Revert "Revert "Merge pull request #13431 from chengduoZH/refine_lod""

This reverts commit a6c8d6b9a2.

paddle/fluid/framework/details/cow_ptr.h

@@ -20,79 +20,41 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-// Change it to thread safe flags if needed.
-class ThreadUnsafeOwnershipFlags {
+template <class T>
+class COWPtr {
  public:
-  explicit ThreadUnsafeOwnershipFlags(bool flag) : flag_(flag) {}
-
-  ThreadUnsafeOwnershipFlags(const ThreadUnsafeOwnershipFlags& other) = delete;
-  ThreadUnsafeOwnershipFlags& operator=(
-      const ThreadUnsafeOwnershipFlags& other) = delete;
-  ThreadUnsafeOwnershipFlags(ThreadUnsafeOwnershipFlags&& other) = default;
+  typedef std::shared_ptr<T> RefPtr;
 
-  void SetOwnership(bool flag) { flag_ = flag; }
+ private:
+  RefPtr m_sp;
 
-  // Invoke the callback if it is not owned.
-  template <typename Callback>
-  void AcquireOwnershipOnce(Callback acquire) {
-    if (!flag_) {
-      acquire();
-      flag_ = true;
+  void detach() {
+    T* tmp = m_sp.get();
+    if (!(tmp == nullptr || m_sp.unique())) {
+      m_sp = RefPtr(new T(*tmp));
     }
   }
 
- private:
-  bool flag_;
-};
-
-// Copy-On-Write pointer.
-// It will hold a T* pointer, and only copy once when `MutableData` is invoked.
-//
-// The template parameter OwnershipFlags should have:
-//   * a constructor takes a bool. True if own.
-//   * SetOwnership(bool flag).
-//   * AcquireOwnershipOnce(Callback). It will invoke the callback if it is not
-//     owned.
-//
-// https://en.wikipedia.org/wiki/Copy-on-write
-template <typename T, typename OwnershipFlags = ThreadUnsafeOwnershipFlags>
-class COWPtr {
  public:
-  // Ctor from raw pointer.
-  explicit COWPtr(T* ptr) : payload_(ptr), ownership_{true} {}
+  COWPtr() : m_sp(nullptr) {}
+  explicit COWPtr(T* t) : m_sp(t) {}
+  explicit COWPtr(const RefPtr& refptr) : m_sp(refptr) {}
 
-  // Move methods. Steal ownership from origin
-  COWPtr(COWPtr&& other)
-      : payload_(other.payload_), ownership_{std::move(other.ownership_)} {}
-  COWPtr& operator=(COWPtr&& origin) = default;
+  const T& Data() const { return operator*(); }
 
-  // Copy methods. Not own payload
-  COWPtr(const COWPtr& other) : payload_(other.payload_), ownership_{false} {}
-  COWPtr& operator=(const COWPtr& other) {
-    payload_ = other.payload_;
-    ownership_.SetOwnership(false);
-    return *this;
-  }
-
-  // Access read only data.
-  const T& Data() const { return *payload_; }
+  T* MutableData() { return operator->(); }
 
-  // Access mutable data. If the data is not owned, the data will be copied
-  // before.
-  T* MutableData() {
-    ownership_.AcquireOwnershipOnce(
-        [this] { payload_.reset(new T(*payload_)); });
-    return payload_.get();
+  const T& operator*() const { return *m_sp; }
+  T& operator*() {
+    detach();
+    return *m_sp;
+  }
+  const T* operator->() const { return m_sp.operator->(); }
+  T* operator->() {
+    detach();
+    return m_sp.operator->();
   }
-
- private:
-  // Actual data pointer.
-  std::shared_ptr<T> payload_;
-
-  // Ownership flag.
-  OwnershipFlags ownership_;
 };
-
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/cow_ptr_test.cc

@@ -30,6 +30,14 @@ TEST(COWPtr, all) {
   ASSERT_EQ(ptr2.Data(), 10);
 }
 
+TEST(COWPtr, change_old) {
+  COWPtr<int> ptr(new int{0});
+  COWPtr<int> ptr2 = ptr;
+  *ptr.MutableData() = 10;
+  ASSERT_EQ(ptr2.Data(), 0);
+  ASSERT_EQ(ptr.Data(), 10);
+}
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/operators/detection_map_op.h

@@ -76,8 +76,8 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     auto ap_type = GetAPType(ctx.Attr<std::string>("ap_type"));
     int class_num = ctx.Attr<int>("class_num");
 
-    auto label_lod = in_label->lod();
-    auto detect_lod = in_detect->lod();
+    auto& label_lod = in_label->lod();
+    auto& detect_lod = in_detect->lod();
     PADDLE_ENFORCE_EQ(label_lod.size(), 1UL,
                       "Only support one level sequence now.");
     PADDLE_ENFORCE_EQ(label_lod[0].size(), detect_lod[0].size(),
@@ -166,11 +166,11 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     auto labels = framework::EigenTensor<T, 2>::From(input_label);
     auto detect = framework::EigenTensor<T, 2>::From(input_detect);
 
-    auto label_lod = input_label.lod();
-    auto detect_lod = input_detect.lod();
+    auto& label_lod = input_label.lod();
+    auto& detect_lod = input_detect.lod();
 
     int batch_size = label_lod[0].size() - 1;
-    auto label_index = label_lod[0];
+    auto& label_index = label_lod[0];
 
     for (int n = 0; n < batch_size; ++n) {
       std::map<int, std::vector<Box>> boxes;
@@ -274,7 +274,6 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
 
     output_true_pos->set_lod(true_pos_lod);
     output_false_pos->set_lod(false_pos_lod);
-    return;
   }
 
   void GetInputPos(const framework::Tensor& input_pos_count,
@@ -292,7 +291,7 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
     auto SetData = [](const framework::LoDTensor& pos_tensor,
                       std::map<int, std::vector<std::pair<T, int>>>& pos) {
       const T* pos_data = pos_tensor.data<T>();
-      auto pos_data_lod = pos_tensor.lod()[0];
+      auto& pos_data_lod = pos_tensor.lod()[0];
       for (size_t i = 0; i < pos_data_lod.size() - 1; ++i) {
         for (size_t j = pos_data_lod[i]; j < pos_data_lod[i + 1]; ++j) {
           T score = pos_data[j * 2];
@@ -317,20 +316,23 @@ class DetectionMAPOpKernel : public framework::OpKernel<T> {
       std::map<int, std::vector<std::pair<T, int>>>* false_pos) const {
     int batch_size = gt_boxes.size();
     for (int n = 0; n < batch_size; ++n) {
-      auto image_gt_boxes = gt_boxes[n];
-      for (auto it = image_gt_boxes.begin(); it != image_gt_boxes.end(); ++it) {
+      auto& image_gt_boxes = gt_boxes[n];
+      for (auto& image_gt_box : image_gt_boxes) {
         size_t count = 0;
-        auto labeled_bboxes = it->second;
+        auto& labeled_bboxes = image_gt_box.second;
         if (evaluate_difficult) {
           count = labeled_bboxes.size();
         } else {
-          for (size_t i = 0; i < labeled_bboxes.size(); ++i)
-            if (!(labeled_bboxes[i].is_difficult)) ++count;
+          for (auto& box : labeled_bboxes) {
+            if (!box.is_difficult) {
+              ++count;
+            }
+          }
         }
         if (count == 0) {
           continue;
         }
-        int label = it->first;
+        int label = image_gt_box.first;
         if (label_pos_count->find(label) == label_pos_count->end()) {
           (*label_pos_count)[label] = count;
         } else {

paddle/fluid/operators/extract_rows_op.cc

@@ -50,7 +50,7 @@ class ExtractRowsOp : public framework::OperatorBase {
     auto &in = scope.FindVar(Input("X"))->Get<framework::SelectedRows>();
     auto out = scope.FindVar(Output("Out"))->GetMutable<framework::LoDTensor>();
 
-    auto in_rows = in.rows();
+    auto &in_rows = in.rows();
     auto out_dim = framework::make_ddim(
         std::vector<int64_t>{static_cast<int64_t>(in_rows.size()), 1});
     auto dst_ptr = out->mutable_data<int64_t>(out_dim, in.place());

paddle/fluid/operators/math/selected_rows_functor.cu

@@ -60,11 +60,9 @@ struct SelectedRowsAdd<platform::CUDADeviceContext, T> {
     auto out_place = context.GetPlace();
     PADDLE_ENFORCE(platform::is_gpu_place(out_place));
 
-    memory::Copy(
-        boost::get<platform::CUDAPlace>(out_place), out_data,
-        boost::get<platform::CUDAPlace>(in1_place), in1_data,
-        in1_value.numel() * sizeof(T),
-        reinterpret_cast<const platform::CUDADeviceContext&>(context).stream());
+    memory::Copy(boost::get<platform::CUDAPlace>(out_place), out_data,
+                 boost::get<platform::CUDAPlace>(in1_place), in1_data,
+                 in1_value.numel() * sizeof(T), context.stream());
 
     auto* in2_data = in2_value.data<T>();
     memory::Copy(boost::get<platform::CUDAPlace>(out_place),
@@ -148,7 +146,7 @@ struct SelectedRowsAddTo<platform::CUDADeviceContext, T> {
     auto in1_height = input1.height();
     PADDLE_ENFORCE_EQ(in1_height, input2->height());
 
-    framework::Vector<int64_t> in1_rows(input1.rows());
+    auto& in1_rows = input1.rows();
     auto& in2_rows = *(input2->mutable_rows());
 
     auto& in1_value = input1.value();

paddle/fluid/operators/sum_op.h

@@ -123,7 +123,6 @@ class SumKernel : public framework::OpKernel<T> {
 
       out_value->Resize(framework::make_ddim(in_dim));
       out_value->mutable_data<T>(context.GetPlace());
-
       // if all the input sparse vars are empty, no need to
       // merge these vars.
       if (first_dim == 0UL) {

python/paddle/fluid/tests/unittests/op_test.py

@@ -345,7 +345,7 @@ class OpTest(unittest.TestCase):
                         actual_t, expect_t, atol=atol, equal_nan=equal_nan),
                     "Output (" + out_name + ") has diff at " + str(place) +
                     "\nExpect " + str(expect_t) + "\n" + "But Got" +
-                    str(actual_t))
+                    str(actual_t) + " in class " + self.__class__.__name__)
                 if isinstance(expect, tuple):
                     self.assertListEqual(actual.recursive_sequence_lengths(),
                                          expect[1], "Output (" + out_name +

python/paddle/fluid/tests/unittests/test_detection_map_op.py

@@ -20,6 +20,7 @@ import six
 import sys
 import collections
 import math
+import paddle.fluid as fluid
 from op_test import OpTest
 
 
@@ -32,7 +33,7 @@ class TestDetectionMAPOp(OpTest):
         self.detect = np.array(self.detect).astype('float32')
         self.mAP = np.array(self.mAP).astype('float32')
 
-        if (len(self.class_pos_count) > 0):
+        if len(self.class_pos_count) > 0:
             self.class_pos_count = np.array(self.class_pos_count).astype(
                 'int32')
             self.true_pos = np.array(self.true_pos).astype('float32')
@@ -273,7 +274,7 @@ class TestDetectionMAPOp11Point(TestDetectionMAPOp):
 class TestDetectionMAPOpMultiBatch(TestDetectionMAPOp):
     def init_test_case(self):
         super(TestDetectionMAPOpMultiBatch, self).init_test_case()
-        self.class_pos_count = [0, 2, 1]
+        self.class_pos_count = [0, 2, 1, 0]
         self.true_pos_lod = [[0, 3, 2]]
         self.true_pos = [[0.7, 1.], [0.3, 0.], [0.2, 1.], [0.8, 0.], [0.1, 1.]]
         self.false_pos_lod = [[0, 3, 2]]