Fix singleton. (#11835)

paddle/fluid/framework/op_info.cc

@@ -21,8 +21,8 @@ namespace framework {
 // a static local variable is already being initialized.
 // https://stackoverflow.com/questions/11711920/how-to-implement-multithread-safe-singleton-in-c11-without-using-mutex
 OpInfoMap& OpInfoMap::Instance() {
-  static OpInfoMap* g_op_info_map = new OpInfoMap();
-  return *g_op_info_map;
+  static OpInfoMap g_op_info_map;
+  return g_op_info_map;
 }
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/memory/detail/buddy_allocator.cc

@@ -19,8 +19,9 @@ namespace paddle {
 namespace memory {
 namespace detail {
 
-BuddyAllocator::BuddyAllocator(SystemAllocator* system_allocator,
-                               size_t min_chunk_size, size_t max_chunk_size)
+BuddyAllocator::BuddyAllocator(
+    std::unique_ptr<SystemAllocator> system_allocator, size_t min_chunk_size,
+    size_t max_chunk_size)
     : min_chunk_size_(min_chunk_size),
       max_chunk_size_(max_chunk_size),
       cache_(system_allocator->UseGpu()),

paddle/fluid/memory/detail/buddy_allocator.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <memory>
 #include <mutex>  // NOLINT
 #include <set>
 #include <tuple>
@@ -32,8 +33,8 @@ namespace detail {
 
 class BuddyAllocator {
  public:
-  BuddyAllocator(SystemAllocator* system_allocator, size_t min_chunk_size,
-                 size_t max_chunk_size);
+  BuddyAllocator(std::unique_ptr<SystemAllocator> system_allocator,
+                 size_t min_chunk_size, size_t max_chunk_size);
 
   ~BuddyAllocator();
 
@@ -103,7 +104,7 @@ class BuddyAllocator {
 
  private:
   /*! Allocate CPU/GPU memory from system */
-  SystemAllocator* system_allocator_;
+  std::unique_ptr<SystemAllocator> system_allocator_;
   std::mutex mutex_;
 };
 

paddle/fluid/memory/malloc.cc

@@ -12,6 +12,8 @@ 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 <vector>
+
 #include "paddle/fluid/memory/malloc.h"
 
 #include "glog/logging.h"
@@ -34,12 +36,15 @@ namespace memory {
 using BuddyAllocator = detail::BuddyAllocator;
 
 BuddyAllocator* GetCPUBuddyAllocator() {
+  static std::once_flag init_flag;
   static detail::BuddyAllocator* a = nullptr;
-  if (a == nullptr) {
-    a = new detail::BuddyAllocator(new detail::CPUAllocator,
-                                   platform::CpuMinChunkSize(),
-                                   platform::CpuMaxChunkSize());
-  }
+
+  std::call_once(init_flag, []() {
+    a = new detail::BuddyAllocator(
+        std::unique_ptr<detail::SystemAllocator>(new detail::CPUAllocator),
+        platform::CpuMinChunkSize(), platform::CpuMaxChunkSize());
+  });
+
   return a;
 }
 
@@ -68,19 +73,22 @@ size_t Used<platform::CPUPlace>(platform::CPUPlace place) {
 #ifdef PADDLE_WITH_CUDA
 
 BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
-  static BuddyAllocator** as = NULL;
-  if (as == NULL) {
+  static std::once_flag init_flag;
+  static detail::BuddyAllocator** a_arr = nullptr;
+
+  std::call_once(init_flag, [gpu_id]() {
     int gpu_num = platform::GetCUDADeviceCount();
-    as = new BuddyAllocator*[gpu_num];
-    for (int gpu = 0; gpu < gpu_num; gpu++) {
-      as[gpu] = nullptr;
-    }
-  }
-  platform::SetDeviceId(gpu_id);
-  if (!as[gpu_id]) {
-    as[gpu_id] = new BuddyAllocator(new detail::GPUAllocator(gpu_id),
-                                    platform::GpuMinChunkSize(),
-                                    platform::GpuMaxChunkSize());
+    PADDLE_ENFORCE(gpu_id < gpu_num, "gpu_id:%d should < gpu_num:%d", gpu_id,
+                   gpu_num);
+
+    a_arr = new BuddyAllocator*[gpu_num];
+    for (int i = 0; i < gpu_num; i++) {
+      a_arr[i] = nullptr;
+      platform::SetDeviceId(i);
+      a_arr[i] = new BuddyAllocator(
+          std::unique_ptr<detail::SystemAllocator>(new detail::GPUAllocator(i)),
+          platform::GpuMinChunkSize(), platform::GpuMaxChunkSize());
+
       VLOG(10) << "\n\nNOTE: each GPU device use "
                << FLAGS_fraction_of_gpu_memory_to_use * 100
                << "% of GPU memory.\n"
@@ -88,7 +96,10 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
                << "FLAGS_fraction_of_gpu_memory_to_use"
                << "' to change the fraction of GPU usage.\n\n";
     }
-  return as[gpu_id];
+  });
+
+  platform::SetDeviceId(gpu_id);
+  return a_arr[gpu_id];
 }
 
 template <>
@@ -125,12 +136,16 @@ void Free<platform::CUDAPlace>(platform::CUDAPlace place, void* p) {
 }
 
 BuddyAllocator* GetCUDAPinnedBuddyAllocator() {
-  static BuddyAllocator* ba = NULL;
-  if (ba == NULL) {
-    ba = new BuddyAllocator(new detail::CUDAPinnedAllocator,
+  static std::once_flag init_flag;
+  static BuddyAllocator* ba = nullptr;
+
+  std::call_once(init_flag, []() {
+    ba = new BuddyAllocator(std::unique_ptr<detail::SystemAllocator>(
+                                new detail::CUDAPinnedAllocator),
                             platform::CUDAPinnedMinChunkSize(),
                             platform::CUDAPinnedMaxChunkSize());
-  }
+  });
+
   return ba;
 }