fix concat synchronization bug

paddle/fluid/operators/math/concat.cu

@@ -177,6 +177,9 @@ class ConcatFunctor<platform::CUDADeviceContext, T> {
           dev_ins_data, dev_ins_col_data, static_cast<int>(inputs_col.size()),
           out_row, out_col, output->data<T>());
     }
+    // Wait() must be called because `inputs_data` may be destructed before
+    // kernel ends
+    context.Wait();
   }
 };
 
@@ -252,6 +255,9 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
           input.data<T>(), in_row, in_col, dev_outs_col_data,
           static_cast<int>(outputs_cols.size()), dev_out_gpu_data);
     }
+    // Wait() must be called because `outputs_data` may be destructed before
+    // kernel ends
+    context.Wait();
   }
 };
 

paddle/fluid/platform/CMakeLists.txt

+if (NOT WIN32)
 proto_library(profiler_proto SRCS profiler.proto DEPS framework_proto)
 py_proto_compile(profiler_py_proto SRCS profiler.proto)
 
@@ -10,6 +11,7 @@ add_custom_command(TARGET profiler_py_proto POST_BUILD
         COMMAND cp *.py ${PADDLE_BINARY_DIR}/python/paddle/fluid/proto/profiler
         COMMENT "Copy generated python proto into directory paddle/fluid/proto/profiler."
         WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+endif(NOT WIN32)
 
 if(WITH_GPU)
   nv_library(enforce SRCS enforce.cc)
@@ -58,9 +60,12 @@ cc_test(init_test SRCS init_test.cc DEPS device_context)
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS memory place device_context)
 
+
+if (NOT WIN32)
 cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS})
 cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
+endif(NOT WIN32)
 
 nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor)
 cc_test(float16_test SRCS float16_test.cc DEPS lod_tensor)

paddle/fluid/platform/cpu_info.cc

@@ -22,9 +22,13 @@ limitations under the License. */
 #ifdef __APPLE__
 #include <sys/sysctl.h>
 #include <sys/types.h>
+
+#elif defined(_WIN32)
+#define NOMINMAX  // msvc max/min macro conflict with std::min/max
+#include <windows.h>
 #else
 #include <unistd.h>
-#endif
+#endif  // _WIN32
 
 #include <algorithm>
 #include "gflags/gflags.h"
@@ -32,16 +36,20 @@ limitations under the License. */
 DEFINE_double(fraction_of_cpu_memory_to_use, 1,
               "Default use 100% of CPU memory for PaddlePaddle,"
               "reserve the rest for page tables, etc");
-
+#if !defined(_WIN32)
 DEFINE_uint64(initial_cpu_memory_in_mb,
 #ifdef PADDLE_WITH_MKLDNN
               /* Aligned with mozga-intel, MKLDNN need at least 5000 MB
                * to obtain the best performance*/
-              5000,
+              5000ul,
 #else
-              500,
+              500ul,
 #endif
               "Initial CPU memory for PaddlePaddle, in MD unit.");
+#else
+DEFINE_uint64(initial_cpu_memory_in_mb, 500ul,
+              "Initial CPU memory for PaddlePaddle, in MD unit.");
+#endif  // !defined(_WIN32)
 
 DEFINE_double(
     fraction_of_cuda_pinned_memory_to_use, 0.5,
@@ -60,6 +68,11 @@ inline size_t CpuTotalPhysicalMemory() {
   size_t len = sizeof(size);
   if (sysctl(mib, 2, &size, &len, NULL, 0) == 0) return (size_t)size;
   return 0L;
+#elif defined(_WIN32)
+  MEMORYSTATUSEX sMeminfo;
+  sMeminfo.dwLength = sizeof(sMeminfo);
+  GlobalMemoryStatusEx(&sMeminfo);
+  return sMeminfo.ullTotalPhys;
 #else
   int64_t pages = sysconf(_SC_PHYS_PAGES);
   int64_t page_size = sysconf(_SC_PAGE_SIZE);

paddle/fluid/platform/device_tracer.h

@@ -13,7 +13,12 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #pragma once
 
+#if !defined(_WIN32)
 #include <sys/time.h>
+#else
+#include <windows.h>
+#endif  // !_WIN32
+
 #include <time.h>
 #include <chrono>  // NOLINT
 #include <string>
@@ -27,12 +32,15 @@ namespace platform {
 ///////////////////////
 // WARN: Under Development. Don't depend on it yet.
 //////////////////////
-
+#if !defined(_WIN32)
 inline uint64_t PosixInNsec() {
   struct timeval tv;
   gettimeofday(&tv, nullptr);
   return 1000 * (static_cast<uint64_t>(tv.tv_sec) * 1000000 + tv.tv_usec);
 }
+#else
+inline uint64_t PosixInNsec() { return static_cast<uint64_t>(0); }
+#endif  // !_WIN32
 
 // DeviceTracer performs the following tasks:
 // 1. Register cuda callbacks for various events: kernel, memcpy, etc.

paddle/fluid/platform/dynload/CMakeLists.txt

@@ -16,7 +16,9 @@ if (CUPTI_FOUND)
     list(APPEND CUDA_SRCS cupti.cc)
 endif(CUPTI_FOUND)
 nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader)
+if (NOT WIN32)
 cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc)
+endif(NOT WIN32)
 if (WITH_MKLML)
     cc_library(dynload_mklml SRCS mklml.cc DEPS dynamic_loader mklml)
 endif()

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -13,8 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 #include "paddle/fluid/platform/dynload/dynamic_loader.h"
 
-#include <dlfcn.h>
-
 #include <memory>
 #include <mutex>  // NOLINT
 #include <string>
@@ -23,6 +21,7 @@ limitations under the License. */
 #include "glog/logging.h"
 #include "paddle/fluid/platform/dynload/cupti_lib_path.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/port.h"
 
 DEFINE_string(cudnn_dir, "",
               "Specify path for loading libcudnn.so. For instance, "

paddle/fluid/platform/enforce.h

@@ -18,6 +18,11 @@ limitations under the License. */
 #include <cxxabi.h>  // for __cxa_demangle
 #endif               // __GNUC__
 
+#if defined(_WIN32)
+#define NOMINMAX  // msvc max/min macro conflict with std::min/max
+#define GLOG_NO_ABBREVIATED_SEVERITIES  // msvc conflict logging with windows.h
+#endif
+
 #ifdef PADDLE_WITH_CUDA
 #include <cublas_v2.h>
 #include <cudnn.h>
@@ -117,7 +122,12 @@ struct EOFException : public std::exception {
 // always forces branch prediction of true.
 // This generates faster binary code. __builtin_expect is since C++11.
 // For more details, please check https://stackoverflow.com/a/43870188/724872.
+#if !defined(_WIN32)
 #define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
+#else
+// there is no equivalent intrinsics in msvc.
+#define UNLIKELY(condition) (condition == 0)
+#endif
 
 template <typename... Args>
 inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
@@ -230,6 +240,7 @@ inline void throw_on_error(T e) {
   throw_on_error(e, "");
 }
 
+#if !defined(_WIN32)
 #define PADDLE_THROW(...)                                              \
   do {                                                                 \
     throw ::paddle::platform::EnforceNotMet(                           \
@@ -248,15 +259,28 @@ inline void throw_on_error(T e) {
                                               __FILE__, __LINE__);      \
     }                                                                   \
   } while (false)
-#else
-#define PADDLE_ENFORCE(...) ::paddle::platform::throw_on_error(__VA_ARGS__);
-#endif
 
 #define PADDLE_THROW_EOF()                                                     \
   do {                                                                         \
     throw ::paddle::platform::EOFException("There is no next data.", __FILE__, \
                                            __LINE__);                          \
   } while (false)
+
+#else
+#define PADDLE_ENFORCE(...) ::paddle::platform::throw_on_error(__VA_ARGS__)
+#endif  // REPLACE_ENFORCE_GLOG
+
+#else  // !_WIN32
+// disable enforce, caused by the varardic macro exception error
+#define PADDLE_THROW(x)                                      \
+  do {                                                       \
+    throw std::make_exception_ptr(                           \
+        std::runtime_error("Windows disable the enforce.")); \
+  } while (false)
+
+#define PADDLE_ENFORCE(x, ...) x
+#endif  // !_WIN32
+
 /*
  * Some enforce helpers here, usage:
  *    int a = 1;

paddle/fluid/platform/profiler.h

@@ -69,6 +69,7 @@ void PushEvent(const std::string& name, const DeviceContext* dev_ctx);
 
 void PopEvent(const std::string& name, const DeviceContext* dev_ctx);
 
+#if !defined(_WIN32)
 struct RecordEvent {
   RecordEvent(const std::string& name, const DeviceContext* dev_ctx);
 
@@ -94,6 +95,15 @@ struct RecordBlock {
   std::string name_;
   uint64_t start_ns_;
 };
+#else
+// windows do not support profiler temporarily.
+struct RecordEvent {
+  RecordEvent(const std::string& name, const DeviceContext* dev_ctx) {}
+};
+struct RecordBlock {
+  explicit RecordBlock(int block_id) {}
+};
+#endif
 
 // Return the event list of all threads. Assumed the returned value calls
 // event_lists, event_lists[i][j] represents the j-th Event of i-th thread.