Refactor CMake scripts. (#4323)

* Refactor CMake scripts.

* Remove CMake CUDA wrapper.
* Bump CMake version for CUDA.
* Use CMake to handle Doxygen.
* Split up CMakeList.
* Export install target.
* Use modern CMake.
* Remove build.sh
* Workaround for gpu_hist test.
* Use cmake 3.12.

* Revert machine.conf.

* Move CLI test to gpu.

* Small cleanup.

* Support using XGBoost as submodule.

* Fix windows

* Fix cpp tests on Windows

* Remove duplicated find_package.

src/CMakeLists.txt

@@ -0,0 +1,107 @@
+file(GLOB_RECURSE CPU_SOURCES *.cc)
+list(REMOVE_ITEM CPU_SOURCES ${PROJECT_SOURCE_DIR}/src/cli_main.cc)
+
+include(CheckCXXSourceCompiles)
+check_cxx_source_compiles("
+#include <xmmintrin.h>
+int main() {
+  char data = 0;
+  const char* address = &data;
+  _mm_prefetch(address, _MM_HINT_NTA);
+  return 0;
+}
+" XGBOOST_MM_PREFETCH_PRESENT)
+check_cxx_source_compiles("
+int main() {
+  char data = 0;
+  const char* address = &data;
+  __builtin_prefetch(address, 0, 0);
+  return 0;
+}
+" XGBOOST_BUILTIN_PREFETCH_PRESENT)
+
+# Add plugins to source files
+if (PLUGIN_LZ4)
+  list(APPEND PLUGINS_SOURCES ${PROJECT_SOURCE_DIR}/plugin/lz4/sparse_page_lz4_format.cc)
+  list(APPEND SRC_LIBS lz4)
+endif (PLUGIN_LZ4)
+if (PLUGIN_DENSE_PARSER)
+  list(APPEND PLUGINS_SOURCES ${PROJECT_SOURCE_DIR}/plugin/dense_parser/dense_libsvm.cc)
+endif (PLUGIN_DENSE_PARSER)
+
+#-- Object library
+# Object library is necessary for jvm-package, which creates its own shared
+# library.
+if (USE_CUDA)
+  file(GLOB_RECURSE CUDA_SOURCES *.cu)
+  add_library(objxgboost OBJECT ${CPU_SOURCES} ${CUDA_SOURCES} ${PLUGINS_SOURCES})
+  target_compile_definitions(objxgboost
+    PRIVATE -DXGBOOST_USE_CUDA=1)
+  target_include_directories(objxgboost PRIVATE ${PROJECT_SOURCE_DIR}/cub/)
+  target_compile_options(objxgboost PRIVATE
+    $<$<COMPILE_LANGUAGE:CUDA>:--expt-extended-lambda>
+    $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>
+    $<$<COMPILE_LANGUAGE:CUDA>:-lineinfo>
+    $<$<COMPILE_LANGUAGE:CUDA>:--std=c++11>
+    $<$<COMPILE_LANGUAGE:CUDA>:${GEN_CODE}>)
+
+  if (USE_NCCL)
+    find_package(Nccl REQUIRED)
+    target_include_directories(objxgboost PRIVATE ${NCCL_INCLUDE_DIR})
+    target_compile_definitions(objxgboost PRIVATE -DXGBOOST_USE_NCCL=1)
+    list(APPEND SRC_LIBS ${NCCL_LIBRARY})
+  endif (USE_NCCL)
+
+  if (USE_NVTX)
+    target_include_directories(objxgboost PRIVATE "${NVTX_HEADER_DIR}")
+    target_compile_definitions(objxgboost PRIVATE -DXGBOOST_USE_NVTX=1)
+  endif (USE_NVTX)
+
+  set_target_properties(objxgboost PROPERTIES
+    CUDA_SEPARABLE_COMPILATION OFF)
+else (USE_CUDA)
+  add_library(objxgboost OBJECT ${CPU_SOURCES} ${PLUGINS_SOURCES})
+endif (USE_CUDA)
+
+target_include_directories(objxgboost
+  PRIVATE
+  ${PROJECT_SOURCE_DIR}/include
+  ${PROJECT_SOURCE_DIR}/dmlc-core/include
+  ${PROJECT_SOURCE_DIR}/rabit/include)
+target_compile_options(objxgboost
+  PRIVATE
+  $<$<AND:$<CXX_COMPILER_ID:MSVC>,$<COMPILE_LANGUAGE:CXX>>:/MP>
+  $<$<COMPILE_LANGUAGE:CXX>:-funroll-loops>)
+if (WIN32 AND MINGW)
+  target_compile_options(objxgboost PUBLIC -static-libstdc++)
+endif (WIN32 AND MINGW)
+
+set_target_properties(objxgboost PROPERTIES
+  POSITION_INDEPENDENT_CODE ON
+  CXX_STANDARD 11
+  CXX_STANDARD_REQUIRED ON)
+target_compile_definitions(objxgboost
+  PRIVATE
+  -DDMLC_LOG_CUSTOMIZE=1  # enable custom logging
+  $<$<NOT:$<CXX_COMPILER_ID:MSVC>>:_MWAITXINTRIN_H_INCLUDED>
+  ${XGBOOST_DEFINITIONS})
+if (XGBOOST_MM_PREFETCH_PRESENT)
+  target_compile_definitions(objxgboost
+    PRIVATE
+    -DXGBOOST_MM_PREFETCH_PRESENT=1)
+endif(XGBOOST_MM_PREFETCH_PRESENT)
+if (XGBOOST_BUILTIN_PREFETCH_PRESENT)
+  target_compile_definitions(objxgboost
+    PRIVATE
+    -DXGBOOST_BUILTIN_PREFETCH_PRESENT=1)
+endif (XGBOOST_BUILTIN_PREFETCH_PRESENT)
+
+if (USE_OPENMP)
+  find_package(OpenMP REQUIRED)
+  if (OpenMP_CXX_FOUND OR OPENMP_FOUND)
+    target_compile_options(objxgboost PRIVATE  $<$<COMPILE_LANGUAGE:CXX>:${OpenMP_CXX_FLAGS}>)
+    list(APPEND SRC_LIBS ${OpenMP_CXX_LIBRARIES})
+    set(LINKED_LIBRARIES_PRIVATE "${LINKED_LIBRARIES_PRIVATE};${SRC_LIBS}" PARENT_SCOPE)
+  endif (OpenMP_CXX_FOUND OR OPENMP_FOUND)
+endif (USE_OPENMP)
+#-- End object library

src/c_api/c_api.cc

@@ -4,8 +4,11 @@
 #include <xgboost/learner.h>
 #include <xgboost/c_api.h>
 #include <xgboost/logging.h>
+
 #include <dmlc/thread_local.h>
 #include <rabit/rabit.h>
+#include <rabit/c_api.h>
+
 #include <cstdio>
 #include <cstring>
 #include <algorithm>

src/tree/updater_gpu_hist.cu

@@ -28,7 +28,9 @@
 namespace xgboost {
 namespace tree {
 
+#if !defined(GTEST_TEST)
 DMLC_REGISTRY_FILE_TAG(updater_gpu_hist);
+#endif  // !defined(GTEST_TEST)
 
 // training parameters specific to this algorithm
 struct GPUHistMakerTrainParam
@@ -47,8 +49,9 @@ struct GPUHistMakerTrainParam
                   "-1 to use all rows assignted to a GPU, and 0 to auto-deduce");
   }
 };
-
+#if !defined(GTEST_TEST)
 DMLC_REGISTER_PARAMETER(GPUHistMakerTrainParam);
+#endif  // !defined(GTEST_TEST)
 
 struct ExpandEntry {
   int nid;
@@ -102,9 +105,10 @@ inline static bool LossGuide(ExpandEntry lhs, ExpandEntry rhs) {
 }
 
 // Find a gidx value for a given feature otherwise return -1 if not found
-__device__ int BinarySearchRow(bst_uint begin, bst_uint end,
-                               common::CompressedIterator<uint32_t> data,
-                               int const fidx_begin, int const fidx_end) {
+__forceinline__ __device__ int BinarySearchRow(
+    bst_uint begin, bst_uint end,
+    common::CompressedIterator<uint32_t> data,
+    int const fidx_begin, int const fidx_end) {
   bst_uint previous_middle = UINT32_MAX;
   while (end != begin) {
     auto middle = begin + (end - begin) / 2;
@@ -466,6 +470,7 @@ struct CalcWeightTrainParam {
 };
 
 // Bin each input data entry, store the bin indices in compressed form.
+template<typename std::enable_if<true,  int>::type = 0>
 __global__ void CompressBinEllpackKernel(
     common::CompressedBufferWriter wr,
     common::CompressedByteT* __restrict__ buffer,  // gidx_buffer
@@ -559,11 +564,11 @@ struct IndicateLeftTransform {
  * segments. Based on a single pass of exclusive scan, uses iterators to
  * redirect inputs and outputs.
  */
-void SortPosition(dh::CubMemory* temp_memory, common::Span<int> position,
-                  common::Span<int> position_out, common::Span<bst_uint> ridx,
-                  common::Span<bst_uint> ridx_out, int left_nidx,
-                  int right_nidx, int64_t* d_left_count,
-                  cudaStream_t stream = nullptr) {
+inline void SortPosition(dh::CubMemory* temp_memory, common::Span<int> position,
+                         common::Span<int> position_out, common::Span<bst_uint> ridx,
+                         common::Span<bst_uint> ridx_out, int left_nidx,
+                         int right_nidx, int64_t* d_left_count,
+                         cudaStream_t stream = nullptr) {
   auto d_position_out = position_out.data();
   auto d_position_in = position.data();
   auto d_ridx_out = ridx_out.data();
@@ -593,7 +598,7 @@ void SortPosition(dh::CubMemory* temp_memory, common::Span<int> position,
 }
 
 /*! \brief Count how many rows are assigned to left node. */
-__device__ void CountLeft(int64_t* d_count, int val, int left_nidx) {
+__forceinline__ __device__ void CountLeft(int64_t* d_count, int val, int left_nidx) {
   unsigned ballot = __ballot(val == left_nidx);
   if (threadIdx.x % 32 == 0) {
     atomicAdd(reinterpret_cast<unsigned long long*>(d_count),    // NOLINT
@@ -1614,8 +1619,11 @@ class GPUHistMaker : public TreeUpdater {
   std::unique_ptr<GPUHistMakerSpecialised<GradientPairPrecise>> double_maker_;
 };
 
+#if !defined(GTEST_TEST)
 XGBOOST_REGISTER_TREE_UPDATER(GPUHistMaker, "grow_gpu_hist")
     .describe("Grow tree with GPU.")
     .set_body([]() { return new GPUHistMaker(); });
+#endif  // !defined(GTEST_TEST)
+
 }  // namespace tree
 }  // namespace xgboost