Merge pull request #14545 from wopeizl/windows/online

Windows/online

cmake/cuda.cmake

@@ -199,8 +199,11 @@ elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
     list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
 endif()
 else(NOT WIN32)
+list(APPEND CUDA_NVCC_FLAGS  "--compiler-options;/bigobj")
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")
   list(APPEND CUDA_NVCC_FLAGS  "-g -G")
+  # match the cl's _ITERATOR_DEBUG_LEVEL
+  list(APPEND CUDA_NVCC_FLAGS  "-D_DEBUG")
 elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
   list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
 else()

cmake/generic.cmake

@@ -349,10 +349,17 @@ function(cc_test TARGET_NAME)
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS ARGS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    if(WIN32)
+      list(APPEND win32_deps shlwapi)
+      if("${cc_test_DEPS};" MATCHES "python;")
+        list(REMOVE_ITEM cc_test_DEPS python)
+        list(APPEND win32_deps ${PYTHON_LIBRARIES})
+      endif()
+    endif(WIN32)
     add_executable(${TARGET_NAME} ${cc_test_SRCS})
     target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     if(WIN32)
-      target_link_libraries(${TARGET_NAME} shlwapi)
+      target_link_libraries(${TARGET_NAME} ${win32_deps})
     endif(WIN32)
     add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_test(NAME ${TARGET_NAME}
@@ -683,7 +690,7 @@ function(py_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS ARGS ENVS)
     cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
-             COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
+             COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
              FLAGS_cpu_deterministic=true
              PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
              ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}

paddle/fluid/framework/CMakeLists.txt

@@ -116,14 +116,9 @@ cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
 cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 
-if (NOT WIN32)
 cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
     shape_inference data_transform lod_tensor profiler transfer_scope_cache)
-else()
-cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform lod_tensor)
-endif(NOT WIN32)
 
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -23,7 +23,7 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
                                      const std::vector<Scope *> &local_scopes,
                                      const std::vector<platform::Place> &places,
@@ -74,7 +74,7 @@ void AllReduceOpHandle::RunImpl() {
     }
 
     if (platform::is_gpu_place(lod_tensors[0]->place())) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
       int dtype = -1;
       size_t numel = 0;

paddle/fluid/framework/details/all_reduce_op_handle.h

@@ -20,7 +20,7 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -29,7 +29,7 @@ namespace framework {
 namespace details {
 
 struct AllReduceOpHandle : public OpHandleBase {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places,
                     const platform::NCCLContextMap *ctxs);
@@ -49,7 +49,7 @@ struct AllReduceOpHandle : public OpHandleBase {
  private:
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   const platform::NCCLContextMap *nccl_ctxs_;
 #endif
 };

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -82,7 +82,7 @@ void BroadcastOpHandle::BroadcastOneVar(
       });
     }
   } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     VarHandle *out_handle = nullptr;
     int root_id = boost::get<platform::CUDAPlace>(in_tensor.place()).device;
     std::vector<std::function<void()>> broadcast_calls;

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -24,7 +24,7 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -34,7 +34,7 @@ namespace details {
 
 struct BroadcastOpHandle : public OpHandleBase {
  public:
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   BroadcastOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                     const std::vector<platform::Place> &places,
                     const platform::NCCLContextMap *nccl_ctxs)
@@ -68,7 +68,7 @@ struct BroadcastOpHandle : public OpHandleBase {
 
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   const platform::NCCLContextMap *nccl_ctxs_;
 #endif
 

paddle/fluid/framework/details/broadcast_op_handle_test.h

@@ -42,7 +42,7 @@ struct TestBroadcastOpHandle {
   std::vector<std::unique_ptr<ir::Node>> nodes_;
   std::vector<p::Place> place_list_;
   bool use_gpu_;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
 
@@ -50,7 +50,7 @@ struct TestBroadcastOpHandle {
     for (size_t j = 0; j < ctxs_.size(); ++j) {
       ctxs_[j]->Wait();
     }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     if (nccl_ctxs_) {
       nccl_ctxs_->WaitAll();
     }
@@ -60,7 +60,7 @@ struct TestBroadcastOpHandle {
   void InitCtxOnGpu(bool use_gpu) {
     use_gpu_ = use_gpu;
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       int count = p::GetCUDADeviceCount();
       if (count <= 1) {
         LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
@@ -84,7 +84,7 @@ struct TestBroadcastOpHandle {
         place_list_.push_back(p);
         ctxs_.emplace_back(new p::CPUDeviceContext(p));
       }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       nccl_ctxs_.reset(nullptr);
 #endif
     }
@@ -106,14 +106,14 @@ struct TestBroadcastOpHandle {
     nodes_.emplace_back(
         ir::CreateNodeForTest("node0", ir::Node::Type::kOperation));
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
                                          place_list_, nccl_ctxs_.get());
 #else
       PADDLE_THROW("CUDA is not support.");
 #endif
     } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
                                          place_list_, nccl_ctxs_.get());
 #else

paddle/fluid/framework/details/build_strategy.cc

@@ -96,7 +96,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &param_names,
     const std::vector<Scope *> &local_scopes,
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const {
 #else
     const bool use_cuda) const {
@@ -118,7 +118,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
       pass->Erase("local_scopes");
       pass->SetNotOwned<const std::vector<Scope *>>("local_scopes",
                                                     &local_scopes);
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
       pass->Erase("nccl_ctxs");
       pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);

paddle/fluid/framework/details/build_strategy.h

@@ -23,7 +23,7 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -98,7 +98,7 @@ struct BuildStrategy {
       const std::string &loss_var_name,
       const std::unordered_set<std::string> &param_names,
       const std::vector<Scope *> &local_scopes,
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const;
 #else
       const bool use_cuda) const;

paddle/fluid/framework/details/data_balance_op_handle.cc

@@ -20,7 +20,7 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 DataBalanceOpHandle::DataBalanceOpHandle(
     ir::Node *node, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,

paddle/fluid/framework/details/data_balance_op_handle.h

@@ -19,7 +19,7 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -29,7 +29,7 @@ namespace details {
 
 struct DataBalanceOpHandle : public OpHandleBase {
  public:
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   DataBalanceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                       const std::vector<platform::Place> &places,
                       const platform::NCCLContextMap *ctxs);

paddle/fluid/framework/details/fused_broadcast_op_handle.h

@@ -25,7 +25,7 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -35,7 +35,7 @@ namespace details {
 
 struct FusedBroadcastOpHandle : public BroadcastOpHandle {
  public:
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   FusedBroadcastOpHandle(ir::Node *node,
                          const std::vector<Scope *> local_scopes,
                          const std::vector<platform::Place> &places,

paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc

@@ -44,14 +44,14 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
     nodes_.emplace_back(
         ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation));
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new FusedBroadcastOpHandle(
           nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
 #else
       PADDLE_THROW("CUDA is not supported.");
 #endif
     } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_ = new FusedBroadcastOpHandle(
           nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
 #else

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -142,7 +142,7 @@ void MultiDevSSAGraphBuilder::Init() const {
   places_ = Get<const std::vector<platform::Place>>(kPlaces);
   local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);
   strategy_ = Get<const BuildStrategy>(kStrategy);
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   nccl_ctxs_ = &Get<platform::NCCLContextMap>("nccl_ctxs");
 #endif
 
@@ -431,7 +431,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
     }
   }
   bool use_gpu = false;
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   use_gpu = nccl_ctxs_ != nullptr;
 #endif
 
@@ -478,7 +478,7 @@ bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const {
 
 void MultiDevSSAGraphBuilder::SetCommunicationContext(
     OpHandleBase *op_handle, const platform::Place &p) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   if (nccl_ctxs_ == nullptr) {
     op_handle->SetDeviceContext(p,
                                 platform::DeviceContextPool::Instance().Get(p));
@@ -492,7 +492,7 @@ void MultiDevSSAGraphBuilder::SetCommunicationContext(
 void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
                                                 const std::string &p_name,
                                                 size_t src_dev_id) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   auto *op_handle = new BroadcastOpHandle(
       result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_);
@@ -522,7 +522,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
 void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
     ir::Graph *result,
     const std::vector<std::unordered_set<std::string>> &bcast_varnames) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   auto *op_handle = new FusedBroadcastOpHandle(
       result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_);
@@ -568,7 +568,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
 
 void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
                                                 const std::string &og) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
       result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_));
@@ -597,7 +597,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
 
 void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
     ir::Graph *result, const std::vector<std::string> &datas) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   result->Get<GraphOps>(kGraphOps).emplace_back(new DataBalanceOpHandle(
       result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_));
@@ -694,7 +694,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
 VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
                                                    const std::string &og,
                                                    int dst_dev_id) const {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   result->Get<GraphOps>(kGraphOps).emplace_back(new ReduceOpHandle(
       result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
       local_scopes_, places_, nccl_ctxs_));

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -40,7 +40,7 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
                          size_t device_id) const;
   void Init() const;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   mutable platform::NCCLContextMap *nccl_ctxs_;
 #endif
 

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -125,7 +125,7 @@ void ReduceOpHandle::RunImpl() {
         }
       });
     } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       auto pre_in = pre_in_var->Get<framework::LoDTensor>();
       VariableVisitor::ShareDimsAndLoD(*pre_in_var, out_var);
       VariableVisitor::GetMutableTensor(out_var).mutable_data(

paddle/fluid/framework/details/reduce_op_handle.h

@@ -23,7 +23,7 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -35,7 +35,7 @@ struct ReduceOpHandle : public OpHandleBase {
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   const platform::NCCLContextMap *nccl_ctxs_;
   ReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
                  const std::vector<platform::Place> &places,

paddle/fluid/framework/details/reduce_op_handle_test.cc

@@ -35,7 +35,7 @@ struct TestReduceOpHandle {
   std::vector<p::Place> gpu_list_;
   std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
 
@@ -43,7 +43,7 @@ struct TestReduceOpHandle {
     for (size_t j = 0; j < ctxs_.size(); ++j) {
       ctxs_[j]->Wait();
     }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     if (nccl_ctxs_) {
       nccl_ctxs_->WaitAll();
     }
@@ -53,7 +53,7 @@ struct TestReduceOpHandle {
   void InitCtxOnGpu(bool use_gpu) {
     use_gpu_ = use_gpu;
     if (use_gpu) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       int count = p::GetCUDADeviceCount();
       if (count <= 1) {
         LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
@@ -77,7 +77,7 @@ struct TestReduceOpHandle {
         gpu_list_.push_back(p);
         ctxs_.emplace_back(new p::CPUDeviceContext(p));
       }
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       nccl_ctxs_.reset(nullptr);
 #endif
     }
@@ -99,14 +99,14 @@ struct TestReduceOpHandle {
 
     nodes.emplace_back(new ir::Node("node"));
     if (use_gpu_) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
                                           gpu_list_, nccl_ctxs_.get()));
 #else
       PADDLE_THROW("CUDA is not support.");
 #endif
     } else {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
                                           gpu_list_, nccl_ctxs_.get()));
 #else

paddle/fluid/framework/ir/is_test_pass_tester.cc

@@ -15,7 +15,10 @@
 #include "paddle/fluid/framework/ir/is_test_pass.h"
 
 #include <gtest/gtest.h>
-
+#ifdef _WIN32
+#undef FALSE
+#undef TRUE
+#endif
 namespace paddle {
 namespace framework {
 namespace ir {

paddle/fluid/framework/lod_tensor.cc

@@ -26,10 +26,8 @@ limitations under the License. */
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/memory/memory.h"
 
-#if !defined(_WIN32)
 #include "paddle/fluid/recordio/scanner.h"
 #include "paddle/fluid/recordio/writer.h"
-#endif  // _WIN32
 
 namespace paddle {
 namespace framework {
@@ -305,7 +303,6 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
   TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
 }
 
-#if !defined(_WIN32)
 void WriteToRecordIO(recordio::Writer *writer,
                      const std::vector<LoDTensor> &tensor,
                      const platform::DeviceContext &dev_ctx) {
@@ -335,19 +332,7 @@ bool ReadFromRecordIO(recordio::Scanner *scanner,
 
   return true;
 }
-#else
-class Writer {};
-class Scanner {};
-void WriteToRecordIO(recordio::Writer *writer,
-                     const std::vector<LoDTensor> &tensor,
-                     const platform::DeviceContext &dev_ctx) {}
-bool ReadFromRecordIO(recordio::Scanner *scanner,
-                      const platform::DeviceContext &dev_ctx,
-                      std::vector<LoDTensor> *result_ptr) {
-  PADDLE_ENFORCE("windows didn't supported recordio!.");
-  return true;
-}
-#endif  // _WIN32
+
 std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
     const std::vector<platform::Place> places) const {
   check_memory_size();

paddle/fluid/framework/lod_tensor_test.cc

@@ -274,7 +274,6 @@ TEST(LoD, ConvertToOffsetBasedLoD) {
   EXPECT_EQ(offset_lod, expected);
 }
 
-#if !defined(_WIN32)
 template <typename T>
 static void TestRecordIO() {
   LoDTensor tensor;
@@ -321,7 +320,6 @@ TEST(LoDTensor, RecordIO) {
   TestRecordIO<float>();
   TestRecordIO<double>();
 }
-#endif  // !defined(_WIN32)
 
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/operator.cc

@@ -149,17 +149,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
 #endif
   }
 
-// The profile has a process-wide mutex, results in serious performance issue
-// in concurrency scenerio. Here use an `if` to fix this issue.
-// Please not remove the `if`, ask @Superjomn if there are any concern.
-#ifndef _WIN32
+  // The profile has a process-wide mutex, results in serious performance issue
+  // in concurrency scenerio. Here use an `if` to fix this issue.
+  // Please not remove the `if`, ask @Superjomn if there are any concern.
   if (platform::IsProfileEnabled()) {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     platform::RecordEvent record_event(Type(), pool.Get(place));
     RunImpl(scope, place);
-  } else
-#endif
-  {
+  } else {
     RunImpl(scope, place);
   }
   VLOG(30) << place << " " << DebugStringEx(&scope);

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -19,6 +19,7 @@
 #include "paddle/fluid/inference/analysis/ut_helper.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
+#include "paddle/fluid/platform/port.h"
 
 namespace paddle {
 namespace inference {
@@ -75,7 +76,7 @@ void TestWord2vecPrediction(const std::string& model_path) {
                      0.000932706};
   const size_t num_elements = outputs.front().data.length() / sizeof(float);
   // The outputs' buffers are in CPU memory.
-  for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
+  for (size_t i = 0; i < std::min((size_t)5UL, num_elements); i++) {
     LOG(INFO) << "data: "
               << static_cast<float*>(outputs.front().data.data())[i];
     PADDLE_ENFORCE(static_cast<float*>(outputs.front().data.data())[i],

paddle/fluid/inference/api/analysis_predictor.cc

@@ -56,7 +56,6 @@ bool AnalysisPredictor::Init(
     const std::shared_ptr<framework::Scope> &parent_scope,
     const std::shared_ptr<framework::ProgramDesc> &program) {
   VLOG(30) << "Predictor::init()";
-#if !defined(_WIN32)
   if (FLAGS_profile) {
     LOG(WARNING) << "Profiler is actived, might affect the performance";
     LOG(INFO) << "You can turn off by set gflags '-profile false'";
@@ -64,7 +63,6 @@ bool AnalysisPredictor::Init(
                                            : platform::ProfilerState::kCPU;
     platform::EnableProfiler(tracking_device);
   }
-#endif
 
   // no matter with or without MKLDNN
   paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
@@ -520,12 +518,10 @@ bool AnalysisPredictor::LoadParameters() {
 }
 
 AnalysisPredictor::~AnalysisPredictor() {
-#if !defined(_WIN32)
   if (FLAGS_profile) {
     platform::DisableProfiler(platform::EventSortingKey::kTotal,
                               "./profile.log");
   }
-#endif
   if (sub_scope_) {
     scope_->DeleteScope(sub_scope_);
   }

paddle/fluid/inference/api/api_impl.cc

@@ -64,7 +64,6 @@ void NativePaddlePredictor::PrepareFeedFetch() {
 bool NativePaddlePredictor::Init(
     std::shared_ptr<framework::Scope> parent_scope) {
   VLOG(3) << "Predictor::init()";
-#if !defined(_WIN32)
   if (FLAGS_profile) {
     LOG(WARNING) << "Profiler is actived, might affect the performance";
     LOG(INFO) << "You can turn off by set gflags '-profile false'";
@@ -73,7 +72,6 @@ bool NativePaddlePredictor::Init(
                                            : platform::ProfilerState::kCPU;
     platform::EnableProfiler(tracking_device);
   }
-#endif
 
   // no matter with or without MKLDNN
   paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
@@ -121,12 +119,10 @@ bool NativePaddlePredictor::Init(
 }
 
 NativePaddlePredictor::~NativePaddlePredictor() {
-#if !defined(_WIN32)
   if (FLAGS_profile) {
     platform::DisableProfiler(platform::EventSortingKey::kTotal,
                               "./profile.log");
   }
-#endif
   if (sub_scope_) {
     scope_->DeleteScope(sub_scope_);
   }

paddle/fluid/inference/api/helper.h

@@ -15,10 +15,6 @@
 #pragma once
 
 #include <glog/logging.h>
-#if !defined(_WIN32)
-#include <sys/time.h>
-#else
-#endif
 
 #include <algorithm>
 #include <chrono>  // NOLINT
@@ -28,6 +24,7 @@
 #include <string>
 #include <vector>
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/string/printf.h"
 
 namespace paddle {

paddle/fluid/inference/tests/api/anakin_rnn1_tester.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gflags/gflags.h>
-#include <sys/time.h>
 #include <time.h>
 #include <algorithm>
 #include <fstream>

paddle/fluid/inference/tests/api/tester_helper.h

@@ -178,11 +178,9 @@ void TestOneThreadPrediction(
     warmup_timer.tic();
     predictor->Run(inputs[0], outputs, batch_size);
     PrintTime(batch_size, 1, 1, 0, warmup_timer.toc(), 1);
-#if !defined(_WIN32)
     if (FLAGS_profile) {
       paddle::platform::ResetProfiler();
     }
-#endif
   }
 
   LOG(INFO) << "Run " << num_times << " times...";
@@ -232,11 +230,9 @@ void TestMultiThreadPrediction(
         warmup_timer.tic();
         predictor->Run(inputs[0], outputs, batch_size);
         PrintTime(batch_size, 1, num_threads, tid, warmup_timer.toc(), 1);
-#if !defined(_WIN32)
         if (FLAGS_profile) {
           paddle::platform::ResetProfiler();
         }
-#endif
       }
 
       LOG(INFO) << "Thread " << tid << " run " << num_times << " times...";

paddle/fluid/inference/tests/book/test_inference_nlp.cc

@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <sys/time.h>
 #include <time.h>
 #include <fstream>
 #include <thread>  // NOLINT

paddle/fluid/inference/tests/test_helper.h

@@ -20,6 +20,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/io.h"
+#include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DECLARE_bool(use_mkldnn);

paddle/fluid/operators/beam_search_op_test.cc

@@ -46,7 +46,7 @@ void CreateInput(LoDTensor* ids, LoDTensor* scores) {
   auto* scores_data = scores->mutable_data<float>(place);
   vector<int64_t> _ids({4, 2, 5, 2, 1, 3, 3, 5, 2, 8, 2, 1});
   vector<float> _scores(
-      {0.5, 0.3, 0.2, 0.6, 0.3, 0.1, 0.9, 0.5, 0.1, 0.7, 0.5, 0.1});
+      {0.5f, 0.3f, 0.2f, 0.6f, 0.3f, 0.1f, 0.9f, 0.5f, 0.1f, 0.7f, 0.5f, 0.1f});
 
   for (int i = 0; i < 12; i++) {
     ids_data[i] = _ids[i];
@@ -80,7 +80,7 @@ TEST(DISABLED_beam_search_op, run) {
   ASSERT_EQ(sids.lod(), sscores.lod());
 
   vector<int> tids({4, 2, 3, 8});
-  vector<float> tscores({0.5, 0.6, 0.9, 0.7});
+  vector<float> tscores({0.5f, 0.6f, 0.9f, 0.7f});
 
   for (int i = 0; i < 4; i++) {
     ASSERT_EQ(tids[i], sids.data<int64_t>()[i]);

paddle/fluid/operators/distributed/grpc_client.cc

@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <sys/time.h>
 #include <limits>
 
 #include "glog/logging.h"  // For VLOG
@@ -20,6 +19,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/distributed/grpc_client.h"
 #include "paddle/fluid/operators/distributed/grpc_serde.h"
 #include "paddle/fluid/operators/distributed/request_handler.h"
+#include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/profiler.h"
 
 DECLARE_bool(rpc_disable_reuse_port);

paddle/fluid/operators/distributed/grpc_serde.cc

@@ -15,7 +15,6 @@ limitations under the License. */
 #ifdef PADDLE_WITH_CUDA
 #include <nccl.h>
 #endif
-#include <sys/time.h>
 #include <thread>  // NOLINT
 
 #include "google/protobuf/io/coded_stream.h"
@@ -26,6 +25,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/distributed/grpc_variable_response.h"
 #include "paddle/fluid/operators/distributed/proto_encoder_helper.h"
 #include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
+#include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {

paddle/fluid/operators/distributed/grpc_serde.h

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include <sys/time.h>
+
 #include <iostream>
 #include <string>
 #include <vector>
@@ -25,6 +25,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/framework/var_type.h"
 #include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
+#include "paddle/fluid/platform/port.h"
 
 #include "paddle/fluid/operators/distributed/send_recv.grpc.pb.h"
 #include "paddle/fluid/operators/distributed/send_recv.pb.h"

paddle/fluid/operators/distributed/sendrecvop_utils.cc

@@ -15,12 +15,12 @@ limitations under the License. */
 #ifdef PADDLE_WITH_CUDA
 #include <nccl.h>
 #endif
-#include <sys/time.h>
 #include <thread>  // NOLINT
 
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/operators/distributed/sendrecvop_utils.h"
 #include "paddle/fluid/operators/distributed/variable_response.h"
+#include "paddle/fluid/platform/port.h"
 
 DEFINE_bool(rpc_disable_reuse_port, false, "Disable SO_REUSEPORT or not.");
 

paddle/fluid/operators/distributed/sendrecvop_utils.h

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include <sys/time.h>
 #include <iostream>
 #include <string>
 #include <vector>
@@ -24,6 +23,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/tensor_util.h"
 #include "paddle/fluid/framework/var_type.h"
+#include "paddle/fluid/platform/port.h"
 
 #include "paddle/fluid/operators/distributed/send_recv.pb.h"
 

paddle/fluid/operators/math/cpu_vec_test.cc

@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <sys/time.h>
 #include <cmath>
 #include <cstring>
 #include <random>
@@ -22,6 +21,7 @@ limitations under the License. */
 #include "gtest/gtest.h"
 
 #include "paddle/fluid/operators/math/cpu_vec.h"
+#include "paddle/fluid/platform/port.h"
 
 inline double GetCurrentUS() {
   struct timeval time;

paddle/fluid/operators/math/im2col_test.cc

@@ -14,9 +14,9 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/math/im2col.h"
 #include <gtest/gtest.h>
-#include <sys/time.h>
 #include <vector>
 #include "paddle/fluid/operators/math/im2col_cfo_cpu.h"
+#include "paddle/fluid/platform/port.h"
 
 template <typename DeviceContext, typename Place>
 void testIm2col() {

paddle/fluid/operators/math/jit_kernel_test.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/operators/math/jit_kernel.h"
-#include <sys/time.h>
 #include <cmath>    // for exp
 #include <cstring>  // for memcpy
 #include <random>
@@ -22,6 +21,7 @@ limitations under the License. */
 #include "gflags/gflags.h"
 #include "glog/logging.h"
 #include "gtest/gtest.h"
+#include "paddle/fluid/platform/port.h"
 
 #ifdef PADDLE_WITH_MKLML
 #include "paddle/fluid/platform/dynload/mklml.h"

paddle/fluid/platform/cudnn_helper.h

@@ -62,7 +62,7 @@ inline const char* cudnnGetErrorString(cudnnStatus_t status) {
 
 #define CUDNN_ENFORCE(condition)                                     \
   do {                                                               \
-    cudnnStatus_t status = condition;                                \
+    auto status = condition;                                         \
     if (UNLIKELY(status != CUDNN_STATUS_SUCCESS)) {                  \
       PADDLE_THROW(::paddle::platform::cudnnGetErrorString(status)); \
     }                                                                \

paddle/fluid/platform/dynload/cudnn.h

@@ -51,7 +51,7 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
 #define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name) \
   struct DynLoad__##__name {                    \
     template <typename... Args>                 \
-    inline cudnnStatus_t operator()(Args... args) { \
+    inline auto operator()(Args... args) {      \
       return ::__name(args...);                 \
     }                                           \
   };                                            \

paddle/fluid/platform/gpu_info.cc

@@ -19,7 +19,16 @@ limitations under the License. */
 #include "gflags/gflags.h"
 #include "paddle/fluid/platform/enforce.h"
 
-DEFINE_double(fraction_of_gpu_memory_to_use, 0.92,
+#ifndef _WIN32
+const float fraction_of_gpu_memory_to_use = 0.92f;
+#else
+// fraction_of_gpu_memory_to_use cannot be too high on windows,
+// since the win32 graphic sub-system can occupy some GPU memory
+// which may lead to insufficient memory left for paddle
+const float fraction_of_gpu_memory_to_use = 0.5f;
+#endif
+
+DEFINE_double(fraction_of_gpu_memory_to_use, fraction_of_gpu_memory_to_use,
               "Allocate a trunk of gpu memory that is this fraction of the "
               "total gpu memory size. Future memory usage will be allocated "
               "from the trunk. If the trunk doesn't have enough gpu memory, "

paddle/fluid/platform/stream_callback_manager.h

@@ -14,11 +14,11 @@
 
 #pragma once
 
+#include <ThreadPool.h>
 #include <cuda.h>
 #include <cuda_runtime.h>
 #include <functional>
 #include <memory>
-#include "ThreadPool.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {

paddle/legacy/cuda/include/hl_warpctc_wrap.h

@@ -12,9 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#ifndef _WIN32
 #ifndef HL_WARPCTC_WRAP_H_
 #define HL_WARPCTC_WRAP_H_
-
 #include "ctc.h"
 #include "hl_base.h"
 
@@ -91,3 +91,4 @@ extern void hl_warpctc_get_workspace_size(const int* cpuLabelLengths,
                                           size_t* bytes);
 
 #endif  // HL_WARPCTC_WRAP_H_
+#endif

paddle/legacy/cuda/src/hl_cuda_device.cc

@@ -132,11 +132,15 @@ inline pid_t gettid() {
   uint64_t tid;
   pthread_threadid_np(NULL, &tid);
 #else
+#ifndef _WIN32
 #ifndef __NR_gettid
 #define __NR_gettid 224
 #endif
   pid_t tid = syscall(__NR_gettid);
 #endif
+#else   // _WIN32
+  pid_t tid = _getpid();
+#endif  // _WIN32
   CHECK_NE((int)tid, -1);
   return tid;
 }

paddle/legacy/utils/ThreadLocal.h

@@ -14,10 +14,12 @@ limitations under the License. */
 
 #pragma once
 
+#ifndef _WIN32
 #include <pthread.h>
 #include <sys/syscall.h>
-#include <sys/types.h>
 #include <unistd.h>
+#endif
+#include <sys/types.h>
 #include <map>
 #include <mutex>
 #include <random>

paddle/legacy/utils/Util.h

@@ -14,7 +14,9 @@ limitations under the License. */
 
 #pragma once
 
+#ifndef _WIN32
 #include <sys/syscall.h>  // for syscall()
+#endif
 #include <sys/types.h>
 #include <algorithm>
 #include <cmath>
@@ -40,6 +42,31 @@ inline int rand_r(unsigned int* seedp) {
 }
 #endif
 
+#ifdef _WIN32
+#define NOMINMAX  // msvc max/min macro conflict with std::min/max
+#include <windows.h>
+
+template <typename T>
+inline int __builtin_clz(const T& value) {
+  DWORD leadning_zero = 0;
+  if (_BitScanReverse(&leadning_zero, value)) {
+    return static_cast<int>(sizeof(T) * 8 - leadning_zero);
+  } else {
+    return static_cast<int>(0);
+  }
+}
+
+inline int __builtin_clzl(const unsigned long& value) {
+  return __builtin_clz(value);
+}
+
+inline int __builtin_clzll(const unsigned long long& value) {
+  return __builtin_clz(value);
+}
+
+#define pid_t int
+#endif
+
 /**
  * Loop over the elements in a container
  * TODO(yuyang18): It's this foreach useful? Why not use C++ 11 foreach,

paddle/testing/CMakeLists.txt

@@ -3,8 +3,10 @@
 if(WITH_TESTING)
   add_library(paddle_test_main STATIC TestMain.cpp)
   add_dependencies(paddle_test_main paddle_proto ${external_project_dependencies})
+  if(NOT WIN32)
     add_library(paddle_test_util STATIC TestUtil.cpp)
     add_dependencies(paddle_test_util paddle_proto ${external_project_dependencies})
+  endif(NOT WIN32)
   if(NOT MOBILE_INFERENCE)
     cc_library(paddle_gtest_main SRCS paddle_gtest_main.cc DEPS device_context memory gtest gflags)
   endif()

python/paddle/fluid/metrics.py

@@ -46,8 +46,8 @@ def _is_numpy_(var):
 
 
 def _is_number_(var):
-    return isinstance(var, int) or isinstance(var, float) or (isinstance(
-        var, np.ndarray) and var.shape == (1, ))
+    return isinstance(var, int) or isinstance(var, np.int64) or isinstance(
+        var, float) or (isinstance(var, np.ndarray) and var.shape == (1, ))
 
 
 def _is_number_or_matrix_(var):