND CUDA_ARCHS "90")
+     endif()
+     list(POP_BACK CUDA_ARCHS CUDA_LAST_SUPPORTED_ARCH)
+     list(APPEND CUDA_ARCHS "${CUDA_LAST_SUPPORTED_ARCH}+PTX")
+-    cuda_select_nvcc_arch_flags(CUDA_ARCH_FLAGS ${CUDA_ARCHS})
+-    string(REPLACE ";" " " CUDA_ARCH_FLAGS "${CUDA_ARCH_FLAGS}")
+-
+-    set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${CUDA_ARCH_FLAGS}")
+     if(USE_DEBUG)
+       set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -g")
+     else()
+@@ -260,7 +256,7 @@ if(USE_CUDA)
+     function(add_histogram hsize hname hadd hconst hdir)
+       add_library(histo${hsize}${hname} OBJECT src/treelearner/kernels/histogram${hsize}.cu)
+       set_target_properties(histo${hsize}${hname} PROPERTIES CUDA_SEPARABLE_COMPILATION ON)
+-      set_target_properties(histo${hsize}${hname} PROPERTIES CUDA_ARCHITECTURES OFF)
++      set_target_properties(histo${hsize}${hname} PROPERTIES CUDA_ARCHITECTURES ${CUDA_ARCHS})
+       if(hadd)
+         list(APPEND histograms histo${hsize}${hname})
+         set(histograms ${histograms} PARENT_SCOPE)
+@@ -600,10 +596,10 @@ if(USE_CUDA)
+   # Disable cmake warning about policy CMP0104. Refer to issue #3754 and PR #4268.
+   # Custom target properties does not propagate, thus we need to specify for
+   # each target that contains or depends on cuda source.
+-  set_target_properties(lightgbm_objs PROPERTIES CUDA_ARCHITECTURES OFF)
+-  set_target_properties(_lightgbm PROPERTIES CUDA_ARCHITECTURES OFF)
++  set_target_properties(lightgbm_objs PROPERTIES CUDA_ARCHITECTURES ${CUDA_ARCHS})
++  set_target_properties(_lightgbm PROPERTIES CUDA_ARCHITECTURES ${CUDA_ARCHS})
+   if(BUILD_CLI)
+-    set_target_properties(lightgbm PROPERTIES CUDA_ARCHITECTURES OFF)
++    set_target_properties(lightgbm PROPERTIES CUDA_ARCHITECTURES ${CUDA_ARCHS})
+   endif()
+
+   set_target_properties(lightgbm_objs PROPERTIES CUDA_SEPARABLE_COMPILATION ON)