Merge pull request #14617 from wopeizl/windows/online

Windows/online

Author: wopeizl

paddle/fluid/framework/parallel_executor.cc

@@ -20,7 +20,7 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/ir/graph.h"
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
 
@@ -54,7 +54,7 @@ class ParallelExecutorPrivate {
   Scope *global_scope_;  // not owned
   std::unique_ptr<details::SSAGraphExecutor> executor_;
 
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
 #endif
   bool own_local_scope_;
@@ -104,7 +104,7 @@ ParallelExecutor::ParallelExecutor(
 
   if (member_->use_cuda_) {
 // Bcast Parameters to all GPUs
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
     auto *nccl_id_var = scope->FindVar(NCCL_ID_VARNAME);
     ncclUniqueId *nccl_id = nullptr;
     if (nccl_id_var != nullptr) {
@@ -124,7 +124,7 @@ ParallelExecutor::ParallelExecutor(
 
 // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
 // ncclOp
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
   std::unique_ptr<ir::Graph> graph = build_strategy.Apply(
       main_program, member_->places_, loss_var_name, params,
       member_->local_scopes_, member_->use_cuda_, member_->nccl_ctxs_.get());
@@ -213,7 +213,7 @@ void ParallelExecutor::BCastParamsToDevices(
     }
     auto &dims = main_tensor.dims();
     if (paddle::platform::is_gpu_place(main_tensor.place())) {
-#ifdef PADDLE_WITH_CUDA
+#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
       std::vector<void *> buffers;
       size_t numel = main_tensor.numel();
       ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());

paddle/fluid/inference/analysis/analyzer_tester.cc

@@ -76,7 +76,8 @@ 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((size_t)5UL, num_elements); i++) {
+  for (size_t i = 0; i < std::min(static_cast<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/memory/detail/system_allocator.cc

@@ -86,7 +86,11 @@ void CPUAllocator::Free(void* p, size_t size, size_t index) {
     munlock(p, size);
 #endif
   }
+#ifdef _WIN32
+  _aligned_free(p);
+#else
   free(p);
+#endif
 }
 
 bool CPUAllocator::UseGpu() const { return false; }

paddle/fluid/operators/dropout_op_test.cc

@@ -12,7 +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
 #include <unistd.h>
+#endif
 
 #include <string>
 #include <thread>  // NOLINT

paddle/fluid/operators/math/sequence_pooling.cu

@@ -16,13 +16,12 @@ limitations under the License. */
 #include "paddle/fluid/operators/math/math_function.h"
 #include "paddle/fluid/operators/math/sequence_pooling.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
+#include "paddle/fluid/platform/macros.h"
 
 namespace paddle {
 namespace operators {
 namespace math {
 
-#define FLT_MAX __FLT_MAX__
-
 template <typename T>
 struct MaxPoolFunctor {
   HOSTDEVICE void operator()(const T* input, const size_t start,

paddle/fluid/platform/gpu_info.cc

@@ -20,12 +20,12 @@ limitations under the License. */
 #include "paddle/fluid/platform/enforce.h"
 
 #ifndef _WIN32
-const float fraction_of_gpu_memory_to_use = 0.92f;
+constexpr static 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;
+constexpr static float fraction_of_gpu_memory_to_use = 0.5f;
 #endif
 
 DEFINE_double(fraction_of_gpu_memory_to_use, fraction_of_gpu_memory_to_use,

paddle/fluid/pybind/protobuf.cc

@@ -29,8 +29,16 @@ limitations under the License. */
 namespace pybind11 {
 namespace detail {
 
+#if !defined(PYBIND11_HIDDEN)
+#ifdef _WIN32
+#define PYBIND11_HIDDEN __declspec(dllexport)
+#else
+#define PYBIND11_HIDDEN __attribute__((visibility("hidden")))
+#endif
+#endif
+
 // Can be replaced by a generic lambda in C++14
-struct __attribute__((visibility("hidden"))) paddle_variant_caster_visitor
+struct PYBIND11_HIDDEN paddle_variant_caster_visitor
     : public boost::static_visitor<handle> {
   return_value_policy policy;
   handle parent;

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -63,7 +63,7 @@ function(py_test_modules TARGET_NAME)
     set(multiValueArgs MODULES DEPS ENVS)
     cmake_parse_arguments(py_test_modules "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
-             COMMAND env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_modules_ENVS}
+             COMMAND ${CMAKE_COMMAND} -E env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_modules_ENVS}
              ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/tools/test_runner.py ${py_test_modules_MODULES}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
     if (py_test_modules_SERIAL)