lod_tensor.h 6.1 KB
Newer Older
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

L
Luo Tao 已提交
3 4 5
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
8

L
Luo Tao 已提交
9 10 11 12 13
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
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. */
14 15 16 17

#pragma once

#include <memory>
18
#ifdef PADDLE_WITH_CUDA
19 20
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
21
#include <thrust/system/cuda/experimental/pinned_allocator.h>
22 23
#endif

24
#include <glog/logging.h>
25 26
#include "paddle/framework/ddim.h"
#include "paddle/framework/tensor.h"
D
dzhwinter 已提交
27
#include "paddle/framework/tensor_util.h"
28
#include "paddle/platform/enforce.h"
29
#include "paddle/platform/place.h"
30 31 32 33

namespace paddle {
namespace framework {

34
#ifndef PADDLE_WITH_CUDA
Q
qijun 已提交
35 36 37 38
template <typename T>
using Vector = std::vector<T>;
#else
template <typename T>
39 40
using Vector = thrust::host_vector<
    T, thrust::system::cuda::experimental::pinned_allocator<T>>;
Q
qijun 已提交
41 42
#endif

43
/*
44
 * LoD is short for Level of Details.
45
 *
46
 * - in a level, each element indicates relative offset of the lower level
47 48 49
 * - the first element should be 0 and that indicates that this sequence start
 * from 0
 * - each sequence's begin and end(no-inclusive) is level[id, id+1]
50 51 52 53 54 55 56
 *
 * For example:
 *    3-level LoD stores
 *
 *    0 2 3
 *    0 2 4 7
 *    0 2 5 7 10 12 15 20
57
 */
58
using LoD = std::vector<Vector<size_t>>;
Q
qijun 已提交
59

60 61
std::ostream& operator<<(std::ostream& os, const LoD& lod);

62 63 64 65 66 67
/*
 * Slice levels from a LoD.
 * NOTE the lowest level should always be the absolute offsets of the underlying
 * tensor instances. So if higher layers are sliced without the lowest level,
 * the lower level of the sliced LoD will be transformed to the absolute offset.
 */
68
LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end);
Q
qijun 已提交
69

70
LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
Q
qijun 已提交
71
                 size_t elem_end);
72 73 74 75
/*
 * Transform an LoD from relative offsets to absolute offsets.
 */
LoD ToAbsOffset(const LoD& in);
Q
qijun 已提交
76

77
bool operator==(const LoD& a, const LoD& b);
Q
qijun 已提交
78

79
/*
80
 * LoDTensor (Level of details Tensor)
81 82
 * see https://en.wikipedia.org/wiki/Level_of_details for reference.
 */
83
class LoDTensor : public Tensor {
84
 public:
85
  LoDTensor() {}
86

87
  explicit LoDTensor(const LoD& lod) : lod_(lod) {}
Q
qijun 已提交
88

89
  void set_lod(const LoD& lod) { lod_ = lod; }
Q
qijun 已提交
90

Y
Yu Yang 已提交
91 92 93
  const LoD& lod() const { return lod_; }

  LoD* mutable_lod() { return &lod_; }
Q
qijun 已提交
94

95
  /*
W
wanghaoshuang 已提交
96
   * Get the start offset and end offset of an  element from LoD.
97
   */
W
wanghaoshuang 已提交
98
  std::pair<size_t, size_t> lod_element(size_t level, size_t elem) const {
99
    PADDLE_ENFORCE_LT(level, NumLevels());
W
wanghaoshuang 已提交
100 101
    PADDLE_ENFORCE_LT(elem, NumElements(level));
    return std::make_pair((lod_)[level][elem], (lod_)[level][elem + 1]);
102 103 104
  }

  /*
105
   * Number of LoDTensor's levels, each level has units of data, for example,
106 107
   * in the sentence's view, article, paragraph, sentence are 3 levels.
   */
108
  size_t NumLevels() const { return lod_.size(); }
109 110 111 112
  /*
   * Number of elements in a level.
   */
  size_t NumElements(size_t level = 0) const {
113
    PADDLE_ENFORCE_LT(level, NumLevels());
114
    // the last offset is the end of last element
Q
qijun 已提交
115
    return (lod_)[level].size() - 1;
116 117
  }

118 119 120 121 122 123 124 125 126 127 128 129
  /*
   * Number of lower-level elements.
   * For example, a 2-level lod-tensor
   *
   * 0-th level   |   |
   * 1-th level   ||  |||
   *
   * NumElements(0, 0) get 2
   * NumElements(0, 1) get 3
   */
  size_t NumElements(size_t level, size_t idx) const;

130 131 132 133 134 135
  /*
   * Get the number of instances in the underlying tensor in the `idx`-th
   * element.
   */
  size_t NumInstancesInElement(size_t level, size_t idx) const;

136
  /*
137
   * Shrink levels[level_begin:level_end]
138
   */
139
  void ShrinkLevels(size_t level_begin, size_t level_end);
140 141

  /*
142
   * Shrink elements of a level, [elem_begin: elem_end]
143
   * @note: low performance in slice lod_.
144
   */
145
  void ShrinkInLevel(size_t level, size_t elem_begin, size_t elem_end);
146

Q
qijun 已提交
147
 private:
148
  LoD lod_;
149
};
150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178

/*
 * Expand the `source` to fit the LoD of `lod`. For example, a `source`
 * LoDTensor is
 *  - LoD: [0, 2]
 *  - tensor: [a0, a1]
 * a `lod` is
 *  - LoD: [0 3 5]
 * returns a new LoDTensor
 *  - [a0 a0 a0 a1 a1]
 */
template <typename T>
LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
                    const platform::Place& place) {
  LoD abs_lod = ToAbsOffset(lod);
  const auto& lod_level = lod[level];
  size_t num_instances = source.dims()[0];

  // new tensor
  LoDTensor tensor;
  tensor.set_lod(lod);
  auto dims = source.dims();
  dims[0] = lod_level.back();
  tensor.Resize(dims);
  tensor.mutable_data<T>(place);

  PADDLE_ENFORCE_EQ(num_instances, lod_level.size() - 1);
  for (size_t ins = 0; ins < num_instances; ins++) {
    for (size_t elem = lod_level[ins]; elem < lod_level[ins + 1]; elem++) {
D
dzhwinter 已提交
179 180 181
      auto slice = tensor.Slice(elem, elem + 1);
      CopyFrom(source.Slice(ins, ins + 1), platform::CPUPlace(),
               platform::CPUDeviceContext(), &slice);
182 183 184 185 186
    }
  }
  return tensor;
}

187 188 189 190 191 192 193 194 195 196 197 198
// Get the absolute offset of a lod[start_level][start_idx:end_idx] and
// relative length of details for every levels(i.e., [start_level: ]).
//
// For example,
//   lod = [[0, 3, 4, 8], [0, 9, 10, 11, 13, 17, 19, 22, 24]]
//   start_level = 0
//   start_idx = 1
//   end_idx = 3
//
// Returns:
//  LoD = [[1, 4], [2, 4, 2, 3, 2]]
//  pair<size_t, size_t> = {11, 24}
199 200
std::pair<LoD, std::pair<size_t, size_t>> GetSubLoDAndAbsoluteOffset(
    const LoD& lod, size_t start_idx, size_t end_idx, size_t start_level);
201

202
void AppendLoD(LoD* lod, const LoD& lod_length);
203

武毅 已提交
204 205 206 207 208 209 210 211 212
/*
 * Serialize/Desiralize LoDTensor to std::ostream
 * You can pass ofstream or ostringstream to serilize to file
 * or to a in memory string. GPU tensor will be copied to CPU.
 */
void SerializeToStream(std::ostream& os, const LoDTensor& tensor,
                       const platform::DeviceContext& dev_ctx);
void DeserializeFromStream(std::istream& is, LoDTensor* tensor);

213 214
}  // namespace framework
}  // namespace paddle