Correct the install command, static library name and typo in nccl.cmake. (#5048)

cmake/external/nccl.cmake

-INCLUDE(ExternalProject)
+include(ExternalProject)
 
-SET(NCCL_SOURCE_DIR ${THIRD_PARTY_PATH}/nccl)
-
-INCLUDE_DIRECTORIES(${NCCL_SOURCE_DIR}/src/extern_nccl/src)
+set(NCCL_SOURCE_DIR ${THIRD_PARTY_PATH}/nccl)
 
+include_directories(${NCCL_SOURCE_DIR}/src/extern_nccl/src)
 
 if(WITH_DSO)
   # If we use DSO, we do not build nccl, just download the dependencies
@@ -12,9 +11,11 @@ if(WITH_DSO)
   set(NCCL_INSTALL_DIR "")
 else()
   # otherwise, we build nccl and link it.
+  set(NCCL_INSTALL_DIR ${THIRD_PARTY_PATH}/install/nccl)
+  # Note: cuda 8.0 is needed to make nccl
+  # When cuda is not installed on the system directory, need to set CUDA_HOME to your cuda root
   set(NCCL_BUILD_COMMAND "make -j 8")
-  set(NCCL_INSTALL_COMMAND  "make install")
-  SET(NCCL_INSTALL_DIR ${THIRD_PARTY_PATH}/install/nccl)
+  set(NCCL_INSTALL_COMMAND  "make install PREFIX=${NCCL_INSTALL_DIR}")
 endif()
 
 ExternalProject_Add(
@@ -31,20 +32,18 @@ ExternalProject_Add(
     TEST_COMMAND      ""
 )
 
-if (WITH_DSO)
-  if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
-    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_any_dummy.c)
-    file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";")
+if(WITH_DSO)
+  if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
+    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_nccl_dummy.c)
+    file(WRITE ${dummyfile} "const char * dummy_nccl = \"${dummyfile}\";")
     add_library(nccl STATIC ${dummyfile})
   else()
     add_library(nccl INTERFACE)
   endif()
 else()
-  ADD_LIBRARY(nccl STATIC IMPORTED GLOBAL)
-  SET_PROPERTY(TARGET nccl PROPERTY IMPORTED_LOCATION
-          ${NCCL_INSTALL_DIR}/lib/libnccl.a)
+  add_library(nccl STATIC IMPORTED GLOBAL)
+  set_property(TARGET nccl PROPERTY IMPORTED_LOCATION
+               ${NCCL_INSTALL_DIR}/lib/libnccl_static.a)
 endif()
 
 add_dependencies(nccl extern_nccl)
-
-LIST(APPEND external_project_dependencies nccl)