cudnn windows support (#12757)

* cudnn widndows

* "add comment"

* "windows support"

* "fix cmake error"

CMakeLists.txt

@@ -212,6 +212,7 @@ elseif()
     set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in GPU only now." FORCE)
 endif()
 
+include(flags)              # set paddle compile flags
 include(cudnn)              # set cudnn libraries, must before configure
 include(cupti)
 include(configure)          # add paddle env configuration
@@ -220,7 +221,6 @@ include(package)            # set paddle packages
 include(ccache)             # set ccache for compilation
 include(util)               # set unittest and link libs
 include(rdma)               # set rdma libraries
-include(flags)              # set paddle compile flags
 include(version)            # set PADDLE_VERSION
 include(coveralls)          # set code coverage
 include(inference_lib)      # add paddle fluid inference libraries

cmake/configure.cmake

@@ -56,10 +56,6 @@ if(NOT CMAKE_CROSSCOMPILING)
         set(SIMD_FLAG ${SSE3_FLAG})
     endif()
 endif()
-if(UNIX AND NOT APPLE)
-  # except apple from nix*Os family
-  set(LINUX TRUE)
-endif(UNIX AND NOT APPLE)
 
 if(NOT WITH_GOLANG)
     add_definitions(-DPADDLE_WITHOUT_GOLANG)

cmake/cudnn.cmake

@@ -25,8 +25,25 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
     $ENV{CUDNN_ROOT}
     $ENV{CUDNN_ROOT}/lib64
     $ENV{CUDNN_ROOT}/lib
-    /usr/lib)
-find_library(CUDNN_LIBRARY NAMES libcudnn.so libcudnn.dylib # libcudnn_static.a
+    /usr/lib
+	${CUDA_TOOLKIT_ROOT_DIR}
+	${CUDA_TOOLKIT_ROOT_DIR}/lib/x64
+	)
+set(CUDNN_LIB_NAME "")
+if (LINUX)
+set(CUDNN_LIB_NAME "libcudnn.so")
+endif(LINUX)
+
+if(WIN32)
+# only support cudnn7
+set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll")
+endif(WIN32)
+
+if(Apple)
+set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so")
+endif(Apple)
+
+find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a
     PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist}
           NO_DEFAULT_PATH
     DOC "Path to cuDNN library.")

cmake/flags.cmake

@@ -142,6 +142,11 @@ else()
         ${GPU_COMMON_FLAGS})
 endif()
 
+if(UNIX AND NOT APPLE)
+  # except apple from nix*Os family
+  set(LINUX TRUE)
+endif(UNIX AND NOT APPLE)
+
 
 foreach(flag ${COMMON_FLAGS})
     safe_set_cflag(CMAKE_C_FLAGS ${flag})

paddle/fluid/framework/CMakeLists.txt

@@ -99,12 +99,13 @@ else()
   cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method)
 endif()
 
-
+if (NOT WIN32)
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
         graph graph_viz_pass multi_devices_graph_pass
         multi_devices_graph_print_pass multi_devices_graph_check_pass
         fast_threaded_ssa_graph_executor)
+endif() # NOT WIN32
 
 cc_library(prune SRCS prune.cc DEPS framework_proto)
 cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)

paddle/fluid/operators/CMakeLists.txt

@@ -84,6 +84,15 @@ function(op_library TARGET)
         message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
     endif()
 
+    #remove windows unsupported op
+    if (WIN32)
+    foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op")
+        if ("${TARGET}" STREQUAL "${windows_unsupport_op}")
+          return()
+        endif()
+    endforeach()
+    endif(WIN32)
+
     list(LENGTH op_library_DEPS op_library_DEPS_len)
     if (${op_library_DEPS_len} GREATER 0)
         set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE)
@@ -181,19 +190,19 @@ function(op_library TARGET)
 endfunction()
 
 add_subdirectory(math)
+if (NOT WIN32)
 add_subdirectory(nccl)
-
 if(WITH_GPU)
     op_library(nccl_op DEPS nccl_common)
     file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
 else()
     set(DEPS_OPS ${DEPS_OPS} nccl_op)
 endif()
+endif() # NOT WIN32
 
 set(DISTRIBUTE_DEPS "")
 if(WITH_DISTRIBUTE)
     add_subdirectory(distributed)
-    
     set(DISTRIBUTE_DEPS "")
     if(WITH_GRPC)
         set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node)
@@ -222,7 +231,7 @@ if(WITH_DISTRIBUTE)
     #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
     #        listen_and_serv_op sum_op executor SERIAL)
-    if(WITH_GPU)
+    if(WITH_GPU AND NOT WIN32)
         set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
         cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS listen_and_serv_op ${DISTRIBUTE_DEPS} executor SERIAL)
         if(WITH_GRPC)
@@ -233,7 +242,7 @@ if(WITH_DISTRIBUTE)
         set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     else()
         set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
-    endif()
+    endif() # WITH_GPU AND NOT WIN32
 else()
     set(DEPS_OPS ${DEPS_OPS}  checkpoint_notify_op prefetch_op recv_op listen_and_serv_op send_op send_barrier_op fetch_barrier_op gen_nccl_id_op)
 endif()
@@ -331,5 +340,7 @@ cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_sea
 cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory)
 cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
 cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
+if(NOT WIN32)
 nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context)
+endif()
 nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor)

paddle/fluid/operators/nccl/CMakeLists.txt

-if(WITH_GPU)
+if(WITH_GPU AND NOT WIN32)
   nv_library(nccl_common SRCS nccl_gpu_common.cc DEPS device_context operator )
 endif()

paddle/fluid/platform/dynload/CMakeLists.txt

@@ -3,7 +3,7 @@ cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce)
 list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc)
 
 # There is no macOS version of NCCL.
-if (NOT APPLE)
+if (NOT APPLE AND NOT WIN32)
   list(APPEND CUDA_SRCS nccl.cc)
 endif()
 

paddle/fluid/platform/enforce.h

@@ -44,7 +44,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/dynload/cublas.h"
 #include "paddle/fluid/platform/dynload/cudnn.h"
 #include "paddle/fluid/platform/dynload/curand.h"
-#ifndef __APPLE__
+#if !defined(__APPLE__) and !defined(_WIN32)
 #include "paddle/fluid/platform/dynload/nccl.h"
 #endif  // __APPLE__
 #endif  // PADDLE_WITH_CUDA
@@ -205,7 +205,7 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
 #endif
 }
 
-#ifndef __APPLE__
+#if !defined(__APPLE__) and !defined(_WIN32)
 template <typename... Args>
 inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
     ncclResult_t stat, const Args&... args) {
@@ -221,7 +221,7 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
 #endif
   }
 }
-#endif  // __APPLE__
+#endif  // __APPLE__ and windows
 #endif  // PADDLE_WITH_CUDA
 
 template <typename T>