FunctionTest.h 10.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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. */

#include "Function.h"
16 17
#include "paddle/math/Matrix.h"
#include "paddle/math/SparseMatrix.h"
18 19
#include "paddle/math/Vector.h"
#include "paddle/math/tests/TensorCheck.h"
H
hedaoyuan 已提交
20
#include "paddle/testing/TestUtil.h"
21 22 23

namespace paddle {

H
hedaoyuan 已提交
24 25
typedef std::shared_ptr<BufferArg> BufferArgPtr;

H
hedaoyuan 已提交
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48
/**
 * \brief A class for comparing CPU and GPU implementations of Function.
 *
 *
 * Use case:
 *  // Initializes a test object, the corresponding cpu and gpu Function
 *  // are constructed according to FunctionName and FuncConfig.
 *  FunctionCompare test(FunctionName, FuncConfig);
 *  // Prepare inputs and outputs arguments.
 *  // Here the input and output can not contain real data,
 *  // only contains the argument type and shape.
 *  test.addInputs(input1);
 *  test.addInputs(input2);
 *  test.addOutputs(output1);
 *  test.addOutputs(output2);
 *  // Run.
 *  // Will according to the type and shape of arguments(inputs_/outputs_),
 *  // automatic initialization cpu and gpu function required arguments
 *  // (cpuInputs_/cpuOutputs_/gpuInputs_/gpuOutputs_).
 *  // Call the CPU and GPU Function calculation results.
 *  // Compares CPU and GPU calculation results for consistency.
 *  test.run();
 */
49 50 51
class FunctionCompare {
public:
  FunctionCompare(const std::string& name, const FuncConfig& config)
H
hedaoyuan 已提交
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66
      : cpuFunc_(FunctionBase::funcRegistrar_.createByType(name + "-CPU")),
        gpuFunc_(FunctionBase::funcRegistrar_.createByType(name + "-GPU")) {
    cpuFunc_->init(config);
    gpuFunc_->init(config);
  }

  ~FunctionCompare() {}

  // input need only contains shape, do not contains data.
  void addInputs(const BufferArg& input) {
    size_t size =
        input.shape().getElements() * sizeOfValuType(input.valueType());
    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));

67 68 69 70 71 72 73 74 75 76 77 78
    cpuInputs_.emplace_back(
        std::make_shared<BufferArg>(cpuMemory_.back()->getBuf(),
                                    input.valueType(),
                                    input.shape(),
                                    UNSPECIFIED,
                                    input.isTransposed()));
    gpuInputs_.emplace_back(
        std::make_shared<BufferArg>(gpuMemory_.back()->getBuf(),
                                    input.valueType(),
                                    input.shape(),
                                    UNSPECIFIED,
                                    input.isTransposed()));
H
hedaoyuan 已提交
79 80 81
  }

  // output need only contains shape, do not contains data.
82
  void addOutputs(const BufferArg& output, ArgType argType = ASSIGN_TO) {
H
hedaoyuan 已提交
83 84 85 86 87
    size_t size =
        output.shape().getElements() * sizeOfValuType(output.valueType());
    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));

88 89 90 91 92 93 94 95 96 97 98 99 100 101
    cpuOutputs_.emplace_back(std::make_shared<BufferArg>(
        cpuMemory_.back()->getBuf(),
        output.valueType(),
        output.shape(),
        // todo(tianbing), argType = output.getArgType(), but default ASSIGN_TO
        argType,
        output.isTransposed()));
    gpuOutputs_.emplace_back(std::make_shared<BufferArg>(
        gpuMemory_.back()->getBuf(),
        output.valueType(),
        output.shape(),
        // todo(tianbing), argType = output.getArgType(), but default ASSIGN_TO
        argType,
        output.isTransposed()));
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
  /// add and init output sparse matrix
  void addOutputs(const SparseMatrixArg& output, ArgType argType = ASSIGN_TO) {
    cpuSparse_ = std::make_shared<CpuSparseMatrix>(output.shape()[0],
                                                   output.shape()[1],
                                                   output.nnz(),
                                                   output.dataType(),
                                                   output.dataFormat(),
                                                   output.isTransposed());

    gpuSparse_ = std::make_shared<GpuSparseMatrix>(output.shape()[0],
                                                   output.shape()[1],
                                                   output.nnz(),
                                                   output.dataType(),
                                                   output.dataFormat(),
                                                   output.isTransposed());

    /// init sparse matrix
    hl_stream_t stream(HPPL_STREAM_1);
    cpuSparse_->randomizeUniform();
    gpuSparse_->copyFrom(*cpuSparse_, stream);
    hl_stream_synchronize(stream);

    cpuOutputs_.emplace_back(
        std::make_shared<SparseMatrixArg>(*cpuSparse_, argType));
    gpuOutputs_.emplace_back(
        std::make_shared<SparseMatrixArg>(*gpuSparse_, argType));
  }

H
hedaoyuan 已提交
132 133 134 135 136 137 138
  void addInputs(const SequenceArg& input) {
    size_t batchSize = input.shape()[0];
    size_t numSeqs = batchSize / 10 + 1;

    size_t sizeId = (numSeqs + 1) * sizeOfValuType(VALUE_TYPE_INT32);
    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(sizeId));
    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(sizeId));
H
hedaoyuan 已提交
139

H
hedaoyuan 已提交
140 141 142 143 144 145 146 147 148 149 150
    TensorShape seqsId({numSeqs + 1});
    // void* cpuBuffer = cpuMemory_.back()->getBuf();
    // void* gpuBuffer = gpuMemory_.back()->getBuf();

    size_t size =
        input.shape().getElements() * sizeOfValuType(input.valueType());
    cpuMemory_.emplace_back(std::make_shared<CpuMemoryHandle>(size));
    gpuMemory_.emplace_back(std::make_shared<GpuMemoryHandle>(size));

    // TODO: need be implemented.
  }
H
hedaoyuan 已提交
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
  void addInputs(const SparseMatrixArg& input) {
    cpuSparse_ = std::make_shared<CpuSparseMatrix>(input.shape()[0],
                                                   input.shape()[1],
                                                   input.nnz(),
                                                   input.dataType(),
                                                   input.dataFormat(),
                                                   input.isTransposed());

    gpuSparse_ = std::make_shared<GpuSparseMatrix>(input.shape()[0],
                                                   input.shape()[1],
                                                   input.nnz(),
                                                   input.dataType(),
                                                   input.dataFormat(),
                                                   input.isTransposed());

    /// init sparse matrix
    hl_stream_t stream(HPPL_STREAM_1);
    cpuSparse_->randomizeUniform();
    gpuSparse_->copyFrom(*cpuSparse_, stream);
    hl_stream_synchronize(stream);

    cpuInputs_.emplace_back(std::make_shared<SparseMatrixArg>(*cpuSparse_));
    gpuInputs_.emplace_back(std::make_shared<SparseMatrixArg>(*gpuSparse_));
  }

H
hedaoyuan 已提交
177 178
  void run() {
    // prepare cpu/gpu arguments
H
hedaoyuan 已提交
179
    initInputs();
H
hedaoyuan 已提交
180

181
    initOutputs();
H
hedaoyuan 已提交
182
    // function calculate
H
hedaoyuan 已提交
183 184 185 186 187 188 189
    auto callFunction = [](FunctionBase* function,
                           std::vector<BufferArgPtr>& inputs,
                           std::vector<BufferArgPtr>& outputs) {
      BufferArgs inArgs;
      BufferArgs outArgs;
      for (auto arg : inputs) {
        inArgs.addArg(*arg);
H
hedaoyuan 已提交
190
      }
H
hedaoyuan 已提交
191 192
      for (auto arg : outputs) {
        outArgs.addArg(*arg);
193
      }
H
hedaoyuan 已提交
194
      function->calc(inArgs, outArgs);
195 196
    };

H
hedaoyuan 已提交
197 198
    callFunction(cpuFunc_.get(), cpuInputs_, cpuOutputs_);
    callFunction(gpuFunc_.get(), gpuInputs_, gpuOutputs_);
199

200
    // check outputs
H
hedaoyuan 已提交
201
    compareOutputs();
202 203
  }

H
hedaoyuan 已提交
204
  std::shared_ptr<FunctionBase> getCpuFunction() const { return cpuFunc_; }
205

H
hedaoyuan 已提交
206
  std::shared_ptr<FunctionBase> getGpuFunction() const { return gpuFunc_; }
207

H
hedaoyuan 已提交
208
protected:
H
hedaoyuan 已提交
209 210
  void initInputs() {
    for (size_t i = 0; i < cpuInputs_.size(); i++) {
211 212 213 214
      if (cpuInputs_[i]->isSparseArg()) {
        continue;  /// sparse matrix already init
      }

H
hedaoyuan 已提交
215
      initArg(*cpuInputs_[i]);
H
hedaoyuan 已提交
216

H
hedaoyuan 已提交
217 218 219 220 221
      // TODO: Need a BufferCopy used to copy from one BufferArg to another.
      CpuVector cpuVector(cpuInputs_[i]->shape().getElements(),
                          (real*)cpuInputs_[i]->data());
      GpuVector gpuVector(gpuInputs_[i]->shape().getElements(),
                          (real*)gpuInputs_[i]->data());
H
hedaoyuan 已提交
222

H
hedaoyuan 已提交
223 224
      gpuVector.copyFrom(cpuVector);
    }
H
hedaoyuan 已提交
225 226
  }

227 228 229
  void initOutputs() {
    for (size_t i = 0; i < cpuOutputs_.size(); i++) {
      if (cpuOutputs_[i]->isSparseArg()) {
230
        continue;  /// sparse matrix already init
231 232 233 234 235 236 237 238 239 240 241 242 243 244
      }

      initArg(*cpuOutputs_[i]);

      // TODO: Need a BufferCopy used to copy from one BufferArg to another.
      CpuVector cpuVector(cpuOutputs_[i]->shape().getElements(),
                          (real*)cpuOutputs_[i]->data());
      GpuVector gpuVector(gpuOutputs_[i]->shape().getElements(),
                          (real*)gpuOutputs_[i]->data());

      gpuVector.copyFrom(cpuVector);
    }
  }

H
hedaoyuan 已提交
245 246 247
  void compareOutputs() {
    for (size_t i = 0; i < cpuOutputs_.size(); i++) {
      // TODO, Need a BufferCheck used to compare the two buffers.
248 249 250 251 252
      const auto cpu = cpuOutputs_[i];
      const auto gpu = gpuOutputs_[i];
      CHECK_EQ(cpu->numElements(), gpu->numElements());
      CpuVector cpuVector(cpu->numElements(), (real*)cpu->data());
      GpuVector gpuVector(gpu->numElements(), (real*)gpu->data());
H
hedaoyuan 已提交
253 254
      autotest::TensorCheckErr(cpuVector, gpuVector);
    }
H
hedaoyuan 已提交
255 256 257 258 259 260 261 262 263 264
  }

  // only init cpu argument, gpu argument copy from cpu argument.
  void initArg(BufferArg& arg) {
    CpuVector vector(arg.shape().getElements(), (real*)arg.data());
    vector.uniform(0.001, 1);
  }

  void initArg(SequenceIdArg& arg, size_t batchSize) {
    size_t numSeqs = arg.numSeqs();
H
hedaoyuan 已提交
265
    int* buf = reinterpret_cast<int*>(arg.data());
H
hedaoyuan 已提交
266 267
    int pos = 0;
    size_t maxLen = 2 * batchSize / numSeqs;
H
hedaoyuan 已提交
268
    for (int i = 0; i < (int)numSeqs; ++i) {
H
hedaoyuan 已提交
269 270 271 272 273 274 275 276 277 278
      int len = uniformRandom(
                    std::min<int64_t>(maxLen, batchSize - pos - numSeqs + i)) +
                1;
      buf[i] = pos;
      pos += len;
      VLOG(1) << " len=" << len;
    }
    buf[numSeqs] = batchSize;
  }

279
protected:
H
hedaoyuan 已提交
280 281
  std::shared_ptr<FunctionBase> cpuFunc_;
  std::shared_ptr<FunctionBase> gpuFunc_;
H
hedaoyuan 已提交
282 283
  std::vector<CpuMemHandlePtr> cpuMemory_;
  std::vector<GpuMemHandlePtr> gpuMemory_;
H
hedaoyuan 已提交
284 285 286 287
  std::vector<BufferArgPtr> cpuInputs_;
  std::vector<BufferArgPtr> cpuOutputs_;
  std::vector<BufferArgPtr> gpuInputs_;
  std::vector<BufferArgPtr> gpuOutputs_;
288 289
  std::shared_ptr<CpuSparseMatrix> cpuSparse_;
  std::shared_ptr<GpuSparseMatrix> gpuSparse_;
290 291 292
};

}  // namespace paddle