Merge pull request #3784 from QiJune/refine_LODTensor

Refine LODTensor using composition instead of derivation

paddle/framework/lod_tensor.cc

@@ -19,25 +19,24 @@
 namespace paddle {
 namespace framework {
 
-LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin,
-                                           size_t level_end) const {
+LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) {
   LOD new_lod;
   new_lod.reserve(level_end - level_begin);
   for (size_t i = level_begin; i < level_end; i++) {
-    new_lod.emplace_back(at(i));
+    new_lod.emplace_back(in.at(i));
   }
   return new_lod;
 }
 
-LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
-                                            size_t elem_end) const {
+LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+                 size_t elem_end) {
   // slice the lod.
   LOD new_lod;
-  new_lod.reserve(size() - level);
-  auto start = this->at(level)[elem_begin];
-  auto end = this->at(level)[elem_end];
+  new_lod.reserve(in.size() - level);
+  auto start = in.at(level)[elem_begin];
+  auto end = in.at(level)[elem_end];
 
-  for (auto it = this->begin() + level; it != this->end(); it++) {
+  for (auto it = in.begin() + level; it != in.end(); it++) {
     auto it_begin = std::find(it->begin(), it->end(), start);
     auto it_end = std::find(it_begin, it->end(), end);
     PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
@@ -49,11 +48,11 @@ LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
                    [start](int v) { return v - start; });
     PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD");
   }
-  PADDLE_ENFORCE_LE(new_lod.size(), this->size());
+  PADDLE_ENFORCE_LE(new_lod.size(), in.size());
   return new_lod;
 }
 
-bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
+bool operator==(const LOD& a, const LOD& b) {
   if (a.size() != b.size()) {
     return false;
   }
@@ -70,9 +69,27 @@ bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
       }
     }
   }
-
   return true;
 }
 
+void LODTensor::SliceLevels(size_t level_begin, size_t level_end) {
+  auto new_lod = framework::SliceLevels(lod_, level_begin, level_end);
+  lod_ = new_lod;
+}
+
+void LODTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) {
+  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
+                 NumLevels());
+  PADDLE_ENFORCE(elem_begin < NumElements(level),
+                 "element begin [%d] out of range [%d]", elem_begin,
+                 NumElements(level));
+  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
+                 "element end [%d] out of range [%d]", elem_end,
+                 NumElements(level));
+
+  auto new_lod = framework::SliceInLevel(lod_, level, elem_begin, elem_end);
+  lod_ = new_lod;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.h

@@ -15,7 +15,7 @@
 #pragma once
 
 #include <memory>
-#if !defined(PADDLE_ONLY_CPU)
+#ifndef PADDLE_ONLY_CPU
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
 #endif
@@ -27,33 +27,39 @@
 namespace paddle {
 namespace framework {
 
+#ifdef PADDLE_ONLY_CPU
+template <typename T>
+using Vector = std::vector<T>;
+#else
+template <typename T>
+using Vector = thrust::host_vector<T>;
+#endif
+
+using LOD = std::vector<Vector<size_t>>;
+
+LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end);
+
+LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+                 size_t elem_end);
+
+bool operator==(const LOD& a, const LOD& b);
+
 /*
  * LODTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LODTensor : public Tensor {
- public:
-// Level save offsets of each unit.
-#ifdef PADDLE_ONLY_CPU
-  template <typename T>
-  using Vector = std::vector<T>;
-#else
-  template <typename T>
-  using Vector = thrust::host_vector<T>;
-#endif
-  // LoD stores offsets of each level of units, the largest units level first,
-  // then the smaller units level. Each Level stores the offsets of units in
-  // Tesor.
-  class LOD : public std::vector<Vector<size_t>> {
+class LODTensor {
  public:
-    LOD SliceLevels(size_t level_begin, size_t level_end) const;
-    LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const;
-  };
-
   LODTensor() {}
-  explicit LODTensor(const LOD &lod) : lod_(lod) {}
+  LODTensor(const LOD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
+
+  void set_lod(const LOD& lod) { lod_ = lod; }
+
+  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
+
+  Tensor& tensor() { return *tensor_; }
 
-  virtual Tensor *Clone() const { return new LODTensor(lod_); }
+  LOD lod() { return lod_; }
 
   /*
    * Get a element from LOD.
@@ -79,71 +85,23 @@ class LODTensor : public Tensor {
     PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
                    NumLevels());
     // the last offset is the end of last element
-    return lod_[level].size() - 1;
+    return (lod_)[level].size() - 1;
   }
 
   /*
-   * Slice of levels[level_begin:level_end], with tensor shared.
+   * Slice of levels[level_begin:level_end]
    */
-  template <typename T>
-  LODTensor SliceLevels(size_t level_begin, size_t level_end) const;
+  void SliceLevels(size_t level_begin, size_t level_end);
 
   /*
-   * Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
+   * Slice of elements of a level, [elem_begin: elem_end]
    * @note: low performance in slice lod_.
    */
-  template <typename T>
-  LODTensor SliceInLevel(size_t level, size_t elem_begin,
-                         size_t elem_end) const;
-
-  /*
-   * Copy other's lod_'s content, free to mutate.
-   */
-  void CopyLOD(const LODTensor &other) { lod_ = other.lod_; }
-  /*
-   * Determine whether LODTensor has a valid LOD info.
-   */
-  const LOD &lod() const { return lod_; }
-  LOD *mutable_lod() { return &lod_; }
-
-  virtual ~LODTensor() {}
+  void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end);
 
  private:
   LOD lod_;
+  Tensor* tensor_;  // not owned
 };
-
-bool operator==(const LODTensor::LOD &a, const LODTensor::LOD &b);
-
-template <typename T>
-LODTensor LODTensor::SliceLevels(size_t level_begin, size_t level_end) const {
-  auto new_lod = lod_.SliceLevels(level_begin, level_end);
-  // slice levels just need to update LOD info, each level will contains the
-  // whole tensor_, so no need to modify tensor_.
-  LODTensor new_tensor(new_lod);
-  new_tensor.ShareDataWith<T>(*this);
-  return new_tensor;
-}
-
-template <typename T>
-LODTensor LODTensor::SliceInLevel(size_t level, size_t elem_begin,
-                                  size_t elem_end) const {
-  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
-                 NumLevels());
-  PADDLE_ENFORCE(elem_begin < NumElements(level),
-                 "element begin [%d] out of range [%d]", elem_begin,
-                 NumElements(level));
-  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
-                 "element end [%d] out of range [%d]", elem_end,
-                 NumElements(level));
-
-  auto new_lod = lod_.SliceInLevel(level, elem_begin, elem_end);
-
-  // slice elements just need to update LOD info, because offsets are not
-  // changed, so the original tensor_ can be reused.
-  LODTensor new_tensor(new_lod);
-  new_tensor.ShareDataWith<T>(*this);
-  return new_tensor;
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor_test.cc

@@ -24,13 +24,12 @@ namespace framework {
 class LODTensorTester : public ::testing::Test {
  public:
   virtual void SetUp() override {
-    lod_tensor.reset(new LODTensor);
     // tensor's batch_size: 30
     // 3 levels
     // 0 10 20
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
-    LODTensor::LOD lod;
+    LOD lod;
     lod.push_back(std::vector<size_t>{0, 10, 20});
     lod.push_back(std::vector<size_t>{0, 5, 10, 15, 20});
     lod.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
@@ -41,75 +40,65 @@ class LODTensorTester : public ::testing::Test {
     // malloc memory
     tensor.mutable_data<float>(place);
 
-    lod_tensor.reset(new LODTensor(lod));
-    lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/});
-
-    lod_tensor->ShareDataWith<float>(tensor);
-    // lod_tensor->ShareDataWith<Tensor>(tensor);
+    lod_tensor.set_lod(lod);
+    lod_tensor.set_tensor(&tensor);
   }
 
  protected:
-  std::unique_ptr<LODTensor> lod_tensor;
   platform::CPUPlace place;
   Tensor tensor;
+  LODTensor lod_tensor;
 };
 
-TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor->NumLevels(), 3UL); }
+TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
 
 TEST_F(LODTensorTester, NumElements) {
-  ASSERT_EQ(lod_tensor->NumElements(0), 2UL);
-  ASSERT_EQ(lod_tensor->NumElements(1), 4UL);
-  ASSERT_EQ(lod_tensor->NumElements(2), 8UL);
+  ASSERT_EQ(lod_tensor.NumElements(0), 2UL);
+  ASSERT_EQ(lod_tensor.NumElements(1), 4UL);
+  ASSERT_EQ(lod_tensor.NumElements(2), 8UL);
 }
 
 TEST_F(LODTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 1);
+    LODTensor new_lod_tensor = lod_tensor;
+    new_lod_tensor.SliceLevels(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
-    // ASSERT_EQ(new_lod_tensor, *lod_tensor);
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+              lod_tensor.tensor().data<float>());
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 2);
+    LODTensor new_lod_tensor = lod_tensor;
+    new_lod_tensor.SliceLevels(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor->NumElements(level));
-    ASSERT_EQ(new_lod_tensor.NumElements(1),
-              lod_tensor->NumElements(level + 1));
-    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.NumElements(1), lod_tensor.NumElements(level + 1));
+    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+              lod_tensor.tensor().data<float>());
   }
 }
 
 TEST_F(LODTensorTester, SliceInLevel) {
   size_t level = 0;
-  auto new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  LODTensor new_lod_tensor = lod_tensor;
+  new_lod_tensor.SliceInLevel(level, 0, 2);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
   EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
   EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
   EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
+  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+            lod_tensor.tensor().data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  new_lod_tensor = lod_tensor;
+  new_lod_tensor.SliceInLevel(level, 0, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
-}
-
-TEST_F(LODTensorTester, ShareLOD) {
-  LODTensor new_lod_tensor;
-  new_lod_tensor.CopyLOD(*lod_tensor);
-  ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod());
-}
-
-TEST_F(LODTensorTester, CopyLOD) {
-  LODTensor new_lod_tensor;
-  new_lod_tensor.CopyLOD(*lod_tensor);
-  bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(),
-                           new_lod_tensor.lod().begin());
-  ASSERT_TRUE(equals);
+  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+            lod_tensor.tensor().data<float>());
 }
 
 }  // namespace framework