Argument.h 10.1 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

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

    http://www.apache.org/licenses/LICENSE-2.0

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. */

#pragma once

#include "hl_gpu.h"

#include "paddle/math/Matrix.h"
#include "paddle/math/Vector.h"
Y
Yu Yang 已提交
21
#include "paddle/parameter/Parameter.h"
Z
zhangjinchao01 已提交
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154
#include "paddle/utils/Locks.h"
#include "paddle/utils/Util.h"

namespace paddle {

// vector of user defined pointers
typedef std::shared_ptr<std::vector<void*>> UserDefinedVectorPtr;
typedef std::shared_ptr<std::vector<std::string>> SVectorPtr;

struct Argument {
  Argument()
      : in(nullptr),
        value(nullptr),
        ids(nullptr),
        grad(nullptr),
        strs(nullptr),
        frameHeight(0),
        frameWidth(0),
        sequenceStartPositions(nullptr),
        subSequenceStartPositions(nullptr),
        cpuSequenceDims(nullptr),
        udp(nullptr),
        deviceId(-1),
        allCount(0),
        valueCount(0),
        gradCount(0),
        dataId(0) {}
  Argument(const Argument& argument) {
    *this = argument;
    valueCount = 0;
    gradCount = 0;
    dataId = argument.dataId;
  }
  ~Argument() {}

  void operator=(const Argument& argument) {
    in = argument.in;
    value = argument.value;
    ids = argument.ids;
    grad = argument.grad;
    strs = argument.strs;
    sequenceStartPositions = argument.sequenceStartPositions;
    subSequenceStartPositions = argument.subSequenceStartPositions;
    cpuSequenceDims = argument.cpuSequenceDims;
    udp = argument.udp;
    deviceId = argument.deviceId;
    allCount = argument.allCount;
    frameHeight = argument.frameHeight;
    frameWidth = argument.frameWidth;
    dataId = argument.dataId;
  }

  MatrixPtr in;  // used if needed
  MatrixPtr value;
  IVectorPtr ids;  // a sequence of ids. Can be use for class id for costLayer
  MatrixPtr grad;  // If empty, gradient is not needed.
  SVectorPtr strs;

  // A dataBatch includes batchSize frames, one frame maybe not only vector
  size_t frameHeight;
  size_t frameWidth;

  // If NULL, each position is treated independently.
  // Otherwise, its size should be #NumberOfSequences + 1.
  // The first position is always 0 and
  // the last position should be equal to batchSize.
  ICpuGpuVectorPtr sequenceStartPositions;

  // If NULL, each sequence has no subsequence.
  // Otherwise, its size should be #NumberOfSubSequences + 1.
  // The first position is always 0 and
  // the last position should be equal to batchSize.
  ICpuGpuVectorPtr subSequenceStartPositions;

  // dimension of sequence, stored only in CPU
  IVectorPtr cpuSequenceDims;

  UserDefinedVectorPtr udp;  // user defined pointer

  int deviceId;            // the GPU device id which the argument in
  int allCount;            // the number of output layers using this argument
  mutable int valueCount;  // waiting this member when layer do forward
  mutable int gradCount;   // waiting this member when layer do backward
  mutable LockedCondition valueReadyCond;
  mutable LockedCondition gradReadyCond;

  int dataId;  // dataProvider id

  /* Increase the reference count of the argument. */
  void countIncrement() { allCount++; }

  int getAllCount() const { return allCount; }

  void waitValueReady() const {
    valueReadyCond.wait([this] { return (valueCount != 0); });

    std::lock_guard<std::mutex> guard(*valueReadyCond.mutex());
    valueCount--;
  }

  void notifyValueReady() const {
    valueReadyCond.notify_all([this] { valueCount = allCount; });
  }

  void waitGradReady() const {
    gradReadyCond.wait([this] { return (gradCount == allCount); });
    gradCount = 0;
  }

  void notifyGradReady() const {
    gradReadyCond.notify_all([this] { gradCount++; });
  }

  int64_t getBatchSize() const {
    if (value) return value->getHeight();
    if (ids) return ids->getSize();
    if (grad) return grad->getHeight();
    if (in) return in->getHeight();
    if (udp) return udp->size();
    if (strs) return strs->size();
    return 0;
  }
  size_t getFrameHeight() const { return frameHeight; }
  size_t getFrameWidth() const { return frameWidth; }
  void setFrameHeight(size_t h) { frameHeight = h; }
  void setFrameWidth(size_t w) { frameWidth = w; }

  int64_t getNumSequences() const {
    return sequenceStartPositions ? sequenceStartPositions->getSize() - 1
                                  : getBatchSize();
  }

  int64_t getNumSubSequences() const {
155 156
    return subSequenceStartPositions ? subSequenceStartPositions->getSize() - 1
                                     : getBatchSize();
Z
zhangjinchao01 已提交
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177
  }

  bool hasSubseq() const { return subSequenceStartPositions != nullptr; }

  const int* getCpuStartPositions() const {
    return hasSubseq() ? subSequenceStartPositions->getData(false)
                       : sequenceStartPositions->getData(false);
  }

  static inline real sumCosts(const std::vector<Argument>& arguments) {
    real cost = 0;
    for (auto& arg : arguments) {
      if (arg.value) {
        SetDevice device(arg.deviceId);
        cost += arg.value->getSum();
      }
    }
    return cost;
  }

  /**
178
   * @brief (value, ids, grad, sequenceStartPositions) of output are subset of
Z
zhangjinchao01 已提交
179 180 181
   *        input. Note that, output share the same memory of input.
   *
   * @param input[in]       input
182
   * @param offset[in]      offset in terms of rows
Z
zhangjinchao01 已提交
183 184 185 186 187 188 189 190
   * @param height[in]      height of output.value
   * @param width[in]       width of output.value
   * @param useGpu[in]
   * @param trans[in]       whether input.value is transform
   * @param seqFlag[in]     whether input has sequenceStartPositions
   * @param seqStart[in]    offset of input.sequenceStartPositions
   * @param seqSize[in]     lenght of output.sequenceStartPositions
   */
191 192 193 194 195 196 197 198
  void subArgFrom(const Argument& input,
                  size_t offset,
                  size_t height,
                  size_t width,
                  bool useGpu,
                  bool trans = false,
                  bool seqFlag = false,
                  size_t seqStart = 0,
Z
zhangjinchao01 已提交
199 200 201 202 203 204 205 206 207 208
                  size_t seqSize = 0);
  /*
   * for sequence input:
   *   startSeq: the sequence id of start
   *   copySize: how many sequences need to copy
   *   return value: how many samples are copied
   * for non-sequence input:
   *   startSeq: the sample id of start
   *   copySize: how many samples need to copy
   *   return value: how many samples are copied
209 210
   * Note that when specifying the stream explicitly in this case,
   * synchronize should also be called somewhere after this function
Z
zhangjinchao01 已提交
211
   */
212 213 214 215 216
  int32_t resizeAndCopyFrom(const Argument& src,
                            int32_t startSeq,
                            int32_t copySize,
                            bool useGpu,
                            hl_stream_t stream);
Z
zhangjinchao01 已提交
217

218 219 220 221 222
  /*
   * same with the above function, except that the stream is
   * HPPL_STREAM_DEFAULT and synchronize is automatically called
   * inside it
   */
223 224 225 226
  int32_t resizeAndCopyFrom(const Argument& src,
                            int32_t startSeq,
                            int32_t copySize,
                            bool useGpu = FLAGS_use_gpu);
227 228 229

  void resizeAndCopyFrom(const Argument& src, bool useGpu, hl_stream_t stream);

230 231 232 233 234
  /*
   * same with the above function, except that the stream is
   * HPPL_STREAM_DEFAULT and synchronize is automatically called
   * inside it
   */
235
  void resizeAndCopyFrom(const Argument& src, bool useGpu = FLAGS_use_gpu);
Z
zhangjinchao01 已提交
236 237 238 239 240 241 242 243 244 245 246 247

  /*
    @brief Concatenate several arguments into one and put the result into it.
    @param args : a vector of argument, each element of which is a frame in a
    batch of sequences.
    @param selectRows : select several row of args to concatenate
    @param seqStartPos : sequence start positions in the final Argument
    @param hl_stream_t : cuda stream
    @param passTyoe : type of task, training or testing
   */
  void concat(const std::vector<Argument>& args,
              const std::vector<int>& selectRows,
248 249 250 251
              const std::vector<int>& seqStartPos,
              bool useGpu,
              hl_stream_t stream,
              PassType passType);
Z
zhangjinchao01 已提交
252 253 254 255

  /*
    Concatenate several args into one and put the result into this.
   */
256 257
  void concat(const std::vector<Argument>& src,
              bool useGpu = FLAGS_use_gpu,
Z
zhangjinchao01 已提交
258 259 260 261 262 263 264 265 266
              hl_stream_t stream = HPPL_STREAM_DEFAULT,
              PassType passType = PASS_TEST);

  /*
   * split vector<Argument> to several vectors according to dataId
   */
  static void splitByDataId(const std::vector<Argument>& argus,
                            std::vector<std::vector<Argument>>* arguGroups);

267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283
  struct SeqInfo {
    // Equal to sequence length for sequence data
    // Equal to number of subsequences for subsequence data
    int topLevelLength;

    int seqStart;
    int seqId;

    // Equal to topLevelLength for sequence data
    // Equal to sum of the length of subsequences for subsequence data
    int subLevelLength;

    // Only used for subsequence data, start position of this sequence
    // is subSequenceStartPositions, i.e.
    // subSequenceStartPositions[subSeqStart] == seqStart
    int subSeqStart;
  };
Z
zhangjinchao01 已提交
284
  /*
285 286 287 288 289
    Get SeqInfo for each sequence of this argument
    Elements in *seqInfo are sorted by topLevelLength in descending order
  */
  void getSeqInfo(std::vector<SeqInfo>* segInfo) const;

Z
zhangjinchao01 已提交
290 291 292 293 294 295 296 297 298 299
  /*
   Check Whether sequenceStartPositions is subset of
   subSequenceStartPositions.
   */
  void checkSubset() const;

  /*
   sequence has sub-sequence degrades to a sequence.
   */
  void degradeSequence(const Argument& input, bool useGpu);
300 301 302 303 304 305 306 307

  /**
   * @brief getValueString will return the argument's output in string. There
   * are several kinds of output. The keys of output dictionary are 'value',
   * 'id', 'sequence pos', 'sub-sequence pos'.
   * @param out [out]: the return values.
   */
  void getValueString(std::unordered_map<std::string, std::string>* out) const;
Z
zhangjinchao01 已提交
308 309 310
};

}  // namespace paddle