Matrix.cpp 154.1 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

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 "Matrix.h"
Q
qijun 已提交
16
#include "MathFunctions.h"
Z
zhangjinchao01 已提交
17 18 19 20 21
#include "SparseMatrix.h"
#include "SparseRowMatrix.h"

#include <float.h>
#include <algorithm>
Q
qijun 已提交
22
#include <cmath>
Z
zhangjinchao01 已提交
23 24

#include <string.h>
L
liaogang 已提交
25
#include "hl_cnn.h"
Z
zhangjinchao01 已提交
26 27 28
#include "hl_gpu.h"
#include "hl_table_apply.h"
#include "hl_top_k.h"
Q
qijun 已提交
29
#include "paddle/utils/Logging.h"
Z
zhangjinchao01 已提交
30

31
#include "paddle/function/GemmFunctor.h"
Z
zhangjinchao01 已提交
32 33 34 35 36 37 38 39 40 41 42 43
#include "paddle/utils/ThreadLocal.h"

#include "SIMDFunctions.h"

namespace paddle {

inline real _pow(real a, real beta) { return std::pow(a, beta); }

inline real _square(real a) { return a * a; }

inline real _safelog(real a) { return a > 0.0f ? std::log(a) : -40.0f; }

44 45 46 47 48
Matrix::Matrix(MemoryHandlePtr memHandle,
               size_t height,
               size_t width,
               bool trans,
               bool use_gpu)
Z
zhangjinchao01 已提交
49
    : BaseMatrix(
50 51
          height,
          width,
Q
qijun 已提交
52
          memHandle ? (reinterpret_cast<real*>(memHandle->getBuf())) : nullptr,
53 54
          trans,
          use_gpu) {
Z
zhangjinchao01 已提交
55 56 57 58
  elementCnt_ = width * height;
  memoryHandle_ = memHandle;
}

59 60
Matrix::Matrix(
    real* data, size_t height, size_t width, bool trans, bool use_gpu)
Z
zhangjinchao01 已提交
61 62 63 64
    : BaseMatrix(height, width, data, trans, use_gpu) {
  elementCnt_ = width * height;
}

65 66 67 68 69 70
Matrix::Matrix(real* data,
               size_t height,
               size_t width,
               size_t stride,
               bool trans,
               bool use_gpu)
Z
zhangjinchao01 已提交
71 72 73 74
    : BaseMatrix(height, width, stride, data, trans, use_gpu) {
  elementCnt_ = width * height;
}

75 76 77 78 79
MatrixPtr Matrix::createSparseMatrix(real* data,
                                     int* row,
                                     int* col,
                                     size_t height,
                                     size_t width,
Z
zhangjinchao01 已提交
80 81
                                     size_t nnz, /* used to allocate space */
                                     SparseValueType valueType, /*value type*/
82 83
                                     SparseFormat format,
                                     bool trans,
Z
zhangjinchao01 已提交
84 85
                                     bool useGpu) {
  if (useGpu) {
86 87
    return std::make_shared<GpuSparseMatrix>(
        data, row, col, height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
88
  } else {
89 90
    return std::make_shared<CpuSparseMatrix>(
        data, row, col, height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
91 92 93
  }
}

94 95
MatrixPtr Matrix::createSparseMatrix(size_t height,
                                     size_t width,
Z
zhangjinchao01 已提交
96 97
                                     size_t nnz, /* used to allocate space */
                                     SparseValueType valueType, /*value type*/
98 99
                                     SparseFormat format,
                                     bool trans,
Z
zhangjinchao01 已提交
100 101
                                     bool useGpu) {
  if (useGpu) {
102 103
    return std::make_shared<GpuSparseMatrix>(
        height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
104
  } else {
105 106
    return std::make_shared<CpuSparseMatrix>(
        height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
107 108 109
  }
}

110 111 112
MatrixPtr Matrix::create(MemoryHandlePtr memHandle,
                         size_t height,
                         size_t width,
Z
zhangjinchao01 已提交
113 114 115 116
                         bool trans) {
  if (auto gpuHandle = std::dynamic_pointer_cast<GpuMemoryHandle>(memHandle)) {
    return std::make_shared<GpuMatrix>(gpuHandle, height, width, trans);
  } else if (auto cpuHandle =
Q
qijun 已提交
117
                 std::dynamic_pointer_cast<CpuMemoryHandle>(memHandle)) {
Z
zhangjinchao01 已提交
118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
    return std::make_shared<CpuMatrix>(cpuHandle, height, width, trans);
  } else {
    LOG(FATAL) << "Wrong";
    return nullptr;
  }
}

MatrixPtr Matrix::create(size_t height, size_t width, bool trans, bool useGpu) {
  if (useGpu) {
    return std::make_shared<GpuMatrix>(height, width, trans);
  } else {
    return std::make_shared<CpuMatrix>(height, width, trans);
  }
}

133 134
MatrixPtr Matrix::create(
    real* data, size_t height, size_t width, bool trans, bool useGpu) {
Z
zhangjinchao01 已提交
135 136 137 138 139 140 141
  if (useGpu) {
    return std::make_shared<GpuMatrix>(data, height, width, trans);
  } else {
    return std::make_shared<CpuMatrix>(data, height, width, trans);
  }
}

142 143 144 145 146 147
MatrixPtr Matrix::create(real* data,
                         size_t height,
                         size_t width,
                         size_t stride,
                         bool trans,
                         bool useGpu) {
Z
zhangjinchao01 已提交
148 149 150 151 152 153 154
  if (useGpu) {
    return std::make_shared<GpuMatrix>(data, height, width, stride, trans);
  } else {
    return std::make_shared<CpuMatrix>(data, height, width, stride, trans);
  }
}

155 156 157 158 159
MatrixPtr Matrix::createSparseMatrix(size_t height,
                                     size_t width,
                                     size_t nnz,
                                     SparseValueType valueType,
                                     bool trans,
Z
zhangjinchao01 已提交
160 161
                                     bool useGpu) {
  if (useGpu) {
162 163
    return std::make_shared<GpuSparseMatrix>(
        height, width, nnz, valueType, SPARSE_CSR, trans);
Z
zhangjinchao01 已提交
164
  } else {
165 166
    return std::make_shared<CpuSparseMatrix>(
        height, width, nnz, valueType, SPARSE_CSR, trans);
Z
zhangjinchao01 已提交
167 168 169
  }
}

170 171
void Matrix::resizeOrCreate(
    MatrixPtr& matrix, size_t height, size_t width, bool trans, bool useGpu) {
Z
zhangjinchao01 已提交
172 173 174
  if (!matrix) {
    matrix = Matrix::create(height, width, trans, useGpu);
  } else {
175
    CHECK_EQ(matrix->useGpu(), useGpu);
Z
zhangjinchao01 已提交
176 177 178 179
    matrix->resize(height, width);
  }
}

180 181 182 183
void Matrix::resizeOrCreateSparseMatrix(MatrixPtr& matrix,
                                        size_t height,
                                        size_t width,
                                        size_t nnz,
Z
zhangjinchao01 已提交
184
                                        SparseValueType valueType,
185 186
                                        SparseFormat format,
                                        bool trans,
Z
zhangjinchao01 已提交
187 188
                                        bool useGpu) {
  if (!matrix) {
189 190
    matrix = Matrix::createSparseMatrix(
        height, width, nnz, valueType, format, trans, useGpu);
Z
zhangjinchao01 已提交
191 192 193
  } else {
    CHECK(dynamic_cast<CpuSparseMatrix*>(matrix.get()) ||
          dynamic_cast<GpuSparseMatrix*>(matrix.get()));
194
    CHECK_EQ(matrix->useGpu(), useGpu);
Z
zhangjinchao01 已提交
195 196 197 198 199 200 201 202 203 204 205 206
    matrix->resize(height, width, nnz, valueType, format);
  }
}

void Matrix::reshape(size_t height, size_t width) {
  CHECK(isContiguous());
  CHECK(height_ * width_ == height * width);
  height_ = height;
  width_ = width;
  stride_ = width_;
}

207 208 209
MatrixPtr Matrix::subMatrix(size_t startRow,
                            size_t endRow,
                            size_t startCol,
Z
zhangjinchao01 已提交
210 211 212 213 214 215 216
                            size_t endCol) {
  CHECK_LE(startRow, endRow);
  CHECK_LE(endRow, getHeight());
  CHECK_LE(startCol, endCol);
  CHECK_LE(endCol, getWidth());

  return Matrix::create(getData() + startRow * getStride() + startCol,
217 218 219 220 221
                        endRow - startRow,
                        endCol - startCol,
                        getStride(),
                        trans_,
                        useGpu_);
Z
zhangjinchao01 已提交
222 223
}

224 225 226 227 228 229 230 231 232
void Matrix::setDiag(real value) {
  CHECK(data_ != NULL);
  CHECK_EQ(height_, width_);

  zeroMem();
  BaseMatrix diag(height_, 1, stride_ + 1, data_, false, useGpu_);
  diag.assign(value);
}

Z
zhangjinchao01 已提交
233 234
GpuMatrix::GpuMatrix(size_t height, size_t width, bool trans)
    : Matrix(std::make_shared<GpuMemoryHandle>(height * width * sizeof(real)),
235 236 237 238
             height,
             width,
             trans,
             true) {}
Z
zhangjinchao01 已提交
239 240 241 242 243 244 245 246 247 248 249 250

GpuMatrix::~GpuMatrix() {}

void GpuMatrix::zeroMem() {
  CHECK(data_ != NULL);
  zero();
}

void GpuMatrix::resetOne() {
  CHECK(data_ != NULL);
  one();
}
251

Z
zhangjinchao01 已提交
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
void GpuMatrix::resize(size_t newHeight, size_t newWidth) {
  size_t newSize = newHeight * newWidth;
  if (NULL == memoryHandle_.get() ||
      newSize * sizeof(real) > memoryHandle_->getAllocSize()) {
    memoryHandle_ = std::make_shared<GpuMemoryHandle>(newSize * sizeof(real));
    data_ = reinterpret_cast<real*>(memoryHandle_->getBuf());
  }
  height_ = newHeight;
  width_ = newWidth;
  elementCnt_ = newSize;
  stride_ = width_;
}

real GpuMatrix::getElement(size_t x, size_t y) const {
  real elem = 0;
  hl_memcpy_device2host(&elem, &data_[x * stride_ + y], sizeof(real));
  return elem;
}

real GpuMatrix::getSum() {
  CHECK(isContiguous());
  real sum = 0.0f;
  hl_vector_sum(data_, &sum, height_ * width_);
  return sum;
}

278 279 280 281 282 283 284 285 286 287 288 289
real GpuMatrix::getMin() {
  CHECK(isContiguous());
  auto vec = GpuVector(height_ * width_, data_);
  return vec.getMin();
}

real GpuMatrix::getMax() {
  CHECK(isContiguous());
  auto vec = GpuVector(height_ * width_, data_);
  return vec.getMax();
}

Z
zhangjinchao01 已提交
290 291 292
void GpuMatrix::accumulateColSum(Matrix& src) {
  CHECK_EQ(getWidth(), src.getWidth());
  CHECK_EQ(getHeight(), (size_t)1);
X
xuwei06 已提交
293
  sumCols(src, 1.0, 1.0);
Z
zhangjinchao01 已提交
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
}

real GpuMatrix::getAbsSum() {
  CHECK(isContiguous());
  real sum = 0.0f;
  hl_vector_abs_sum(data_, &sum, height_ * width_);
  return sum;
}

void GpuMatrix::copyFrom(const Matrix& src) {
  CHECK(isContiguous());
  CHECK(src.isContiguous());
  CHECK(elementCnt_ == src.getElementCnt());

  if (typeid(src) == typeid(CpuMatrix)) {
309 310
    hl_memcpy_host2device(
        data_, const_cast<real*>(src.getData()), sizeof(real) * elementCnt_);
Z
zhangjinchao01 已提交
311
  } else if (typeid(src) == typeid(GpuMatrix)) {
312 313
    hl_memcpy_device2device(
        data_, const_cast<real*>(src.getData()), sizeof(real) * elementCnt_);
Z
zhangjinchao01 已提交
314 315 316 317 318 319 320 321 322
  } else {
    LOG(FATAL) << "Wrong";
  }
}

void GpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) {
  CHECK(isContiguous());
  CHECK(src.isContiguous());
  CHECK(elementCnt_ == src.getElementCnt());
323 324 325 326
  hl_memcpy_async(this->getData(),
                  const_cast<real*>(src.getData()),
                  sizeof(real) * elementCnt_,
                  stream);
Z
zhangjinchao01 已提交
327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
}

void GpuMatrix::copyFrom(const real* hostSrc, size_t size) {
  CHECK(isContiguous());
  CHECK(size <= elementCnt_);
  hl_memcpy_host2device(data_, const_cast<real*>(hostSrc), sizeof(real) * size);
}

void GpuMatrix::copyFrom(const real* hostSrc, const int64_t* seq) {
  LOG(FATAL) << "not implemented";
}

void GpuMatrix::copyFrom(const IVector& src) {
  CHECK(isContiguous());
  CpuMatrix matrix(src.getSize(), 1, false);
  matrix.copyFrom(src);
  copyFrom(matrix);
}

346
void GpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) {
Z
zhangjinchao01 已提交
347 348 349 350 351
  size_t height = getHeight();
  size_t width = getWidth();
  CHECK_EQ(b.getWidth(), width);
  real* dst = getData();
  real* src = b.getData();
352
  const int* index = rowIndex.getData();
Z
zhangjinchao01 已提交
353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376
  hl_sequence2batch_copy(dst, src, index, width, height, true);
}

MatrixPtr GpuMatrix::clone(size_t height, size_t width, bool useGpu) {
  CHECK(isContiguous());

  if (height == 0 && width == 0) {
    height = height_;
    width = width_;
  }

  CHECK(width && height);

  if (useGpu) {
    return std::make_shared<GpuMatrix>(height, width);
  } else {
    return std::make_shared<CpuMatrix>(height, width);
  }
}

MatrixPtr GpuMatrix::getTranspose() {
  if (memoryHandle_.get() != NULL) {
    MatrixPtr copy_T(
        new GpuMatrix(std::dynamic_pointer_cast<GpuMemoryHandle>(memoryHandle_),
377 378 379
                      height_,
                      width_,
                      true));
Z
zhangjinchao01 已提交
380 381 382 383 384 385 386
    return copy_T;
  } else {
    MatrixPtr copy_T(new GpuMatrix(data_, height_, width_, true));
    return copy_T;
  }
}

387
void GpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
Z
zhangjinchao01 已提交
388 389 390 391
  if (memAlloc) {
    matTrans = std::make_shared<GpuMatrix>(width_, height_);
  } else {
    CHECK(matTrans != NULL);
H
Haonan 已提交
392 393
    CHECK_EQ(matTrans->getHeight(), width_);
    CHECK_EQ(matTrans->getWidth(), height_);
Z
zhangjinchao01 已提交
394 395 396 397 398 399 400 401 402
  }
  real* dataTrans = matTrans->getData();
  real* data = getData();
  int lda = getStride();
  int ldc = matTrans->getStride();

  hl_matrix_transpose(data, dataTrans, height_, width_, lda, ldc);
}

403 404 405 406 407
void GpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) {
  if (memAlloc) {
    matRot = std::make_shared<GpuMatrix>(width_, height_);
  } else {
    CHECK(matRot != NULL);
H
Haonan 已提交
408 409
    CHECK_EQ(matRot->getHeight(), width_);
    CHECK_EQ(matRot->getWidth(), height_);
410 411
  }

H
Haonan 已提交
412 413 414
  real* dataRot = matRot->getData();
  real* data = getData();
  hl_matrix_rotate(data, dataRot, height_, width_, clockWise);
415 416
}

L
lzhao4ever 已提交
417 418 419 420 421 422
MatrixPtr GpuMatrix::getInverse() {
  MatrixPtr matInv;
  inverse(matInv, true);
  return matInv;
}

423
void GpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) {
L
lzhao4ever 已提交
424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
  CHECK_EQ(height_, width_);

  if (memAlloc) {
    matInv = std::make_shared<GpuMatrix>(height_, width_);
  } else {
    CHECK(matInv != NULL);
  }

  real* data = getData();
  real* dataInv = matInv->getData();
  int lda = getStride();
  int ldc = matInv->getStride();

  hl_matrix_inverse(data, dataInv, height_, lda, ldc);
}

Z
zhangjinchao01 已提交
440 441 442 443 444
void GpuMatrix::addBias(Matrix& b, real scale) {
  CHECK(b.getHeight() == 1) << "the Bias should be a vector";
  BaseMatrix::addBias(b, scale);
}

445 446 447 448
void GpuMatrix::addSharedBias(Matrix& b, real scale) {
  CHECK(b.getHeight() == 1) << "the Bias should be a vector";
  CHECK_LE(b.getWidth(), getWidth());
  CHECK_EQ(getWidth() % b.getWidth(), 0UL);
449 450
  hl_matrix_add_shared_bias(
      getData(), b.getData(), b.getWidth(), getHeight(), getWidth(), scale);
451 452
}

Z
zhangjinchao01 已提交
453
void GpuMatrix::collectBias(Matrix& a, real scale) {
454
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
455 456
  CHECK_EQ(getHeight(), (size_t)1);
  CHECK_EQ(width_, a.getWidth());
Q
qijun 已提交
457
  GpuSparseMatrix* sMatPtr = dynamic_cast<GpuSparseMatrix*>(&a);
Z
zhangjinchao01 已提交
458
  if (!sMatPtr) {
459
    sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
460 461 462
  } else {
    real* data = getData();
    hl_sparse_matrix_s A_d = sMatPtr->sMatrix_.get();
Q
qijun 已提交
463
    hl_sparse_matrix_column_sum(data, A_d, sMatPtr->getHeight(), width_, scale);
Z
zhangjinchao01 已提交
464
  }
465
#endif
Z
zhangjinchao01 已提交
466 467
}

468 469 470
void GpuMatrix::collectSharedBias(Matrix& a, real scale) {
  CHECK_EQ(getHeight(), (size_t)1);
  CHECK_EQ(a.getWidth() % getWidth(), 0UL);
471 472
  hl_matrix_collect_shared_bias(
      getData(), a.getData(), getWidth(), a.getHeight(), a.getWidth(), scale);
473 474
}

Z
zhangjinchao01 已提交
475 476 477 478 479 480 481 482 483 484 485 486 487 488
void GpuMatrix::sequenceAvgForward(Matrix& a,
                                   const IVector& startsPos,
                                   int mode) {
  size_t height = getHeight();
  size_t width = getWidth();
  CHECK_EQ(height, startsPos.getSize() - 1);
  CHECK_EQ(width, a.getWidth());
  real* dst = getData();
  real* src = a.getData();
  const int* starts = startsPos.getData();

  hl_sequence_avg_forward(dst, src, starts, height, width, mode);
}

L
Luo Tao 已提交
489 490 491 492 493 494 495 496 497 498 499 500 501 502
void GpuMatrix::sequenceAvgBackward(Matrix& a,
                                    const IVector& startsPos,
                                    int mode) {
  size_t height = a.getHeight();
  size_t width = getWidth();
  CHECK_EQ(height, startsPos.getSize() - 1);
  CHECK_EQ(width, a.getWidth());
  real* dst = getData();
  real* src = a.getData();
  const int* starts = startsPos.getData();

  hl_sequence_avg_backward(dst, src, starts, height, width, mode);
}

Z
zhangjinchao01 已提交
503
/* this = scaleAB*(a*b) +  scaleT*this */
504 505 506
void GpuMatrix::mul(const GpuMatrix& a,
                    const GpuMatrix& b,
                    real scaleAB,
Z
zhangjinchao01 已提交
507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537
                    real scaleT) {
  CHECK(!isTransposed()) << "Not supported";

  if (!a.isTransposed() && !b.isTransposed()) {
    CHECK_EQ(width_, b.width_);
    CHECK_EQ(height_, a.height_);
    CHECK_EQ(a.width_, b.height_);
  } else if (a.isTransposed() && !b.isTransposed()) {
    CHECK_EQ(width_, b.width_);
    CHECK_EQ(height_, a.width_);
    CHECK_EQ(a.height_, b.height_);
  } else if (!a.isTransposed() && b.isTransposed()) {
    CHECK_EQ(width_, b.height_);
    CHECK_EQ(height_, a.height_);
    CHECK_EQ(a.width_, b.width_);
  } else {
    LOG(FATAL) << "Is not supported";
  }

  real* A_d = a.data_;
  real* B_d = b.data_;
  real* C_d = data_;
  int dimM = getHeight();
  int dimN = getWidth();
  int dimK = !a.isTransposed() ? a.width_ : a.height_;
  int lda = a.getStride();
  int ldb = b.getStride();
  int ldc = getStride();
  hl_trans_op_t transa = !a.isTransposed() ? HPPL_OP_N : HPPL_OP_T;
  hl_trans_op_t transb = !b.isTransposed() ? HPPL_OP_N : HPPL_OP_T;

538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555
  hl_matrix_mul(A_d,
                transa,
                B_d,
                transb,
                C_d,
                dimM,
                dimN,
                dimK,
                scaleAB,
                scaleT,
                lda,
                ldb,
                ldc);
}

void GpuMatrix::mul(const GpuSparseMatrix& a,
                    const GpuMatrix& b,
                    real scaleAB,
Z
zhangjinchao01 已提交
556
                    real scaleT) {
557
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573
  CHECK(isContiguous());
  CHECK(b.isContiguous());
  CHECK(b.useGpu_ == true) << "Matrix type are not equal";
  CHECK(!trans_ && !b.trans_) << "not supported";

  if (!a.trans_) {
    CHECK(width_ == b.width_ && height_ == a.height_ && a.width_ == b.height_)
        << "Matrix dimensions are not equal";
  } else {
    CHECK(width_ == b.width_ && height_ == a.width_ && a.height_ == b.height_)
        << "Matrix dimensions are not equal";
  }
  hl_trans_op_t transA = a.trans_ ? HPPL_OP_T : HPPL_OP_N;
  hl_sparse_matrix_s A_d = a.sMatrix_.get();
  real* B_d = b.data_;
  real* C_d = data_;
574 575 576 577 578 579 580 581 582 583
  hl_matrix_csr_mul_dense(A_d,
                          transA,
                          B_d,
                          HPPL_OP_N,
                          C_d,
                          height_,
                          width_,
                          b.height_,
                          scaleAB,
                          scaleT);
584
#endif
585 586 587 588 589
}

void GpuMatrix::mul(const GpuMatrix& a,
                    const GpuSparseMatrix& b,
                    real scaleAB,
Z
zhangjinchao01 已提交
590
                    real scaleT) {
591
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
  CHECK(isContiguous());
  CHECK(a.isContiguous());
  CHECK(a.useGpu_ == true) << "Matrix type are not equal";

  hl_sparse_matrix_s B_d = b.sMatrix_.get();
  real* A_d = a.data_;
  real* C_d = data_;
  hl_trans_op_t transB = b.trans_ ? HPPL_OP_T : HPPL_OP_N;
  if (!b.trans_) {
    CHECK(width_ == b.width_ && height_ == a.height_ && a.width_ == b.height_)
        << "Matrix dimensions are not equal";
  } else {
    CHECK(width_ == b.height_ && height_ == a.height_ && a.width_ == b.width_)
        << "Matrix dimensions are not equal";
  }
  if (b.format_ == SPARSE_CSC) {
608 609 610 611 612 613 614 615 616 617
    hl_matrix_dense_mul_csc(A_d,
                            HPPL_OP_N,
                            B_d,
                            transB,
                            C_d,
                            height_,
                            width_,
                            a.width_,
                            scaleAB,
                            scaleT);
Z
zhangjinchao01 已提交
618
  } else {
619 620 621 622 623 624 625 626 627 628
    hl_matrix_dense_mul_csr(A_d,
                            HPPL_OP_N,
                            B_d,
                            transB,
                            C_d,
                            height_,
                            width_,
                            a.width_,
                            scaleAB,
                            scaleT);
Z
zhangjinchao01 已提交
629
  }
630
#endif
Z
zhangjinchao01 已提交
631 632 633
}

/* this = a*b */
634
void GpuMatrix::mul(const Matrix& a, const Matrix& b) { mul(a, b, 1.0, 0.0); }
Z
zhangjinchao01 已提交
635

636 637
void GpuMatrix::mul(const Matrix& a,
                    const Matrix& b,
638
                    real scaleAB,
Z
zhangjinchao01 已提交
639
                    real scaleT) {
640 641 642 643
  const auto a_ptr = dynamic_cast<const GpuMatrix*>(&a);
  const auto b_ptr = dynamic_cast<const GpuMatrix*>(&b);
  const auto a_ptr_s = dynamic_cast<const GpuSparseMatrix*>(&a);
  const auto b_ptr_s = dynamic_cast<const GpuSparseMatrix*>(&b);
Z
zhangjinchao01 已提交
644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678

  if (a_ptr && b_ptr) {
    mul(*a_ptr, *b_ptr, scaleAB, scaleT);
  } else if (a_ptr_s && b_ptr) {
    mul(*a_ptr_s, *b_ptr, scaleAB, scaleT);
  } else if (a_ptr && b_ptr_s) {
    mul(*a_ptr, *b_ptr_s, scaleAB, scaleT);
  } else {
    LOG(FATAL) << "Not supported";
  }
}

/* this = this* b */
void GpuMatrix::rightMul(Matrix& b) { rightMul(b, 1.0, 0.0); }

/* this = scaleAB*(this*b) +  scaleT*this */
void GpuMatrix::rightMul(Matrix& b, real scaleAB, real scaleT) {
  CHECK(dynamic_cast<GpuMatrix*>(&b));
  CHECK(!isTransposed()) << "Not supported";
  CHECK(!b.isTransposed()) << "Not supported";
  mul(*this, *dynamic_cast<GpuMatrix*>(&b), scaleAB, scaleT);
}

/* this = a*this */
void GpuMatrix::leftMul(Matrix& a) { leftMul(a, 1.0, 0.0); }

/* this = scaleAB*(a*this) +  scaleT*this */
void GpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) {
  CHECK(dynamic_cast<GpuMatrix*>(&a));
  CHECK(!isTransposed()) << "Not supported";
  CHECK(!a.isTransposed()) << "Not supported";
  mul(*dynamic_cast<GpuMatrix*>(&a), *this, scaleAB, scaleT);
}

void GpuMatrix::selectRows(Matrix& table, IVector& ids) {
679
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
680 681 682 683 684 685 686 687 688 689 690
  CHECK(dynamic_cast<GpuMatrix*>(&table));
  CHECK(table.useGpu());
  CHECK(ids.useGpu());
  CHECK_EQ(getHeight(), ids.getSize());
  CHECK_EQ(getWidth(), table.getWidth());
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  real* a = getData();
  size_t tableSize = table.getHeight();
  int* index = ids.getData();

691 692 693 694 695 696 697 698
  hl_matrix_select_rows(a,
                        stride_,
                        table.getData(),
                        table.stride_,
                        index,
                        numSamples,
                        tableSize,
                        dim);
Z
zhangjinchao01 已提交
699 700 701 702
#endif
}

void GpuMatrix::addToRows(Matrix& table, IVector& ids) {
703
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
704 705 706 707 708 709 710 711 712 713 714
  CHECK(dynamic_cast<GpuMatrix*>(&table));
  CHECK(table.useGpu());
  CHECK(ids.useGpu());
  CHECK_EQ(getHeight(), ids.getSize());
  CHECK_EQ(getWidth(), table.getWidth());
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  real* a = getData();
  size_t tableSize = table.getHeight();
  int* index = ids.getData();

715 716 717 718 719 720 721 722
  hl_matrix_add_to_rows(table.getData(),
                        table.stride_,
                        a,
                        stride_,
                        index,
                        numSamples,
                        tableSize,
                        dim);
Z
zhangjinchao01 已提交
723 724 725 726 727 728
#endif
}

void GpuMatrix::colMerge(Matrix& src) {
  CHECK(src.height_ == height_);
  if (!trans_ && !src.trans_) {
729
    sumRows(src, /* scaleSum= */ 1, /* scaleDest= */ 0);
Z
zhangjinchao01 已提交
730 731 732 733 734 735 736 737 738
  } else {
    LOG(FATAL) << "Is not supported";
  }
}

void GpuMatrix::rowSum(Matrix& sum) {
  CHECK_EQ(sum.getHeight(), getHeight());
  CHECK_EQ(sum.getWidth(), (size_t)1);

739
  sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0);
Z
zhangjinchao01 已提交
740 741 742 743 744 745 746 747 748 749
}

void GpuMatrix::rowMax(Matrix& max) {
  CHECK_EQ(max.getHeight(), getHeight());
  CHECK_EQ(max.getWidth(), (size_t)1);

  max.maxRows(*this);
}

void GpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
750
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
751 752 753 754 755
  CHECK(maxIds.useGpu() && maxVal.useGpu()) << "Matrix type are not equal";
  size_t numSamples = getHeight();
  size_t beam = maxVal.getWidth();
  CHECK_EQ(maxIds.getSize(), numSamples * beam);
  CHECK_EQ(maxVal.getHeight(), numSamples);
L
Liang Zhao 已提交
756
  CHECK_EQ(maxVal.getWidth(), beam);
Z
zhangjinchao01 已提交
757

758 759 760 761 762 763 764
  hl_matrix_top_k(maxVal.getData(),
                  maxVal.getStride(),
                  maxIds.getData(),
                  this->getData(),
                  this->getStride(),
                  this->getWidth(),
                  beam,
Z
zhangjinchao01 已提交
765 766 767 768 769 770 771 772 773 774 775
                  numSamples);
#endif
}

void GpuMatrix::colMax(Matrix& max) {
  CHECK_EQ(max.getWidth(), getWidth());
  CHECK_EQ(max.getHeight(), (size_t)1);

  max.maxCols(*this);
}

776 777 778 779
void GpuMatrix::colMax(IVector& maxIds, Matrix& maxVal) {
  LOG(FATAL) << "Is not supported";
}

780 781 782
void GpuMatrix::maxoutForward(Matrix& a,
                              IVector& id,
                              size_t channels,
783 784 785 786 787 788 789
                              size_t groups) {
  CHECK(dynamic_cast<GpuMatrix*>(&a));
  CHECK(dynamic_cast<GpuIVector*>(&id));
  CHECK_EQ(a.getHeight(), getHeight());

  size_t size = getWidth();
  size_t batchSize = getHeight();
Q
qijun 已提交
790
  const real* input = a.getData();
791 792 793
  real* output = getData();
  int* idForGpu = id.getData();

794 795
  hl_maxout_forward(
      input, output, idForGpu, batchSize, size, size / channels, groups);
796 797
}

798 799 800
void GpuMatrix::maxoutBackward(Matrix& a,
                               IVector& id,
                               size_t channels,
801 802 803 804 805 806 807
                               size_t groups) {
  CHECK(dynamic_cast<GpuMatrix*>(&a));
  CHECK(dynamic_cast<GpuIVector*>(&id));
  CHECK_EQ(a.getHeight(), getHeight());

  size_t size = a.getWidth();
  size_t batchSize = getHeight();
Q
qijun 已提交
808
  real* input = getData();
809 810 811
  const real* output = a.getData();
  const int* idForGpu = id.getData();

812 813
  hl_maxout_backward(
      input, output, idForGpu, batchSize, size, size / channels, groups);
814 815
}

Z
zhangjinchao01 已提交
816
/*calulate the error of classification */
817 818 819 820 821 822 823 824 825 826 827 828 829
void GpuMatrix::classificationError(Matrix& output,
                                    IVector& label,
                                    size_t topkSize) {
  auto gpuOutput = dynamic_cast<GpuMatrix*>(&output);
  auto gpuLabel = dynamic_cast<GpuIVector*>(&label);
  size_t numSamples = this->getHeight();
  GpuMatrixPtr gpuTopVal = std::make_shared<GpuMatrix>(numSamples, topkSize);
  GpuIVectorPtr gpuTopIds = std::make_shared<GpuIVector>(numSamples * topkSize);

  CHECK(gpuOutput && gpuLabel) << "Invalid argument pointer";
  CHECK(gpuTopVal && gpuTopIds) << "Allocate GPU memory failed";
  CHECK(gpuLabel->getSize() == numSamples) << "Vector size is not equal";
  CHECK(numSamples == gpuOutput->getHeight() && this->getWidth() == 1)
Z
zhangjinchao01 已提交
830 831
      << "Matrix dimensions are not equal";

832 833 834 835 836 837 838 839 840 841 842
  size_t dim = gpuOutput->getWidth();
  hl_matrix_classification_error(gpuTopVal->getData(),
                                 gpuTopVal->getStride(),
                                 gpuTopIds->getData(),
                                 gpuOutput->getData(),
                                 gpuOutput->getStride(),
                                 dim,
                                 topkSize,
                                 numSamples,
                                 gpuLabel->getData(),
                                 this->getData());
Z
zhangjinchao01 已提交
843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878
}

/* copy -log(output[i * width + label]) to this->data[i] */
void GpuMatrix::oneHotCrossEntropy(Matrix& output, IVector& label) {
  GpuMatrix* output_ptr = dynamic_cast<GpuMatrix*>(&output);
  GpuIVector* label_ptr = dynamic_cast<GpuIVector*>(&label);

  CHECK(output_ptr && label_ptr) << "Invalid argument pointer";

  CHECK(height_ == label.getSize() && width_ == 1 && height_ == output.height_)
      << "Matrix dimensions are not equal";

  real* A_d = output_ptr->data_;
  real* C_d = data_;
  int* label_d = label_ptr->getData();

  hl_matrix_cross_entropy(A_d, C_d, label_d, height_, output.width_);
}

/* calculate the error of outputV according to label */
void GpuMatrix::oneHotCrossEntropyBp(Matrix& outputV, IVector& label) {
  GpuMatrix* output_ptr = dynamic_cast<GpuMatrix*>(&outputV);
  GpuIVector* label_ptr = dynamic_cast<GpuIVector*>(&label);

  CHECK(output_ptr && label_ptr) << "Invalid argument pointer";

  CHECK(height_ == output_ptr->height_ && width_ == output_ptr->width_)
      << "Matrix dimensions are not equal";

  real* output_d = output_ptr->data_;
  real* grad_d = data_;
  int* label_d = label_ptr->getData();

  hl_matrix_cross_entropy_bp(grad_d, output_d, label_d, height_, width_);
}

879 880
void GpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output,
                                               IVector& label,
Z
zhangjinchao01 已提交
881 882 883 884 885
                                               real alpha) {
  LOG(FATAL) << "Not implemented";
}

void GpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& outputV,
886 887
                                                 IVector& label,
                                                 real alpha) {
Z
zhangjinchao01 已提交
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
  LOG(FATAL) << "Not implemented";
}

void GpuMatrix::softmax(Matrix& output) {
  CHECK(output.useGpu()) << "Matrix type are not equal";

  size_t height = getHeight();
  size_t width = getWidth();
  CHECK(height == output.getHeight() && width == output.getWidth())
      << "Matrix dimensions are not equal";

  real* inputData = getData();
  real* outputData = output.getData();
  hl_matrix_softmax(inputData, outputData, height, width);
}

void GpuMatrix::sequenceSoftmax(Matrix& output, const IVector& index) {
  CHECK_EQ(getWidth(), 1UL);
  CHECK_EQ(output.getWidth(), 1UL);
  CHECK(isContiguous());

  real* inputData = getData();
  real* outputData = output.getData();
  auto starts = index.getData();
  int numSequences = index.getSize() - 1;
Q
qijun 已提交
913
  hl_sequence_softmax_forward(inputData, outputData, starts, numSequences);
Z
zhangjinchao01 已提交
914 915 916 917 918 919 920 921 922 923 924 925 926
}

void GpuMatrix::softmaxDerivative(Matrix& output, Matrix& sftmaxSum) {
  CHECK(output.useGpu_ == true && sftmaxSum.useGpu_ == true)
      << "Matrix type are not equal";

  CHECK(height_ == output.height_ && width_ == output.width_ &&
        height_ == sftmaxSum.height_)
      << "Matrix dimensions are not equal";

  real* output_d = output.data_;
  real* sftmaxSum_d = sftmaxSum.data_;
  real* grad_d = data_;
Q
qijun 已提交
927
  hl_matrix_softmax_derivative(grad_d, output_d, sftmaxSum_d, height_, width_);
Z
zhangjinchao01 已提交
928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953
}

void GpuMatrix::softmaxBackward(Matrix& outputV) {
  CHECK(outputV.useGpu()) << "Matrix type are not equal";

  size_t height = getHeight();
  size_t width = getWidth();
  CHECK(height == outputV.getHeight() && width == outputV.getWidth())
      << "Matrix dimensions are not equal";

  real* output_grad = getData();
  real* output_value = outputV.getData();
  hl_softmax_backward(output_value, output_grad, height, width);
}

void GpuMatrix::sumOfSquares(Matrix& output, Matrix& label) {
  CHECK_EQ(label.getHeight(), height_);
  CHECK_EQ(output.getHeight(), height_);
  CHECK_EQ(label.getWidth(), output.getWidth());
  CHECK_EQ((size_t)1, width_);

  auto labelptr = dynamic_cast<GpuSparseMatrix*>(&label);
  if (labelptr) {
    LOG(FATAL) << "not supported: GpuSparseMatrix as label";
  }

954 955 956 957
  BaseMatrix::sumOfSquaredDiffs(output,
                                label,
                                /* scaleSum= */ 1,
                                /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
}

void GpuMatrix::sumOfSquaresBp(Matrix& outputV, Matrix& label) {
  add2(outputV, label, 1, 2, -2);
}

void GpuMatrix::tanh(Matrix& output) { BaseMatrix::tanh(output); }

void GpuMatrix::tanhDerivative(Matrix& output) {
  BaseMatrix::tanhDerivative(output);
}

void GpuMatrix::softrelu(Matrix& output) { BaseMatrix::softrelu(output); }

void GpuMatrix::softreluDerivative(Matrix& output) {
  BaseMatrix::softreluDerivative(output);
}

void GpuMatrix::scaledTanh(Matrix& output, real p1, real p2) {
  BaseMatrix::scaledTanh(output, p1, p2);
}

void GpuMatrix::randomizeUniform() {
  CHECK(isContiguous());
  real* data = data_;
  size_t size = height_ * width_;

  hl_rand(data, size);
}

void GpuMatrix::print(std::ostream& os) const {
  CHECK(isContiguous());
  CpuMatrix cpuMat(getHeight(), getWidth());
  cpuMat.copyFrom(*this);
  cpuMat.print(os);
}

void GpuMatrix::print(std::ostream& os, size_t height, size_t width) const {
  CHECK(isContiguous());
  CpuMatrix cpuMat(getHeight(), getWidth());
  cpuMat.copyFrom(*this);
  cpuMat.print(os, height, width);
}

void GpuMatrix::check(std::ostream& os, Matrix& refMat, bool printDiff) {
  CHECK(isContiguous());
  CHECK(height_ == refMat.getHeight());
  CHECK(width_ == refMat.getWidth());
  CpuMatrix cpuRef(height_, width_);
  GpuMatrix gpuRef(height_, width_);
  cpuRef.copyFrom(refMat);
  gpuRef.copyFrom(*this);
  size_t diffCnt = 0;
  for (size_t i = 0; i < height_; ++i) {
    for (size_t j = 0; j < width_; ++j) {
      real a = gpuRef.getElement(i, j);
      real b = cpuRef.getElement(i, j);
      if (fabs(a - b) > 0.00001) {
        ++diffCnt;
        if (printDiff) {
          os << "ref= " << a << "  check= " << b << std::endl;
        }
      }
    }
  }
  LOG(INFO) << "the  diffCnt is " << diffCnt;
}

X
xzl 已提交
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 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
void GpuMatrix::upsampleForward(Matrix& input,
                    Matrix& mask,
                    size_t imgSizeH,
                    size_t imgSizeW,
                    size_t channels,
                    size_t outputH,
                    size_t outputW) {
    CHECK(input.useGpu_ == true) << "Matrix type are not equal";
    CHECK(mask.useGpu_ == true) << "Matrix type are not equal";

    real *inputData = input.getData();
    real *maskData = mask.getData();
    real *outData = data_;

    size_t batch = input.getHeight();

    CHECK(imgSizeH * imgSizeW * channels == input.getWidth());
    CHECK(imgSizeH * imgSizeW * channels == mask.getWidth());
    CHECK_EQ(batch, this->getHeight());
    CHECK(width_ == outputH * outputW * channels);
    hl_upsample_forward(inputData, maskData,
                        batch,
                        imgSizeH,
                        imgSizeW,
                        channels,
                        outputH,
                        outputW,
                        outData);
}

void GpuMatrix::upsampleBackward(Matrix& outputGrad,
                    Matrix& mask,
                    size_t imgSizeH,
                    size_t imgSizeW,
                    size_t channels,
                    size_t outputH,
                    size_t outputW) {
    CHECK(outputGrad.useGpu_ == true) << "Matrix type are not equal";
    CHECK(mask.useGpu_ == true) << "Matrix type are not equal";

    real *outputGradData = outputGrad.getData();
    real *maskData = mask.getData();
    real *inputGradData = data_;
    size_t batch = outputGrad.getHeight();

    CHECK(imgSizeH * imgSizeW == this->getWidth()/channels);
    CHECK_EQ(batch, this->getHeight());
    CHECK_EQ(channels * outputH * outputW, outputGrad.getWidth());
    hl_upsample_backward(outputGradData, maskData,
                        batch,
                        imgSizeH,
                        imgSizeW,
                        channels,
                        outputH,
                        outputW,
                        inputGradData);
}

X
xzl 已提交
1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095
void GpuMatrix::maxPoolForward(Matrix& inputMat,
                               size_t imgSizeH,
                               size_t imgSizeW,
                               size_t channels,
                               size_t sizeX,
                               size_t sizeY,
                               size_t strideH,
                               size_t strideW,
                               size_t outputH,
                               size_t outputW,
                               size_t paddingH,
                               size_t paddingW,
1096
                               MatrixPtr maskMatP) {
Z
zhangjinchao01 已提交
1097 1098 1099
  CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal";

  real* inputData = inputMat.getData();
X
xzl 已提交
1100
  real* maskData = NULL;
Z
zhangjinchao01 已提交
1101
  size_t frameNum = inputMat.getHeight();
1102
  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
Z
zhangjinchao01 已提交
1103 1104 1105
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputH * outputW * channels);

1106
  if (maskMatP != NULL) {
X
xzl 已提交
1107 1108 1109 1110 1111
    CHECK(maskMatP->useGpu_ == true) << "Matrix type are not equal";
    CHECK(outputH * outputW * channels == maskMatP->getWidth());
    maskData = maskMatP->getData();
  }

1112 1113 1114
  hl_maxpool_forward(frameNum,
                     inputData,
                     channels,
1115 1116
                     imgSizeH,
                     imgSizeW,
1117 1118 1119 1120 1121 1122 1123 1124 1125
                     outputH,
                     outputW,
                     sizeX,
                     sizeY,
                     strideH,
                     strideW,
                     paddingH,
                     paddingW,
                     data_,
X
xzl 已提交
1126
                     getStride(),
1127
                     maskData);
1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144
}

void GpuMatrix::maxPoolBackward(Matrix& inputMat,
                                size_t imgSizeH,
                                size_t imgSizeW,
                                Matrix& outGrad,
                                Matrix& outV,
                                size_t sizeX,
                                size_t sizeY,
                                size_t strideH,
                                size_t strideW,
                                size_t outputH,
                                size_t outputW,
                                real scaleTargets,
                                real scaleOutput,
                                size_t paddingH,
                                size_t paddingW) {
Z
zhangjinchao01 已提交
1145 1146 1147 1148 1149 1150 1151 1152 1153
  CHECK(inputMat.useGpu_ == true && outGrad.useGpu_ == true &&
        outV.useGpu_ == true)
      << "Matrix type are not equal";

  real* inputData = inputMat.getData();
  real* outData = outV.getData();
  real* outDiff = outGrad.getData();
  size_t frameNum = inputMat.getHeight();
  size_t channels = outV.getWidth() / outputH / outputW;
1154
  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
Z
zhangjinchao01 已提交
1155 1156 1157 1158
  CHECK(height_ == inputMat.getHeight());
  CHECK(outGrad.getHeight() == outV.getHeight() &&
        outGrad.getWidth() == outV.getWidth());

1159 1160 1161 1162 1163
  hl_maxpool_backward(frameNum,
                      inputData,
                      outData,
                      outDiff,
                      channels,
1164 1165
                      imgSizeH,
                      imgSizeW,
1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176
                      outputH,
                      outputW,
                      sizeX,
                      sizeY,
                      strideH,
                      strideW,
                      paddingH,
                      paddingW,
                      scaleTargets,
                      scaleOutput,
                      data_,
Q
qijun 已提交
1177
                      outGrad.getStride());
Z
zhangjinchao01 已提交
1178 1179
}

1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190
void GpuMatrix::avgPoolForward(Matrix& inputMat,
                               size_t imgSizeH,
                               size_t imgSizeW,
                               size_t channels,
                               size_t sizeX,
                               size_t sizeY,
                               size_t strideH,
                               size_t strideW,
                               size_t outputH,
                               size_t outputW,
                               size_t paddingH,
Q
qijun 已提交
1191
                               size_t paddingW) {
Z
zhangjinchao01 已提交
1192 1193 1194 1195
  CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal";

  real* inputData = inputMat.getData();
  size_t frameNum = inputMat.getHeight();
1196
  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
Z
zhangjinchao01 已提交
1197 1198 1199
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputH * outputW * channels);

1200 1201 1202
  hl_avgpool_forward(frameNum,
                     inputData,
                     channels,
1203 1204
                     imgSizeH,
                     imgSizeW,
1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228
                     outputH,
                     outputW,
                     sizeX,
                     sizeY,
                     strideH,
                     strideW,
                     paddingH,
                     paddingW,
                     data_,
                     getStride());
}

void GpuMatrix::avgPoolBackward(Matrix& outGrad,
                                size_t imgSizeH,
                                size_t imgSizeW,
                                size_t sizeX,
                                size_t sizeY,
                                size_t strideH,
                                size_t strideW,
                                size_t outputH,
                                size_t outputW,
                                real scaleTargets,
                                real scaleOutput,
                                size_t paddingH,
Q
qijun 已提交
1229
                                size_t paddingW) {
Z
zhangjinchao01 已提交
1230 1231 1232 1233 1234
  CHECK(outGrad.useGpu_ == true) << "Matrix type are not equal";

  real* outDiff = outGrad.getData();
  size_t frameNum = outGrad.getHeight();
  size_t channels = outGrad.getWidth() / outputH / outputW;
1235
  CHECK(imgSizeH * imgSizeW * channels == width_);
Z
zhangjinchao01 已提交
1236 1237 1238
  CHECK(height_ == outGrad.getHeight());
  CHECK(outGrad.getWidth() == outputH * outputW * channels);

1239 1240 1241
  hl_avgpool_backward(frameNum,
                      outDiff,
                      channels,
1242 1243
                      imgSizeH,
                      imgSizeW,
1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254
                      outputH,
                      outputW,
                      sizeX,
                      sizeY,
                      strideH,
                      strideW,
                      paddingH,
                      paddingW,
                      scaleTargets,
                      scaleOutput,
                      data_,
Q
qijun 已提交
1255
                      outGrad.getStride());
Z
zhangjinchao01 已提交
1256 1257
}

C
chengduoZH 已提交
1258
void GpuMatrix::maxPool3DForward(Matrix& inputMat,
C
chengduoZH 已提交
1259
                                 Matrix& maxPoolIdx,
C
chengduoZH 已提交
1260
                                 size_t channels,
C
chengduoZH 已提交
1261 1262 1263
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
1264 1265 1266
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
1267 1268 1269 1270 1271 1272 1273 1274 1275
                                 size_t sizeZ,
                                 size_t sizeY,
                                 size_t sizeX,
                                 size_t strideD,
                                 size_t strideH,
                                 size_t strideW,
                                 size_t paddingD,
                                 size_t paddingH,
                                 size_t paddingW) {
C
chengduoZH 已提交
1276
  CHECK(inputMat.useGpu_) << "Matrix type are not correct";
C
chengduoZH 已提交
1277 1278

  real* inputData = inputMat.getData();
C
chengduoZH 已提交
1279
  real* maxPoolIdxData = maxPoolIdx.getData();
C
chengduoZH 已提交
1280
  size_t num = inputMat.getHeight();
1281
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
C
chengduoZH 已提交
1282 1283 1284 1285 1286 1287
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputD * outputH * outputW * channels);

  hl_maxpool3D_forward(num,
                       inputData,
                       channels,
1288 1289 1290
                       imgSizeD,
                       imgSizeH,
                       imgSizeW,
C
chengduoZH 已提交
1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302
                       outputD,
                       outputH,
                       outputW,
                       sizeZ,
                       sizeY,
                       sizeX,
                       strideD,
                       strideH,
                       strideW,
                       paddingD,
                       paddingH,
                       paddingW,
C
chengduoZH 已提交
1303
                       getData(),
C
chengduoZH 已提交
1304
                       maxPoolIdxData,
C
chengduoZH 已提交
1305 1306 1307
                       getStride());
}

C
chengduoZH 已提交
1308 1309
void GpuMatrix::maxPool3DBackward(Matrix& outGrad,
                                  Matrix& maxPoolIdx,
C
chengduoZH 已提交
1310 1311 1312
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
1313 1314 1315
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
1316 1317 1318 1319 1320 1321 1322 1323
                                  size_t sizeZ,
                                  size_t sizeY,
                                  size_t sizeX,
                                  size_t strideD,
                                  size_t strideH,
                                  size_t strideW,
                                  size_t paddingD,
                                  size_t paddingH,
C
chengduoZH 已提交
1324 1325 1326
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
C
chengduoZH 已提交
1327
  CHECK(outGrad.useGpu_ && maxPoolIdx.useGpu_) << "Matrix type are not equal";
C
chengduoZH 已提交
1328 1329

  real* outDiff = outGrad.getData();
C
chengduoZH 已提交
1330 1331 1332
  real* maxPoolIdxData = maxPoolIdx.getData();
  size_t frameNum = getHeight();
  size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
1333
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
C
chengduoZH 已提交
1334 1335
  CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() &&
        outGrad.getWidth() == maxPoolIdx.getWidth());
C
chengduoZH 已提交
1336 1337 1338 1339

  hl_maxpool3D_backward(frameNum,
                        outDiff,
                        channels,
1340 1341 1342
                        imgSizeD,
                        imgSizeH,
                        imgSizeW,
C
chengduoZH 已提交
1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356
                        outputD,
                        outputH,
                        outputW,
                        sizeZ,
                        sizeY,
                        sizeX,
                        strideD,
                        strideH,
                        strideW,
                        paddingD,
                        paddingH,
                        paddingW,
                        scaleTargets,
                        scaleOutput,
C
chengduoZH 已提交
1357
                        getData(),
C
chengduoZH 已提交
1358
                        maxPoolIdxData,
C
chengduoZH 已提交
1359 1360 1361 1362
                        outGrad.getStride());
}

void GpuMatrix::avgPool3DForward(Matrix& inputMat,
C
chengduoZH 已提交
1363
                                 size_t channels,
C
chengduoZH 已提交
1364 1365 1366
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
1367 1368 1369
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
1370 1371 1372 1373 1374 1375 1376 1377 1378
                                 size_t sizeZ,
                                 size_t sizeY,
                                 size_t sizeX,
                                 size_t strideD,
                                 size_t strideH,
                                 size_t strideW,
                                 size_t paddingD,
                                 size_t paddingH,
                                 size_t paddingW) {
C
chengduoZH 已提交
1379
  CHECK(inputMat.useGpu_) << "Matrix type are not equal";
C
chengduoZH 已提交
1380 1381 1382

  real* inputData = inputMat.getData();
  size_t frameNum = inputMat.getHeight();
1383
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
C
chengduoZH 已提交
1384 1385 1386 1387 1388 1389
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputD * outputH * outputW * channels);

  hl_avgpool3D_forward(frameNum,
                       inputData,
                       channels,
1390 1391 1392
                       imgSizeD,
                       imgSizeH,
                       imgSizeW,
C
chengduoZH 已提交
1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404
                       outputD,
                       outputH,
                       outputW,
                       sizeZ,
                       sizeY,
                       sizeX,
                       strideD,
                       strideH,
                       strideW,
                       paddingD,
                       paddingH,
                       paddingW,
C
chengduoZH 已提交
1405
                       getData(),
C
chengduoZH 已提交
1406 1407 1408 1409 1410 1411 1412
                       getStride());
}

void GpuMatrix::avgPool3DBackward(Matrix& outGrad,
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
1413 1414 1415
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
1416 1417 1418 1419 1420 1421 1422 1423
                                  size_t sizeZ,
                                  size_t sizeY,
                                  size_t sizeX,
                                  size_t strideD,
                                  size_t strideH,
                                  size_t strideW,
                                  size_t paddingD,
                                  size_t paddingH,
C
chengduoZH 已提交
1424 1425 1426 1427
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
  CHECK(outGrad.useGpu_) << "Matrix type are not equal";
C
chengduoZH 已提交
1428 1429 1430 1431

  real* outDiff = outGrad.getData();
  size_t frameNum = outGrad.getHeight();
  size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
1432
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_);
C
chengduoZH 已提交
1433 1434 1435 1436 1437 1438
  CHECK(height_ == outGrad.getHeight());
  CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels);

  hl_avgpool3D_backward(frameNum,
                        outDiff,
                        channels,
1439 1440 1441
                        imgSizeD,
                        imgSizeH,
                        imgSizeW,
C
chengduoZH 已提交
1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455
                        outputD,
                        outputH,
                        outputW,
                        sizeZ,
                        sizeY,
                        sizeX,
                        strideD,
                        strideH,
                        strideW,
                        paddingD,
                        paddingH,
                        paddingW,
                        scaleTargets,
                        scaleOutput,
C
chengduoZH 已提交
1456
                        getData(),
C
chengduoZH 已提交
1457 1458 1459
                        outGrad.getStride());
}

1460 1461
void GpuMatrix::maxSequenceForward(Matrix& input,
                                   const IVector& sequence,
Z
zhangjinchao01 已提交
1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477
                                   IVector& index) {
  CHECK(dynamic_cast<GpuMatrix*>(&input));
  CHECK(dynamic_cast<const GpuIVector*>(&sequence));
  CHECK(dynamic_cast<GpuIVector*>(&index));

  real* outData = getData();
  real* inputData = input.getData();
  const int* starts = sequence.getData();
  int* maxIndex = index.getData();
  size_t numSequences = getHeight();
  size_t dim = getWidth();

  CHECK_EQ(dim, input.getWidth());
  CHECK_EQ(numSequences, sequence.getSize() - 1);
  CHECK_EQ(numSequences * dim, index.getSize());

1478 1479
  hl_max_sequence_forward(
      inputData, starts, outData, maxIndex, numSequences, dim);
Z
zhangjinchao01 已提交
1480 1481
}

1482 1483
void GpuMatrix::maxSequenceBackward(Matrix& outputGrad,
                                    const IVector& sequence,
Z
zhangjinchao01 已提交
1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508
                                    IVector& index) {
  CHECK(dynamic_cast<GpuMatrix*>(&outputGrad));
  CHECK(dynamic_cast<const GpuIVector*>(&sequence));
  CHECK(dynamic_cast<GpuIVector*>(&index));

  real* inputGrad = getData();
  real* outGrad = outputGrad.getData();
  int* maxIndex = index.getData();
  size_t dim = getWidth();
  size_t numSequences = sequence.getSize() - 1;

  CHECK_EQ(dim, outputGrad.getWidth());
  CHECK_EQ(numSequences, outputGrad.getHeight());
  CHECK_EQ(numSequences * dim, index.getSize());

  hl_max_sequence_backward(outGrad, maxIndex, inputGrad, numSequences, dim);
}

void GpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
  CHECK(data.useGpu_ == true && W.useGpu_ == true)
      << "Matrix type are not equal";
  real* input = data.getData();
  real* w = W.getData();
  size_t numElements = data.getWidth();
  size_t numSamples = data.getHeight();
H
hedaoyuan 已提交
1509 1510 1511
  size_t paraSize = W.getHeight() * W.getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
Z
zhangjinchao01 已提交
1512
  real* output = getData();
Q
qijun 已提交
1513
  hl_param_relu_forward(output, input, w, numElements, numSamples, partial_sum);
Z
zhangjinchao01 已提交
1514 1515 1516 1517 1518 1519 1520 1521 1522 1523
}

void GpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) {
  CHECK(oGrad.useGpu_ == true && data.useGpu_ == true)
      << "Matrix type are not equal";
  real* ograd = oGrad.getData();
  real* input = data.getData();
  real* wgrad = data_;
  size_t numElements = data.getWidth();
  size_t numSamples = data.getHeight();
H
hedaoyuan 已提交
1524 1525 1526
  size_t paraSize = this->getHeight() * this->getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
1527 1528
  hl_param_relu_backward_w(
      wgrad, ograd, input, numElements, numSamples, partial_sum);
Z
zhangjinchao01 已提交
1529 1530 1531 1532 1533 1534 1535 1536 1537
}

void GpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) {
  real* diff = data_;
  real* input = data.getData();
  real* ograd = oGrad.getData();
  real* w = W.getData();
  size_t numElements = data.getWidth();
  size_t numSamples = data.getHeight();
H
hedaoyuan 已提交
1538 1539 1540
  size_t paraSize = W.getHeight() * W.getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
1541 1542
  hl_param_relu_backward_diff(
      ograd, input, w, diff, numElements, numSamples, partial_sum);
Z
zhangjinchao01 已提交
1543 1544 1545 1546 1547 1548
}

void GpuMatrix::addColumnVector(const Matrix& b) {
  BaseMatrix::addColVector(const_cast<Matrix&>(b));
}

L
liaogang 已提交
1549 1550 1551 1552 1553
void GpuMatrix::bilinearForward(const Matrix& in,
                                const size_t inImgH,
                                const size_t inImgW,
                                const size_t outImgH,
                                const size_t outImgW,
L
liaogang 已提交
1554 1555 1556
                                const size_t numChannels,
                                const real ratioH,
                                const real ratioW) {
L
liaogang 已提交
1557 1558 1559 1560 1561 1562 1563 1564
  CHECK(dynamic_cast<const GpuMatrix*>(&in));

  const size_t outputW = getWidth();
  const size_t outputH = getHeight();
  const size_t inputW = in.getWidth();
  const size_t inputH = in.getHeight();

  real* outData = getData();
1565
  const real* inData = in.getData();
L
liaogang 已提交
1566 1567 1568 1569

  if (inImgH == outImgW && inImgW == outImgW) {
    this->copyFrom(in);
  } else {
1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582
    hl_bilinear_forward(inData,
                        inImgH,
                        inImgW,
                        inputH,
                        inputW,
                        outData,
                        outImgH,
                        outImgW,
                        outputH,
                        outputW,
                        numChannels,
                        ratioH,
                        ratioW);
L
liaogang 已提交
1583 1584 1585 1586 1587 1588 1589 1590
  }
}

void GpuMatrix::bilinearBackward(const Matrix& out,
                                 const size_t outImgH,
                                 const size_t outImgW,
                                 const size_t inImgH,
                                 const size_t inImgW,
L
liaogang 已提交
1591 1592 1593
                                 const size_t numChannels,
                                 const real ratioH,
                                 const real ratioW) {
L
liaogang 已提交
1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604
  CHECK(dynamic_cast<const GpuMatrix*>(&out));

  const size_t inputW = getWidth();
  const size_t inputH = getHeight();
  const size_t outputW = out.getWidth();
  const size_t outputH = out.getHeight();

  real* inGrad = getData();
  const real* outGrad = out.getData();

  if (outImgH == inImgH && outImgW == inImgW) {
L
liaogang 已提交
1605
    this->add(const_cast<Matrix&>(out));
L
liaogang 已提交
1606
  } else {
1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619
    hl_bilinear_backward(inGrad,
                         inImgH,
                         inImgW,
                         inputH,
                         inputW,
                         outGrad,
                         outImgH,
                         outImgW,
                         outputH,
                         outputW,
                         numChannels,
                         ratioH,
                         ratioW);
L
liaogang 已提交
1620 1621 1622
  }
}

1623
void GpuMatrix::multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) {
1624
#ifdef PADDLE_WITH_CUDA
1625 1626 1627 1628 1629 1630 1631 1632 1633
  GpuMatrix* outputPtr = dynamic_cast<GpuMatrix*>(&output);
  auto labelPtr = dynamic_cast<GpuSparseMatrix*>(&label);

  CHECK(outputPtr && labelPtr) << "Invalid argument pointer";
  CHECK(labelPtr->format_ == SPARSE_CSR) << "Matrix format not supported";
  CHECK(height_ == outputPtr->height_ && width_ == 1 &&
        outputPtr->width_ == labelPtr->getWidth() &&
        outputPtr->height_ == labelPtr->getHeight())
      << "Matrix dimensions are not equal";
1634

1635 1636 1637 1638 1639
  real* output_d = outputPtr->data_;
  real* entropy_d = data_;
  hl_sparse_matrix_s mat_d = labelPtr->sMatrix_.get();
  hl_matrix_multi_binary_cross_entropy(
      output_d, entropy_d, mat_d, height_, outputPtr->width_);
1640
#endif
1641 1642
}

1643
void GpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) {
1644
#ifdef PADDLE_WITH_CUDA
1645 1646
  GpuMatrix* outputPtr = dynamic_cast<GpuMatrix*>(&output);
  auto labelPtr = dynamic_cast<GpuSparseMatrix*>(&label);
H
Haonan 已提交
1647

1648 1649 1650 1651 1652 1653
  CHECK(outputPtr && labelPtr) << "Invalid argument pointer";
  CHECK(labelPtr->format_ == SPARSE_CSR) << "Matrix format not supported";
  CHECK(height_ == outputPtr->height_ && width_ == outputPtr->width_ &&
        outputPtr->width_ == labelPtr->getWidth() &&
        outputPtr->height_ == labelPtr->getHeight())
      << "Matrix dimensions are not equal";
1654

1655 1656 1657 1658 1659
  real* output_d = outputPtr->data_;
  real* grad_d = data_;
  hl_sparse_matrix_s mat_d = labelPtr->sMatrix_.get();
  hl_matrix_multi_binary_cross_entropy_bp(
      output_d, grad_d, mat_d, height_, width_);
1660
#endif
1661 1662
}

C
chengduoZH 已提交
1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723
void GpuMatrix::vol2Col(real* dataSrc,
                        int channels,
                        int depth,
                        int height,
                        int width,
                        int filterD,
                        int filterH,
                        int filterW,
                        int strideD,
                        int strideH,
                        int strideW,
                        int paddingD,
                        int paddingH,
                        int paddingW) {
  hl_matrix_vol2Col(dataSrc,
                    channels,
                    depth,
                    height,
                    width,
                    filterD,
                    filterH,
                    filterW,
                    strideD,
                    strideH,
                    strideW,
                    paddingD,
                    paddingH,
                    paddingW,
                    getData());
}

void GpuMatrix::col2Vol(real* dataDst,
                        int channels,
                        int depth,
                        int height,
                        int width,
                        int filterD,
                        int filterH,
                        int filterW,
                        int strideD,
                        int strideH,
                        int strideW,
                        int paddingD,
                        int paddingH,
                        int paddingW,
                        real alpha,
                        real beta) {
  hl_matrix_col2Vol(dataDst,
                    channels,
                    depth,
                    height,
                    width,
                    filterD,
                    filterH,
                    filterW,
                    strideD,
                    strideH,
                    strideW,
                    paddingD,
                    paddingH,
                    paddingW,
C
chengduoZH 已提交
1724
                    getData(),
C
chengduoZH 已提交
1725 1726 1727
                    alpha,
                    beta);
}
C
chengduoZH 已提交
1728

Z
zhangjinchao01 已提交
1729 1730 1731 1732 1733 1734
/**
 * CpuMatrix
 */

CpuMatrix::CpuMatrix(size_t height, size_t width, bool trans)
    : Matrix(std::make_shared<CpuMemoryHandle>(height * width * sizeof(real)),
1735 1736 1737 1738
             height,
             width,
             trans,
             false) {}
Z
zhangjinchao01 已提交
1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759

CpuMatrix::~CpuMatrix() {}

void CpuMatrix::zeroMem() {
  CHECK(data_ != NULL);
  if (isContiguous()) {
    memset(data_, 0, height_ * width_ * sizeof(real));
  } else {
    BaseMatrix::zero();
  }
}
void CpuMatrix::resetOne() {
  CHECK(data_ != NULL);
  BaseMatrix::one();
}

void CpuMatrix::copyFrom(const Matrix& src) {
  CHECK(isContiguous());
  if (typeid(src) == typeid(GpuMatrix)) {
    CHECK(src.isContiguous());
    CHECK(elementCnt_ == src.getElementCnt());
1760 1761
    hl_memcpy_device2host(
        data_, const_cast<real*>(src.getData()), sizeof(real) * elementCnt_);
1762 1763
  } else if (typeid(src) == typeid(CpuMatrix) ||
             typeid(src) == typeid(SharedCpuMatrix)) {
Z
zhangjinchao01 已提交
1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825
    CHECK(src.isContiguous());
    CHECK(elementCnt_ == src.getElementCnt());
    memcpy(data_, src.getData(), sizeof(real) * elementCnt_);
  } else if (typeid(src) == typeid(CpuSparseMatrix)) {
    CHECK_GE(elementCnt_, src.getElementCnt());
    copyFrom(dynamic_cast<CpuSparseMatrix&>(const_cast<Matrix&>(src)));
  } else {
    LOG(FATAL) << "Wrong";
  }
}

void CpuMatrix::copyFrom(CpuSparseMatrix& src) {
  CHECK(isContiguous());
  CHECK(height_ == src.getHeight());
  CHECK(width_ == src.getWidth());
  memset(data_, 0, sizeof(real) * height_ * width_);
  if (src.getValueType() == FLOAT_VALUE) {
    if (src.getFormat() == SPARSE_CSC) {
      int* rows = src.getRows();
      real* vals = src.getValue();
      for (size_t i = 0; i < width_; i++) {
        for (size_t j = src.getColStartIdx(i); j < src.getColStartIdx(i + 1);
             j++) {
          data_[rows[j] * width_ + i] = vals[j];
        }
      }
    } else {
      int* cols = src.getCols();
      real* vals = src.getValue();
      for (size_t i = 0; i < height_; i++) {
        for (size_t j = src.getRowStartIdx(i); j < src.getRowStartIdx(i + 1);
             j++) {
          data_[i * width_ + cols[j]] = vals[j];
        }
      }
    }
  } else {
    if (src.getFormat() == SPARSE_CSC) {
      int* rows = src.getRows();
      for (size_t i = 0; i < width_; i++) {
        for (size_t j = src.getColStartIdx(i); j < src.getColStartIdx(i + 1);
             j++) {
          data_[rows[j] * width_ + i] = 1.0;
        }
      }
    } else {
      int* cols = src.getCols();
      for (size_t i = 0; i < height_; i++) {
        for (size_t j = src.getRowStartIdx(i); j < src.getRowStartIdx(i + 1);
             j++) {
          data_[i * width_ + cols[j]] = 1.0;
        }
      }
    }
  }
}

void CpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) {
  CHECK(isContiguous());
  CHECK(src.isContiguous());
  CHECK(elementCnt_ == src.getElementCnt());
  if (typeid(src) == typeid(GpuMatrix)) {
1826 1827 1828 1829
    hl_memcpy_async(this->getData(),
                    const_cast<real*>(src.getData()),
                    sizeof(real) * elementCnt_,
                    stream);
1830 1831
    // There is a need to add synchronization to ensure that the data is copied.
    hl_stream_synchronize(stream);
Z
zhangjinchao01 已提交
1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869
  } else if (typeid(src) == typeid(CpuMatrix)) {
    memcpy(data_, src.getData(), sizeof(real) * elementCnt_);
  } else {
    LOG(FATAL) << "Wrong";
  }
}

void CpuMatrix::copyFrom(const real* cpuSrc, size_t size) {
  CHECK(isContiguous());
  CHECK(size <= elementCnt_);
  memcpy(data_, cpuSrc, sizeof(real) * size);
}

void CpuMatrix::copyFrom(const real* cpuSrc, const int64_t* seq) {
  CHECK(isContiguous());
  for (size_t i = 0; i < height_; i++) {
    memcpy(data_ + i * width_, cpuSrc + seq[i] * width_, sizeof(real) * width_);
  }
}

void CpuMatrix::copyFrom(const IVector& src) {
  CHECK(isContiguous());
  CHECK(elementCnt_ == src.getSize())
      << "the src and dst should have same size.";
  const int* cpuSrc = NULL;
  IVectorPtr tmp;
  if (src.useGpu()) {
    CpuIVector tmp(src.getSize());
    tmp.copyFrom(src);
    cpuSrc = tmp.getData();
  } else {
    cpuSrc = src.getData();
  }
  for (size_t i = 0; i < elementCnt_; ++i) {
    data_[i] = cpuSrc[i];
  }
}

1870
void CpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) {
Z
zhangjinchao01 已提交
1871 1872 1873
  size_t height = getHeight();
  size_t width = getWidth();
  CHECK_EQ(b.getWidth(), width);
1874
  const int* index = rowIndex.getData();
Z
zhangjinchao01 已提交
1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932
  for (size_t i = 0; i < height; i++) {
    CHECK_LT(static_cast<size_t>(index[i]), b.getHeight());
    real* src = b.getData() + index[i] * width;
    real* dst = getData() + i * width;
    memcpy(dst, src, sizeof(real) * width);
  }
}

MatrixPtr CpuMatrix::clone(size_t height, size_t width, bool useGpu) {
  CHECK(isContiguous());

  if (height == 0 && width == 0) {
    height = height_;
    width = width_;
  }

  CHECK(width && height);

  if (useGpu) {
    return std::make_shared<GpuMatrix>(height, width);
  } else {
    return std::make_shared<CpuMatrix>(height, width);
  }
}

void CpuMatrix::resize(size_t newHeight, size_t newWidth) {
  size_t newSize = newHeight * newWidth;
  if (NULL == memoryHandle_.get() ||
      newSize * sizeof(real) > memoryHandle_->getAllocSize()) {
    memoryHandle_ = std::make_shared<CpuMemoryHandle>(newSize * sizeof(real));
    data_ = reinterpret_cast<real*>(memoryHandle_->getBuf());
  }

  height_ = newHeight;
  width_ = newWidth;
  elementCnt_ = newSize;
  stride_ = width_;
}

real CpuMatrix::getElement(size_t x, size_t y) const {
  return data_[x * stride_ + y];
}

real CpuMatrix::getSum() {
  CHECK(isContiguous());
  double sum = 0;
  for (size_t i = 0; i < height_; ++i) {
    for (size_t j = 0; j < width_; ++j) {
      sum += data_[i * width_ + j];
    }
  }
  return sum;
}

void CpuMatrix::accumulateColSum(Matrix& src) {
  CHECK_EQ(getWidth(), src.getWidth());
  CHECK_EQ(getHeight(), (size_t)1);

1933
  sumCols(src, /* scaleSum= */ 1, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949
}

real CpuMatrix::getAbsSum() {
  CHECK(isContiguous());
  double sum = 0;
  for (size_t i = 0; i < height_; ++i) {
    for (size_t j = 0; j < width_; ++j) {
      sum += fabs(data_[i * width_ + j]);
    }
  }
  return sum;
}

MatrixPtr CpuMatrix::getTranspose() {
  if (memoryHandle_.get() != NULL) {
    return std::make_shared<CpuMatrix>(
1950 1951 1952 1953
        std::dynamic_pointer_cast<CpuMemoryHandle>(memoryHandle_),
        height_,
        width_,
        true);
Z
zhangjinchao01 已提交
1954 1955 1956 1957 1958 1959
  } else {
    MatrixPtr copy_T(new CpuMatrix(data_, height_, width_, true));
    return copy_T;
  }
}

1960
void CpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
Z
zhangjinchao01 已提交
1961 1962 1963 1964
  if (memAlloc) {
    matTrans = std::make_shared<CpuMatrix>(width_, height_);
  } else {
    CHECK(matTrans != NULL);
H
Haonan 已提交
1965 1966
    CHECK_EQ(matTrans->getHeight(), width_);
    CHECK_EQ(matTrans->getWidth(), height_);
Z
zhangjinchao01 已提交
1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979
  }
  real* dataTrans = matTrans->getData();
  real* data = getData();
  int lda = getStride();
  int ldc = matTrans->getStride();

  for (size_t i = 0; i < height_; i++) {
    for (size_t j = 0; j < width_; j++) {
      dataTrans[j * ldc + i] = data[i * lda + j];
    }
  }
}

1980 1981 1982 1983 1984
void CpuMatrix::rotate(MatrixPtr& matRot, bool memAlloc, bool clockWise) {
  if (memAlloc) {
    matRot = std::make_shared<CpuMatrix>(width_, height_);
  } else {
    CHECK(matRot != NULL);
H
Haonan 已提交
1985 1986
    CHECK_EQ(matRot->getHeight(), width_);
    CHECK_EQ(matRot->getWidth(), height_);
1987 1988 1989 1990 1991 1992 1993
  }
  real* dataRot = matRot->getData();
  real* data = getData();

  for (size_t i = 0; i < height_; i++) {
    for (size_t j = 0; j < width_; j++) {
      if (clockWise) {
H
Haonan 已提交
1994
        dataRot[j * height_ + i] = data[(height_ - i - 1) * width_ + j];
1995
      } else {
H
Haonan 已提交
1996
        dataRot[j * height_ + i] = data[i * width_ + (width_ - j - 1)];
1997 1998 1999 2000 2001
      }
    }
  }
}

L
lzhao4ever 已提交
2002 2003 2004 2005 2006 2007
MatrixPtr CpuMatrix::getInverse() {
  MatrixPtr matInv;
  inverse(matInv, true);
  return matInv;
}

2008
void CpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) {
L
lzhao4ever 已提交
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041
  CHECK_EQ(height_, width_);

  if (memAlloc) {
    matInv = std::make_shared<CpuMatrix>(height_, width_);
  } else {
    CHECK(matInv != NULL);
  }

  CHECK_EQ(height_, matInv->getHeight());
  CHECK_EQ(width_, matInv->getWidth());
  matInv->copyFrom(*this);

  real* data = getData();
  real* dataInv = matInv->getData();
  int ldc = matInv->getStride();

  if (height_ == 1) {
    CHECK_NE(*data, 0);
    *dataInv = 1.0 / (*data);
    return;
  }

  /* Compute the LU decomposition of the matrix */
  std::vector<int> ipiv(height_);
  CBLAS_ORDER order = (matInv->isTransposed() ? CblasColMajor : CblasRowMajor);
  int info = getrf<real>(order, height_, height_, dataInv, ldc, ipiv.data());
  CHECK_EQ(info, 0);

  /* Compute the inverse of the matrix given its LU decompsotion */
  info = getri<real>(order, height_, dataInv, ldc, ipiv.data());
  CHECK_EQ(info, 0);
}

X
xzl 已提交
2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109
void CpuMatrix::upsampleForward(Matrix& input,
                    Matrix& mask,
                    size_t imgSizeH,
                    size_t imgSizeW,
                    size_t channels,
                    size_t outputH,
                    size_t outputW) {
    real *inputData = input.getData();
    real *maskData = mask.getData();
    real *outData = data_;
    size_t inLength = imgSizeH * imgSizeW;
    size_t outLength = outputH * outputW;
    size_t batch = input.getHeight();
    CHECK(inLength == input.getWidth() / channels);
    CHECK_EQ(batch, this->getHeight());
    CHECK_EQ(channels * outLength, this->getWidth());

    for (size_t k = 0; k < batch; k++) {
        for (size_t c = 0; c < channels; c++) {
            for (size_t i = 0; i < inLength; i++) {
                size_t out_index = static_cast<int>(maskData[i]);
                if (out_index >= outLength) {
                    LOG(FATAL) << "upsample index " << out_index
                        << " out of range.";
                }
                outData[out_index] = inputData[i];
            }
            inputData += inLength;
            maskData += inLength;
            outData += outLength;
        }
    }
}

void CpuMatrix::upsampleBackward(Matrix& outputGrad,
                    Matrix& mask,
                    size_t imgSizeH,
                    size_t imgSizeW,
                    size_t channels,
                    size_t outputH,
                    size_t outputW) {
    real *outputGradData = outputGrad.getData();
    real *maskData = mask.getData();
    real *inputGradData = data_;
    size_t inLength = imgSizeH * imgSizeW;
    size_t outLength = outputH * outputW;
    size_t batch = outputGrad.getHeight();
    CHECK(inLength == this->getWidth()/channels);
    CHECK_EQ(batch, this->getHeight());
    CHECK_EQ(channels * outLength, outputGrad.getWidth());

    for (size_t k = 0; k < batch; k++) {
        for (size_t c = 0; c < channels; c++) {
            for (size_t i = 0; i < inLength; i++) {
                size_t out_index = static_cast<int>(maskData[i]);
                if (out_index >= outLength) {
                    LOG(FATAL) << "upsample index " << out_index
                        << " out of range.";
                }
                inputGradData[i] = outputGradData[out_index];
            }
            inputGradData += inLength;
            maskData += inLength;
            outputGradData += outLength;
        }
    }
}

X
xzl 已提交
2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121
void CpuMatrix::maxPoolForward(Matrix& inputMat,
                               size_t imgSizeH,
                               size_t imgSizeW,
                               size_t channels,
                               size_t sizeX,
                               size_t sizeY,
                               size_t strideH,
                               size_t strideW,
                               size_t outputH,
                               size_t outputW,
                               size_t paddingH,
                               size_t paddingW,
2122
                               MatrixPtr maskMatP) {
Z
zhangjinchao01 已提交
2123 2124
  real* inputData = inputMat.getData();
  real* outData = data_;
X
xzl 已提交
2125
  real* maskData = NULL;
Z
zhangjinchao01 已提交
2126
  size_t num = inputMat.getHeight();
2127 2128 2129
  size_t inLength = imgSizeH * imgSizeW;
  size_t outLength = outputH * outputW;
  CHECK(inLength == inputMat.getWidth() / channels);
2130
  CHECK_EQ(num, this->getHeight());
2131
  CHECK_EQ(channels * outLength, this->getWidth());
Q
qijun 已提交
2132
  size_t outStride = getStride();
Z
zhangjinchao01 已提交
2133

2134
  if (maskMatP != NULL) {
X
xzl 已提交
2135 2136 2137 2138
    maskData = maskMatP->getData();
    CHECK_EQ(channels * outLength, maskMatP->getWidth());
  }

Z
zhangjinchao01 已提交
2139
  /* initialize the data_ */
Q
qijun 已提交
2140 2141
  for (size_t i = 0; i < height_; i++) {
    for (size_t j = 0; j < width_; j++) {
Q
qijun 已提交
2142
      outData[i * outStride + j] = -(real)FLT_MAX;
Q
qijun 已提交
2143
    }
Z
zhangjinchao01 已提交
2144 2145 2146
  }

  /* pool max one by one */
Q
qijun 已提交
2147 2148
  for (size_t n = 0; n < num; ++n) {  // frame by frame
    if (!isContiguous()) {
Q
qijun 已提交
2149
      outData = data_ + n * outStride;
Q
qijun 已提交
2150
    }
Z
zhangjinchao01 已提交
2151 2152
    for (size_t c = 0; c < channels; ++c) {  // channel by channel
      for (size_t ph = 0; ph < outputH; ++ph) {
2153 2154 2155
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2156
        for (size_t pw = 0; pw < outputW; ++pw) {
2157
          int wstart = pw * strideW - paddingW;
2158
          int wend = std::min(wstart + sizeX, imgSizeW);
2159
          wstart = std::max(wstart, 0);
X
xzl 已提交
2160
          if (maskData == NULL) {
X
xzl 已提交
2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174
            for (int h = hstart; h < hend; ++h) {
              for (int w = wstart; w < wend; ++w) {
                outData[ph * outputW + pw] = std::max(
                    outData[ph * outputW + pw], inputData[h * imgSizeW + w]);
              }
            }
          } else {
            for (int h = hstart; h < hend; ++h) {
              for (int w = wstart; w < wend; ++w) {
                if (outData[ph * outputW + pw] < inputData[h * imgSizeW + w]) {
                  outData[ph * outputW + pw] = inputData[h * imgSizeW + w];
                  maskData[ph * outputW + pw] = h * imgSizeW + w;
                }
              }
Z
zhangjinchao01 已提交
2175 2176 2177 2178 2179
            }
          }
        }
      }
      // compute offset
2180 2181
      inputData += inLength;
      outData += outLength;
X
xzl 已提交
2182

X
xzl 已提交
2183
      if (maskData != NULL) maskData += outLength;
Z
zhangjinchao01 已提交
2184 2185 2186 2187
    }
  }
}

2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202
void CpuMatrix::maxPoolBackward(Matrix& image,
                                size_t imgSizeH,
                                size_t imgSizeW,
                                Matrix& outGrad,
                                Matrix& outV,
                                size_t sizeX,
                                size_t sizeY,
                                size_t strideH,
                                size_t strideW,
                                size_t outputH,
                                size_t outputW,
                                real scaleTargets,
                                real scaleOutput,
                                size_t paddingH,
                                size_t paddingW) {
Z
zhangjinchao01 已提交
2203
  size_t num = image.getHeight();
2204 2205 2206 2207
  size_t inLength = imgSizeH * imgSizeW;
  size_t outLength = outputH * outputW;
  size_t channels = size_t(width_ / inLength);
  CHECK(image.getWidth() == inLength * channels);
Z
zhangjinchao01 已提交
2208 2209 2210 2211 2212 2213 2214 2215
  CHECK(image.getHeight() == height_ && image.getWidth() == width_);
  CHECK(outV.getHeight() == outGrad.getHeight() &&
        outV.getWidth() == outGrad.getWidth());

  real* tgtGrad = data_;
  real* inData = image.getData();
  real* otData = outV.getData();
  real* otGrad = outGrad.getData();
Q
qijun 已提交
2216 2217 2218 2219 2220

  size_t outStride = outV.getStride();
  real* origOutData = otData;
  real* origOutGrad = otGrad;

Z
zhangjinchao01 已提交
2221
  for (size_t n = 0; n < num; ++n) {
Q
qijun 已提交
2222
    if (!outV.isContiguous()) {
Q
qijun 已提交
2223 2224
      otData = origOutData + n * outStride;
      otGrad = origOutGrad + n * outStride;
Q
qijun 已提交
2225
    }
Z
zhangjinchao01 已提交
2226 2227
    for (size_t c = 0; c < channels; ++c) {
      for (size_t ph = 0; ph < outputH; ++ph) {
2228 2229 2230
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2231
        for (size_t pw = 0; pw < outputW; ++pw) {
2232 2233 2234 2235 2236
          int wstart = pw * strideW - paddingW;
          int wend = std::min(wstart + sizeX, imgSizeW);
          wstart = std::max(wstart, 0);
          for (int h = hstart; h < hend; ++h) {
            for (int w = wstart; w < wend; ++w) {
Z
zhangjinchao01 已提交
2237 2238 2239
              tgtGrad[h * imgSizeW + w] =
                  scaleTargets * tgtGrad[h * imgSizeW + w] +
                  scaleOutput * otGrad[ph * outputW + pw] *
2240
                      (inData[h * imgSizeW + w] == otData[ph * outputW + pw]);
Z
zhangjinchao01 已提交
2241 2242 2243 2244 2245
            }
          }
        }
      }
      // offset
2246 2247 2248 2249
      inData += inLength;
      tgtGrad += inLength;
      otData += outLength;
      otGrad += outLength;
Z
zhangjinchao01 已提交
2250 2251 2252 2253
    }
  }
}

2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264
void CpuMatrix::avgPoolForward(Matrix& input,
                               size_t imgSizeH,
                               size_t imgSizeW,
                               size_t channels,
                               size_t sizeX,
                               size_t sizeY,
                               size_t strideH,
                               size_t strideW,
                               size_t outputH,
                               size_t outputW,
                               size_t paddingH,
Q
qijun 已提交
2265
                               size_t paddingW) {
Z
zhangjinchao01 已提交
2266 2267
  // The main loop
  size_t num = input.getHeight();
2268 2269 2270 2271
  size_t inLength = imgSizeH * imgSizeW;
  size_t outLength = outputH * outputW;
  CHECK(inLength * channels == input.getWidth());
  CHECK(outLength * channels * num == height_ * width_);
Z
zhangjinchao01 已提交
2272 2273 2274 2275
  real* tgtData = data_;
  real* inData = input.getData();

  for (size_t n = 0; n < num; ++n) {
Q
qijun 已提交
2276 2277 2278
    if (!isContiguous()) {
      tgtData = data_ + n * getStride();
    }
Z
zhangjinchao01 已提交
2279 2280
    for (size_t c = 0; c < channels; ++c) {
      for (size_t ph = 0; ph < outputH; ++ph) {
2281 2282 2283
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2284
        for (size_t pw = 0; pw < outputW; ++pw) {
2285
          int wstart = pw * strideW - paddingW;
2286
          int wend = std::min(wstart + sizeX, imgSizeW);
2287
          wstart = std::max(wstart, 0);
Z
zhangjinchao01 已提交
2288
          tgtData[ph * outputW + pw] = 0;  // clear
2289 2290
          for (int h = hstart; h < hend; ++h) {
            for (int w = wstart; w < wend; ++w) {
2291
              tgtData[ph * outputW + pw] += inData[h * imgSizeW + w];
Z
zhangjinchao01 已提交
2292 2293
            }
          }
2294 2295
          int poolSize = (hend - hstart) * (wend - wstart);
          CHECK(poolSize);
2296
          tgtData[ph * outputW + pw] /= poolSize;
Z
zhangjinchao01 已提交
2297 2298 2299
        }
      }
      // compute offset
2300 2301
      inData += inLength;
      tgtData += outLength;
Z
zhangjinchao01 已提交
2302 2303 2304 2305
    }
  }
}

2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318
void CpuMatrix::avgPoolBackward(Matrix& input,
                                size_t imgSizeH,
                                size_t imgSizeW,
                                size_t sizeX,
                                size_t sizeY,
                                size_t strideH,
                                size_t strideW,
                                size_t outputH,
                                size_t outputW,
                                real scaleTargets,
                                real scaleOutput,
                                size_t paddingH,
                                size_t paddingW) {
Z
zhangjinchao01 已提交
2319 2320
  size_t num = input.getHeight();
  size_t channels = input.getWidth() / outputH / outputW;
2321 2322 2323
  size_t inLength = imgSizeH * imgSizeW;
  size_t outLength = outputH * outputW;
  CHECK(inLength * channels == getWidth());
Z
zhangjinchao01 已提交
2324 2325 2326 2327
  real* inData = input.getData();
  real* outData = getData();

  for (size_t n = 0; n < num; ++n) {
Q
qijun 已提交
2328 2329 2330
    if (!input.isContiguous()) {
      inData = input.getData() + n * input.getStride();
    }
Z
zhangjinchao01 已提交
2331 2332
    for (size_t c = 0; c < channels; ++c) {
      for (size_t ph = 0; ph < outputH; ++ph) {
2333 2334 2335
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2336
        for (size_t pw = 0; pw < outputW; ++pw) {
2337
          int wstart = pw * strideW - paddingW;
2338
          int wend = std::min(wstart + sizeX, imgSizeW);
2339
          wstart = std::max(wstart, 0);
2340
          int poolSize = (hend - hstart) * (wend - wstart);
2341 2342 2343 2344 2345
          CHECK(poolSize);

          for (int h = hstart; h < hend; ++h) {
            for (int w = wstart; w < wend; ++w) {
              outData[h * imgSizeW + w] += inData[ph * outputW + pw] / poolSize;
Z
zhangjinchao01 已提交
2346 2347 2348 2349 2350
            }
          }
        }
      }
      // offset
2351 2352
      outData += inLength;
      inData += outLength;
Z
zhangjinchao01 已提交
2353 2354 2355 2356
    }
  }
}

C
chengduoZH 已提交
2357
void CpuMatrix::maxPool3DForward(Matrix& inputMat,
C
chengduoZH 已提交
2358
                                 Matrix& maxPoolIdx,
C
chengduoZH 已提交
2359
                                 size_t channels,
C
chengduoZH 已提交
2360 2361 2362
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
2363 2364 2365
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
2366 2367 2368 2369 2370 2371 2372 2373 2374 2375
                                 size_t sizeZ,
                                 size_t sizeY,
                                 size_t sizeX,
                                 size_t strideD,
                                 size_t strideH,
                                 size_t strideW,
                                 size_t paddingD,
                                 size_t paddingH,
                                 size_t paddingW) {
  real* inputData = inputMat.getData();
C
chengduoZH 已提交
2376
  real* outData = getData();
C
chengduoZH 已提交
2377
  real* maxPoolIdxData = maxPoolIdx.getData();
C
chengduoZH 已提交
2378
  size_t num = inputMat.getHeight();
2379 2380 2381
  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
  size_t outLength = outputH * outputW * outputD;
  CHECK(inLength == inputMat.getWidth() / channels);
C
chengduoZH 已提交
2382
  CHECK_EQ(num, this->getHeight());
2383
  CHECK_EQ(channels * outLength, this->getWidth());
C
chengduoZH 已提交
2384 2385 2386 2387 2388 2389
  size_t outStride = getStride();

  /* initialize the data_ */
  for (size_t i = 0; i < height_; i++) {
    for (size_t j = 0; j < width_; j++) {
      outData[(i)*outStride + j] = -(real)FLT_MAX;
C
chengduoZH 已提交
2390
      maxPoolIdxData[(i)*outStride + j] = -1;
C
chengduoZH 已提交
2391 2392 2393 2394 2395 2396
    }
  }

  /* pool max one by one */
  for (size_t n = 0; n < num; ++n) {  // frame by frame
    if (!isContiguous()) {
C
chengduoZH 已提交
2397
      outData = getData() + n * outStride;
C
chengduoZH 已提交
2398
      maxPoolIdxData = maxPoolIdx.getData() + n * outStride;
C
chengduoZH 已提交
2399 2400 2401
    }
    for (size_t c = 0; c < channels; ++c) {  // channel by channel
      for (size_t pd = 0; pd < outputD; ++pd) {
2402 2403 2404
        int dstart = pd * strideD - paddingD;
        int dend = std::min(dstart + sizeZ, imgSizeD);
        dstart = std::max(dstart, 0);
C
chengduoZH 已提交
2405
        for (size_t ph = 0; ph < outputH; ++ph) {
2406 2407 2408
          int hstart = ph * strideH - paddingH;
          int hend = std::min(hstart + sizeY, imgSizeH);
          hstart = std::max(hstart, 0);
C
chengduoZH 已提交
2409 2410
          for (size_t pw = 0; pw < outputW; ++pw) {
            int wstart = pw * strideW - paddingW;
2411
            int wend = std::min(wstart + sizeX, imgSizeW);
C
chengduoZH 已提交
2412
            wstart = std::max(wstart, 0);
C
chengduoZH 已提交
2413
            int maxIdx = -1;
C
chengduoZH 已提交
2414
            real maxOutData = outData[(pd * outputH + ph) * outputW + pw];
C
chengduoZH 已提交
2415 2416 2417
            for (int d = dstart; d < dend; ++d) {
              for (int h = hstart; h < hend; ++h) {
                for (int w = wstart; w < wend; ++w) {
C
chengduoZH 已提交
2418
                  if (maxOutData <
2419 2420 2421
                      inputData[(d * imgSizeH + h) * imgSizeW + w]) {
                    maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w];
                    maxIdx = (d * imgSizeH + h) * imgSizeW + w;
C
chengduoZH 已提交
2422
                  }
C
chengduoZH 已提交
2423 2424 2425
                }
              }
            }
C
chengduoZH 已提交
2426
            outData[(pd * outputH + ph) * outputW + pw] = maxOutData;
C
chengduoZH 已提交
2427
            maxPoolIdxData[(pd * outputH + ph) * outputW + pw] = maxIdx;
C
chengduoZH 已提交
2428 2429 2430 2431
          }
        }
      }
      // compute offset
2432 2433 2434
      inputData += inLength;
      outData += outLength;
      maxPoolIdxData += outLength;
C
chengduoZH 已提交
2435 2436 2437 2438
    }
  }
}

C
chengduoZH 已提交
2439 2440
void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
                                  Matrix& maxPoolIdx,
C
chengduoZH 已提交
2441 2442 2443
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
2444 2445 2446
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
2447 2448 2449 2450 2451 2452 2453 2454
                                  size_t sizeZ,
                                  size_t sizeY,
                                  size_t sizeX,
                                  size_t strideD,
                                  size_t strideH,
                                  size_t strideW,
                                  size_t paddingD,
                                  size_t paddingH,
C
chengduoZH 已提交
2455 2456 2457
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
C
chengduoZH 已提交
2458
  size_t num = getHeight();
2459 2460 2461
  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
  size_t outLength = outputH * outputW * outputD;
  size_t channels = size_t(width_ / inLength);
C
chengduoZH 已提交
2462 2463
  CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() &&
        maxPoolIdx.getWidth() == outGrad.getWidth());
C
chengduoZH 已提交
2464

C
chengduoZH 已提交
2465
  real* tgtGrad = getData();
C
chengduoZH 已提交
2466
  real* otGrad = outGrad.getData();
C
chengduoZH 已提交
2467 2468
  real* maxPoolIdxData = maxPoolIdx.getData();
  size_t outStride = outGrad.getStride();
C
chengduoZH 已提交
2469 2470

  for (size_t n = 0; n < num; ++n) {
C
chengduoZH 已提交
2471
    if (!outGrad.isContiguous()) {
C
chengduoZH 已提交
2472
      otGrad = outGrad.getData() + n * outStride;
C
chengduoZH 已提交
2473
      maxPoolIdxData = maxPoolIdx.getData() + n * outStride;
C
chengduoZH 已提交
2474 2475 2476 2477 2478
    }
    for (size_t c = 0; c < channels; ++c) {
      for (size_t pd = 0; pd < outputD; ++pd) {
        for (size_t ph = 0; ph < outputH; ++ph) {
          for (size_t pw = 0; pw < outputW; ++pw) {
C
chengduoZH 已提交
2479 2480 2481 2482
            const size_t index = (pd * outputH + ph) * outputW + pw;
            const size_t tgtIdx = static_cast<size_t>(maxPoolIdxData[index]);
            tgtGrad[tgtIdx] =
                scaleTargets * tgtGrad[tgtIdx] + scaleOutput * otGrad[index];
C
chengduoZH 已提交
2483 2484 2485 2486
          }
        }
      }
      // offset
2487 2488 2489
      tgtGrad += inLength;
      otGrad += outLength;
      maxPoolIdxData += outLength;
C
chengduoZH 已提交
2490 2491 2492 2493 2494
    }
  }
}

void CpuMatrix::avgPool3DForward(Matrix& input,
C
chengduoZH 已提交
2495
                                 size_t channels,
C
chengduoZH 已提交
2496 2497 2498
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
2499 2500 2501
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512
                                 size_t sizeZ,
                                 size_t sizeY,
                                 size_t sizeX,
                                 size_t strideD,
                                 size_t strideH,
                                 size_t strideW,
                                 size_t paddingD,
                                 size_t paddingH,
                                 size_t paddingW) {
  // The main loop
  size_t num = input.getHeight();
2513 2514 2515 2516
  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
  size_t outLength = outputH * outputW * outputD;
  CHECK(inLength * channels == input.getWidth());
  CHECK(outLength * channels * num == height_ * width_);
C
chengduoZH 已提交
2517
  real* tgtData = getData();
C
chengduoZH 已提交
2518 2519 2520 2521 2522 2523 2524 2525
  real* inData = input.getData();

  for (size_t n = 0; n < num; ++n) {
    if (!isContiguous()) {
      tgtData = data_ + n * getStride();
    }
    for (size_t c = 0; c < channels; ++c) {
      for (size_t pd = 0; pd < outputD; ++pd) {
2526 2527 2528
        int dstart = pd * strideD - paddingD;
        int dend = std::min(dstart + sizeZ, imgSizeD);
        dstart = std::max(dstart, 0);
C
chengduoZH 已提交
2529
        for (size_t ph = 0; ph < outputH; ++ph) {
2530 2531 2532
          int hstart = ph * strideH - paddingH;
          int hend = std::min(hstart + sizeY, imgSizeH);
          hstart = std::max(hstart, 0);
C
chengduoZH 已提交
2533 2534
          for (size_t pw = 0; pw < outputW; ++pw) {
            int wstart = pw * strideW - paddingW;
2535
            int wend = std::min(wstart + sizeX, imgSizeW);
C
chengduoZH 已提交
2536 2537 2538 2539 2540 2541 2542
            wstart = std::max(wstart, 0);

            tgtData[(pd * outputH + ph) * outputW + pw] = 0;  // clear
            for (int d = dstart; d < dend; ++d) {
              for (int h = hstart; h < hend; ++h) {
                for (int w = wstart; w < wend; ++w) {
                  tgtData[(pd * outputH + ph) * outputW + pw] +=
2543
                      inData[(d * imgSizeH + h) * imgSizeW + w];
C
chengduoZH 已提交
2544 2545 2546
                }
              }
            }
2547 2548
            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
            CHECK(poolSize);
C
chengduoZH 已提交
2549 2550 2551 2552 2553
            tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize;
          }
        }
      }
      // compute offset
2554 2555
      inData += inLength;
      tgtData += outLength;
C
chengduoZH 已提交
2556 2557 2558 2559 2560 2561 2562 2563
    }
  }
}

void CpuMatrix::avgPool3DBackward(Matrix& input,
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
2564 2565 2566
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
2567 2568 2569 2570 2571 2572 2573 2574
                                  size_t sizeZ,
                                  size_t sizeY,
                                  size_t sizeX,
                                  size_t strideD,
                                  size_t strideH,
                                  size_t strideW,
                                  size_t paddingD,
                                  size_t paddingH,
C
chengduoZH 已提交
2575 2576 2577
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
C
chengduoZH 已提交
2578
  size_t num = input.getHeight();
2579 2580 2581 2582
  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
  size_t outLength = outputH * outputW * outputD;
  size_t channels = input.getWidth() / outLength;
  CHECK(inLength * channels == getWidth());
C
chengduoZH 已提交
2583 2584 2585 2586 2587 2588 2589 2590 2591
  real* inData = input.getData();
  real* outData = getData();

  for (size_t n = 0; n < num; ++n) {
    if (!input.isContiguous()) {
      inData = input.getData() + n * input.getStride();
    }
    for (size_t c = 0; c < channels; ++c) {
      for (size_t pd = 0; pd < outputD; ++pd) {
2592 2593 2594
        int dstart = pd * strideD - paddingD;
        int dend = std::min(dstart + sizeZ, imgSizeD);
        dstart = std::max(dstart, 0);
C
chengduoZH 已提交
2595
        for (size_t ph = 0; ph < outputH; ++ph) {
2596 2597 2598
          int hstart = ph * strideH - paddingH;
          int hend = std::min(hstart + sizeY, imgSizeH);
          hstart = std::max(hstart, 0);
C
chengduoZH 已提交
2599 2600
          for (size_t pw = 0; pw < outputW; ++pw) {
            int wstart = pw * strideW - paddingW;
2601
            int wend = std::min(wstart + sizeX, imgSizeW);
C
chengduoZH 已提交
2602
            wstart = std::max(wstart, 0);
2603
            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
C
chengduoZH 已提交
2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616
            CHECK(poolSize);
            for (int d = dstart; d < dend; ++d) {
              for (int h = hstart; h < hend; ++h) {
                for (int w = wstart; w < wend; ++w) {
                  outData[(d * imgSizeH + h) * imgSizeW + w] +=
                      inData[(pd * outputH + ph) * outputW + pw] / poolSize;
                }
              }
            }
          }
        }
      }
      // offset
2617 2618
      outData += inLength;
      inData += outLength;
C
chengduoZH 已提交
2619 2620 2621 2622
    }
  }
}

Z
zhangjinchao01 已提交
2623 2624 2625 2626 2627
/**
 * Input: one or more sequences. Each sequence contains some instances.
 * Output: output size is the number of input sequences (NOT input instances).
 * output[i] is set to max_{for each instance in this sequence}{input[i]}
 */
2628 2629
void CpuMatrix::maxSequenceForward(Matrix& input,
                                   const IVector& sequence,
Z
zhangjinchao01 已提交
2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669
                                   IVector& index) {
  CHECK(dynamic_cast<CpuMatrix*>(&input));
  CHECK(dynamic_cast<const CpuIVector*>(&sequence));
  CHECK(dynamic_cast<CpuIVector*>(&index));

  real* outData = getData();
  real* inputData = input.getData();
  const int* starts = sequence.getData();
  int* maxIndex = index.getData();
  size_t numSequences = getHeight();
  size_t dim = getWidth();

  CHECK_EQ(dim, input.getWidth());
  CHECK_EQ(numSequences, sequence.getSize() - 1);
  CHECK_EQ(starts[numSequences], (int)input.getHeight());
  CHECK_EQ(numSequences * dim, index.getSize());

  for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) {
    // current sequence, loop for each input instance
    // (1) first instance: do not need compare, copy value to outV directly
    for (size_t k = 0; k < dim; ++k) {
      outData[sequenceId * dim + k] = inputData[starts[sequenceId] * dim + k];
      maxIndex[sequenceId * dim + k] = starts[sequenceId];
    }
    // (2) other instance in same sequence
    for (int insId = starts[sequenceId] + 1; insId < starts[sequenceId + 1];
         ++insId) {
      // insId is the index on all instances
      for (size_t k = 0; k < dim; ++k) {
        // for each dim
        if (inputData[insId * dim + k] > outData[sequenceId * dim + k]) {
          // update max value and record index
          outData[sequenceId * dim + k] = inputData[insId * dim + k];
          maxIndex[sequenceId * dim + k] = insId;
        }
      }
    }
  }
}

2670 2671
void CpuMatrix::maxSequenceBackward(Matrix& outputGrad,
                                    const IVector& sequence,
Z
zhangjinchao01 已提交
2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708
                                    IVector& index) {
  CHECK(dynamic_cast<CpuMatrix*>(&outputGrad));
  CHECK(dynamic_cast<const CpuIVector*>(&sequence));
  CHECK(dynamic_cast<CpuIVector*>(&index));

  real* inputGrad = getData();
  real* outGrad = outputGrad.getData();
  int* maxIndex = index.getData();
  size_t dim = getWidth();
  size_t numSequences = sequence.getSize() - 1;

  CHECK_EQ(dim, outputGrad.getWidth());
  CHECK_EQ(numSequences, outputGrad.getHeight());
  CHECK_EQ(numSequences * dim, index.getSize());

  for (size_t sequenceId = 0; sequenceId < numSequences; ++sequenceId) {
    // current sequence
    for (size_t j = 0; j < dim; ++j) {
      // each dim
      int insId = maxIndex[sequenceId * dim + j];
      inputGrad[insId * dim + j] += outGrad[sequenceId * dim + j];
    }
  }
}

inline void vecAddTo(real* a, const real* b, size_t len) {
  for (unsigned int i = 0; i < len; ++i) {
    a[i] += b[i];
  }
}

inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) {
  for (unsigned int i = 0; i < len; ++i) {
    a[i] += scaleB * b[i];
  }
}

2709 2710
inline void colVecAddTo(
    real* a, const real* b, size_t len, size_t aWidth, size_t bWidth) {
Z
zhangjinchao01 已提交
2711 2712 2713 2714 2715
  for (unsigned int i = 0; i < len; ++i) {
    a[i * aWidth] += b[i * bWidth];
  }
}

2716 2717
inline void colVecAddTo(
    real* a, real* b, real c, size_t len, size_t aWidth, size_t bWidth) {
Z
zhangjinchao01 已提交
2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749
  for (unsigned int i = 0; i < len; ++i) {
    a[i * aWidth] += b[i * bWidth] * c;
  }
}

void CpuMatrix::addBias(Matrix& b, real scale) {
  CHECK(b.useGpu_ == false) << "Matrix type are not equal";

  CHECK_EQ(b.getHeight(), (size_t)1);
  CHECK_EQ(width_, b.getWidth());
  real* aData = getData();
  real* bData = b.getData();
  size_t numSamples = getHeight();
  size_t dim = getWidth();

  if (scale == 1 && getStride() % 32 == 0) {  // use libaddto
    // @TODO(yuyang18) Make input addr can be unaligned.
    // So merge this if and else
    CHECK_EQ((size_t)aData % 32, 0UL);
    CHECK_EQ((size_t)bData % 32, 0UL);
    for (size_t i = 0; i < numSamples; i++) {
      simd::addTo(aData + i * getStride(), bData, dim);
    }
  } else {
    for (size_t i = 0; i < numSamples; i++) {
      for (size_t j = 0; j < dim; j++) {
        aData[i * getStride() + j] += scale * bData[j];
      }
    }
  }
}

2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767
void CpuMatrix::addSharedBias(Matrix& b, real scale) {
  CHECK_EQ(b.getHeight(), (size_t)1);
  real* aData = getData();
  real* bData = b.getData();
  size_t numSamples = getHeight();
  size_t channel = b.getWidth();
  CHECK_EQ(getWidth() % channel, 0UL);
  size_t dim = getWidth() / channel;

  for (size_t i = 0; i < numSamples; i++) {
    for (size_t c = 0; c < channel; c++) {
      for (size_t j = 0; j < dim; j++) {
        aData[i * getStride() + c * dim + j] += scale * bData[c];
      }
    }
  }
}

Z
zhangjinchao01 已提交
2768 2769 2770 2771 2772
void CpuMatrix::collectBias(Matrix& a, real scale) {
  CHECK_EQ(getHeight(), (size_t)1);
  CHECK_EQ(width_, a.getWidth());
  CpuSparseMatrix* aptr = dynamic_cast<CpuSparseMatrix*>(&a);
  if (!aptr) {
2773
    sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784
  } else {
    size_t nnz = aptr->getElementCnt();
    int* cols = aptr->getCols();
    real* A = aptr->getValue();
    real* B = getData();
    for (size_t i = 0; i < nnz; i++) {
      B[cols[i]] += scale * A[i];
    }
  }
}

2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801
void CpuMatrix::collectSharedBias(Matrix& a, real scale) {
  CHECK_EQ(getHeight(), (size_t)1);
  real* B = getData();
  real* A = a.getData();
  size_t numSamples = a.getHeight();
  size_t channel = getWidth();
  CHECK_EQ(a.getWidth() % channel, 0UL);
  size_t dim = a.getWidth() / channel;
  for (size_t i = 0; i < numSamples; i++) {
    for (size_t c = 0; c < channel; c++) {
      for (size_t j = 0; j < dim; j++) {
        B[c] += scale * A[i * channel * dim + c * dim + j];
      }
    }
  }
}

Z
zhangjinchao01 已提交
2802 2803 2804 2805 2806 2807 2808 2809 2810 2811
void CpuMatrix::sequenceAvgForward(Matrix& a,
                                   const IVector& startsPos,
                                   int mode) {
  size_t height = getHeight();
  size_t width = getWidth();
  CHECK_EQ(height, startsPos.getSize() - 1);
  CHECK_EQ(width, a.getWidth());
  real* dst = getData();
  real* src = a.getData();
  const int* starts = startsPos.getData();
X
xuwei06 已提交
2812
  MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false);
Z
zhangjinchao01 已提交
2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823
  MatrixPtr dataMtx = Matrix::create(nullptr, 1, width, false, false);
  for (size_t i = 0; i < height; i++) {
    int sequenceLength = starts[i + 1] - starts[i];
    if (0 == sequenceLength) {
      // empty sequence
      continue;
    }
    outMtx->setData(dst + i * width);
    dataMtx->setData(src + starts[i] * width, sequenceLength, width);
    if (mode == 0) {
      // plain average
2824 2825 2826
      outMtx->sumCols(*dataMtx,
                      (real)1 / (real)sequenceLength,
                      /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2827 2828
    } else if (mode == 1) {
      // sum instead of average
2829
      outMtx->sumCols(*dataMtx, /* scaleSum= */ 1, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2830 2831
    } else if (mode == 2) {
      // divide by square root of sequenceLength
2832 2833 2834
      outMtx->sumCols(*dataMtx,
                      (real)1 / std::sqrt(sequenceLength),
                      /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2835 2836 2837 2838 2839 2840
    } else {
      LOG(FATAL) << "should not reach here";
    }
  }
}

L
Luo Tao 已提交
2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875
void CpuMatrix::sequenceAvgBackward(Matrix& a,
                                    const IVector& startsPos,
                                    int mode) {
  size_t height = a.getHeight();
  size_t width = getWidth();
  CHECK_EQ(height, startsPos.getSize() - 1);
  CHECK_EQ(width, a.getWidth());
  real* dst = getData();
  real* src = a.getData();
  const int* starts = startsPos.getData();
  MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false);
  MatrixPtr dataMtx = Matrix::create(nullptr, 1, width, false, false);
  for (size_t i = 0; i < height; ++i) {
    int sequenceLength = starts[i + 1] - starts[i];
    if (0 == sequenceLength) {
      // empty sequence
      continue;
    }
    outMtx->setData(dst + starts[i] * width, sequenceLength, width);
    dataMtx->setData(src + i * width);
    if (mode == 0) {
      // plain average
      outMtx->addBias(*dataMtx, 1.0f / sequenceLength);
    } else if (mode == 1) {
      // sum instead of average
      outMtx->addBias(*dataMtx, 1.0f);
    } else if (mode == 2) {
      // divide by square root of sequenceLength
      outMtx->addBias(*dataMtx, 1.0f / std::sqrt(sequenceLength));
    } else {
      LOG(FATAL) << "should not reach here";
    }
  }
}

Z
zhangjinchao01 已提交
2876
/* this = scaleAB*(a*b) + scaleT*this*/
2877 2878
void CpuMatrix::mul(const Matrix& a,
                    const Matrix& b,
2879
                    real scaleAB,
Z
zhangjinchao01 已提交
2880 2881
                    real scaleT) {
  CHECK(!isTransposed()) << "Not supported";
2882 2883 2884 2885
  const auto a_ptr = dynamic_cast<const CpuMatrix*>(&a);
  const auto b_ptr = dynamic_cast<const CpuMatrix*>(&b);
  const auto a_ptr_s = dynamic_cast<const CpuSparseMatrix*>(&a);
  const auto b_ptr_s = dynamic_cast<const CpuSparseMatrix*>(&b);
Z
zhangjinchao01 已提交
2886

2887 2888 2889 2890 2891 2892
  if (a_ptr && b_ptr) {
    mul((CpuMatrix*)a_ptr, (CpuMatrix*)b_ptr, scaleAB, scaleT);
  } else if (a_ptr_s && b_ptr) {
    mul((CpuSparseMatrix*)a_ptr_s, (CpuMatrix*)b_ptr, scaleAB, scaleT);
  } else if (a_ptr && b_ptr_s) {
    mul((CpuMatrix*)a_ptr, (CpuSparseMatrix*)b_ptr_s, scaleAB, scaleT);
Z
zhangjinchao01 已提交
2893 2894 2895 2896 2897
  } else {
    LOG(FATAL) << "Not supported";
  }
}

2898 2899 2900
void CpuMatrix::mul(CpuSparseMatrix* a,
                    CpuMatrix* b,
                    real scaleAB,
Z
zhangjinchao01 已提交
2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914
                    real scaleT) {
  if (dynamic_cast<CacheRowCpuMatrix*>(b)) {
    return mul(a, dynamic_cast<CacheRowCpuMatrix*>(b), this, scaleAB, scaleT);
  } else if (dynamic_cast<SparseRowCpuMatrix*>(b)) {
    return mul(a, dynamic_cast<SparseRowCpuMatrix*>(b), this, scaleAB, scaleT);
  } else {
    return mul(a, b, this, scaleAB, scaleT);
  }
}

void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
  CHECK(!isTransposed()) << "Not supported";

  size_t a_col, b_col, a_row, b_row;
2915
  bool a_trans, b_trans;
Z
zhangjinchao01 已提交
2916 2917 2918
  if (!a->isTransposed()) {
    a_col = a->getWidth();
    a_row = a->getHeight();
2919
    a_trans = false;
Z
zhangjinchao01 已提交
2920 2921 2922
  } else {
    a_col = a->getHeight();
    a_row = a->getWidth();
2923
    a_trans = true;
Z
zhangjinchao01 已提交
2924 2925 2926 2927
  }
  if (!b->isTransposed()) {
    b_col = b->getWidth();
    b_row = b->getHeight();
2928
    b_trans = false;
Z
zhangjinchao01 已提交
2929 2930 2931
  } else {
    b_col = b->getHeight();
    b_row = b->getWidth();
2932
    b_trans = true;
Z
zhangjinchao01 已提交
2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948
  }

  CHECK_EQ(a_col, b_row);
  CHECK_EQ(a_row, getHeight());
  CHECK_EQ(b_col, getWidth());

  real* A = a->getData();
  real* B = b->getData();
  real* C = getData();

  int M = getHeight();
  int N = getWidth();
  int K = a_col;
  int lda = a->getStride();
  int ldb = b->getStride();
  int ldc = getStride();
2949
  BlasGemm<DEVICE_TYPE_CPU, real>::compute(
L
Liu Yiqun 已提交
2950
      a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc);
Z
zhangjinchao01 已提交
2951 2952
}

2953 2954
void CpuMatrix::mul(
    CpuMatrix* a, CpuMatrix* b, CpuSparseMatrix* c, real scaleAB, real scaleT) {
Z
zhangjinchao01 已提交
2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060
  CHECK(!c->isTransposed()) << "Not supported";
  CHECK_EQ(c->getValueType(), FLOAT_VALUE);

  real* A = a->getData();
  real* B = b->getData();
  real* C = c->getValue();
  int* rows = c->getRows();
  int* cols = c->getCols();
  size_t height = c->getHeight();
  size_t width = c->getWidth();
  if (scaleT == 0) {
    c->zeroMem();
  }

  if (!a->isTransposed() && !b->isTransposed()) {
    size_t m = a->getWidth();
    CHECK_EQ(b->getHeight(), m);
    CHECK_EQ(a->getHeight(), height);
    CHECK_EQ(b->getWidth(), width);
    if (c->getFormat() == SPARSE_CSC) {
      for (size_t i = 0; i < width; i++) {
        size_t start = c->getColStartIdx(i);
        size_t end = c->getColStartIdx(i + 1);
        for (size_t j = start; j < end; j++) {
          real sum = 0;
          size_t rowIdx = rows[j];
          for (size_t k = 0; k < m; k++) {
            sum += A[rowIdx * m + k] * B[k * width + i];
          }
          C[j] = scaleAB * sum + scaleT * C[j];
        }
      }
    } else {
      for (size_t i = 0; i < height; i++) {
        size_t start = c->getRowStartIdx(i);
        size_t end = c->getRowStartIdx(i + 1);
        for (size_t j = start; j < end; j++) {
          real sum = 0;
          size_t colIdx = cols[j];
          for (size_t k = 0; k < m; k++) {
            sum += A[i * m + k] * B[k * width + colIdx];
          }
          C[j] = scaleAB * sum + scaleT * C[j];
        }
      }
    }
  } else if (a->isTransposed() && !b->isTransposed()) {
    size_t m = a->getHeight();
    CHECK_EQ(m, b->getHeight());
    CHECK_EQ(b->getWidth(), width);
    CHECK_EQ(a->getWidth(), height);

    if (c->getFormat() == SPARSE_CSC) {
      for (size_t i = 0; i < width; i++) {
        size_t start = c->getColStartIdx(i);
        size_t end = c->getColStartIdx(i + 1);
        for (size_t j = start; j < end; j++) {
          real sum = 0;
          size_t rowIdx = rows[j];
          for (size_t k = 0; k < m; k++) {
            sum += A[k * height + rowIdx] * B[k * width + i];
          }
          C[j] = scaleAB * sum + scaleT * C[j];
        }
      }
    } else {
      for (size_t i = 0; i < height; i++) {
        int start = c->getRowStartIdx(i);
        int end = c->getRowStartIdx(i + 1);
        for (int j = start; j < end; j++) {
          real sum = 0;
          size_t colIdx = cols[j];
          for (size_t k = 0; k < m; k++) {
            sum += A[k * height + i] * B[k * width + colIdx];
          }
          C[j] = scaleAB * sum + scaleT * C[j];
        }
      }
    }
  } else if (!a->isTransposed() && b->isTransposed()) {
    size_t m = a->getWidth();
    CHECK_EQ(b->getWidth(), m);
    CHECK_EQ(a->getHeight(), height);
    CHECK_EQ(b->getHeight(), width);
    if (c->getFormat() == SPARSE_CSR) {
      for (size_t i = 0; i < height; i++) {
        size_t start = c->getRowStartIdx(i);
        size_t end = c->getRowStartIdx(i + 1);
        for (size_t j = start; j < end; j++) {
          real sum = 0;
          size_t colIdx = cols[j];
          for (size_t k = 0; k < m; k++) {
            sum += A[i * m + k] * B[colIdx * m + k];
          }
          C[j] = scaleAB * sum + scaleT * C[j];
        }
      }
    } else {
      LOG(FATAL) << "Not supported csc format "
                    "when a is not trans and b is trans";
    }
  } else {
    LOG(FATAL) << "Not supported";
  }
}

3061 3062 3063
void CpuMatrix::mul(CpuMatrix* a,
                    CpuSparseMatrix* b,
                    real scaleAB,
Z
zhangjinchao01 已提交
3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100
                    real scaleT) {
  CHECK(!trans_) << "Not supported";
  CHECK(!a->isTransposed()) << "Not supported";
  CHECK(scaleT == 0 || scaleT == 1);

  // TODO(yuyang18): Maybe bug implementation here
  CHECK_EQ(scaleAB, static_cast<real>(1.0));

  real* A = a->getData();
  real* B = b->getValue();
  real* C = getData();
  int* rows = b->getRows();
  int* cols = b->getCols();

  if (scaleT == 0) {
    zeroMem();
  }
  if (b->getFormat() == SPARSE_CSC) {
    if (!b->isTransposed()) {
      size_t m = a->getWidth();
      CHECK_EQ(b->getHeight(), m);
      CHECK_EQ(a->getHeight(), height_);
      CHECK_EQ(b->getWidth(), width_);

      if (b->getValueType() == NO_VALUE) {
        for (size_t j = 0; j < b->getWidth(); ++j) {
          int start = b->getColStartIdx(j);
          int end = b->getColStartIdx(j + 1);
          for (int i = start; i < end; ++i) {
            colVecAddTo(C + j, A + rows[i], height_, width_, a->getWidth());
          }
        }
      } else if (b->getValueType() == FLOAT_VALUE) {
        for (size_t j = 0; j < b->getWidth(); ++j) {
          int start = b->getColStartIdx(j);
          int end = b->getColStartIdx(j + 1);
          for (int i = start; i < end; ++i) {
3101 3102
            colVecAddTo(
                C + j, A + rows[i], B[i], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123
          }
        }
      }
    } else /*if (b->isTransposed())*/ {
      size_t m = a->getWidth();
      CHECK_EQ(b->getHeight(), width_);
      CHECK_EQ(a->getHeight(), height_);
      CHECK_EQ(b->getWidth(), m);
      if (b->getValueType() == NO_VALUE) {
        for (size_t i = 0; i < b->getWidth(); ++i) {
          int start = b->getColStartIdx(i);
          int end = b->getColStartIdx(i + 1);
          for (int j = start; j < end; ++j) {
            colVecAddTo(C + rows[j], A + i, height_, width_, a->getWidth());
          }
        }
      } else if (b->getValueType() == FLOAT_VALUE) {
        for (size_t i = 0; i < b->getWidth(); ++i) {
          int start = b->getColStartIdx(i);
          int end = b->getColStartIdx(i + 1);
          for (int j = start; j < end; ++j) {
3124 3125
            colVecAddTo(
                C + rows[j], A + i, B[j], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149
          }
        }
      }
    }
  } else {
    if (!b->isTransposed()) {
      size_t m = a->getWidth();
      CHECK_EQ(b->getHeight(), m);
      CHECK_EQ(a->getHeight(), height_);
      CHECK_EQ(b->getWidth(), width_);

      if (b->getValueType() == NO_VALUE) {
        for (size_t j = 0; j < b->getHeight(); ++j) {
          int start = b->getRowStartIdx(j);
          int end = b->getRowStartIdx(j + 1);
          for (int i = start; i < end; ++i) {
            colVecAddTo(C + cols[i], A + j, height_, width_, a->getWidth());
          }
        }
      } else if (b->getValueType() == FLOAT_VALUE) {
        for (size_t j = 0; j < b->getHeight(); ++j) {
          int start = b->getRowStartIdx(j);
          int end = b->getRowStartIdx(j + 1);
          for (int i = start; i < end; ++i) {
3150 3151
            colVecAddTo(
                C + cols[i], A + j, B[i], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172
          }
        }
      }
    } else /*if (b->isTransposed())*/ {
      size_t m = a->getWidth();
      CHECK_EQ(b->getHeight(), width_);
      CHECK_EQ(a->getHeight(), height_);
      CHECK_EQ(b->getWidth(), m);
      if (b->getValueType() == NO_VALUE) {
        for (size_t i = 0; i < b->getHeight(); ++i) {
          int start = b->getRowStartIdx(i);
          int end = b->getRowStartIdx(i + 1);
          for (int j = start; j < end; ++j) {
            colVecAddTo(C + i, A + cols[j], height_, width_, a->getWidth());
          }
        }
      } else if (b->getValueType() == FLOAT_VALUE) {
        for (size_t i = 0; i < b->getHeight(); ++i) {
          int start = b->getRowStartIdx(i);
          int end = b->getRowStartIdx(i + 1);
          for (int j = start; j < end; ++j) {
3173 3174
            colVecAddTo(
                C + i, A + cols[j], B[j], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272
          }
        }
      }
    }
  }
}

void CpuMatrix::selectRows(Matrix& table, IVector& ids) {
  if (dynamic_cast<CacheRowCpuMatrix*>(&table)) {
    selectRowsImp(*dynamic_cast<CacheRowCpuMatrix*>(&table), ids);
  } else if (dynamic_cast<SparseRowCpuMatrix*>(&table)) {
    selectRowsImp(*dynamic_cast<SparseRowCpuMatrix*>(&table), ids);
  } else {
    CHECK(table.isContiguous());
    selectRowsImp(*dynamic_cast<CpuMatrix*>(&table), ids);
  }
}

void CpuMatrix::selectElements(Matrix& table, IVector& ids) {
  CHECK_EQ(table.getHeight(), ids.getSize());
  CHECK_EQ(getHeight(), ids.getSize());
  CHECK_EQ(getWidth(), 1U);
  real* tableData = table.getData();
  int* idsData = ids.getData();
  for (size_t i = 0; i < table.getHeight(); i++) {
    data_[i] += tableData[i * table.getWidth() + idsData[i]];
  }
}

void CpuMatrix::addElements(Matrix& table, IVector& ids) {
  CHECK_EQ(table.getHeight(), ids.getSize());
  CHECK_EQ(getHeight(), ids.getSize());
  CHECK_EQ(getWidth(), 1U);
  real* tableData = table.getData();
  int* idsData = ids.getData();
  for (size_t i = 0; i < table.getHeight(); i++) {
    tableData[i * table.getWidth() + idsData[i]] += data_[i];
  }
}

// this.row[i] += table.row[ids[i]]
template <typename TableMatType>
void CpuMatrix::selectRowsImp(TableMatType& table, IVector& ids) {
  CHECK(!table.useGpu());
  CHECK(!ids.useGpu());
  CHECK_EQ(getHeight(), ids.getSize());
  CHECK_EQ(getWidth(), table.getWidth());
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  real* a = getData();
  size_t tableSize = table.getHeight();
  int* index = ids.getData();

  for (size_t i = 0; i < numSamples; ++i) {
    if (index[i] == -1) continue;
    CHECK_LT(index[i], (int)tableSize);
    CHECK_GE(index[i], 0);
    vecAddTo(a + i * stride_, table.getRow(index[i]), dim);
  }
}

void CpuMatrix::addToRows(Matrix& table, IVector& ids) {
  if (dynamic_cast<CacheRowCpuMatrix*>(&table)) {
    addToRowsImp(*dynamic_cast<CacheRowCpuMatrix*>(&table), ids);
  } else if (dynamic_cast<SparseAutoGrowRowCpuMatrix*>(&table)) {
    addToRowsImp(*dynamic_cast<SparseAutoGrowRowCpuMatrix*>(&table), ids);
  } else if (dynamic_cast<SparseRowCpuMatrix*>(&table)) {
    addToRowsImp(*dynamic_cast<SparseRowCpuMatrix*>(&table), ids);
  } else {
    CHECK(table.isContiguous());
    addToRowsImp(*dynamic_cast<CpuMatrix*>(&table), ids);
  }
}

// table.row[ids[i]] += this.row[i]
template <typename TableMatType>
void CpuMatrix::addToRowsImp(TableMatType& table, IVector& ids) {
  CHECK(!table.useGpu());
  CHECK(!ids.useGpu());
  CHECK_EQ(getHeight(), ids.getSize());
  CHECK_EQ(getWidth(), table.getWidth());
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  real* a = getData();
  size_t tableSize = table.getHeight();
  int* index = ids.getData();

  for (size_t i = 0; i < numSamples; ++i) {
    if (index[i] == -1) continue;
    CHECK_LT(index[i], (int)tableSize);
    CHECK_GE(index[i], 0);
    vecAddTo(table.getRow(index[i]), a + i * stride_, dim);
  }
}

static ThreadLocal<std::vector<const real*>> threadLocalColArray;

template <typename MatBType, typename MatCType>
3273 3274
void CpuMatrix::mul(
    CpuSparseMatrix* a, MatBType* b, MatCType* c, real scaleAB, real scaleT) {
Z
zhangjinchao01 已提交
3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376
  CHECK(!c->isTransposed()) << "Not supported";
  CHECK(!b->isTransposed()) << "Not supported";
  // TODO(yuyang18): Maybe bug implementation here.
  CHECK(scaleAB == 1) << "Not supported";
  CHECK(scaleT == 0 || scaleT == 1) << "Not supported";
  CHECK_EQ(a->getFormat(), SPARSE_CSR) << "Not supported";

  real* B = b->getData();
  real* C = c->getData();
  size_t height = c->getHeight();
  size_t width = c->getWidth();
  int* cols = a->getCols();
  real* values = a->getValue();

  if (scaleT == 0) {
    c->zeroMem();
  }

  if (!a->isTransposed()) {
    size_t m = a->getWidth();
    CHECK_EQ(b->getHeight(), m);
    CHECK_EQ(a->getHeight(), height);
    CHECK_EQ(b->getWidth(), width);

    if (a->getValueType() == NO_VALUE) {
      if (width % 32 == 0) {  // use libaddto
        // @TODO(yuyang18) Make input addr can be unaligned.
        // So merge this if and else
        CHECK_EQ((size_t)B % 32, 0UL);
        CHECK_EQ((size_t)C % 32, 0UL);
        auto& colArray = *threadLocalColArray;
        for (size_t i = 0; i < a->getHeight(); ++i) {
          const int start = a->getRowStartIdx(i);
          const int end = a->getRowStartIdx(i + 1);
          size_t colNum = end - start;
          colArray.resize(colNum);
          for (int j = 0; j < end - start; ++j) {
            colArray[j] = b->getRow(cols[j + start]);
          }
          simd::batchAddTo(c->getRow(i), &colArray[0], colNum, width);
        }

      } else {
        for (size_t i = 0; i < a->getHeight(); ++i) {
          const int start = a->getRowStartIdx(i);
          const int end = a->getRowStartIdx(i + 1);
          for (int j = start; j < end; ++j) {
            vecAddTo(c->getRow(i), b->getRow(cols[j]), width);
          }
        }
      }
    } else if (a->getValueType() == FLOAT_VALUE) {
      for (size_t i = 0; i < a->getHeight(); ++i) {
        const int start = a->getRowStartIdx(i);
        const int end = a->getRowStartIdx(i + 1);
        for (int j = start; j < end; ++j) {
          vecAddTo(c->getRow(i), b->getRow(cols[j]), values[j], width);
        }
      }
    }
  } else /*if (a->isTransposed())*/ {
    size_t m = a->getHeight();
    CHECK_EQ(b->getHeight(), m);
    CHECK_EQ(a->getWidth(), height);
    CHECK_EQ(b->getWidth(), width);
    if (a->getValueType() == NO_VALUE) {
      if (width % 32 == 0) {  // use libaddto
        // @TODO(yuyang18) Make input addr can be unaligned.
        // So merge this if and else
        CHECK_EQ((size_t)B % 32, 0UL);
        CHECK_EQ((size_t)C % 32, 0UL);
        for (size_t i = 0; i < a->getHeight(); ++i) {
          const int start = a->getRowStartIdx(i);
          const int end = a->getRowStartIdx(i + 1);
          for (int j = start; j < end; ++j) {
            simd::addTo(c->getRow(cols[j]), b->getRow(i), width);
          }
        }

      } else {
        for (size_t i = 0; i < a->getHeight(); ++i) {
          const int start = a->getRowStartIdx(i);
          const int end = a->getRowStartIdx(i + 1);
          for (int j = start; j < end; ++j) {
            vecAddTo(c->getRow(cols[j]), b->getRow(i), width);
          }
        }
      }
    } else if (a->getValueType() == FLOAT_VALUE) {
      for (size_t i = 0; i < a->getHeight(); ++i) {
        const int start = a->getRowStartIdx(i);
        const int end = a->getRowStartIdx(i + 1);
        for (int j = start; j < end; ++j) {
          vecAddTo(c->getRow(cols[j]), b->getRow(i), values[j], width);
        }
      }
    }
  }
}

// instantiation mul() called in SparseRowMatrix.cpp
template void CpuMatrix::mul<CpuMatrix, SparseRowCpuMatrix>(
3377 3378 3379 3380
    CpuSparseMatrix* a,
    CpuMatrix* b,
    SparseRowCpuMatrix* c,
    real scaleAB,
Z
zhangjinchao01 已提交
3381 3382
    real scaleT);
template void CpuMatrix::mul<CpuMatrix, SparseAutoGrowRowCpuMatrix>(
3383 3384 3385 3386 3387
    CpuSparseMatrix* a,
    CpuMatrix* b,
    SparseAutoGrowRowCpuMatrix* c,
    real scaleAB,
    real scaleT);
Z
zhangjinchao01 已提交
3388 3389 3390 3391 3392 3393
template void CpuMatrix::mul<CpuMatrix, CacheRowCpuMatrix>(CpuSparseMatrix* a,
                                                           CpuMatrix* b,
                                                           CacheRowCpuMatrix* c,
                                                           real scaleAB,
                                                           real scaleT);

3394
#ifndef PADDLE_MOBILE_INFERENCE
3395 3396 3397
void SharedCpuMatrix::mul(CpuSparseMatrix* a,
                          CpuMatrix* b,
                          real scaleAB,
Z
zhangjinchao01 已提交
3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436
                          real scaleT) {
  CHECK(!isTransposed()) << "Not supported";
  CHECK(!b->isTransposed()) << "Not supported";
  CHECK_EQ(scaleAB, 1) << "Not supported";
  CHECK_EQ(scaleT, 1) << "Not supported";
  CHECK_EQ(a->getFormat(), SPARSE_CSR) << "not supported";

  real* B = b->getData();
  real* C = getData();
  size_t height = getHeight();
  size_t width = getWidth();

  // get real trans
  MatrixPtr aTrans;
  if (a->isTransposed()) {
    aTrans = a->getTmpSparseMatrix(a->getWidth(), a->getHeight());
    a->transpose(aTrans, false);
  }
  a = dynamic_cast<CpuSparseMatrix*>(aTrans.get());

  size_t m = a->getWidth();
  CHECK_EQ(b->getHeight(), m);
  CHECK_EQ(a->getHeight(), height);
  CHECK_EQ(b->getWidth(), width);

  size_t blockSize = (height / blockNum_) + 1;
  CpuMatrixPtr localBuf = *localBuf_;
  if (!localBuf) {
    localBuf = std::make_shared<CpuMatrix>(blockSize, width);
  } else {
    localBuf->resize(blockSize, width);
  }
  localBuf->zeroMem();
  real* localC = localBuf->getData();
  std::vector<int>& blockSeq = *blockSeq_;
  if (blockSeq.size() == 0) {
    for (int k = 0; k < blockNum_; ++k) {
      blockSeq.push_back(k);
    }
3437 3438
    std::shuffle(
        blockSeq.begin(), blockSeq.end(), ThreadLocalRandomEngine::get());
Z
zhangjinchao01 已提交
3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475
  }
  std::vector<int>& localBufRows = *localBufRows_;
  int* cols = a->getCols();
  real* value = a->getValue();

  for (int k = 0; k < blockNum_; ++k) {
    int blockId = blockSeq[k];
    size_t blockBegin = blockId * blockSize;
    size_t blockEnd = (blockId + 1) * blockSize;
    if (blockId == blockNum_ - 1) {
      blockEnd = height;
    }
    if (a->getValueType() == NO_VALUE) {
      for (size_t i = blockBegin; i < blockEnd; ++i) {
        int start = a->getRowStartIdx(i);
        int end = a->getRowStartIdx(i);
        size_t colNum = a->getColNum(i);
        if (colNum == 0) {
          continue;
        }  // skip empty row
        localBufRows.push_back(i);
        size_t bufPos = localBufRows.size() - 1;
        for (int j = start; j < end; ++j) {
          vecAddTo(localC + bufPos * width, B + cols[j] * width, width);
        }
      }
    } else if (a->getValueType() == FLOAT_VALUE) {
      for (size_t i = blockBegin; i < blockEnd; ++i) {
        int start = a->getRowStartIdx(i);
        int end = a->getRowStartIdx(i);
        size_t colNum = a->getColNum(i);
        if (colNum == 0) {
          continue;
        }  // skip empty row
        localBufRows.push_back(i);
        size_t bufPos = localBufRows.size() - 1;
        for (int j = start; j < end; ++j) {
3476 3477
          vecAddTo(
              localC + bufPos * width, B + cols[j] * width, value[j], width);
Z
zhangjinchao01 已提交
3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521 3522
        }
      }
    }

    {
      std::lock_guard<std::mutex> guard(*blockLocks_[blockId]);
      for (size_t i = 0; i < localBufRows.size(); ++i) {
        vecAddTo(C + localBufRows[i] * width, localC + i * width, width);
      }
    }
    memset(localC, 0, localBufRows.size() * width * sizeof(real));
    localBufRows.clear();
  }

  VLOG(2) << " B[0]=" << B[0] << " B[1]=" << B[1] << " C[0]=" << C[0]
          << " C[1]=" << C[1];
}

void SharedCpuMatrix::add(Matrix& b, real p1, real p2) {
  CHECK_EQ(blockNum_, 1);
  std::lock_guard<std::mutex> guard(*blockLocks_[0]);
  CpuMatrix::add(b, p1, p2);
}

void SharedCpuMatrix::add(real p1, real p2) {
  CHECK_EQ(blockNum_, 1);
  std::lock_guard<std::mutex> guard(*blockLocks_[0]);
  CpuMatrix::add(p1, p2);
}

void SharedCpuMatrix::initShared(int blockNum) {
  CHECK_GT(height_ * width_, 1UL * 1024 * 1024)
      << "should not share small matrix";
  initBlock(blockNum);
}

void SharedCpuMatrix::initBlock(int blockNum) {
  CHECK_LE(blockNum, 200) << "should not use large block number";
  blockNum_ = blockNum;
  blockLocks_.resize(blockNum);
  for (auto& locker : blockLocks_) {
    locker.reset(new std::mutex);
  }
}

3523
#endif
Z
zhangjinchao01 已提交
3524 3525 3526 3527 3528 3529
/* Add a (column) vector b to matrix a, column by column */
void CpuMatrix::addColumnVector(const Matrix& b) {
  BaseMatrix::addColVector(const_cast<Matrix&>(b));
}

/* this = a*b */
3530
void CpuMatrix::mul(const Matrix& a, const Matrix& b) {
Z
zhangjinchao01 已提交
3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561
  return mul(a, b, 1.0, 0.0);
}

/* this = scaleAB*(this*b) +  scaleT*this */
void CpuMatrix::rightMul(Matrix& b, real scaleAB, real scaleT) {
  (void)b;
  (void)scaleAB;
  (void)scaleT;
  LOG(FATAL) << "Not implemented";
}

/* this = this* b */
void CpuMatrix::rightMul(Matrix& b) { return rightMul(b, 1.0, 0.0); }

/* this = scaleAB*(a*this) +  scaleT*this */
void CpuMatrix::leftMul(Matrix& a, real scaleAB, real scaleT) {
  (void)a;
  (void)scaleAB;
  (void)scaleT;
  LOG(FATAL) << "Not implemented";
}

/* this = a*this) */
void CpuMatrix::leftMul(Matrix& a) { return leftMul(a, 1.0, 0.0); }

void CpuMatrix::colMerge(Matrix& src) { src.rowSum(*this); }

void CpuMatrix::rowSum(Matrix& sum) {
  CHECK_EQ(sum.getHeight(), getHeight());
  CHECK_EQ(sum.getWidth(), (size_t)1);

3562
  sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0);
Z
zhangjinchao01 已提交
3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593
}

void CpuMatrix::rowMaxId(IVector& maxIds) {
  CHECK(!maxIds.useGpu()) << "Matrix type are not equal";

  size_t numSamples = getHeight();
  CHECK_EQ(maxIds.getSize(), numSamples);

  real* a = getData();
  int* s = maxIds.getData();
  size_t dim = getWidth();

  for (size_t i = 0; i < numSamples; i++) {
    real sm = a[i * dim];
    int maxId = 0;
    for (size_t j = 1; j < dim; j++) {
      if (a[i * dim + j] > sm) {
        maxId = j;
        sm = a[i * dim + j];
      }
    }
    s[i] = maxId;
  }
}

void CpuMatrix::rowMax(Matrix& max) {
  CHECK_EQ(max.getHeight(), getHeight());
  CHECK_EQ(max.getWidth(), (size_t)1);
  max.maxRows(*this);
}

L
Liang Zhao 已提交
3594
/* Get the top k elements of each row of this matrix */
Z
zhangjinchao01 已提交
3595 3596 3597 3598 3599 3600 3601
void CpuMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
  CHECK(isContiguous());
  CHECK(!maxIds.useGpu() && !maxVal.useGpu()) << "Matrix type are not equal";
  size_t numSamples = getHeight();
  size_t beam = maxVal.getWidth();
  CHECK_EQ(maxIds.getSize(), numSamples * beam);
  CHECK_EQ(maxVal.getHeight(), numSamples);
L
Liang Zhao 已提交
3602
  CHECK_EQ(maxVal.getWidth(), beam);
Z
zhangjinchao01 已提交
3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614

  real* a = getData();
  int* s = maxIds.getData();
  real* t = maxVal.getData();
  size_t dim = getWidth();
  for (size_t i = 0; i < numSamples; i++) {
    std::vector<std::pair<real, size_t>> vec;
    for (size_t j = 0; j < dim; j++) {
      vec.push_back(std::pair<real, size_t>(a[i * dim + j], j));
    }

    std::partial_sort(
3615 3616 3617
        vec.begin(),
        vec.begin() + beam,
        vec.end(),
Z
zhangjinchao01 已提交
3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633
        [](const std::pair<real, size_t>& l, const std::pair<real, size_t>& r) {
          return l.first > r.first;
        });
    for (size_t j = 0; j < beam; j++) {
      t[i * beam + j] = vec[j].first;
      s[i * beam + j] = vec[j].second;
    }
  }
}

void CpuMatrix::colMax(Matrix& max) {
  CHECK_EQ(max.getWidth(), getWidth());
  CHECK_EQ(max.getHeight(), (size_t)1);
  max.maxCols(*this);
}

3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652
void CpuMatrix::colMax(IVector& maxIds, Matrix& maxVal) {
  CHECK(isContiguous());
  CHECK(!maxIds.useGpu() && !maxVal.useGpu()) << "Matrix type are not equal";
  size_t numSamples = getWidth();
  size_t beam = maxVal.getHeight();
  CHECK_EQ(maxIds.getSize(), numSamples * beam);
  CHECK_EQ(maxVal.getWidth(), numSamples);

  real* a = getData();
  int* s = maxIds.getData();
  real* t = maxVal.getData();
  size_t dim = getHeight();
  for (size_t i = 0; i < numSamples; i++) {
    std::vector<std::pair<real, size_t>> vec;
    for (size_t j = 0; j < dim; j++) {
      vec.push_back(std::pair<real, size_t>(a[i + j * numSamples], j));
    }

    std::partial_sort(
3653 3654 3655
        vec.begin(),
        vec.begin() + beam,
        vec.end(),
3656 3657 3658 3659 3660 3661 3662 3663 3664 3665
        [](const std::pair<real, size_t>& l, const std::pair<real, size_t>& r) {
          return l.first > r.first;
        });
    for (size_t j = 0; j < beam; j++) {
      t[i + j * numSamples] = vec[j].first;
      s[i + j * numSamples] = vec[j].second;
    }
  }
}

3666 3667 3668
void CpuMatrix::maxoutForward(Matrix& a,
                              IVector& id,
                              size_t channels,
3669 3670 3671 3672 3673 3674 3675 3676
                              size_t groups) {
  CHECK(dynamic_cast<CpuMatrix*>(&a));
  CHECK(dynamic_cast<CpuIVector*>(&id));
  CHECK_EQ(a.getHeight(), getHeight());

  size_t size = getWidth();
  size_t batchSize = getHeight();
  size_t featLen = size / channels;
Q
qijun 已提交
3677
  const real* input = a.getData();
3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700
  int* idForCpu = id.getData();

  MatrixPtr maxInMat, maxOutMat;
  Matrix::resizeOrCreate(maxInMat, groups, size, false, false);
  Matrix::resizeOrCreate(maxOutMat, 1, size, false, false);

  for (size_t batch_idx = 0; batch_idx < batchSize; ++batch_idx) {
    size_t newIndex = batch_idx * size;
    IVectorPtr tmpId = IVector::create(idForCpu + newIndex, size, false);

    for (size_t i = 0; i < channels; ++i) {
      size_t newFeatLen = i * featLen;
      for (size_t j = 0; j < groups; ++j) {
        maxInMat->subMatrix(j, j + 1, newFeatLen, newFeatLen + featLen)
            ->copyFrom(input + (newIndex + newFeatLen) * groups + j * featLen,
                       featLen);
      }
    }
    maxInMat->colMax(*tmpId, *maxOutMat);
    this->subRowMatrix(batch_idx, batch_idx + 1)->copyFrom(*maxOutMat);
  }
}

3701 3702 3703
void CpuMatrix::maxoutBackward(Matrix& a,
                               IVector& id,
                               size_t channels,
3704 3705 3706 3707 3708 3709 3710 3711 3712
                               size_t groups) {
  CHECK(dynamic_cast<CpuMatrix*>(&a));
  CHECK(dynamic_cast<CpuIVector*>(&id));
  CHECK_EQ(a.getHeight(), getHeight());

  size_t size = a.getWidth();
  size_t batchSize = getHeight();
  size_t featLen = size / channels;
  size_t newFeatLen = groups * featLen;
Q
qijun 已提交
3713 3714
  real* inputG = getData();
  const real* outG = a.getData();
3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728
  int* idForCpu = id.getData();

  for (size_t batch_idx = 0; batch_idx < batchSize; ++batch_idx) {
    size_t newIndex = batch_idx * size;
    int* idData = idForCpu + newIndex;

    for (size_t i = 0; i < size; ++i) {
      int gradIdx =
          idData[i] * featLen + (i / featLen) * newFeatLen + i % featLen;
      (inputG + newIndex * groups)[gradIdx] += (outG + newIndex)[i];
    }
  }
}

Z
zhangjinchao01 已提交
3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753
void CpuMatrix::rowNormalizeL1(Matrix& out) {
  CHECK(!out.useGpu());

  size_t numSamples = getHeight();
  size_t dim = getWidth();
  CHECK_EQ(out.getHeight(), numSamples);
  CHECK_EQ(out.getWidth(), dim);
  real* a = getData();
  real* b = out.getData();
  for (size_t i = 0; i < numSamples; ++i) {
    real s = 0;
    for (size_t j = 0; j < dim; ++j) {
      s += a[i * dim + j];
    }
    // Right now, we just bet that sum won't be zero. If this really happens,
    // we will figure out what should be done then.
    CHECK_GT(s, 0);
    s = 1 / s;
    for (size_t j = 0; j < dim; ++j) {
      b[i * dim + j] = s * a[i * dim + j];
    }
  }
}

/* calulate classification error */
3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767
void CpuMatrix::classificationError(Matrix& output,
                                    IVector& label,
                                    size_t topkSize) {
  size_t numSamples = this->getHeight();
  auto cpuOutput = dynamic_cast<CpuMatrix*>(&output);
  auto cpuLabel = dynamic_cast<CpuIVector*>(&label);
  IVectorPtr cpuTopIds = std::make_shared<CpuIVector>(numSamples * topkSize);
  MatrixPtr cpuTopVal = std::make_shared<CpuMatrix>(numSamples, topkSize);

  CHECK(cpuOutput && cpuLabel) << "Invalid argument pointer";
  CHECK(cpuTopIds && cpuTopVal) << "Allocate cpu memory failed";
  CHECK(cpuLabel->getSize() == numSamples) << "Vector size is not equal";
  CHECK(cpuOutput->getHeight() == numSamples && this->getWidth() == 1)
      << "Matrix dimensions are not equal";
Z
zhangjinchao01 已提交
3768

3769 3770
  // top k matrix classification
  cpuOutput->rowMax(*cpuTopIds, *cpuTopVal);
Z
zhangjinchao01 已提交
3771

3772 3773 3774 3775
  size_t dim = cpuOutput->getWidth();
  real* result = this->getData();
  int* ids = cpuTopIds->getData();
  int* lbl = cpuLabel->getData();
Z
zhangjinchao01 已提交
3776 3777 3778
  for (size_t i = 0; i < numSamples; ++i) {
    CHECK_GE(lbl[i], 0);
    CHECK_LT((size_t)lbl[i], dim);
3779 3780 3781 3782 3783

    for (size_t j = 0; j < topkSize; ++j) {
      if (ids[j + i * topkSize] == lbl[i]) {
        result[i] = 0;
        break;
Z
zhangjinchao01 已提交
3784
      }
3785
      result[i] = 1.0f;
Z
zhangjinchao01 已提交
3786 3787 3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830
    }
  }
}

/* copy -log(output[label]) to this->data[i] */
void CpuMatrix::oneHotCrossEntropy(Matrix& output, IVector& label) {
  CHECK(dynamic_cast<CpuMatrix*>(&output));
  CHECK(dynamic_cast<CpuIVector*>(&label));

  size_t numSamples = getHeight();
  size_t dim = output.getWidth();
  CHECK_EQ(label.getSize(), numSamples);
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(getWidth(), (size_t)1);

  real* out = output.getData();
  real* cost = getData();
  int* lbl = label.getData();
  for (size_t i = 0; i < numSamples; ++i, out += dim) {
    CHECK_GE(lbl[i], 0);
    CHECK_LT((size_t)lbl[i], dim);
    cost[i] = -std::log(out[lbl[i]]);
  }
}

/* calculate the error of outputV according to label */
void CpuMatrix::oneHotCrossEntropyBp(Matrix& output, IVector& label) {
  CHECK(dynamic_cast<CpuMatrix*>(&output));
  CHECK(dynamic_cast<CpuIVector*>(&label));
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  CHECK_EQ(output.getWidth(), dim);
  real* out = output.getData();
  real* grad = getData();
  int* lbl = label.getData();
  for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) {
    grad[lbl[i]] -= 1 / out[lbl[i]];
  }
}

/*
    We implement the matrix functionality in CostLayer.cpp,
    but we define the scalar function here for sanity check
    deletion of the function does not affect anything neverthelss
*/
3831 3832
void CpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output,
                                               IVector& label,
Z
zhangjinchao01 已提交
3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862
                                               real alpha) {
  CHECK(dynamic_cast<CpuMatrix*>(&output));
  CHECK(dynamic_cast<CpuIVector*>(&label));

  size_t numSamples = getHeight();
  size_t dim = output.getWidth();
  CHECK_EQ(label.getSize(), numSamples);
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(getWidth(), (size_t)1);

  real* out = output.getData();
  real* cost = getData();
  int* lbl = label.getData();
  for (size_t i = 0; i < numSamples; ++i, out += dim) {
    CHECK_GE(lbl[i], 0);
    CHECK_LT((size_t)lbl[i], dim);
    real sum = 0;
    for (size_t j = 0; j < dim; ++j) {
      sum += out[j];
    }
    sum = _safelog(sum);
    cost[i] = -_safelog(out[lbl[i]]) + sum + alpha * _square(sum);
  }
}

/*
    We implement the matrix functionality in CostLayer.cpp,
    but we define the scalar function here for sanity check
    deletion of the function does not affect anything neverthelss
*/
3863 3864
void CpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& output,
                                                 IVector& label,
Z
zhangjinchao01 已提交
3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944
                                                 real alpha) {
  CHECK(dynamic_cast<CpuMatrix*>(&output));
  CHECK(dynamic_cast<CpuIVector*>(&label));
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  CHECK_EQ(output.getWidth(), dim);
  real* out = output.getData();
  real* grad = getData();
  int* lbl = label.getData();

  for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) {
    grad[lbl[i]] -= 1 / out[lbl[i]];
    real sum = 0;
    for (size_t j = 0; j < dim; ++j) {
      sum += out[j];
    }
    for (size_t j = 0; j < dim; ++j) {
      if (j == (size_t)lbl[i]) {
        grad[j] += -1 / out[j];
      }
      grad[j] += 1 / sum + 2 * alpha * _safelog(sum) / sum;
    }
  }
}

#define FORWARD_LOOP()                      \
  size_t numSamples = getHeight();          \
  size_t dim = getWidth();                  \
  CHECK_EQ(output.getHeight(), numSamples); \
  CHECK_EQ(output.getWidth(), dim);         \
  const real* in = getData();               \
  real* out = output.getData();             \
  for (size_t i = 0; i < numSamples; ++i, in += dim, out += dim)

#define BACKWARD_LOOP()                     \
  size_t numSamples = getHeight();          \
  size_t dim = getWidth();                  \
  CHECK_EQ(output.getHeight(), numSamples); \
  CHECK_EQ(output.getWidth(), dim);         \
  real* grad = getData();                   \
  real* out = output.getData();             \
  for (size_t i = 0; i < numSamples; ++i, grad += dim, out += dim)

void CpuMatrix::softmax(Matrix& output) {
  CHECK(!output.useGpu());

  const float THRESHOLD = -64.0;

  FORWARD_LOOP() {
    real max = -1.0e20;
    for (size_t j = 0; j < dim; ++j) {
      if (in[j] > max) {
        max = in[j];
      }
    }
    for (size_t j = 0; j < dim; ++j) {
      real a = in[j] - max;
      if (a < THRESHOLD) {
        a = THRESHOLD;
      }
      out[j] = a;
    }
    vExp(dim, out, out);

    real sum = 0;
    for (size_t j = 0; j < dim; ++j) {
      sum += out[j];
    }
    sum = 1 / sum;
    for (size_t j = 0; j < dim; ++j) {
      out[j] *= sum;
    }
  }
}

void CpuMatrix::sequenceSoftmax(Matrix& output, const IVector& index) {
  CHECK_EQ(getWidth(), 1UL);
  CHECK_EQ(output.getWidth(), 1UL);
  CHECK(isContiguous());

3945 3946 3947 3948 3949 3950 3951 3952 3953 3954
  MatrixPtr inTmp = Matrix::create(nullptr,
                                   /* height= */ 1,
                                   1,
                                   /* trans= */ false,
                                   false);
  MatrixPtr outTmp = Matrix::create(nullptr,
                                    /* height= */ 1,
                                    1,
                                    /* trans= */ false,
                                    false);
Z
zhangjinchao01 已提交
3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006
  size_t numSequences = index.getSize() - 1;
  auto starts = index.getData();
  for (size_t i = 0; i < numSequences; ++i) {
    size_t offset = starts[i];
    size_t size = starts[i + 1] - starts[i];
    inTmp->setData(getData() + offset, 1UL, size);
    outTmp->setData(output.getData() + offset, 1UL, size);
    inTmp->softmax(*outTmp);
  }
}

void CpuMatrix::softmaxDerivative(Matrix& output, Matrix& sftmaxSum) {
  CHECK(output.useGpu_ == false) << "Matrix type are not equal";
  CHECK_EQ(getHeight(), sftmaxSum.getHeight());

  real* sums = sftmaxSum.getData();

  BACKWARD_LOOP() {
    real sum = sums[i];
    for (size_t j = 0; j < dim; ++j) {
      grad[j] = out[j] * (grad[j] - sum);
    }
  }
}

void CpuMatrix::sumOfSquares(Matrix& output, Matrix& label) {
  CHECK(output.useGpu_ == false && label.useGpu_ == false)
      << "Matrix type are not equal";

  size_t numSamples = getHeight();
  size_t dim = output.getWidth();
  CHECK_EQ(label.getHeight(), numSamples);
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(label.getWidth(), dim);
  CHECK_EQ(getWidth(), (size_t)1);
  real* out = output.getData();
  real* cost = getData();

  auto labelptr = dynamic_cast<CpuSparseMatrix*>(&label);
  if (labelptr) {
    // it is a CpuSparseMatrix
    if (labelptr->getFormat() == SPARSE_CSR) {
      // treat label as a SparseMatrix
      for (size_t i = 0; i < numSamples; ++i) {
        for (size_t j = 0; j < dim; ++j) {
          cost[i] += _square(out[i * dim + j]);
        }
      }
      if (labelptr->getValueType() == NO_VALUE) {
        int* cols = labelptr->getCols();
        for (size_t i = 0; i < numSamples; ++i) {
          for (size_t j = labelptr->getRowStartIdx(i);
4007 4008
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023
            cost[i] += 1.0 - 2.0 * out[i * dim + cols[j]];
            /*
             * explanation of above line: original codes are follows:
             * cost[i] -= _square(out[i * dim + feature.col]);
             * cost[i] += _square(1.0 - out[i * dim + feature.col]);
             */
          }
        }
      } else if (labelptr->getValueType() == FLOAT_VALUE) {
        int* cols = labelptr->getCols();
        real* values = labelptr->getValue();
        for (size_t i = 0; i < numSamples; ++i) {
          real sum1 = 0;
          real sum2 = 0;
          for (size_t j = labelptr->getRowStartIdx(i);
4024 4025
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046
            sum1 += values[j] * values[j];
            sum2 += values[j] * out[i * dim + cols[j]];
            /*
             * explanation of above line: original codes are follows:
             * cost[i] -= _square(out[i * dim + feature.col]);
             * cost[i] += _square(value.col - out[i * dim + feature.col]);
             */
          }
          cost[i] += sum1 - 2.0 * sum2;
        }
      } else {
        LOG(FATAL) << "unsupported sparse matrix value type in sumOfSquares";
        return;
      }
      return;
    } else {
      LOG(FATAL) << "unsupported sparse matrix format in sumOfSquares";
      return;
    }
  }

4047 4048 4049 4050
  BaseMatrix::sumOfSquaredDiffs(output,
                                label,
                                /* scaleSum= */ 1,
                                /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
4051 4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079
}

/* calculate the error of outputV according to label */
void CpuMatrix::sumOfSquaresBp(Matrix& output, Matrix& label) {
  CHECK(output.useGpu_ == false && label.useGpu_ == false)
      << "Matrix type are not equal";

  size_t numSamples = getHeight();
  size_t dim = getWidth();
  CHECK_EQ(output.getWidth(), dim);
  CHECK_EQ(label.getWidth(), dim);

  real* out = output.getData();
  real* grad = getData();

  auto labelptr = dynamic_cast<CpuSparseMatrix*>(&label);
  if (labelptr) {
    // it is a CpuSparseMatrix
    if (labelptr->getFormat() == SPARSE_CSR) {
      // treat label as a SparseMatrix
      for (size_t i = 0; i < numSamples; ++i) {
        for (size_t j = 0; j < dim; ++j) {
          grad[i * dim + j] += 2.0 * out[i * dim + j];
        }
      }
      if (labelptr->getValueType() == NO_VALUE) {
        int* cols = labelptr->getCols();
        for (size_t i = 0; i < numSamples; ++i) {
          for (size_t j = labelptr->getRowStartIdx(i);
4080 4081
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095
            grad[i * dim + cols[j]] -= 2.0;
            /*
             * explanation of above line: original codes are follows:
             * grad[i * dim + feature.col] -= 2.0 * out[i * dim + feature.col];
             * grad[i * dim + feature.col] += 2.0 * (out[i * dim + feature.col]
             * - 1);
             */
          }
        }
      } else if (labelptr->getValueType() == FLOAT_VALUE) {
        int* cols = labelptr->getCols();
        real* values = labelptr->getValue();
        for (size_t i = 0; i < numSamples; ++i) {
          for (size_t j = labelptr->getRowStartIdx(i);
4096 4097
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
4098 4099 4100 4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127 4128 4129 4130
            grad[i * dim + cols[j]] -= 2.0 * values[j];
            /*
             * explanation of above line: original codes are follows:
             * grad[i * dim + feature.col] -= 2.0 * out[i * dim + feature.col];
             * grad[i * dim + feature.col] += 2.0 * (out[i * dim + feature.col]
             * - value.col);
             */
          }
        }
      } else {
        LOG(FATAL) << "unsupported sparse matrix value type in sumOfSquares";
        return;
      }
      return;
    } else {
      LOG(FATAL) << "unsupported sparse matrix format in sumOfSquares";
      return;
    }
  }

  real* lbl = label.getData();
  size_t ld = getStride();
  size_t outLd = output.getStride();
  size_t lblLd = label.getStride();
  CHECK(lbl);
  for (size_t i = 0; i < numSamples;
       ++i, out += outLd, lbl += lblLd, grad += ld) {
    for (size_t j = 0; j < dim; ++j) {
      grad[j] += 2.0 * (out[j] - lbl[j]);  // positive gradient;
    }
  }
}

4131
void CpuMatrix::smoothL1(Matrix& output, Matrix& label, real destScale) {
G
gaoyuan 已提交
4132 4133 4134 4135 4136 4137 4138 4139 4140
  CHECK(output.useGpu_ == false && label.useGpu_ == false)
      << "Matrix type are not equal";

  size_t numSamples = getHeight();
  size_t dim = output.getWidth();
  CHECK_EQ(label.getHeight(), numSamples);
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(label.getWidth(), dim);
  CHECK_EQ(getWidth(), (size_t)1);
D
dangqingqing 已提交
4141

G
gaoyuan 已提交
4142
  real* cost = getData();
D
dangqingqing 已提交
4143
  real* out = output.getData();
G
gaoyuan 已提交
4144 4145
  real* lbl = label.getData();

D
dangqingqing 已提交
4146
  for (size_t i = 0; i < numSamples; ++i, out += dim, lbl += dim) {
G
gaoyuan 已提交
4147
    for (size_t j = 0; j < dim; ++j) {
D
dangqingqing 已提交
4148
      real absVal = std::fabs(out[j] - lbl[j]);
4149
      cost[i] *= destScale;
D
dangqingqing 已提交
4150 4151
      if (absVal < 1.0)
        cost[i] += 0.5 * absVal * absVal;
G
gaoyuan 已提交
4152
      else
D
dangqingqing 已提交
4153
        cost[i] += absVal - 0.5;
G
gaoyuan 已提交
4154 4155 4156 4157
    }
  }
}

4158
void CpuMatrix::smoothL1Bp(Matrix& output, Matrix& label, real destScale) {
G
gaoyuan 已提交
4159 4160 4161 4162 4163 4164 4165 4166
  CHECK(output.useGpu_ == false && label.useGpu_ == false)
      << "Matrix type are not equal";

  size_t numSamples = getHeight();
  size_t dim = output.getWidth();
  CHECK_EQ(label.getHeight(), numSamples);
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(label.getWidth(), dim);
D
dangqingqing 已提交
4167 4168
  CHECK_EQ(getWidth(), dim);

G
gaoyuan 已提交
4169 4170
  real* out = output.getData();
  real* lbl = label.getData();
D
dangqingqing 已提交
4171
  real* grad = getData();
G
gaoyuan 已提交
4172

D
dangqingqing 已提交
4173
  for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim, lbl += dim) {
G
gaoyuan 已提交
4174
    for (size_t j = 0; j < dim; ++j) {
D
dangqingqing 已提交
4175
      real val = out[j] - lbl[j];
4176
      grad[j] *= destScale;
D
dangqingqing 已提交
4177 4178 4179 4180 4181
      if (std::fabs(val) < 1) {
        grad[j] += val;
      } else {
        grad[j] += (real(0) < val) - (val < real(0));
      }
G
gaoyuan 已提交
4182 4183 4184 4185
    }
  }
}

Z
zhangjinchao01 已提交
4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239 4240 4241 4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269 4270 4271 4272 4273 4274 4275 4276 4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287
void CpuMatrix::tanh(Matrix& output) {
  CHECK(isContiguous());
  CHECK(output.isContiguous());
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(output.getWidth(), dim);
  vTanh(numSamples * dim, getData(), output.getData());
}

void CpuMatrix::tanhDerivative(Matrix& output) {
  BaseMatrix::tanhDerivative(output);
}

void CpuMatrix::softrelu(Matrix& output) {
  CHECK(isContiguous());
  CHECK(output.isContiguous());
  const real THRESHOLD = 40.0;
  FORWARD_LOOP() {  // TODO(yuyang18): SIMD it?
    for (size_t j = 0; j < dim; ++j) {
      real x = in[j];
      if (x > THRESHOLD) {
        x = THRESHOLD;
      } else if (x < -THRESHOLD) {
        x = -THRESHOLD;
      }
      out[j] = x;
    }
  }
  vExp(numSamples * dim, output.getData(), output.getData());
  vLog1p(numSamples * dim, output.getData(), output.getData());
}

void CpuMatrix::softreluDerivative(Matrix& output) {
  CHECK(isContiguous());
  CHECK(output.isContiguous());
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  size_t size = numSamples * dim;
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(output.getWidth(), dim);
  real* grad = getData();
  MatrixPtr tmpMat = Matrix::create(numSamples, dim);
  real* tmp = tmpMat->getData();

  vExp(size, output.getData(), tmpMat->getData());

  for (size_t i = 0; i < size; ++i) {
    grad[i] *= (1.0 - 1.0 / tmp[i]);
  }
}

void CpuMatrix::scaledTanh(Matrix& output, real p1, real p2) {
  CHECK(isContiguous());
  CHECK(output.isContiguous());
  size_t numSamples = getHeight();
  size_t dim = getWidth();
  CHECK_EQ(output.getHeight(), numSamples);
  CHECK_EQ(output.getWidth(), dim);

  const real* in = getData();
  real* out = output.getData();

  // out = p2*in
  for (size_t i = 0; i < numSamples * dim; ++i) {
    out[i] = p2 * in[i];
  }

  vTanh(numSamples * dim, out, out);

  // out = p1 * out
  for (size_t i = 0; i < numSamples * dim; ++i) {
    out[i] = p1 * out[i];
  }
}

/* uniform randomization, minimize precision = 1e-5 */
void CpuMatrix::randomizeUniform() {
  CHECK(isContiguous());
  real* data = getData();
  unsigned int* randSeed = ThreadLocalRand::getSeed();
  real recipRandMax = 1.0f / (real)RAND_MAX;
  for (size_t i = 0; i < elementCnt_; ++i) {
    *data++ = rand_r(randSeed) * recipRandMax;
  }
}

void CpuMatrix::print(std::ostream& os) const {
  CHECK(isContiguous());
  for (size_t i = 0; i < height_; ++i) {
    for (size_t j = 0; j < width_; ++j) {
      os << data_[i * width_ + j] << " ";
    }
    os << std::endl;
  }
}

void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
  real* input = data.getData();
  real* w = W.getData();
  size_t numElements = data.getWidth();
  size_t numSamples = data.getHeight();
H
hedaoyuan 已提交
4288 4289 4290
  size_t paraSize = W.getHeight() * W.getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
Z
zhangjinchao01 已提交
4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303
  for (size_t n = 0, k = 0; n < numSamples; ++n) {
    for (size_t i = 0; i < numElements; ++i, ++k) {
      data_[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
    }
  }
}

void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) {
  real* ograd = oGrad.getData();
  real* input = data.getData();
  real* wgrad = data_;
  size_t numElements = data.getWidth();
  size_t numSamples = data.getHeight();
H
hedaoyuan 已提交
4304 4305 4306
  size_t paraSize = this->getHeight() * this->getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
Z
zhangjinchao01 已提交
4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320
  for (size_t n = 0, k = 0; n < numSamples; ++n) {
    for (size_t i = 0; i < numElements; ++i, ++k) {
      wgrad[i / partial_sum] += ograd[k] * (input[k] > 0 ? 0 : input[k]);
    }
  }
}

void CpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) {
  real* diff = data_;
  real* input = data.getData();
  real* ograd = oGrad.getData();
  real* w = W.getData();
  size_t numElements = data.getWidth();
  size_t numSamples = data.getHeight();
H
hedaoyuan 已提交
4321 4322 4323
  size_t paraSize = W.getHeight() * W.getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
Z
zhangjinchao01 已提交
4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432
  for (size_t n = 0, k = 0; n < numSamples; ++n) {
    for (size_t i = 0; i < numElements; ++i, ++k) {
      diff[k] += ograd[k] * (input[k] > 0 ? 1 : w[i / partial_sum]);
    }
  }
}

void CpuMatrix::print(std::ostream& os, size_t height, size_t width) const {
  CHECK(isContiguous());
  size_t h = height_ < height ? height_ : height;
  size_t w = width_ < width ? width_ : width;
  os.setf(std::ostream::scientific);
  os << "[";
  for (size_t i = 0; i < h; ++i) {
    for (size_t j = 0; j < w; ++j) {
      os << data_[i * width_ + j] << " ";
    }
    if (i == h - 1) {
      os << "]";
    }
    os << std::endl;
  }
}

void CpuMatrix::printOneRow(std::ostream& os, size_t idx) const {
  CHECK_LT(idx, height_);
  size_t offset = idx * stride_;
  os << data_[offset];
  for (size_t i = 1; i < width_; ++i) {
    os << " " << data_[offset + i];
  }
  os << ";";
}

void CpuMatrix::check(std::ostream& os, Matrix& refMat, bool printDiff) {
  CHECK(isContiguous());
  CHECK(height_ == refMat.getHeight());
  CHECK(width_ == refMat.getWidth());
  CpuMatrix cpuRef(height_, width_);
  cpuRef.copyFrom(refMat);
  size_t diffCnt = 0;
  for (size_t i = 0; i < height_; ++i) {
    for (size_t j = 0; j < width_; ++j) {
      real a = getElement(i, j);
      real b = cpuRef.getElement(i, j);
      if (fabs(a - b) > 0.00001) {
        ++diffCnt;
        if (printDiff) {
          os << "ref= " << a << "  check= " << b << std::endl;
        }
      }
    }
  }
  LOG(INFO) << "the  diffCnt is " << diffCnt;
}

real CpuMatrix::getMin() {
  size_t size = getHeight() * getWidth();
  real* data = getData();
  real res = data[0];
  for (size_t i = 1; i < size; ++i) {
    if (res > data[i]) {
      res = data[i];
    }
  }
  return res;
}

real CpuMatrix::getMax() {
  size_t size = getHeight() * getWidth();
  real* data = getData();
  real res = data[0];
  for (size_t i = 1; i < size; ++i) {
    if (res < data[i]) {
      res = data[i];
    }
  }
  return res;
}

void CpuMatrix::circularConv(Matrix& in0, Matrix& in1) {
  size_t height = this->getHeight();
  size_t width0 = this->getWidth();
  size_t width1 = in1.getWidth();

  CHECK_EQ(height, in0.getHeight());
  CHECK_EQ(width0, in0.getWidth());
  CHECK_EQ(height, in1.getHeight());

  CHECK_EQ(width1 % 2, 1U);

  real* outV = this->getData();
  real* inV0 = in0.getData();
  real* inV1 = in1.getData();

  int leftCtxLen = (width1 - 1) / 2;
  for (size_t x = 0; x < height;
       ++x, outV += width0, inV0 += width0, inV1 += width1) {
    for (size_t i = 0; i < width0; ++i) {  // each dimension of output
      for (size_t j = 0; j < width1; ++j) {
        // iterate over all dimentions of inV1
        int index = i + j - leftCtxLen;
        index = (index + width0) % width0;
        outV[i] += inV0[index] * inV1[j];
      }
    }
  }
}

4433 4434
void CpuMatrix::circularConvDerivative(
    Matrix& outG, Matrix& in0, Matrix& in1, Matrix& inG0, Matrix& inG1) {
Z
zhangjinchao01 已提交
4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453
  size_t height = in0.getHeight();
  size_t width0 = in0.getWidth();
  size_t width1 = in1.getWidth();

  CHECK_EQ(height, in1.getHeight());
  CHECK_EQ(height, inG0.getHeight());
  CHECK_EQ(width0, inG0.getWidth());
  CHECK_EQ(height, inG1.getHeight());
  CHECK_EQ(width1, inG1.getWidth());
  CHECK_EQ(height, outG.getHeight());
  CHECK_EQ(width0, outG.getWidth());

  real* outGV = outG.getData();
  real* inV0 = in0.getData();
  real* inV1 = in1.getData();
  real* inGV0 = inG0.getData();
  real* inGV1 = inG1.getData();

  int leftCtxLen = (width1 - 1) / 2;
4454 4455 4456 4457 4458 4459
  for (size_t x = 0; x < height; ++x,
              outGV += width0,
              inV0 += width0,
              inV1 += width1,
              inGV0 += width0,
              inGV1 += width1) {
Z
zhangjinchao01 已提交
4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494 4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527
    for (size_t j = 0; j < width1; ++j) {  // iterate over width1
      for (size_t i = 0; i < width0; ++i) {
        // such over all dimensions of outG
        int index = i + j - leftCtxLen;
        index = (index + width0) % width0;
        inGV0[index] += outGV[i] * inV1[j];
        inGV1[j] += outGV[i] * inV0[index];
      }
    }
  }
}

void CpuMatrix::multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) {
  CHECK(dynamic_cast<CpuMatrix*>(&output));
  auto labelPtr = dynamic_cast<CpuSparseMatrix*>(&label);
  CHECK(labelPtr);

  size_t numSamples = getHeight();
  size_t dim = output.getWidth();
  CHECK_EQ(numSamples, output.getHeight());
  CHECK_EQ(numSamples, labelPtr->getHeight());
  CHECK_EQ(dim, labelPtr->getWidth());

  real* out = output.getData();
  real* cost = getData();
  for (size_t i = 0; i < numSamples; ++i, out += dim) {
    for (size_t j = 0; j < dim; ++j) {
      CHECK(out[j] > 0 && out[j] < 1.0);
      cost[i] -= std::log(1 - out[j]);
    }

    const int* cols = labelPtr->getRowCols(i);
    for (size_t j = 0; j < labelPtr->getColNum(i); ++j) {
      CHECK_LT(size_t(cols[j]), dim);
      cost[i] -= std::log(out[cols[j]] / (1 - out[cols[j]]));
    }
  }
}

void CpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) {
  CHECK(dynamic_cast<CpuMatrix*>(&output));
  auto labelPtr = dynamic_cast<CpuSparseMatrix*>(&label);
  CHECK(labelPtr);

  size_t numSamples = getHeight();
  size_t dim = getWidth();
  CHECK_EQ(numSamples, output.getHeight());
  CHECK_EQ(numSamples, labelPtr->getHeight());
  CHECK_EQ(dim, output.getWidth());
  CHECK_EQ(dim, labelPtr->getWidth());

  real* out = output.getData();
  real* grad = getData();
  for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim) {
    for (size_t j = 0; j < dim; ++j) {
      CHECK(out[j] > 0 && out[j] < 1.0);
      grad[j] += 1.0 / (1 - out[j]);
    }

    const int* cols = labelPtr->getRowCols(i);
    for (size_t j = 0; j < labelPtr->getColNum(i); ++j) {
      CHECK_LT(size_t(cols[j]), dim);
      grad[cols[j]] -= 1.0 / (out[cols[j]] * (1 - out[cols[j]]));
    }
  }
}

/* calculate the classification error for multi binary label */
4528 4529
void CpuMatrix::classificationErrorMulti(Matrix& output,
                                         Matrix& label,
Z
zhangjinchao01 已提交
4530 4531 4532 4533 4534 4535 4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563
                                         real threshold) {
  CHECK(dynamic_cast<CpuMatrix*>(&output));
  auto labelPtr = dynamic_cast<CpuSparseMatrix*>(&label);
  CHECK(labelPtr);

  size_t numSamples = getHeight();
  size_t dim = output.getWidth();
  CHECK_EQ(numSamples, output.getHeight());
  CHECK_EQ(numSamples, labelPtr->getHeight());
  CHECK_EQ(dim, labelPtr->getWidth());

  real* out = output.getData();
  real* result = getData();
  for (size_t i = 0; i < numSamples; ++i, out += dim) {
    real sum = 0.0;
    for (size_t j = 0; j < dim; ++j) {
      if (out[j] >= threshold) {
        sum += 1.0;
      }
    }

    const int* cols = labelPtr->getRowCols(i);
    for (size_t j = 0; j < labelPtr->getColNum(i); ++j) {
      CHECK_LT(size_t(cols[j]), dim);
      if (out[cols[j]] < threshold) {
        sum += 1.0;
      } else {
        sum -= 1.0;
      }
    }
    result[i] = sum / dim;
  }
}

L
liaogang 已提交
4564 4565 4566 4567 4568
void CpuMatrix::bilinearForward(const Matrix& in,
                                const size_t inImgH,
                                const size_t inImgW,
                                const size_t outImgH,
                                const size_t outImgW,
L
liaogang 已提交
4569 4570 4571
                                const size_t numChannels,
                                const real ratioH,
                                const real ratioW) {
L
liaogang 已提交
4572 4573 4574
  CHECK(dynamic_cast<const CpuMatrix*>(&in));

  size_t outputW = getWidth();
L
liaogang 已提交
4575
  size_t batchSize = getHeight();
L
liaogang 已提交
4576 4577
  size_t inputW = in.getWidth();
  size_t inputH = in.getHeight();
L
liaogang 已提交
4578 4579
  size_t inPosOffset = inImgH * inImgW;
  size_t outPosOffset = outImgH * outImgW;
L
liaogang 已提交
4580
  (void)(inputH);
L
liaogang 已提交
4581 4582

  real* outData = getData();
4583
  const real* inData = in.getData();
L
liaogang 已提交
4584 4585 4586 4587

  if (inImgH == outImgH && inImgW == outImgW) {
    this->copyFrom(in);
  } else {
4588
    for (size_t k = 0; k < batchSize; ++k) {  // loop for batches
L
liaogang 已提交
4589 4590 4591
      for (size_t i = 0; i < outImgH; ++i) {  // loop for images
        size_t h = ratioH * i;
        size_t hid = (h < inImgH - 1) ? 1 : 0;
L
liaogang 已提交
4592 4593
        real h1lambda = ratioH * i - h;
        real h2lambda = 1 - h1lambda;
L
liaogang 已提交
4594

L
liaogang 已提交
4595 4596 4597
        for (size_t j = 0; j < outImgW; ++j) {
          size_t w = ratioW * j;
          size_t wid = (w < inImgW - 1) ? 1 : 0;
L
liaogang 已提交
4598 4599
          real w1lambda = ratioW * j - w;
          real w2lambda = 1 - w1lambda;
L
liaogang 已提交
4600 4601 4602
          // calculate four position for bilinear interpolation
          const real* inPos = &inData[k * inputW + h * inImgW + w];
          real* outPos = &outData[k * outputW + i * outImgW + j];
L
liaogang 已提交
4603
          for (size_t c = 0; c < numChannels; ++c) {  // loop for channels
L
liaogang 已提交
4604
            // bilinear interpolation
L
liaogang 已提交
4605
            outPos[0] =
4606 4607 4608
                h2lambda * (w2lambda * inPos[0] + w1lambda * inPos[wid]) +
                h1lambda * (w2lambda * inPos[hid * inImgW] +
                            w1lambda * inPos[hid * inImgW + wid]);
L
liaogang 已提交
4609 4610
            inPos += inPosOffset;
            outPos += outPosOffset;
L
liaogang 已提交
4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622
          }
        }
      }
    }
  }
}

void CpuMatrix::bilinearBackward(const Matrix& out,
                                 const size_t outImgH,
                                 const size_t outImgW,
                                 const size_t inImgH,
                                 const size_t inImgW,
L
liaogang 已提交
4623 4624 4625
                                 const size_t numChannels,
                                 const real ratioH,
                                 const real ratioW) {
L
liaogang 已提交
4626 4627 4628 4629 4630
  CHECK(dynamic_cast<const CpuMatrix*>(&out));

  size_t inputW = getWidth();
  size_t inputH = getHeight();
  size_t outputW = out.getWidth();
L
liaogang 已提交
4631
  size_t batchSize = out.getHeight();
L
liaogang 已提交
4632 4633
  size_t inPosOffset = inImgH * inImgW;
  size_t outPosOffset = outImgH * outImgW;
L
liaogang 已提交
4634
  (void)(inputH);
L
liaogang 已提交
4635 4636 4637 4638 4639

  real* inGrad = getData();
  const real* outGrad = out.getData();

  if (inImgH == outImgH && inImgW == outImgW) {
L
liaogang 已提交
4640
    this->add(const_cast<Matrix&>(out));
L
liaogang 已提交
4641
  } else {
4642
    for (size_t k = 0; k < batchSize; ++k) {  // loop for batches
L
liaogang 已提交
4643 4644 4645
      for (size_t i = 0; i < outImgH; ++i) {  // loop for images
        size_t h = ratioH * i;
        size_t hid = (h < inImgH - 1) ? 1 : 0;
L
liaogang 已提交
4646 4647
        real h1lambda = ratioH * i - h;
        real h2lambda = 1 - h1lambda;
L
liaogang 已提交
4648 4649 4650
        for (size_t j = 0; j < outImgW; ++j) {
          size_t w = ratioW * j;
          size_t wid = (w < inImgW - 1) ? 1 : 0;
L
liaogang 已提交
4651 4652
          real w1lambda = ratioW * j - w;
          real w2lambda = 1 - w1lambda;
L
liaogang 已提交
4653 4654 4655

          real* inPos = &inGrad[k * inputW + h * inImgW + w];
          const real* outPos = &outGrad[k * outputW + i * outImgW + j];
L
liaogang 已提交
4656
          for (size_t c = 0; c < numChannels; ++c) {  // loop for channels
L
liaogang 已提交
4657 4658 4659 4660
            inPos[0] += h2lambda * w2lambda * outPos[0];
            inPos[wid] += h2lambda * w1lambda * outPos[0];
            inPos[hid * inImgW] += h1lambda * w2lambda * outPos[0];
            inPos[hid * inImgW + wid] += h1lambda * w1lambda * outPos[0];
L
liaogang 已提交
4661 4662
            inPos += inPosOffset;
            outPos += outPosOffset;
L
liaogang 已提交
4663 4664 4665 4666 4667 4668 4669
          }
        }
      }
    }
  }
}

C
chengduoZH 已提交
4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730
void CpuMatrix::vol2Col(real* data,
                        int channels,
                        int depth,
                        int height,
                        int width,
                        int filterD,
                        int filterH,
                        int filterW,
                        int strideD,
                        int strideH,
                        int strideW,
                        int paddingD,
                        int paddingH,
                        int paddingW) {
  real* outData = getData();
  int outHeight = (height + 2 * paddingH - filterH) / strideH + 1;
  int outWidth = (width + 2 * paddingW - filterW) / strideW + 1;
  int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1;

  int channelsCol = channels * filterD * filterH * filterW;
  for (int c = 0; c < channelsCol; ++c) {
    int wOffset = c % filterW;
    int hOffset = (c / filterW) % filterH;
    int dOffset = (c / filterW / filterH) % filterD;
    int cIn = c / filterW / filterH / filterD;
    for (int d = 0; d < outDepth; ++d) {
      for (int h = 0; h < outHeight; ++h) {
        for (int w = 0; w < outWidth; ++w) {
          int dPad = d * strideD - paddingD + dOffset;
          int hPad = h * strideH - paddingH + hOffset;
          int wPad = w * strideW - paddingW + wOffset;

          if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width &&
              dPad >= 0 && dPad < depth)
            outData[((c * outDepth + d) * outHeight + h) * outWidth + w] =
                data[((cIn * depth + dPad) * height + hPad) * width + wPad];
          else
            outData[((c * outDepth + d) * outHeight + h) * outWidth + w] = 0;
        }
      }
    }
  }
}

void CpuMatrix::col2Vol(real* trg,
                        int channels,
                        int depth,
                        int height,
                        int width,
                        int filterD,
                        int filterH,
                        int filterW,
                        int strideD,
                        int strideH,
                        int strideW,
                        int paddingD,
                        int paddingH,
                        int paddingW,
                        real alpha,
                        real beta) {
  real* src = getData();
C
chengduoZH 已提交
4731
  int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1;
C
chengduoZH 已提交
4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758
  int outHeight = (height + 2 * paddingH - filterH) / strideH + 1;
  int outWidth = (width + 2 * paddingW - filterW) / strideW + 1;
  int channelsCol = channels * filterD * filterH * filterW;
  for (int c = 0; c < channelsCol; ++c) {
    int wOffset = c % filterW;
    int hOffset = (c / filterW) % filterH;
    int dOffset = (c / filterW / filterH) % filterD;
    int cIm = c / filterW / filterH / filterD;
    for (int d = 0; d < outDepth; ++d) {
      for (int h = 0; h < outHeight; ++h) {
        for (int w = 0; w < outWidth; ++w) {
          int dPad = d * strideD - paddingD + dOffset;
          int hPad = h * strideH - paddingH + hOffset;
          int wPad = w * strideW - paddingW + wOffset;
          if (hPad >= 0 && hPad < height && wPad >= 0 && wPad < width &&
              dPad >= 0 && dPad < depth)
            trg[((cIm * depth + dPad) * height + hPad) * width + wPad] =
                alpha *
                    src[((c * outDepth + d) * outHeight + h) * outWidth + w] +
                beta *
                    trg[((cIm * depth + dPad) * height + hPad) * width + wPad];
        }
      }
    }
  }
}

Z
zhangjinchao01 已提交
4759 4760 4761 4762 4763 4764 4765 4766
////////////////////////////////////////////////////////////////
//               functions executed via cpu                   //
////////////////////////////////////////////////////////////////

void GpuMatrix::selectElements(Matrix& table, IVector& ids) {
  execViaCpu2(&CpuMatrix::selectElements, *this, table, ids);
}
}  // namespace paddle