[Bugfix] Include NCCL as a submodule (#2934)

* Add NCCL as a submodule

* Allow using third_party/nccl or system nccl

* Add nccl_external as a dependency

* Fix conditional
Co-authored-by: Quan (Andy) Gan <coin2028@hotmail.com>

.gitmodules

@@ -32,3 +32,6 @@
 [submodule "third_party/nanoflann"]
 	path = third_party/nanoflann
 	url = https://github.com/jlblancoc/nanoflann
+[submodule "third_party/nccl"]
+	path = third_party/nccl
+	url = https://github.com/nvidia/nccl

CMakeLists.txt

@@ -23,6 +23,7 @@ endif()
 # and add set(OPTION VALUE) to override these build options.
 # Alernatively, use cmake -DOPTION=VALUE through command-line.
 dgl_option(USE_CUDA "Build with CUDA" OFF)
+dgl_option(USE_SYSTEM_NCCL "Build using system's NCCL library" OFF)
 dgl_option(USE_OPENMP "Build with OpenMP" ON)
 dgl_option(USE_AVX "Build with AVX optimization" ON)
 dgl_option(USE_FP16 "Build with fp16 support to enable mixed precision training" OFF)
@@ -159,13 +160,21 @@ if(USE_CUDA)
   dgl_config_cuda(DGL_CUDA_SRC)
   list(APPEND DGL_SRC ${DGL_CUDA_SRC})
 
-  include(cmake/util/FindNccl.cmake)
+  if (USE_SYSTEM_NCCL)
+    include(cmake/util/FindNccl.cmake)
+  else()
+    include(cmake/modules/NCCL.cmake)
+  endif()
+
   include_directories(${NCCL_INCLUDE_DIR})
   list(APPEND DGL_LINKER_LIBS ${NCCL_LIBRARY})
 endif(USE_CUDA)
 
 if(USE_CUDA)
   cuda_add_library(dgl SHARED ${DGL_SRC})
+  if (NOT USE_SYSTEM_NCCL)
+    add_dependencies(dgl nccl_external)
+  endif()
 else(USE_CUDA)
   add_library(dgl SHARED ${DGL_SRC})
 endif(USE_CUDA)

cmake/modules/NCCL.cmake

+include(ExternalProject)
+
+# set path to submodule
+set(NCCL_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/nccl")
+
+# NCCL doesn't have CMAKE, so build externally
+ExternalProject_Add(nccl_external
+  SOURCE_DIR ${PROJECT_SOURCE_DIR}/third_party/nccl
+  BUILD_IN_SOURCE 1
+  CONFIGURE_COMMAND ""
+  BUILD_COMMAND
+    env
+    make
+    "src.build"
+    "-j"
+    "BUILDDIR=${NCCL_BUILD_DIR}"
+  BUILD_BYPRODUCTS "${NCCL_BUILD_DIR}/lib/libnccl_static.a"
+  INSTALL_COMMAND ""
+  )
+
+# set output variables
+set(NCCL_FOUND TRUE)
+set(NCCL_LIBRARY "${NCCL_BUILD_DIR}/lib/libnccl_static.a")
+set(NCCL_INCLUDE_DIR "${NCCL_BUILD_DIR}/include")

nccl @ 911d61f2

+Subproject commit 911d61f214d45c98df1ee8c0ac23c33fb94b63de