test_matrixCompare.cpp 37.7 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

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. */

#ifndef PADDLE_ONLY_CPU
/// This unittest checks GpuMatrix/CpuMatrix get same result, so disable when
/// only cpu version.

#include <gtest/gtest.h>
Y
Yu Yang 已提交
20 21 22
#include "TensorCheck.h"
#include "paddle/math/Matrix.h"
#include "paddle/math/SparseMatrix.h"
23
#include "paddle/testing/TestUtil.h"
24
#include "paddle/utils/Stat.h"
Y
Yu Yang 已提交
25
#include "paddle/utils/Util.h"
26

Z
zhangjinchao01 已提交
27 28
using namespace paddle;  // NOLINT
using namespace std;     // NOLINT
29 30
using autotest::TensorCheckEqual;
using autotest::TensorCheckErr;
L
liaogang 已提交
31

Z
zhangjinchao01 已提交
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
void testMatrixMaxSequence(int batchSize, int inputDim) {
  // forward
  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(batchSize, inputDim);
  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(batchSize, inputDim);
  cpuInput->randomizeUniform();
  gpuInput->copyFrom(*cpuInput);

  IVectorPtr cpuSequence;
  generateSequenceStartPositions(batchSize, cpuSequence);
  IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true);
  gpuSequence->copyFrom(*cpuSequence);

  int newBatchSize = cpuSequence->getSize() - 1;
  MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(newBatchSize, inputDim);
  MatrixPtr gpuOutput = std::make_shared<GpuMatrix>(newBatchSize, inputDim);
  cpuOutput->zero();
  gpuOutput->zero();

  IVectorPtr cpuIndex = nullptr;
  IVectorPtr gpuIndex = nullptr;
  IVector::resizeOrCreate(cpuIndex, newBatchSize * inputDim, false);
  IVector::resizeOrCreate(gpuIndex, newBatchSize * inputDim, true);
  cpuIndex->zeroMem();
  gpuIndex->zeroMem();

  cpuOutput->maxSequenceForward(*cpuInput, *cpuSequence, *cpuIndex);
  gpuOutput->maxSequenceForward(*gpuInput, *gpuSequence, *gpuIndex);

60 61
  TensorCheckEqual(*cpuOutput, *gpuOutput);
  TensorCheckEqual(*cpuIndex, *gpuIndex);
Z
zhangjinchao01 已提交
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76

  // backward
  MatrixPtr cpuOutputGrad = std::make_shared<CpuMatrix>(newBatchSize, inputDim);
  MatrixPtr gpuOutputGrad = std::make_shared<GpuMatrix>(newBatchSize, inputDim);
  cpuOutputGrad->randomizeUniform();
  gpuOutputGrad->copyFrom(*cpuOutputGrad);

  MatrixPtr cpuInputGrad = std::make_shared<CpuMatrix>(batchSize, inputDim);
  MatrixPtr gpuInputGrad = std::make_shared<GpuMatrix>(batchSize, inputDim);
  cpuInputGrad->randomizeUniform();
  gpuInputGrad->copyFrom(*cpuInputGrad);

  cpuInputGrad->maxSequenceBackward(*cpuOutputGrad, *cpuSequence, *cpuIndex);
  gpuInputGrad->maxSequenceBackward(*gpuOutputGrad, *gpuSequence, *gpuIndex);

77
  TensorCheckEqual(*cpuInputGrad, *gpuInputGrad);
Z
zhangjinchao01 已提交
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
}

TEST(Matrix, maxSequence) {
  for (auto batchSize : {1, 10, 128, 1000, 6000}) {
    for (auto inputDim : {1, 32, 100, 512}) {
      VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim;
      testMatrixMaxSequence(batchSize, inputDim);
    }
  }
}

void testMatrixGetSum(int height, int width) {
  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(height, width);
  cpuInput->randomizeUniform();
  gpuInput->copyFrom(*cpuInput);

#ifndef PADDLE_TYPE_DOUBLE
  int x = log10(height * width);
  real err = 1e-6 * pow(10, x);
#else
  real err = 1e-8;
#endif

  real cpuSum = cpuInput->getSum();
  real gpuSum = gpuInput->getSum();

  EXPECT_LE(fabs(cpuSum - gpuSum), err);
}

void testMatrixZeroAtOffset(int height, int width) {
  MatrixPtr cpuA = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr gpuA = std::make_shared<GpuMatrix>(height, width);
  MatrixPtr cpuTest = std::make_shared<CpuMatrix>(height, width);

  cpuA->randomizeUniform();
  gpuA->copyFrom(*cpuA);
  cpuTest->copyFrom(*cpuA);

  int columnOffset = rand() % width;  // NOLINT we just use rand() for test.
  int numColumns = rand() % (width - columnOffset);  // NOLINT

120 121
  if (numColumns == 0) return;

Z
zhangjinchao01 已提交
122 123 124 125 126 127 128 129 130 131 132
  cpuA->zeroAtOffset(columnOffset, numColumns);
  gpuA->zeroAtOffset(columnOffset, numColumns);

  /* cpuTest */
  real* a = cpuTest->getData() + columnOffset;
  for (int64_t i = 0; i < height; ++i) {
    for (int64_t j = 0; j < numColumns; ++j) {
      a[i * width + j] = 0;
    }
  }

133 134
  TensorCheckEqual(*cpuA, *gpuA);
  TensorCheckEqual(*cpuA, *cpuTest);
Z
zhangjinchao01 已提交
135 136
}

X
xutianbing 已提交
137 138 139 140 141 142 143 144 145 146 147 148 149 150
void testMatrixDeepSwap(int height, int width) {
  MatrixPtr cpuA = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr cpuB = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr cpuCopyA = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr cpuCopyB = std::make_shared<CpuMatrix>(height, width);

  cpuA->randomizeUniform();
  cpuB->randomizeUniform();
  cpuCopyA->copyFrom(*cpuA);
  cpuCopyB->copyFrom(*cpuB);

  // swap matrix cpuA and cpuB
  cpuA->deepSwap(*cpuB);

H
hedaoyuan 已提交
151 152
  TensorCheckEqual(*cpuA, *cpuCopyB);
  TensorCheckEqual(*cpuB, *cpuCopyA);
X
xutianbing 已提交
153 154
}

Z
zhangjinchao01 已提交
155 156 157 158 159 160 161 162 163 164 165
void testMatrixTranspose(int height, int width) {
  MatrixPtr cpu = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr gpu = std::make_shared<GpuMatrix>(height, width);
  MatrixPtr cpuT = std::make_shared<CpuMatrix>(width, height);
  MatrixPtr gpuT = std::make_shared<GpuMatrix>(width, height);

  cpu->randomizeUniform();
  gpu->copyFrom(*cpu);
  cpu->transpose(cpuT, false);
  gpu->transpose(gpuT, false);

166
  TensorCheckEqual(*cpuT, *gpuT);
Z
zhangjinchao01 已提交
167 168
}

L
lzhao4ever 已提交
169 170 171 172 173 174
void testMatrixInverse(int height) {
  MatrixPtr cpu = std::make_shared<CpuMatrix>(height, height);
  MatrixPtr gpu = std::make_shared<GpuMatrix>(height, height);
  MatrixPtr cpuI = std::make_shared<CpuMatrix>(height, height);
  MatrixPtr gpuI = std::make_shared<GpuMatrix>(height, height);

175
  /* Make matrix well conditioned: cpu * cpuT + Identity */
L
lzhao4ever 已提交
176
  cpu->randomizeUniform();
177 178
  MatrixPtr cpuT = cpu->getTranspose();
  MatrixPtr outputCheck = std::make_shared<CpuMatrix>(height, height);
179
  outputCheck->mul(*cpu, *cpuT);
180 181 182
  cpu->setDiag(1.0);
  cpu->add(*outputCheck);

L
lzhao4ever 已提交
183 184 185 186
  gpu->copyFrom(*cpu);
  cpu->inverse(cpuI, false);
  gpu->inverse(gpuI, false);

187
  TensorCheckErr(*cpuI, *gpuI);
L
lzhao4ever 已提交
188

189
  outputCheck->mul(*cpu, *cpuI);
190
  cpu->setDiag(1.0);
191
  TensorCheckErr(*cpu, *outputCheck);
L
lzhao4ever 已提交
192 193
}

Z
zhangjinchao01 已提交
194
TEST(Matrix, unary) {
L
lzhao4ever 已提交
195 196
  for (auto height : {1, 3, 11, 73, 128, 200, 330}) {
    for (auto width : {1, 3, 32, 100, 512, 1000, 3210}) {
Z
zhangjinchao01 已提交
197 198
      VLOG(3) << " height=" << height << " width=" << width;

199
      testMatrixDeepSwap(height, width);
200
      testMatrixZeroAtOffset(height, width);
Z
zhangjinchao01 已提交
201 202 203
      testMatrixGetSum(height, width);
      testMatrixTranspose(height, width);
    }
L
lzhao4ever 已提交
204 205
    // inverse
    testMatrixInverse(height);
Z
zhangjinchao01 已提交
206 207 208
  }
}

209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224
void testMatrixSoftmax(int height, int width) {
  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(height, width);
  MatrixPtr gpuOutput = std::make_shared<GpuMatrix>(height, width);

  cpuInput->randomizeUniform();
  gpuInput->copyFrom(*cpuInput);
  cpuOutput->zero();
  gpuOutput->zero();
  cpuInput->softmax(*cpuOutput);
  gpuInput->softmax(*gpuOutput);

  TensorCheckErr(*cpuOutput, *gpuOutput);
}

Z
zhangjinchao01 已提交
225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240
void testSequenceSoftmax(int batchSize) {
  // forward
  int inputDim = 1;
  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(batchSize, inputDim);
  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(batchSize, inputDim);
  cpuInput->randomizeUniform();
  gpuInput->copyFrom(*cpuInput);

  IVectorPtr cpuSequence;
  generateSequenceStartPositions(batchSize, cpuSequence);
  IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true);
  gpuSequence->copyFrom(*cpuSequence);

  cpuInput->sequenceSoftmax(*cpuInput, *cpuSequence);
  gpuInput->sequenceSoftmax(*gpuInput, *gpuSequence);

241
  TensorCheckErr(*cpuInput, *gpuInput);
Z
zhangjinchao01 已提交
242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
}

void testMatrixSoftmaxThreshold(int height, int width) {
  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(height, width);
  MatrixPtr gpuOutput = std::make_shared<GpuMatrix>(height, width);

  cpuInput->randomizeUniform();
  cpuInput->getData()[0] = 100.0;
  gpuInput->copyFrom(*cpuInput);
  cpuOutput->zero();
  gpuOutput->zero();
  cpuInput->softmax(*cpuOutput);
  gpuInput->softmax(*gpuOutput);

  MatrixPtr outputCheck = std::make_shared<CpuMatrix>(height, width);
  outputCheck->copyFrom(*gpuOutput);
  // check output zero
  int cpuCount = 0;
  int gpuCount = 0;
  auto zeroNum = [](MatrixPtr out, int& count) {
    for (size_t i = 0; i < out->getHeight(); i++) {
      for (size_t j = 0; j < out->getWidth(); j++) {
        if (out->getElement(i, j) == 0) count++;
      }
    }
  };
  zeroNum(cpuOutput, cpuCount);
  zeroNum(outputCheck, gpuCount);
  EXPECT_EQ(cpuCount, 0) << "Cpu softmax output value 0";
  EXPECT_EQ(gpuCount, 0) << "Gpu softmax output value 0";
}

void testMatrixSoftmaxBp(int height, int width) {
  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(height, width);
  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(height, width);
  MatrixPtr gpuOutput = std::make_shared<GpuMatrix>(height, width);

  cpuInput->randomizeUniform();
  gpuInput->copyFrom(*cpuInput);
  cpuOutput->randomizeUniform();
  gpuOutput->copyFrom(*cpuOutput);
  gpuOutput->softmaxBackward(*gpuInput);

  MatrixPtr sftMaxSum = std::make_shared<CpuMatrix>(height, 1);
  MatrixPtr sftMaxDot = std::make_shared<CpuMatrix>(height, width);
  sftMaxDot->dotMul(*cpuOutput, *cpuInput);
  sftMaxSum->colMerge(*sftMaxDot);
  cpuOutput->softmaxDerivative(*cpuInput, *sftMaxSum);

294
  TensorCheckErr(*cpuOutput, *gpuOutput);
Z
zhangjinchao01 已提交
295 296 297 298 299 300 301
}

TEST(Matrix, softmax) {
  for (auto height : {1, 11, 73, 128, 200}) {
    for (auto width : {1, 32, 100, 512, 1000}) {
      VLOG(3) << " height=" << height << " width=" << width;

302
      testMatrixSoftmax(height, width);
Z
zhangjinchao01 已提交
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330
      testMatrixSoftmaxBp(height, width);
      testMatrixSoftmaxThreshold(height, width);
    }
    testSequenceSoftmax(height);
  }
}

void testMatrixAddToRows(int numSamples, int tableSize, int inputDim) {
  MatrixPtr cpuTable = std::make_shared<CpuMatrix>(tableSize, inputDim);
  MatrixPtr gpuTable = std::make_shared<GpuMatrix>(tableSize, inputDim);
  cpuTable->randomizeUniform();
  gpuTable->copyFrom(*cpuTable);

  IVectorPtr cpuIds;
  IVectorPtr gpuIds;
  cpuIds = VectorT<int>::create(numSamples, false);
  gpuIds = VectorT<int>::create(numSamples, true);
  cpuIds->rand(tableSize);
  gpuIds->copyFrom(*cpuIds);

  MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(numSamples, inputDim);
  MatrixPtr gpuOutput = std::make_shared<GpuMatrix>(numSamples, inputDim);
  cpuOutput->randomizeUniform();
  gpuOutput->copyFrom(*cpuOutput);

  cpuOutput->addToRows(*cpuTable, *cpuIds);
  gpuOutput->addToRows(*gpuTable, *gpuIds);

331
  TensorCheckErr(*cpuTable, *gpuTable);
Z
zhangjinchao01 已提交
332 333 334 335 336 337 338
}

TEST(Matrix, tableProjection) {
  for (auto numSamples : {10, 100, 1000, 10000, 80000}) {
    for (auto tableSize : {10, 100}) {
      for (auto inputDim : {20, 50}) {
        VLOG(3) << " numSamples=" << numSamples << " tableSize=" << tableSize
339
                << " inputDim=" << inputDim;
Z
zhangjinchao01 已提交
340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369
        testMatrixAddToRows(numSamples, tableSize, inputDim);
      }
    }
  }
}

void testMatrixMul(bool transa, bool transb, int dimM, int dimN, int dimK) {
  int heightA = transa == false ? dimM : dimK;
  int widthA = transa == false ? dimK : dimM;
  int heightB = transb == false ? dimK : dimN;
  int widthB = transb == false ? dimN : dimK;
  int heightC = dimM;
  int widthC = dimN;

  MatrixPtr cpuA = std::make_shared<CpuMatrix>(heightA, widthA, transa);
  MatrixPtr cpuB = std::make_shared<CpuMatrix>(heightB, widthB, transb);
  MatrixPtr cpuC = std::make_shared<CpuMatrix>(heightC, widthC);
  MatrixPtr gpuA = std::make_shared<GpuMatrix>(heightA, widthA, transa);
  MatrixPtr gpuB = std::make_shared<GpuMatrix>(heightB, widthB, transb);
  MatrixPtr gpuC = std::make_shared<GpuMatrix>(heightC, widthC);

  real alpha = 1.5;
  real beta = 2.0;
  cpuA->randomizeUniform();
  cpuB->randomizeUniform();
  cpuC->randomizeUniform();
  gpuA->copyFrom(*cpuA);
  gpuB->copyFrom(*cpuB);
  gpuC->copyFrom(*cpuC);

370 371
  cpuC->mul(*cpuA, *cpuB, alpha, beta);
  gpuC->mul(*gpuA, *gpuB, alpha, beta);
Z
zhangjinchao01 已提交
372

373
  TensorCheckErr(*cpuC, *gpuC);
Z
zhangjinchao01 已提交
374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
}

void testSubMatrixMul(bool transa, bool transb, int dimM, int dimN, int dimK) {
  int heightA = transa == false ? dimM : dimK;
  int widthA = transa == false ? dimK : dimM;
  int heightB = transb == false ? dimK : dimN;
  int widthB = transb == false ? dimN : dimK;
  int heightC = dimM;
  int widthC = dimN;

  MatrixPtr cpuA = std::make_shared<CpuMatrix>(heightA, widthA, transa);
  MatrixPtr cpuB = std::make_shared<CpuMatrix>(heightB, widthB, transb);
  MatrixPtr cpuC = std::make_shared<CpuMatrix>(heightC, widthC);
  MatrixPtr gpuA = std::make_shared<GpuMatrix>(heightA, widthA, transa);
  MatrixPtr gpuB = std::make_shared<GpuMatrix>(heightB, widthB, transb);
  MatrixPtr gpuC = std::make_shared<GpuMatrix>(heightC, widthC);

  real alpha = 1.5;
  real beta = 2.0;
  cpuA->randomizeUniform();
  cpuB->randomizeUniform();
  cpuC->randomizeUniform();
  gpuA->copyFrom(*cpuA);
  gpuB->copyFrom(*cpuB);
  gpuC->copyFrom(*cpuC);

  auto subSize = [](int& start, int& end, int dim) {
    if (dim == 1) {
      start = 0;
      end = dim;
    } else {
      int subDim = rand() % (dim - 1) + 1;  // NOLINT
      start = rand() % (dim - subDim);      // NOLINT
      end = start + subDim;
    }
  };

411 412 413 414 415 416
  auto subMatrix = [](MatrixPtr& sub,
                      MatrixPtr matrix,
                      size_t startRow,
                      size_t endRow,
                      size_t startCol,
                      size_t endCol) {
Z
zhangjinchao01 已提交
417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441
    if (!matrix->isTransposed()) {
      sub = matrix->subMatrix(startRow, endRow, startCol, endCol);
    } else {
      sub = matrix->subMatrix(startCol, endCol, startRow, endRow);
    }
  };

  int startM, endM;
  int startN, endN;
  int startK, endK;
  subSize(startM, endM, dimM);
  subSize(startN, endN, dimN);
  subSize(startK, endK, dimK);

  MatrixPtr subCpuA;
  MatrixPtr subCpuB;
  MatrixPtr subGpuA;
  MatrixPtr subGpuB;
  subMatrix(subCpuA, cpuA, startM, endM, startK, endK);
  subMatrix(subGpuA, gpuA, startM, endM, startK, endK);
  subMatrix(subCpuB, cpuB, startK, endK, startN, endN);
  subMatrix(subGpuB, gpuB, startK, endK, startN, endN);
  MatrixPtr subCpuC = cpuC->subMatrix(startM, endM, startN, endN);
  MatrixPtr subGpuC = gpuC->subMatrix(startM, endM, startN, endN);

442 443
  subCpuC->mul(*subCpuA, *subCpuB, alpha, beta);
  subGpuC->mul(*subGpuA, *subGpuB, alpha, beta);
Z
zhangjinchao01 已提交
444

445
  TensorCheckErr(*cpuC, *gpuC);
Z
zhangjinchao01 已提交
446 447 448 449 450 451 452 453 454 455 456 457
}

TEST(Matrix, mul) {
  for (auto transa : {false, true}) {
    for (auto transb : {false, true}) {
      for (auto dimM : {1, 9, 53, 127, 345, 1023, 2135}) {
        for (auto dimN : {1, 5, 37, 256, 1024}) {
          for (auto dimK : {8, 45, 346, 784, 1025}) {
            if (true == transa && true == transb) {
              continue;
            }
            VLOG(3) << setiosflags(ios::left) << setfill(' ')
458 459 460
                    << " transa=" << transa << " transb=" << transb
                    << " dimM=" << setw(5) << dimM << " dimN=" << setw(5)
                    << dimN << " dimK=" << setw(5) << dimK;
Z
zhangjinchao01 已提交
461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489

            testMatrixMul(transa, transb, dimM, dimN, dimK);
            testSubMatrixMul(transa, transb, dimM, dimN, dimK);
          }
        }
      }
    }
  }
}

void testVectorRowFunc(int size) {
  CpuVectorPtr cpu = std::make_shared<CpuVectorT<real>>(size);
  GpuVectorPtr gpu = std::make_shared<GpuVectorT<real>>(size);

  cpu->rand();
  gpu->copyFrom(*cpu);

  EXPECT_EQ(cpu->getMax(), gpu->getMax());
  EXPECT_EQ(cpu->getMin(), gpu->getMin());
  EXPECT_EQ(cpu->getAbsMax(), gpu->getAbsMax());
}

TEST(Vector, rowFunc) {
  for (auto size : {1, 5, 31, 90, 150, 500, 1000, 4000}) {
    VLOG(3) << " size=" << size;
    testVectorRowFunc(size);
  }
}

490
template <class T>
Z
zhangjinchao01 已提交
491 492 493 494 495 496 497 498
void testVectorReset(int size) {
  std::shared_ptr<CpuVectorT<T>> cpu = std::make_shared<CpuVectorT<T>>(size);
  std::shared_ptr<GpuVectorT<T>> gpu = std::make_shared<GpuVectorT<T>>(size);

  T value = (T)((int)rand() % 100 + 1.0f / ((int)rand() % 100));
  cpu->reset(value);
  gpu->reset(value);

499
  TensorCheckEqual(*cpu, *gpu);
Z
zhangjinchao01 已提交
500 501
}

502
template <class T>
Z
zhangjinchao01 已提交
503 504 505
void testVecortSelectFrom(int size) {
  std::shared_ptr<CpuVectorT<T>> cpuDst = std::make_shared<CpuVectorT<T>>(size);
  std::shared_ptr<GpuVectorT<T>> gpuDst = std::make_shared<GpuVectorT<T>>(size);
506 507 508 509
  std::shared_ptr<CpuVectorT<T>> cpuSrc =
      std::make_shared<CpuVectorT<T>>(size * 2);
  std::shared_ptr<GpuVectorT<T>> gpuSrc =
      std::make_shared<GpuVectorT<T>>(size * 2);
Z
zhangjinchao01 已提交
510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
  CpuIVectorPtr cpuIds = std::make_shared<CpuVectorT<int>>(size);
  GpuIVectorPtr gpuIds = std::make_shared<GpuVectorT<int>>(size);

  if (std::is_same<T, real>::value) {
    cpuSrc->rand();
  } else {
    cpuSrc->rand(100000);
  }
  gpuSrc->copyFrom(*cpuSrc);
  cpuIds->rand(size);
  gpuIds->copyFrom(*cpuIds);

  cpuDst->selectFrom(*cpuSrc, *cpuIds);
  gpuDst->selectFrom(*gpuSrc, *gpuIds);

525
  TensorCheckEqual(*cpuDst, *gpuDst);
Z
zhangjinchao01 已提交
526 527
}

528
template <class T>
Z
zhangjinchao01 已提交
529 530 531 532 533 534 535
void testVecotrZeroMem(int size) {
  std::shared_ptr<CpuVectorT<T>> cpu = std::make_shared<CpuVectorT<T>>(size);
  std::shared_ptr<GpuVectorT<T>> gpu = std::make_shared<GpuVectorT<T>>(size);

  cpu->zeroMem();
  gpu->zeroMem();

536
  TensorCheckEqual(*cpu, *gpu);
Z
zhangjinchao01 已提交
537 538
}

539
template <class T>
Z
zhangjinchao01 已提交
540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556
void testVectorIsEqual(int size) {
  std::shared_ptr<CpuVectorT<T>> cpuA = std::make_shared<CpuVectorT<T>>(size);
  std::shared_ptr<CpuVectorT<T>> cpuB = std::make_shared<CpuVectorT<T>>(size);
  std::shared_ptr<GpuVectorT<T>> gpuA = std::make_shared<GpuVectorT<T>>(size);
  std::shared_ptr<GpuVectorT<T>> gpuB = std::make_shared<GpuVectorT<T>>(size);

  if (std::is_same<T, real>::value) {
    cpuB->rand();
  } else {
    cpuB->rand(100000);
  }
  gpuB->copyFrom(*cpuB);

  T value = (T)((int)rand() % 100 + 1.0f / ((int)rand() % 100));
  cpuA->isEqualTo(*cpuB, value);
  gpuA->isEqualTo(*gpuB, value);

557
  TensorCheckEqual(*cpuA, *gpuA);
Z
zhangjinchao01 已提交
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587
}

TEST(Vector, Equal) {
  for (auto size : {1, 5, 31, 90, 150, 500, 1000, 4000}) {
    VLOG(3) << " size=" << size;
    testVectorReset<int>(size);
    testVectorReset<real>(size);
    testVecortSelectFrom<int>(size);
    testVecortSelectFrom<real>(size);
    testVecotrZeroMem<int>(size);
    testVecotrZeroMem<real>(size);
    testVectorIsEqual<int>(size);
    testVectorIsEqual<real>(size);
  }
}

void testMatrixTopK(int samples, int dim, int beamSize) {
  MatrixPtr cpuSrc = std::make_shared<CpuMatrix>(samples, dim);
  MatrixPtr gpuSrc = std::make_shared<GpuMatrix>(samples, dim);
  MatrixPtr cpuVal = std::make_shared<CpuMatrix>(samples, beamSize);
  MatrixPtr gpuVal = std::make_shared<GpuMatrix>(samples, beamSize);
  IVectorPtr cpuIds = std::make_shared<CpuIVector>(samples * beamSize);
  IVectorPtr gpuIds = std::make_shared<GpuIVector>(samples * beamSize);

  cpuSrc->randomizeUniform();
  gpuSrc->copyFrom(*cpuSrc);

  cpuSrc->rowMax(*cpuIds, *cpuVal);
  gpuSrc->rowMax(*gpuIds, *gpuVal);

588
  TensorCheckEqual(*cpuVal, *gpuVal);
Z
zhangjinchao01 已提交
589 590 591 592
}

TEST(Matrix, topK) {
  for (auto samples : {1, 5, 31, 90, 150, 500}) {
593 594
    for (auto dim :
         {1, 5, 8, 10, 15, 64, 80, 120, 256, 300, 1280, 5120, 50000}) {
Z
zhangjinchao01 已提交
595 596
      for (auto beamSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) {
        if (beamSize > dim) continue;
597
        VLOG(3) << " samples=" << samples << " beamSize=" << beamSize
Z
zhangjinchao01 已提交
598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623
                << " dim=" << dim;
        testMatrixTopK(samples, dim, beamSize);
      }
    }
  }
}

void testSMatrixTopK(int samples, int dim, int beamSize, real ratio) {
  int nnz = samples * dim * ratio;
  MatrixPtr cpuSrc = std::make_shared<CpuSparseMatrix>(samples, dim, nnz);
  MatrixPtr gpuSrc = std::make_shared<GpuSparseMatrix>(samples, dim, nnz);
  MatrixPtr cpuVal = std::make_shared<CpuMatrix>(samples, beamSize);
  MatrixPtr gpuVal = std::make_shared<GpuMatrix>(samples, beamSize);
  IVectorPtr cpuIds = std::make_shared<CpuIVector>(samples * beamSize);
  IVectorPtr gpuIds = std::make_shared<GpuIVector>(samples * beamSize);

  cpuSrc->randomizeUniform();
  gpuSrc->copyFrom(*cpuSrc);
  cpuVal->zero();
  cpuIds->zero();
  gpuVal->zero();
  gpuIds->zero();

  cpuSrc->rowMax(*cpuIds, *cpuVal);
  gpuSrc->rowMax(*gpuIds, *gpuVal);

624
  TensorCheckEqual(*cpuVal, *gpuVal);
Z
zhangjinchao01 已提交
625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644

  IVectorPtr outCheckIds = std::make_shared<CpuIVector>(samples * beamSize);
  outCheckIds->copyFrom(*gpuIds);

  const int* data1 = cpuIds->getData();
  const int* data2 = outCheckIds->getData();
  size_t size = cpuIds->getSize();
  for (size_t i = 0; i < size; i++) {
    if (data1[i] == -1 && data1[i] != data2[i]) {
      EXPECT_EQ(data1[i], data2[i]);
    }
  }
}

TEST(SMatrix, topK) {
  for (auto samples : {1, 5, 100}) {
    for (auto dim : {10000, 10000, 50000}) {
      for (auto beamSize : {1, 5, 40, 100, 500}) {
        for (auto ratio : {0.01, 0.001}) {
          if (beamSize > dim) continue;
645 646
          VLOG(3) << " samples=" << samples << " beamSize=" << beamSize
                  << " dim=" << dim << " ratio=" << ratio;
Z
zhangjinchao01 已提交
647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673
          testSMatrixTopK(samples, dim, beamSize, ratio);
        }
      }
    }
  }
}

void testMatrixSequenceAvgForward(int batchSize, int inputDim, int mode) {
  MatrixPtr cpuInput = std::make_shared<CpuMatrix>(batchSize, inputDim);
  MatrixPtr gpuInput = std::make_shared<GpuMatrix>(batchSize, inputDim);
  cpuInput->randomizeUniform();
  gpuInput->copyFrom(*cpuInput);

  IVectorPtr cpuSequence;
  generateSequenceStartPositions(batchSize, cpuSequence);
  IVectorPtr gpuSequence = IVector::create(cpuSequence->getSize(), true);
  gpuSequence->copyFrom(*cpuSequence);

  int newBatchSize = cpuSequence->getSize() - 1;
  MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(newBatchSize, inputDim);
  MatrixPtr gpuOutput = std::make_shared<GpuMatrix>(newBatchSize, inputDim);
  cpuOutput->zero();
  gpuOutput->zero();

  cpuOutput->sequenceAvgForward(*cpuInput, *cpuSequence, mode);
  gpuOutput->sequenceAvgForward(*gpuInput, *gpuSequence, mode);

674
  TensorCheckErr(*cpuOutput, *gpuOutput);
Z
zhangjinchao01 已提交
675 676 677 678 679 680 681 682 683 684 685 686 687 688
}

TEST(Matrix, sequenceAvgForward) {
  for (auto batchSize : {10, 128, 6000}) {
    for (auto inputDim : {32, 100, 512}) {
      for (auto mode : {0, 1, 2}) {
        VLOG(3) << " batchSize=" << batchSize << " inputDim=" << inputDim
                << " mode=" << mode;
        testMatrixSequenceAvgForward(batchSize, inputDim, mode);
      }
    }
  }
}

689
void testCosSimDerivate(int heightX, int heightY, int width, real scale) {
Z
zhangjinchao01 已提交
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717
  MatrixPtr prevOutX = CpuMatrix::create(heightX, width, false, false);
  MatrixPtr prevOutY = CpuMatrix::create(heightY, width, false, false);
  MatrixPtr grad = CpuMatrix::create(heightX, 1, false, false);
  MatrixPtr output = CpuMatrix::create(heightX, 1, false, false);
  MatrixPtr prevGradX = CpuMatrix::create(heightX, width, false, false);
  MatrixPtr prevGradY = CpuMatrix::create(heightY, width, false, false);

  prevOutX->randomizeUniform();
  prevOutY->randomizeUniform();
  grad->randomizeUniform();
  output->randomizeUniform();
  prevGradX->randomizeUniform();
  prevGradY->randomizeUniform();

  MatrixPtr prevOutXGpu = GpuMatrix::create(heightX, width, false, true);
  MatrixPtr prevOutYGpu = GpuMatrix::create(heightY, width, false, true);
  MatrixPtr gradGpu = GpuMatrix::create(heightX, 1, false, true);
  MatrixPtr outputGpu = GpuMatrix::create(heightX, 1, false, true);
  MatrixPtr prevGradXGpu = GpuMatrix::create(heightX, width, false, true);
  MatrixPtr prevGradYGpu = GpuMatrix::create(heightY, width, false, true);

  prevOutXGpu->copyFrom(*prevOutX);
  prevOutYGpu->copyFrom(*prevOutY);
  gradGpu->copyFrom(*grad);
  outputGpu->copyFrom(*output);
  prevGradXGpu->copyFrom(*prevGradX);
  prevGradYGpu->copyFrom(*prevGradY);

718 719
  grad->cosSimDerivative(
      *output, *prevOutX, *prevOutY, *prevGradX, *prevGradY, scale);
Z
zhangjinchao01 已提交
720 721 722 723 724 725 726 727

  gradGpu->cosSimDerivative(*outputGpu,
                            *prevOutXGpu,
                            *prevOutYGpu,
                            *prevGradXGpu,
                            *prevGradYGpu,
                            scale);

728 729
  TensorCheckErr(*prevGradX, *prevGradXGpu);
  TensorCheckErr(*prevGradY, *prevGradYGpu);
Z
zhangjinchao01 已提交
730 731 732 733 734 735 736 737 738 739 740 741 742 743
}

TEST(Matrix, cosSimDerivate) {
  for (auto heightX : {1, 10, 100}) {
    for (auto heightY : {1, heightX}) {
      for (auto width : {1, 10, 100}) {
        for (auto scale : {1.0, 2.0}) {
          testCosSimDerivate(heightX, heightY, width, scale);
        }
      }
    }
  }
}

744 745 746 747
void testParamReluBackwardDiff(int height,
                               int width,
                               int w_height,
                               int w_width) {
Z
zhangjinchao01 已提交
748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771
  MatrixPtr oGrad = CpuMatrix::create(height, width, false, false);
  MatrixPtr input = CpuMatrix::create(height, width, false, false);
  MatrixPtr diff = CpuMatrix::create(height, width, false, false);
  MatrixPtr w = CpuMatrix::create(w_height, w_width, false, false);

  oGrad->randomizeUniform();
  input->randomizeUniform();
  w->randomizeUniform();
  diff->randomizeUniform();
  input->add(-0.5);

  MatrixPtr oGradGpu = GpuMatrix::create(height, width, false, true);
  MatrixPtr inputGpu = GpuMatrix::create(height, width, false, true);
  MatrixPtr diffGpu = CpuMatrix::create(height, width, false, true);
  MatrixPtr wGpu = GpuMatrix::create(w_height, w_width, false, true);

  oGradGpu->copyFrom(*oGrad);
  inputGpu->copyFrom(*input);
  wGpu->copyFrom(*w);
  diffGpu->copyFrom(*diff);

  diff->paramReluBackwardDiff(*oGrad, *input, *w);
  diffGpu->paramReluBackwardDiff(*oGradGpu, *inputGpu, *wGpu);

772
  TensorCheckErr(*diff, *diffGpu);
Z
zhangjinchao01 已提交
773 774 775 776 777 778 779 780 781 782 783 784 785 786
}

TEST(Matrix, paramReluBackwardDiff) {
  for (auto height : {10, 100}) {
    for (auto width : {10, 100}) {
      for (auto w_height : {1, 2}) {
        for (auto w_width : {1, 2}) {
          testParamReluBackwardDiff(height, width, w_height, w_width);
        }
      }
    }
  }
}

H
He 已提交
787 788 789 790 791 792 793 794 795 796 797 798 799
void testClassificationError(int numSamples, int dim) {
  MatrixPtr cpuError = std::make_shared<CpuMatrix>(numSamples, 1);
  MatrixPtr gpuError = std::make_shared<GpuMatrix>(numSamples, 1);
  MatrixPtr cpuOutput = std::make_shared<CpuMatrix>(numSamples, dim);
  MatrixPtr gpuOutput = std::make_shared<GpuMatrix>(numSamples, dim);
  IVectorPtr cpuLabel = std::make_shared<CpuIVector>(numSamples);
  IVectorPtr gpuLabel = std::make_shared<GpuIVector>(numSamples);

  cpuOutput->randomizeUniform();
  cpuLabel->rand(dim);
  gpuOutput->copyFrom(*cpuOutput);
  gpuLabel->copyFrom(*cpuLabel);

800 801
  cpuError->classificationError(*cpuOutput, *cpuLabel);
  gpuError->classificationError(*gpuOutput, *gpuLabel);
H
He 已提交
802

803
  TensorCheckEqual(*cpuError, *gpuError);
H
He 已提交
804 805 806 807 808 809 810 811 812 813 814
}

TEST(Matrix, classificationError) {
  for (auto numSamples : {1, 10, 100, 1000, 70000}) {
    for (auto dim : {1, 10, 100, 1000}) {
      VLOG(3) << " numSamples=" << numSamples << " dim=" << dim;
      testClassificationError(numSamples, dim);
    }
  }
}

815 816 817 818 819 820 821 822 823 824
void testMaxPoolFwdBwd(int numSamples,
                       int channels,
                       int imgSizeH,
                       int imgSizeW,
                       int ksizeH,
                       int ksizeW,
                       int strideH,
                       int strideW,
                       int padH,
                       int padW) {
825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841
  int outH = 0, outW = 0;
  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;

  int inWidth = imgSizeH * imgSizeW * channels;
  MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
  MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true);

  int outWidth = channels * outH * outW;
  MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false);
  MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true);

  input->randomizeUniform();
  target->randomizeUniform();
  inputGpu->copyFrom(*input);
  targetGpu->copyFrom(*target);

842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865
  target->maxPoolForward(*input,
                         imgSizeH,
                         imgSizeW,
                         channels,
                         ksizeW,
                         ksizeH,
                         strideH,
                         strideW,
                         outH,
                         outW,
                         padH,
                         padW);
  targetGpu->maxPoolForward(*inputGpu,
                            imgSizeH,
                            imgSizeW,
                            channels,
                            ksizeW,
                            ksizeH,
                            strideH,
                            strideW,
                            outH,
                            outW,
                            padH,
                            padW);
866 867 868 869 870 871 872
  MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false);
  targetCheck->copyFrom(*targetGpu);
  checkMatrixEqual(target, targetCheck);

  MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false);
  MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true);
  MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false);
873 874
  MatrixPtr targetGpuGrad =
      GpuMatrix::create(numSamples, outWidth, false, true);
875 876 877 878 879 880

  inputGrad->randomizeUniform();
  targetGrad->randomizeUniform();
  inputGpuGrad->copyFrom(*inputGrad);
  targetGpuGrad->copyFrom(*targetGrad);

881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912
  inputGrad->maxPoolBackward(*input,
                             imgSizeH,
                             imgSizeW,
                             *targetGrad,
                             *target,
                             ksizeW,
                             ksizeH,
                             strideH,
                             strideW,
                             outH,
                             outW,
                             1.0,
                             1.0,
                             padH,
                             padW);
  inputGpuGrad->maxPoolBackward(*inputGpu,
                                imgSizeH,
                                imgSizeW,
                                *targetGpuGrad,
                                *targetGpu,
                                ksizeW,
                                ksizeH,
                                strideH,
                                strideW,
                                outH,
                                outW,
                                1.0,
                                1.0,
                                padH,
                                padW);
  MatrixPtr targetBwdCheck =
      CpuMatrix::create(numSamples, inWidth, false, false);
913 914 915 916
  targetBwdCheck->copyFrom(*inputGpuGrad);
  checkMatrixEqual(inputGrad, targetBwdCheck);
}

917 918 919 920 921 922 923 924 925 926
void testAvgPoolFwdBwd(int numSamples,
                       int channels,
                       int imgSizeH,
                       int imgSizeW,
                       int ksizeH,
                       int ksizeW,
                       int strideH,
                       int strideW,
                       int padH,
                       int padW) {
927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943
  int outH = 0, outW = 0;
  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;

  int inWidth = imgSizeH * imgSizeW * channels;
  MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
  MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true);

  int outWidth = channels * outH * outW;
  MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false);
  MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true);

  input->randomizeUniform();
  target->randomizeUniform();
  inputGpu->copyFrom(*input);
  targetGpu->copyFrom(*target);

944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967
  target->avgPoolForward(*input,
                         imgSizeH,
                         imgSizeW,
                         channels,
                         ksizeW,
                         ksizeH,
                         strideH,
                         strideW,
                         outH,
                         outW,
                         padH,
                         padW);
  targetGpu->avgPoolForward(*inputGpu,
                            imgSizeH,
                            imgSizeW,
                            channels,
                            ksizeW,
                            ksizeH,
                            strideH,
                            strideW,
                            outH,
                            outW,
                            padH,
                            padW);
968 969

  TensorCheckErr(*target, *targetGpu);
970 971 972 973

  MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false);
  MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true);
  MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false);
974 975
  MatrixPtr targetGpuGrad =
      GpuMatrix::create(numSamples, outWidth, false, true);
976 977 978 979 980 981

  inputGrad->randomizeUniform();
  targetGrad->randomizeUniform();
  inputGpuGrad->copyFrom(*inputGrad);
  targetGpuGrad->copyFrom(*targetGrad);

982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
  inputGrad->avgPoolBackward(*targetGrad,
                             imgSizeH,
                             imgSizeW,
                             ksizeW,
                             ksizeH,
                             strideH,
                             strideW,
                             outH,
                             outW,
                             1.0,
                             1.0,
                             padH,
                             padW);
  inputGpuGrad->avgPoolBackward(*targetGpuGrad,
                                imgSizeH,
                                imgSizeW,
                                ksizeW,
                                ksizeH,
                                strideH,
                                strideW,
                                outH,
                                outW,
                                1.0,
                                1.0,
                                padH,
                                padW);
1008 1009

  TensorCheckErr(*inputGrad, *inputGpuGrad);
1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020
}

TEST(Matrix, PoolFwdBwd) {
  for (auto numSamples : {5, 32}) {
    for (auto channels : {1, 9, 32}) {
      for (auto imgSizeH : {14, 28}) {
        for (auto imgSizeW : {16, 30}) {
          for (auto sizeX : {2, 5}) {
            for (auto sizeY : {2, 5}) {
              for (auto sH : {1, 2}) {
                for (auto sW : {1, 2}) {
1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051
                  for (auto pH : {0, (sizeY - 1) / 2}) {
                    for (auto pW : {0, (sizeX - 1) / 2}) {
                      VLOG(3) << " numSamples=" << numSamples
                              << " channels=" << channels
                              << " imgSizeH=" << imgSizeH
                              << " imgSizeW=" << imgSizeW << " sizeX=" << sizeX
                              << " sizeY=" << sizeY << " strideH=" << sH
                              << " strideW=" << sW << " padingH=" << pH
                              << " padingW=" << pW;
                      testMaxPoolFwdBwd(numSamples,
                                        channels,
                                        imgSizeH,
                                        imgSizeW,
                                        sizeX,
                                        sizeY,
                                        sH,
                                        sW,
                                        pH,
                                        pW);
                      testAvgPoolFwdBwd(numSamples,
                                        channels,
                                        imgSizeH,
                                        imgSizeW,
                                        sizeX,
                                        sizeY,
                                        sH,
                                        sW,
                                        pH,
                                        pW);
                    }
                  }
1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
                }
              }
            }
          }
        }
      }
    }
  }
}

1062 1063
void testMaxOutFwdBwd(
    int numSamples, int imgSizeH, int imgSizeW, int channels, int groups) {
1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
  int inWidth = imgSizeH * imgSizeW * channels;
  int outChannels = channels / groups;
  int outWidth = imgSizeH * imgSizeW * outChannels;

  // forward
  MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
  MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true);

  MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false);
  MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true);

  IVectorPtr id = CpuIVector::create(numSamples * outWidth, false);
  IVectorPtr idGpu = GpuIVector::create(numSamples * outWidth, true);

  input->randomizeUniform();
  inputGpu->copyFrom(*input);

  target->maxoutForward(*input, *id, outChannels, groups);
  targetGpu->maxoutForward(*inputGpu, *idGpu, outChannels, groups);

1084 1085
  TensorCheckErr(*target, *targetGpu);
  TensorCheckEqual(*id, *idGpu);
1086 1087 1088 1089 1090 1091

  // backward
  MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false);
  MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true);

  MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false);
1092 1093
  MatrixPtr targetGpuGrad =
      GpuMatrix::create(numSamples, outWidth, false, true);
1094 1095 1096 1097 1098 1099 1100 1101 1102

  inputGrad->randomizeUniform();
  targetGrad->randomizeUniform();
  inputGpuGrad->copyFrom(*inputGrad);
  targetGpuGrad->copyFrom(*targetGrad);

  inputGrad->maxoutBackward(*targetGrad, *id, outChannels, groups);
  inputGpuGrad->maxoutBackward(*targetGpuGrad, *idGpu, outChannels, groups);

1103
  TensorCheckErr(*inputGrad, *inputGpuGrad);
1104 1105 1106 1107 1108 1109 1110 1111
}

TEST(Matrix, MaxOutFwdBwd) {
  for (auto numSamples : {5, 10}) {
    for (auto channels : {8, 16}) {
      for (auto imgSizeH : {14, 28}) {
        for (auto imgSizeW : {16, 30}) {
          for (auto groups : {2, 4}) {
1112 1113
            VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
                    << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW
1114 1115 1116 1117 1118 1119 1120 1121 1122
                    << " groups=" << groups;
            testMaxOutFwdBwd(numSamples, imgSizeH, imgSizeW, channels, groups);
          }
        }
      }
    }
  }
}

Z
zhangjinchao01 已提交
1123
#endif