Matrix.cpp 150.6 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

X
xzl 已提交
31
#include "NEONFunctions.h"
32
#include "paddle/function/GemmFunctor.h"
Z
zhangjinchao01 已提交
33 34 35 36 37 38 39 40 41 42 43 44
#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; }

45 46 47 48 49
Matrix::Matrix(MemoryHandlePtr memHandle,
               size_t height,
               size_t width,
               bool trans,
               bool use_gpu)
Z
zhangjinchao01 已提交
50
    : BaseMatrix(
51 52
          height,
          width,
Q
qijun 已提交
53
          memHandle ? (reinterpret_cast<real*>(memHandle->getBuf())) : nullptr,
54 55
          trans,
          use_gpu) {
Z
zhangjinchao01 已提交
56 57 58 59
  elementCnt_ = width * height;
  memoryHandle_ = memHandle;
}

60 61
Matrix::Matrix(
    real* data, size_t height, size_t width, bool trans, bool use_gpu)
Z
zhangjinchao01 已提交
62 63 64 65
    : BaseMatrix(height, width, data, trans, use_gpu) {
  elementCnt_ = width * height;
}

66 67 68 69 70 71
Matrix::Matrix(real* data,
               size_t height,
               size_t width,
               size_t stride,
               bool trans,
               bool use_gpu)
Z
zhangjinchao01 已提交
72 73 74 75
    : BaseMatrix(height, width, stride, data, trans, use_gpu) {
  elementCnt_ = width * height;
}

76 77 78 79 80
MatrixPtr Matrix::createSparseMatrix(real* data,
                                     int* row,
                                     int* col,
                                     size_t height,
                                     size_t width,
Z
zhangjinchao01 已提交
81 82
                                     size_t nnz, /* used to allocate space */
                                     SparseValueType valueType, /*value type*/
83 84
                                     SparseFormat format,
                                     bool trans,
Z
zhangjinchao01 已提交
85 86
                                     bool useGpu) {
  if (useGpu) {
87 88
    return std::make_shared<GpuSparseMatrix>(
        data, row, col, height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
89
  } else {
90 91
    return std::make_shared<CpuSparseMatrix>(
        data, row, col, height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
92 93 94
  }
}

95 96
MatrixPtr Matrix::createSparseMatrix(size_t height,
                                     size_t width,
Z
zhangjinchao01 已提交
97 98
                                     size_t nnz, /* used to allocate space */
                                     SparseValueType valueType, /*value type*/
99 100
                                     SparseFormat format,
                                     bool trans,
Z
zhangjinchao01 已提交
101 102
                                     bool useGpu) {
  if (useGpu) {
103 104
    return std::make_shared<GpuSparseMatrix>(
        height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
105
  } else {
106 107
    return std::make_shared<CpuSparseMatrix>(
        height, width, nnz, valueType, format, trans);
Z
zhangjinchao01 已提交
108 109 110
  }
}

111 112 113
MatrixPtr Matrix::create(MemoryHandlePtr memHandle,
                         size_t height,
                         size_t width,
Z
zhangjinchao01 已提交
114 115 116 117
                         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
118
                 std::dynamic_pointer_cast<CpuMemoryHandle>(memHandle)) {
Z
zhangjinchao01 已提交
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
    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);
  }
}

134 135
MatrixPtr Matrix::create(
    real* data, size_t height, size_t width, bool trans, bool useGpu) {
Z
zhangjinchao01 已提交
136 137 138 139 140 141 142
  if (useGpu) {
    return std::make_shared<GpuMatrix>(data, height, width, trans);
  } else {
    return std::make_shared<CpuMatrix>(data, height, width, trans);
  }
}

143 144 145 146 147 148
MatrixPtr Matrix::create(real* data,
                         size_t height,
                         size_t width,
                         size_t stride,
                         bool trans,
                         bool useGpu) {
Z
zhangjinchao01 已提交
149 150 151 152 153 154 155
  if (useGpu) {
    return std::make_shared<GpuMatrix>(data, height, width, stride, trans);
  } else {
    return std::make_shared<CpuMatrix>(data, height, width, stride, trans);
  }
}

156 157 158 159 160
MatrixPtr Matrix::createSparseMatrix(size_t height,
                                     size_t width,
                                     size_t nnz,
                                     SparseValueType valueType,
                                     bool trans,
Z
zhangjinchao01 已提交
161 162
                                     bool useGpu) {
  if (useGpu) {
163 164
    return std::make_shared<GpuSparseMatrix>(
        height, width, nnz, valueType, SPARSE_CSR, trans);
Z
zhangjinchao01 已提交
165
  } else {
166 167
    return std::make_shared<CpuSparseMatrix>(
        height, width, nnz, valueType, SPARSE_CSR, trans);
Z
zhangjinchao01 已提交
168 169 170
  }
}

171 172
void Matrix::resizeOrCreate(
    MatrixPtr& matrix, size_t height, size_t width, bool trans, bool useGpu) {
Z
zhangjinchao01 已提交
173 174 175
  if (!matrix) {
    matrix = Matrix::create(height, width, trans, useGpu);
  } else {
176
    CHECK_EQ(matrix->useGpu(), useGpu);
Z
zhangjinchao01 已提交
177 178 179 180
    matrix->resize(height, width);
  }
}

181 182 183 184
void Matrix::resizeOrCreateSparseMatrix(MatrixPtr& matrix,
                                        size_t height,
                                        size_t width,
                                        size_t nnz,
Z
zhangjinchao01 已提交
185
                                        SparseValueType valueType,
186 187
                                        SparseFormat format,
                                        bool trans,
Z
zhangjinchao01 已提交
188 189
                                        bool useGpu) {
  if (!matrix) {
190 191
    matrix = Matrix::createSparseMatrix(
        height, width, nnz, valueType, format, trans, useGpu);
Z
zhangjinchao01 已提交
192 193 194
  } else {
    CHECK(dynamic_cast<CpuSparseMatrix*>(matrix.get()) ||
          dynamic_cast<GpuSparseMatrix*>(matrix.get()));
195
    CHECK_EQ(matrix->useGpu(), useGpu);
Z
zhangjinchao01 已提交
196 197 198 199 200 201 202 203 204 205 206 207
    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_;
}

208 209 210
MatrixPtr Matrix::subMatrix(size_t startRow,
                            size_t endRow,
                            size_t startCol,
Z
zhangjinchao01 已提交
211 212 213 214 215 216 217
                            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,
218 219 220 221 222
                        endRow - startRow,
                        endCol - startCol,
                        getStride(),
                        trans_,
                        useGpu_);
Z
zhangjinchao01 已提交
223 224
}

225 226 227 228 229 230 231 232 233
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
234 235
GpuMatrix::GpuMatrix(size_t height, size_t width, bool trans)
    : Matrix(std::make_shared<GpuMemoryHandle>(height * width * sizeof(real)),
236 237 238 239
             height,
             width,
             trans,
             true) {}
Z
zhangjinchao01 已提交
240 241 242 243 244 245 246 247 248 249 250 251

GpuMatrix::~GpuMatrix() {}

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

void GpuMatrix::resetOne() {
  CHECK(data_ != NULL);
  one();
}
252

Z
zhangjinchao01 已提交
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278
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;
}

279 280 281 282 283 284 285 286 287 288 289 290
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
291 292 293
void GpuMatrix::accumulateColSum(Matrix& src) {
  CHECK_EQ(getWidth(), src.getWidth());
  CHECK_EQ(getHeight(), (size_t)1);
X
xuwei06 已提交
294
  sumCols(src, 1.0, 1.0);
Z
zhangjinchao01 已提交
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
}

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)) {
310 311
    hl_memcpy_host2device(
        data_, const_cast<real*>(src.getData()), sizeof(real) * elementCnt_);
Z
zhangjinchao01 已提交
312
  } else if (typeid(src) == typeid(GpuMatrix)) {
313 314
    hl_memcpy_device2device(
        data_, const_cast<real*>(src.getData()), sizeof(real) * elementCnt_);
Z
zhangjinchao01 已提交
315 316 317 318 319 320 321 322 323
  } else {
    LOG(FATAL) << "Wrong";
  }
}

void GpuMatrix::copyFrom(const Matrix& src, hl_stream_t stream) {
  CHECK(isContiguous());
  CHECK(src.isContiguous());
  CHECK(elementCnt_ == src.getElementCnt());
324 325 326 327
  hl_memcpy_async(this->getData(),
                  const_cast<real*>(src.getData()),
                  sizeof(real) * elementCnt_,
                  stream);
Z
zhangjinchao01 已提交
328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346
}

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);
}

347
void GpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) {
Z
zhangjinchao01 已提交
348 349 350 351 352
  size_t height = getHeight();
  size_t width = getWidth();
  CHECK_EQ(b.getWidth(), width);
  real* dst = getData();
  real* src = b.getData();
353
  const int* index = rowIndex.getData();
Z
zhangjinchao01 已提交
354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377
  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_),
378 379 380
                      height_,
                      width_,
                      true));
Z
zhangjinchao01 已提交
381 382 383 384 385 386 387
    return copy_T;
  } else {
    MatrixPtr copy_T(new GpuMatrix(data_, height_, width_, true));
    return copy_T;
  }
}

388
void GpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
Z
zhangjinchao01 已提交
389 390 391 392
  if (memAlloc) {
    matTrans = std::make_shared<GpuMatrix>(width_, height_);
  } else {
    CHECK(matTrans != NULL);
H
Haonan 已提交
393 394
    CHECK_EQ(matTrans->getHeight(), width_);
    CHECK_EQ(matTrans->getWidth(), height_);
Z
zhangjinchao01 已提交
395 396 397 398 399 400 401 402 403
  }
  real* dataTrans = matTrans->getData();
  real* data = getData();
  int lda = getStride();
  int ldc = matTrans->getStride();

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

404 405 406 407 408
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
409 410
    CHECK_EQ(matRot->getHeight(), width_);
    CHECK_EQ(matRot->getWidth(), height_);
411 412
  }

H
Haonan 已提交
413 414 415
  real* dataRot = matRot->getData();
  real* data = getData();
  hl_matrix_rotate(data, dataRot, height_, width_, clockWise);
416 417
}

L
lzhao4ever 已提交
418 419 420 421 422 423
MatrixPtr GpuMatrix::getInverse() {
  MatrixPtr matInv;
  inverse(matInv, true);
  return matInv;
}

424
void GpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) {
L
lzhao4ever 已提交
425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
  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
441 442 443 444 445
void GpuMatrix::addBias(Matrix& b, real scale) {
  CHECK(b.getHeight() == 1) << "the Bias should be a vector";
  BaseMatrix::addBias(b, scale);
}

446 447 448 449
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);
450 451
  hl_matrix_add_shared_bias(
      getData(), b.getData(), b.getWidth(), getHeight(), getWidth(), scale);
452 453
}

Z
zhangjinchao01 已提交
454
void GpuMatrix::collectBias(Matrix& a, real scale) {
455
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
456 457
  CHECK_EQ(getHeight(), (size_t)1);
  CHECK_EQ(width_, a.getWidth());
Q
qijun 已提交
458
  GpuSparseMatrix* sMatPtr = dynamic_cast<GpuSparseMatrix*>(&a);
Z
zhangjinchao01 已提交
459
  if (!sMatPtr) {
460
    sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
461 462 463
  } else {
    real* data = getData();
    hl_sparse_matrix_s A_d = sMatPtr->sMatrix_.get();
Q
qijun 已提交
464
    hl_sparse_matrix_column_sum(data, A_d, sMatPtr->getHeight(), width_, scale);
Z
zhangjinchao01 已提交
465
  }
466
#endif
Z
zhangjinchao01 已提交
467 468
}

469 470 471
void GpuMatrix::collectSharedBias(Matrix& a, real scale) {
  CHECK_EQ(getHeight(), (size_t)1);
  CHECK_EQ(a.getWidth() % getWidth(), 0UL);
472 473
  hl_matrix_collect_shared_bias(
      getData(), a.getData(), getWidth(), a.getHeight(), a.getWidth(), scale);
474 475
}

Z
zhangjinchao01 已提交
476 477 478 479 480 481 482 483 484 485 486 487 488 489
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
490 491 492 493 494 495 496 497 498 499 500 501 502 503
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
504
/* this = scaleAB*(a*b) +  scaleT*this */
505 506 507
void GpuMatrix::mul(const GpuMatrix& a,
                    const GpuMatrix& b,
                    real scaleAB,
Z
zhangjinchao01 已提交
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538
                    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;

539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556
  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
557
                    real scaleT) {
558
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574
  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_;
575 576 577 578 579 580 581 582 583 584
  hl_matrix_csr_mul_dense(A_d,
                          transA,
                          B_d,
                          HPPL_OP_N,
                          C_d,
                          height_,
                          width_,
                          b.height_,
                          scaleAB,
                          scaleT);
585
#endif
586 587 588 589 590
}

void GpuMatrix::mul(const GpuMatrix& a,
                    const GpuSparseMatrix& b,
                    real scaleAB,
Z
zhangjinchao01 已提交
591
                    real scaleT) {
592
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608
  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) {
609 610 611 612 613 614 615 616 617 618
    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
619
  } else {
620 621 622 623 624 625 626 627 628 629
    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
630
  }
631
#endif
Z
zhangjinchao01 已提交
632 633 634
}

/* this = a*b */
635
void GpuMatrix::mul(const Matrix& a, const Matrix& b) { mul(a, b, 1.0, 0.0); }
Z
zhangjinchao01 已提交
636

637 638
void GpuMatrix::mul(const Matrix& a,
                    const Matrix& b,
639
                    real scaleAB,
Z
zhangjinchao01 已提交
640
                    real scaleT) {
641 642 643 644
  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
645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679

  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) {
680
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
681 682 683 684 685 686 687 688 689 690 691
  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();

692 693 694 695 696 697 698 699
  hl_matrix_select_rows(a,
                        stride_,
                        table.getData(),
                        table.stride_,
                        index,
                        numSamples,
                        tableSize,
                        dim);
Z
zhangjinchao01 已提交
700 701 702 703
#endif
}

void GpuMatrix::addToRows(Matrix& table, IVector& ids) {
704
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
705 706 707 708 709 710 711 712 713 714 715
  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();

716 717 718 719 720 721 722 723
  hl_matrix_add_to_rows(table.getData(),
                        table.stride_,
                        a,
                        stride_,
                        index,
                        numSamples,
                        tableSize,
                        dim);
Z
zhangjinchao01 已提交
724 725 726 727 728 729
#endif
}

void GpuMatrix::colMerge(Matrix& src) {
  CHECK(src.height_ == height_);
  if (!trans_ && !src.trans_) {
730
    sumRows(src, /* scaleSum= */ 1, /* scaleDest= */ 0);
Z
zhangjinchao01 已提交
731 732 733 734 735 736 737 738 739
  } else {
    LOG(FATAL) << "Is not supported";
  }
}

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

740
  sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0);
Z
zhangjinchao01 已提交
741 742 743 744 745 746 747 748 749 750
}

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) {
751
#ifdef PADDLE_WITH_CUDA
Z
zhangjinchao01 已提交
752 753 754 755 756
  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
757
  CHECK_EQ(maxVal.getWidth(), beam);
Z
zhangjinchao01 已提交
758

759 760 761 762 763 764 765
  hl_matrix_top_k(maxVal.getData(),
                  maxVal.getStride(),
                  maxIds.getData(),
                  this->getData(),
                  this->getStride(),
                  this->getWidth(),
                  beam,
Z
zhangjinchao01 已提交
766 767 768 769 770 771 772 773 774 775 776
                  numSamples);
#endif
}

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

  max.maxCols(*this);
}

777 778 779 780
void GpuMatrix::colMax(IVector& maxIds, Matrix& maxVal) {
  LOG(FATAL) << "Is not supported";
}

781 782 783
void GpuMatrix::maxoutForward(Matrix& a,
                              IVector& id,
                              size_t channels,
784 785 786 787 788 789 790
                              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
791
  const real* input = a.getData();
792 793 794
  real* output = getData();
  int* idForGpu = id.getData();

795 796
  hl_maxout_forward(
      input, output, idForGpu, batchSize, size, size / channels, groups);
797 798
}

799 800 801
void GpuMatrix::maxoutBackward(Matrix& a,
                               IVector& id,
                               size_t channels,
802 803 804 805 806 807 808
                               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
809
  real* input = getData();
810 811 812
  const real* output = a.getData();
  const int* idForGpu = id.getData();

813 814
  hl_maxout_backward(
      input, output, idForGpu, batchSize, size, size / channels, groups);
815 816
}

Z
zhangjinchao01 已提交
817
/*calulate the error of classification */
818 819 820 821 822 823 824 825 826 827 828 829 830
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
831 832
      << "Matrix dimensions are not equal";

833 834 835 836 837 838 839 840 841 842 843
  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
844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879
}

/* 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_);
}

880 881
void GpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output,
                                               IVector& label,
Z
zhangjinchao01 已提交
882 883 884 885 886
                                               real alpha) {
  LOG(FATAL) << "Not implemented";
}

void GpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& outputV,
887 888
                                                 IVector& label,
                                                 real alpha) {
Z
zhangjinchao01 已提交
889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913
  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
914
  hl_sequence_softmax_forward(inputData, outputData, starts, numSequences);
Z
zhangjinchao01 已提交
915 916 917 918 919 920 921 922 923 924 925 926 927
}

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
928
  hl_matrix_softmax_derivative(grad_d, output_d, sftmaxSum_d, height_, width_);
Z
zhangjinchao01 已提交
929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954
}

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";
  }

955 956 957 958
  BaseMatrix::sumOfSquaredDiffs(output,
                                label,
                                /* scaleSum= */ 1,
                                /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
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 1026
}

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
1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
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,
1039
                               MatrixPtr maskMatP) {
Z
zhangjinchao01 已提交
1040 1041 1042
  CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal";

  real* inputData = inputMat.getData();
X
xzl 已提交
1043
  real* maskData = NULL;
Z
zhangjinchao01 已提交
1044
  size_t frameNum = inputMat.getHeight();
1045
  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
Z
zhangjinchao01 已提交
1046 1047 1048
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputH * outputW * channels);

1049
  if (maskMatP != NULL) {
X
xzl 已提交
1050 1051 1052 1053 1054
    CHECK(maskMatP->useGpu_ == true) << "Matrix type are not equal";
    CHECK(outputH * outputW * channels == maskMatP->getWidth());
    maskData = maskMatP->getData();
  }

1055 1056 1057
  hl_maxpool_forward(frameNum,
                     inputData,
                     channels,
1058 1059
                     imgSizeH,
                     imgSizeW,
1060 1061 1062 1063 1064 1065 1066 1067 1068
                     outputH,
                     outputW,
                     sizeX,
                     sizeY,
                     strideH,
                     strideW,
                     paddingH,
                     paddingW,
                     data_,
X
xzl 已提交
1069
                     getStride(),
1070
                     maskData);
1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087
}

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
1088 1089 1090 1091 1092 1093 1094 1095 1096
  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;
1097
  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
Z
zhangjinchao01 已提交
1098 1099 1100 1101
  CHECK(height_ == inputMat.getHeight());
  CHECK(outGrad.getHeight() == outV.getHeight() &&
        outGrad.getWidth() == outV.getWidth());

1102 1103 1104 1105 1106
  hl_maxpool_backward(frameNum,
                      inputData,
                      outData,
                      outDiff,
                      channels,
1107 1108
                      imgSizeH,
                      imgSizeW,
1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119
                      outputH,
                      outputW,
                      sizeX,
                      sizeY,
                      strideH,
                      strideW,
                      paddingH,
                      paddingW,
                      scaleTargets,
                      scaleOutput,
                      data_,
Q
qijun 已提交
1120
                      outGrad.getStride());
Z
zhangjinchao01 已提交
1121 1122
}

1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133
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,
1134 1135
                               size_t paddingW,
                               bool excludeMode) {
Z
zhangjinchao01 已提交
1136 1137 1138 1139
  CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal";

  real* inputData = inputMat.getData();
  size_t frameNum = inputMat.getHeight();
1140
  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
Z
zhangjinchao01 已提交
1141 1142 1143
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputH * outputW * channels);

1144 1145 1146
  hl_avgpool_forward(frameNum,
                     inputData,
                     channels,
1147 1148
                     imgSizeH,
                     imgSizeW,
1149 1150 1151 1152 1153 1154 1155 1156 1157
                     outputH,
                     outputW,
                     sizeX,
                     sizeY,
                     strideH,
                     strideW,
                     paddingH,
                     paddingW,
                     data_,
1158 1159
                     getStride(),
                     excludeMode);
1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173
}

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,
1174 1175
                                size_t paddingW,
                                bool excludeMode) {
Z
zhangjinchao01 已提交
1176 1177 1178 1179 1180
  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;
1181
  CHECK(imgSizeH * imgSizeW * channels == width_);
Z
zhangjinchao01 已提交
1182 1183 1184
  CHECK(height_ == outGrad.getHeight());
  CHECK(outGrad.getWidth() == outputH * outputW * channels);

1185 1186 1187
  hl_avgpool_backward(frameNum,
                      outDiff,
                      channels,
1188 1189
                      imgSizeH,
                      imgSizeW,
1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200
                      outputH,
                      outputW,
                      sizeX,
                      sizeY,
                      strideH,
                      strideW,
                      paddingH,
                      paddingW,
                      scaleTargets,
                      scaleOutput,
                      data_,
1201 1202
                      outGrad.getStride(),
                      excludeMode);
Z
zhangjinchao01 已提交
1203 1204
}

C
chengduoZH 已提交
1205
void GpuMatrix::maxPool3DForward(Matrix& inputMat,
C
chengduoZH 已提交
1206
                                 Matrix& maxPoolIdx,
C
chengduoZH 已提交
1207
                                 size_t channels,
C
chengduoZH 已提交
1208 1209 1210
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
1211 1212 1213
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
1214 1215 1216 1217 1218 1219 1220 1221 1222
                                 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
1223
  CHECK(inputMat.useGpu_) << "Matrix type are not correct";
C
chengduoZH 已提交
1224 1225

  real* inputData = inputMat.getData();
C
chengduoZH 已提交
1226
  real* maxPoolIdxData = maxPoolIdx.getData();
C
chengduoZH 已提交
1227
  size_t num = inputMat.getHeight();
1228
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
C
chengduoZH 已提交
1229 1230 1231 1232 1233 1234
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputD * outputH * outputW * channels);

  hl_maxpool3D_forward(num,
                       inputData,
                       channels,
1235 1236 1237
                       imgSizeD,
                       imgSizeH,
                       imgSizeW,
C
chengduoZH 已提交
1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249
                       outputD,
                       outputH,
                       outputW,
                       sizeZ,
                       sizeY,
                       sizeX,
                       strideD,
                       strideH,
                       strideW,
                       paddingD,
                       paddingH,
                       paddingW,
C
chengduoZH 已提交
1250
                       getData(),
C
chengduoZH 已提交
1251
                       maxPoolIdxData,
C
chengduoZH 已提交
1252 1253 1254
                       getStride());
}

C
chengduoZH 已提交
1255 1256
void GpuMatrix::maxPool3DBackward(Matrix& outGrad,
                                  Matrix& maxPoolIdx,
C
chengduoZH 已提交
1257 1258 1259
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
1260 1261 1262
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
1263 1264 1265 1266 1267 1268 1269 1270
                                  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
1271 1272 1273
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
C
chengduoZH 已提交
1274
  CHECK(outGrad.useGpu_ && maxPoolIdx.useGpu_) << "Matrix type are not equal";
C
chengduoZH 已提交
1275 1276

  real* outDiff = outGrad.getData();
C
chengduoZH 已提交
1277 1278 1279
  real* maxPoolIdxData = maxPoolIdx.getData();
  size_t frameNum = getHeight();
  size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
1280
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
C
chengduoZH 已提交
1281 1282
  CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() &&
        outGrad.getWidth() == maxPoolIdx.getWidth());
C
chengduoZH 已提交
1283 1284 1285 1286

  hl_maxpool3D_backward(frameNum,
                        outDiff,
                        channels,
1287 1288 1289
                        imgSizeD,
                        imgSizeH,
                        imgSizeW,
C
chengduoZH 已提交
1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303
                        outputD,
                        outputH,
                        outputW,
                        sizeZ,
                        sizeY,
                        sizeX,
                        strideD,
                        strideH,
                        strideW,
                        paddingD,
                        paddingH,
                        paddingW,
                        scaleTargets,
                        scaleOutput,
C
chengduoZH 已提交
1304
                        getData(),
C
chengduoZH 已提交
1305
                        maxPoolIdxData,
C
chengduoZH 已提交
1306 1307 1308 1309
                        outGrad.getStride());
}

void GpuMatrix::avgPool3DForward(Matrix& inputMat,
C
chengduoZH 已提交
1310
                                 size_t channels,
C
chengduoZH 已提交
1311 1312 1313
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
1314 1315 1316
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
1317 1318 1319 1320 1321 1322 1323 1324 1325
                                 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
1326
  CHECK(inputMat.useGpu_) << "Matrix type are not equal";
C
chengduoZH 已提交
1327 1328 1329

  real* inputData = inputMat.getData();
  size_t frameNum = inputMat.getHeight();
1330
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
C
chengduoZH 已提交
1331 1332 1333 1334 1335 1336
  CHECK(height_ == inputMat.getHeight());
  CHECK(width_ == outputD * outputH * outputW * channels);

  hl_avgpool3D_forward(frameNum,
                       inputData,
                       channels,
1337 1338 1339
                       imgSizeD,
                       imgSizeH,
                       imgSizeW,
C
chengduoZH 已提交
1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351
                       outputD,
                       outputH,
                       outputW,
                       sizeZ,
                       sizeY,
                       sizeX,
                       strideD,
                       strideH,
                       strideW,
                       paddingD,
                       paddingH,
                       paddingW,
C
chengduoZH 已提交
1352
                       getData(),
C
chengduoZH 已提交
1353 1354 1355 1356 1357 1358 1359
                       getStride());
}

void GpuMatrix::avgPool3DBackward(Matrix& outGrad,
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
1360 1361 1362
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
1363 1364 1365 1366 1367 1368 1369 1370
                                  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
1371 1372 1373 1374
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
  CHECK(outGrad.useGpu_) << "Matrix type are not equal";
C
chengduoZH 已提交
1375 1376 1377 1378

  real* outDiff = outGrad.getData();
  size_t frameNum = outGrad.getHeight();
  size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
1379
  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_);
C
chengduoZH 已提交
1380 1381 1382 1383 1384 1385
  CHECK(height_ == outGrad.getHeight());
  CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels);

  hl_avgpool3D_backward(frameNum,
                        outDiff,
                        channels,
1386 1387 1388
                        imgSizeD,
                        imgSizeH,
                        imgSizeW,
C
chengduoZH 已提交
1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402
                        outputD,
                        outputH,
                        outputW,
                        sizeZ,
                        sizeY,
                        sizeX,
                        strideD,
                        strideH,
                        strideW,
                        paddingD,
                        paddingH,
                        paddingW,
                        scaleTargets,
                        scaleOutput,
C
chengduoZH 已提交
1403
                        getData(),
C
chengduoZH 已提交
1404 1405 1406
                        outGrad.getStride());
}

1407 1408
void GpuMatrix::maxSequenceForward(Matrix& input,
                                   const IVector& sequence,
Z
zhangjinchao01 已提交
1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424
                                   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());

1425 1426
  hl_max_sequence_forward(
      inputData, starts, outData, maxIndex, numSequences, dim);
Z
zhangjinchao01 已提交
1427 1428
}

1429 1430
void GpuMatrix::maxSequenceBackward(Matrix& outputGrad,
                                    const IVector& sequence,
Z
zhangjinchao01 已提交
1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455
                                    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
1456 1457 1458
  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
1459
  real* output = getData();
Q
qijun 已提交
1460
  hl_param_relu_forward(output, input, w, numElements, numSamples, partial_sum);
Z
zhangjinchao01 已提交
1461 1462 1463 1464 1465 1466 1467 1468 1469 1470
}

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
1471 1472 1473
  size_t paraSize = this->getHeight() * this->getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
1474 1475
  hl_param_relu_backward_w(
      wgrad, ograd, input, numElements, numSamples, partial_sum);
Z
zhangjinchao01 已提交
1476 1477 1478 1479 1480 1481 1482 1483 1484
}

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
1485 1486 1487
  size_t paraSize = W.getHeight() * W.getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
  size_t partial_sum = numElements / paraSize;
1488 1489
  hl_param_relu_backward_diff(
      ograd, input, w, diff, numElements, numSamples, partial_sum);
Z
zhangjinchao01 已提交
1490 1491 1492 1493 1494 1495
}

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

L
liaogang 已提交
1496 1497 1498 1499 1500
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
1501 1502 1503
                                const size_t numChannels,
                                const real ratioH,
                                const real ratioW) {
L
liaogang 已提交
1504 1505 1506 1507 1508 1509 1510 1511
  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();
1512
  const real* inData = in.getData();
L
liaogang 已提交
1513 1514 1515 1516

  if (inImgH == outImgW && inImgW == outImgW) {
    this->copyFrom(in);
  } else {
1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529
    hl_bilinear_forward(inData,
                        inImgH,
                        inImgW,
                        inputH,
                        inputW,
                        outData,
                        outImgH,
                        outImgW,
                        outputH,
                        outputW,
                        numChannels,
                        ratioH,
                        ratioW);
L
liaogang 已提交
1530 1531 1532 1533 1534 1535 1536 1537
  }
}

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
1538 1539 1540
                                 const size_t numChannels,
                                 const real ratioH,
                                 const real ratioW) {
L
liaogang 已提交
1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551
  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
1552
    this->add(const_cast<Matrix&>(out));
L
liaogang 已提交
1553
  } else {
1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566
    hl_bilinear_backward(inGrad,
                         inImgH,
                         inImgW,
                         inputH,
                         inputW,
                         outGrad,
                         outImgH,
                         outImgW,
                         outputH,
                         outputW,
                         numChannels,
                         ratioH,
                         ratioW);
L
liaogang 已提交
1567 1568 1569
  }
}

1570
void GpuMatrix::multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) {
1571
#ifdef PADDLE_WITH_CUDA
1572 1573 1574 1575 1576 1577 1578 1579 1580
  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";
1581

1582 1583 1584 1585 1586
  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_);
1587
#endif
1588 1589
}

1590
void GpuMatrix::multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) {
1591
#ifdef PADDLE_WITH_CUDA
1592 1593
  GpuMatrix* outputPtr = dynamic_cast<GpuMatrix*>(&output);
  auto labelPtr = dynamic_cast<GpuSparseMatrix*>(&label);
H
Haonan 已提交
1594

1595 1596 1597 1598 1599 1600
  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";
1601

1602 1603 1604 1605 1606
  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_);
1607
#endif
1608 1609
}

C
chengduoZH 已提交
1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670
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
1671
                    getData(),
C
chengduoZH 已提交
1672 1673 1674
                    alpha,
                    beta);
}
C
chengduoZH 已提交
1675

Z
zhangjinchao01 已提交
1676 1677 1678 1679 1680 1681
/**
 * CpuMatrix
 */

CpuMatrix::CpuMatrix(size_t height, size_t width, bool trans)
    : Matrix(std::make_shared<CpuMemoryHandle>(height * width * sizeof(real)),
1682 1683 1684 1685
             height,
             width,
             trans,
             false) {}
Z
zhangjinchao01 已提交
1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706

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());
1707 1708
    hl_memcpy_device2host(
        data_, const_cast<real*>(src.getData()), sizeof(real) * elementCnt_);
1709 1710
  } else if (typeid(src) == typeid(CpuMatrix) ||
             typeid(src) == typeid(SharedCpuMatrix)) {
Z
zhangjinchao01 已提交
1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772
    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)) {
1773 1774 1775 1776
    hl_memcpy_async(this->getData(),
                    const_cast<real*>(src.getData()),
                    sizeof(real) * elementCnt_,
                    stream);
1777 1778
    // There is a need to add synchronization to ensure that the data is copied.
    hl_stream_synchronize(stream);
Z
zhangjinchao01 已提交
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
  } 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];
  }
}

1817
void CpuMatrix::copyByRowIndex(Matrix& b, const IVector& rowIndex) {
Z
zhangjinchao01 已提交
1818 1819 1820
  size_t height = getHeight();
  size_t width = getWidth();
  CHECK_EQ(b.getWidth(), width);
1821
  const int* index = rowIndex.getData();
Z
zhangjinchao01 已提交
1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 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 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879
  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);

1880
  sumCols(src, /* scaleSum= */ 1, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896
}

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>(
1897 1898 1899 1900
        std::dynamic_pointer_cast<CpuMemoryHandle>(memoryHandle_),
        height_,
        width_,
        true);
Z
zhangjinchao01 已提交
1901 1902 1903 1904 1905 1906
  } else {
    MatrixPtr copy_T(new CpuMatrix(data_, height_, width_, true));
    return copy_T;
  }
}

1907
void CpuMatrix::transpose(MatrixPtr& matTrans, bool memAlloc) {
Z
zhangjinchao01 已提交
1908 1909 1910 1911
  if (memAlloc) {
    matTrans = std::make_shared<CpuMatrix>(width_, height_);
  } else {
    CHECK(matTrans != NULL);
H
Haonan 已提交
1912 1913
    CHECK_EQ(matTrans->getHeight(), width_);
    CHECK_EQ(matTrans->getWidth(), height_);
Z
zhangjinchao01 已提交
1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926
  }
  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];
    }
  }
}

1927 1928 1929 1930 1931
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
1932 1933
    CHECK_EQ(matRot->getHeight(), width_);
    CHECK_EQ(matRot->getWidth(), height_);
1934 1935 1936 1937 1938 1939 1940
  }
  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
1941
        dataRot[j * height_ + i] = data[(height_ - i - 1) * width_ + j];
1942
      } else {
H
Haonan 已提交
1943
        dataRot[j * height_ + i] = data[i * width_ + (width_ - j - 1)];
1944 1945 1946 1947 1948
      }
    }
  }
}

L
lzhao4ever 已提交
1949 1950 1951 1952 1953 1954
MatrixPtr CpuMatrix::getInverse() {
  MatrixPtr matInv;
  inverse(matInv, true);
  return matInv;
}

1955
void CpuMatrix::inverse(MatrixPtr& matInv, bool memAlloc) {
L
lzhao4ever 已提交
1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988
  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
1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000
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,
2001
                               MatrixPtr maskMatP) {
Z
zhangjinchao01 已提交
2002 2003
  real* inputData = inputMat.getData();
  real* outData = data_;
X
xzl 已提交
2004
  real* maskData = NULL;
Z
zhangjinchao01 已提交
2005
  size_t num = inputMat.getHeight();
2006 2007 2008
  size_t inLength = imgSizeH * imgSizeW;
  size_t outLength = outputH * outputW;
  CHECK(inLength == inputMat.getWidth() / channels);
2009
  CHECK_EQ(num, this->getHeight());
2010
  CHECK_EQ(channels * outLength, this->getWidth());
Q
qijun 已提交
2011
  size_t outStride = getStride();
Z
zhangjinchao01 已提交
2012

2013
  if (maskMatP != NULL) {
X
xzl 已提交
2014 2015 2016 2017
    maskData = maskMatP->getData();
    CHECK_EQ(channels * outLength, maskMatP->getWidth());
  }

Z
zhangjinchao01 已提交
2018
  /* initialize the data_ */
Q
qijun 已提交
2019 2020
  for (size_t i = 0; i < height_; i++) {
    for (size_t j = 0; j < width_; j++) {
Q
qijun 已提交
2021
      outData[i * outStride + j] = -(real)FLT_MAX;
Q
qijun 已提交
2022
    }
Z
zhangjinchao01 已提交
2023 2024 2025
  }

  /* pool max one by one */
Q
qijun 已提交
2026 2027
  for (size_t n = 0; n < num; ++n) {  // frame by frame
    if (!isContiguous()) {
Q
qijun 已提交
2028
      outData = data_ + n * outStride;
Q
qijun 已提交
2029
    }
Z
zhangjinchao01 已提交
2030 2031
    for (size_t c = 0; c < channels; ++c) {  // channel by channel
      for (size_t ph = 0; ph < outputH; ++ph) {
2032 2033 2034
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2035
        for (size_t pw = 0; pw < outputW; ++pw) {
2036
          int wstart = pw * strideW - paddingW;
2037
          int wend = std::min(wstart + sizeX, imgSizeW);
2038
          wstart = std::max(wstart, 0);
X
xzl 已提交
2039
          if (maskData == NULL) {
X
xzl 已提交
2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053
            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
2054 2055 2056 2057 2058
            }
          }
        }
      }
      // compute offset
2059 2060
      inputData += inLength;
      outData += outLength;
X
xzl 已提交
2061

X
xzl 已提交
2062
      if (maskData != NULL) maskData += outLength;
Z
zhangjinchao01 已提交
2063 2064 2065 2066
    }
  }
}

2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081
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
2082
  size_t num = image.getHeight();
2083 2084 2085 2086
  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
2087 2088 2089 2090 2091 2092 2093 2094
  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
2095 2096 2097 2098 2099

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

Z
zhangjinchao01 已提交
2100
  for (size_t n = 0; n < num; ++n) {
Q
qijun 已提交
2101
    if (!outV.isContiguous()) {
Q
qijun 已提交
2102 2103
      otData = origOutData + n * outStride;
      otGrad = origOutGrad + n * outStride;
Q
qijun 已提交
2104
    }
Z
zhangjinchao01 已提交
2105 2106
    for (size_t c = 0; c < channels; ++c) {
      for (size_t ph = 0; ph < outputH; ++ph) {
2107 2108 2109
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2110
        for (size_t pw = 0; pw < outputW; ++pw) {
2111 2112 2113 2114 2115
          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
2116 2117 2118
              tgtGrad[h * imgSizeW + w] =
                  scaleTargets * tgtGrad[h * imgSizeW + w] +
                  scaleOutput * otGrad[ph * outputW + pw] *
2119
                      (inData[h * imgSizeW + w] == otData[ph * outputW + pw]);
Z
zhangjinchao01 已提交
2120 2121 2122 2123 2124
            }
          }
        }
      }
      // offset
2125 2126 2127 2128
      inData += inLength;
      tgtGrad += inLength;
      otData += outLength;
      otGrad += outLength;
Z
zhangjinchao01 已提交
2129 2130 2131 2132
    }
  }
}

2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143
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,
2144 2145
                               size_t paddingW,
                               bool excludeMode) {
Z
zhangjinchao01 已提交
2146 2147
  // The main loop
  size_t num = input.getHeight();
2148 2149 2150 2151
  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
2152 2153 2154 2155
  real* tgtData = data_;
  real* inData = input.getData();

  for (size_t n = 0; n < num; ++n) {
Q
qijun 已提交
2156 2157 2158
    if (!isContiguous()) {
      tgtData = data_ + n * getStride();
    }
Z
zhangjinchao01 已提交
2159 2160
    for (size_t c = 0; c < channels; ++c) {
      for (size_t ph = 0; ph < outputH; ++ph) {
2161 2162 2163
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2164
        for (size_t pw = 0; pw < outputW; ++pw) {
2165
          int wstart = pw * strideW - paddingW;
2166
          int wend = std::min(wstart + sizeX, imgSizeW);
2167
          wstart = std::max(wstart, 0);
Z
zhangjinchao01 已提交
2168
          tgtData[ph * outputW + pw] = 0;  // clear
2169 2170
          for (int h = hstart; h < hend; ++h) {
            for (int w = wstart; w < wend; ++w) {
2171
              tgtData[ph * outputW + pw] += inData[h * imgSizeW + w];
Z
zhangjinchao01 已提交
2172 2173
            }
          }
2174 2175
          int poolSize =
              excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX;
2176
          CHECK(poolSize);
2177
          tgtData[ph * outputW + pw] /= poolSize;
Z
zhangjinchao01 已提交
2178 2179 2180
        }
      }
      // compute offset
2181 2182
      inData += inLength;
      tgtData += outLength;
Z
zhangjinchao01 已提交
2183 2184 2185 2186
    }
  }
}

2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198
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,
2199 2200
                                size_t paddingW,
                                bool excludeMode) {
Z
zhangjinchao01 已提交
2201 2202
  size_t num = input.getHeight();
  size_t channels = input.getWidth() / outputH / outputW;
2203 2204 2205
  size_t inLength = imgSizeH * imgSizeW;
  size_t outLength = outputH * outputW;
  CHECK(inLength * channels == getWidth());
Z
zhangjinchao01 已提交
2206 2207 2208 2209
  real* inData = input.getData();
  real* outData = getData();

  for (size_t n = 0; n < num; ++n) {
Q
qijun 已提交
2210 2211 2212
    if (!input.isContiguous()) {
      inData = input.getData() + n * input.getStride();
    }
Z
zhangjinchao01 已提交
2213 2214
    for (size_t c = 0; c < channels; ++c) {
      for (size_t ph = 0; ph < outputH; ++ph) {
2215 2216 2217
        int hstart = ph * strideH - paddingH;
        int hend = std::min(hstart + sizeY, imgSizeH);
        hstart = std::max(hstart, 0);
Z
zhangjinchao01 已提交
2218
        for (size_t pw = 0; pw < outputW; ++pw) {
2219
          int wstart = pw * strideW - paddingW;
2220
          int wend = std::min(wstart + sizeX, imgSizeW);
2221
          wstart = std::max(wstart, 0);
2222 2223
          int poolSize =
              excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX;
2224 2225 2226 2227 2228
          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
2229 2230 2231 2232 2233
            }
          }
        }
      }
      // offset
2234 2235
      outData += inLength;
      inData += outLength;
Z
zhangjinchao01 已提交
2236 2237 2238 2239
    }
  }
}

C
chengduoZH 已提交
2240
void CpuMatrix::maxPool3DForward(Matrix& inputMat,
C
chengduoZH 已提交
2241
                                 Matrix& maxPoolIdx,
C
chengduoZH 已提交
2242
                                 size_t channels,
C
chengduoZH 已提交
2243 2244 2245
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
2246 2247 2248
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
2249 2250 2251 2252 2253 2254 2255 2256 2257 2258
                                 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
2259
  real* outData = getData();
C
chengduoZH 已提交
2260
  real* maxPoolIdxData = maxPoolIdx.getData();
C
chengduoZH 已提交
2261
  size_t num = inputMat.getHeight();
2262 2263 2264
  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
  size_t outLength = outputH * outputW * outputD;
  CHECK(inLength == inputMat.getWidth() / channels);
C
chengduoZH 已提交
2265
  CHECK_EQ(num, this->getHeight());
2266
  CHECK_EQ(channels * outLength, this->getWidth());
C
chengduoZH 已提交
2267 2268 2269 2270 2271 2272
  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
2273
      maxPoolIdxData[(i)*outStride + j] = -1;
C
chengduoZH 已提交
2274 2275 2276 2277 2278 2279
    }
  }

  /* pool max one by one */
  for (size_t n = 0; n < num; ++n) {  // frame by frame
    if (!isContiguous()) {
C
chengduoZH 已提交
2280
      outData = getData() + n * outStride;
C
chengduoZH 已提交
2281
      maxPoolIdxData = maxPoolIdx.getData() + n * outStride;
C
chengduoZH 已提交
2282 2283 2284
    }
    for (size_t c = 0; c < channels; ++c) {  // channel by channel
      for (size_t pd = 0; pd < outputD; ++pd) {
2285 2286 2287
        int dstart = pd * strideD - paddingD;
        int dend = std::min(dstart + sizeZ, imgSizeD);
        dstart = std::max(dstart, 0);
C
chengduoZH 已提交
2288
        for (size_t ph = 0; ph < outputH; ++ph) {
2289 2290 2291
          int hstart = ph * strideH - paddingH;
          int hend = std::min(hstart + sizeY, imgSizeH);
          hstart = std::max(hstart, 0);
C
chengduoZH 已提交
2292 2293
          for (size_t pw = 0; pw < outputW; ++pw) {
            int wstart = pw * strideW - paddingW;
2294
            int wend = std::min(wstart + sizeX, imgSizeW);
C
chengduoZH 已提交
2295
            wstart = std::max(wstart, 0);
C
chengduoZH 已提交
2296
            int maxIdx = -1;
C
chengduoZH 已提交
2297
            real maxOutData = outData[(pd * outputH + ph) * outputW + pw];
C
chengduoZH 已提交
2298 2299 2300
            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
2301
                  if (maxOutData <
2302 2303 2304
                      inputData[(d * imgSizeH + h) * imgSizeW + w]) {
                    maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w];
                    maxIdx = (d * imgSizeH + h) * imgSizeW + w;
C
chengduoZH 已提交
2305
                  }
C
chengduoZH 已提交
2306 2307 2308
                }
              }
            }
C
chengduoZH 已提交
2309
            outData[(pd * outputH + ph) * outputW + pw] = maxOutData;
C
chengduoZH 已提交
2310
            maxPoolIdxData[(pd * outputH + ph) * outputW + pw] = maxIdx;
C
chengduoZH 已提交
2311 2312 2313 2314
          }
        }
      }
      // compute offset
2315 2316 2317
      inputData += inLength;
      outData += outLength;
      maxPoolIdxData += outLength;
C
chengduoZH 已提交
2318 2319 2320 2321
    }
  }
}

C
chengduoZH 已提交
2322 2323
void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
                                  Matrix& maxPoolIdx,
C
chengduoZH 已提交
2324 2325 2326
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
2327 2328 2329
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
2330 2331 2332 2333 2334 2335 2336 2337
                                  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
2338 2339 2340
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
C
chengduoZH 已提交
2341
  size_t num = getHeight();
2342 2343 2344
  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
  size_t outLength = outputH * outputW * outputD;
  size_t channels = size_t(width_ / inLength);
C
chengduoZH 已提交
2345 2346
  CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() &&
        maxPoolIdx.getWidth() == outGrad.getWidth());
C
chengduoZH 已提交
2347

C
chengduoZH 已提交
2348
  real* tgtGrad = getData();
C
chengduoZH 已提交
2349
  real* otGrad = outGrad.getData();
C
chengduoZH 已提交
2350 2351
  real* maxPoolIdxData = maxPoolIdx.getData();
  size_t outStride = outGrad.getStride();
C
chengduoZH 已提交
2352 2353

  for (size_t n = 0; n < num; ++n) {
C
chengduoZH 已提交
2354
    if (!outGrad.isContiguous()) {
C
chengduoZH 已提交
2355
      otGrad = outGrad.getData() + n * outStride;
C
chengduoZH 已提交
2356
      maxPoolIdxData = maxPoolIdx.getData() + n * outStride;
C
chengduoZH 已提交
2357 2358 2359 2360 2361
    }
    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
2362 2363 2364 2365
            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
2366 2367 2368 2369
          }
        }
      }
      // offset
2370 2371 2372
      tgtGrad += inLength;
      otGrad += outLength;
      maxPoolIdxData += outLength;
C
chengduoZH 已提交
2373 2374 2375 2376 2377
    }
  }
}

void CpuMatrix::avgPool3DForward(Matrix& input,
C
chengduoZH 已提交
2378
                                 size_t channels,
C
chengduoZH 已提交
2379 2380 2381
                                 size_t imgSizeD,
                                 size_t imgSizeH,
                                 size_t imgSizeW,
C
chengduoZH 已提交
2382 2383 2384
                                 size_t outputD,
                                 size_t outputH,
                                 size_t outputW,
C
chengduoZH 已提交
2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395
                                 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();
2396 2397 2398 2399
  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
2400
  real* tgtData = getData();
C
chengduoZH 已提交
2401 2402 2403 2404 2405 2406 2407 2408
  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) {
2409 2410 2411
        int dstart = pd * strideD - paddingD;
        int dend = std::min(dstart + sizeZ, imgSizeD);
        dstart = std::max(dstart, 0);
C
chengduoZH 已提交
2412
        for (size_t ph = 0; ph < outputH; ++ph) {
2413 2414 2415
          int hstart = ph * strideH - paddingH;
          int hend = std::min(hstart + sizeY, imgSizeH);
          hstart = std::max(hstart, 0);
C
chengduoZH 已提交
2416 2417
          for (size_t pw = 0; pw < outputW; ++pw) {
            int wstart = pw * strideW - paddingW;
2418
            int wend = std::min(wstart + sizeX, imgSizeW);
C
chengduoZH 已提交
2419 2420 2421 2422 2423 2424 2425
            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] +=
2426
                      inData[(d * imgSizeH + h) * imgSizeW + w];
C
chengduoZH 已提交
2427 2428 2429
                }
              }
            }
2430 2431
            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
            CHECK(poolSize);
C
chengduoZH 已提交
2432 2433 2434 2435 2436
            tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize;
          }
        }
      }
      // compute offset
2437 2438
      inData += inLength;
      tgtData += outLength;
C
chengduoZH 已提交
2439 2440 2441 2442 2443 2444 2445 2446
    }
  }
}

void CpuMatrix::avgPool3DBackward(Matrix& input,
                                  size_t imgSizeD,
                                  size_t imgSizeH,
                                  size_t imgSizeW,
C
chengduoZH 已提交
2447 2448 2449
                                  size_t outputD,
                                  size_t outputH,
                                  size_t outputW,
C
chengduoZH 已提交
2450 2451 2452 2453 2454 2455 2456 2457
                                  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
2458 2459 2460
                                  size_t paddingW,
                                  real scaleTargets,
                                  real scaleOutput) {
C
chengduoZH 已提交
2461
  size_t num = input.getHeight();
2462 2463 2464 2465
  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
2466 2467 2468 2469 2470 2471 2472 2473 2474
  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) {
2475 2476 2477
        int dstart = pd * strideD - paddingD;
        int dend = std::min(dstart + sizeZ, imgSizeD);
        dstart = std::max(dstart, 0);
C
chengduoZH 已提交
2478
        for (size_t ph = 0; ph < outputH; ++ph) {
2479 2480 2481
          int hstart = ph * strideH - paddingH;
          int hend = std::min(hstart + sizeY, imgSizeH);
          hstart = std::max(hstart, 0);
C
chengduoZH 已提交
2482 2483
          for (size_t pw = 0; pw < outputW; ++pw) {
            int wstart = pw * strideW - paddingW;
2484
            int wend = std::min(wstart + sizeX, imgSizeW);
C
chengduoZH 已提交
2485
            wstart = std::max(wstart, 0);
2486
            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
C
chengduoZH 已提交
2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499
            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
2500 2501
      outData += inLength;
      inData += outLength;
C
chengduoZH 已提交
2502 2503 2504 2505
    }
  }
}

Z
zhangjinchao01 已提交
2506 2507 2508 2509 2510
/**
 * 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]}
 */
2511 2512
void CpuMatrix::maxSequenceForward(Matrix& input,
                                   const IVector& sequence,
Z
zhangjinchao01 已提交
2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552
                                   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;
        }
      }
    }
  }
}

2553 2554
void CpuMatrix::maxSequenceBackward(Matrix& outputGrad,
                                    const IVector& sequence,
Z
zhangjinchao01 已提交
2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591
                                    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];
  }
}

2592 2593
inline void colVecAddTo(
    real* a, const real* b, size_t len, size_t aWidth, size_t bWidth) {
Z
zhangjinchao01 已提交
2594 2595 2596 2597 2598
  for (unsigned int i = 0; i < len; ++i) {
    a[i * aWidth] += b[i * bWidth];
  }
}

2599 2600
inline void colVecAddTo(
    real* a, real* b, real c, size_t len, size_t aWidth, size_t bWidth) {
Z
zhangjinchao01 已提交
2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632
  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];
      }
    }
  }
}

2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650
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
2651 2652 2653 2654 2655
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) {
2656
    sumCols(a, /* scaleSum= */ scale, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667
  } 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];
    }
  }
}

2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684
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
2685 2686 2687 2688 2689 2690 2691 2692 2693 2694
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
2695
  MatrixPtr outMtx = Matrix::create(nullptr, 1, width, false, false);
Z
zhangjinchao01 已提交
2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706
  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
2707 2708 2709
      outMtx->sumCols(*dataMtx,
                      (real)1 / (real)sequenceLength,
                      /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2710 2711
    } else if (mode == 1) {
      // sum instead of average
2712
      outMtx->sumCols(*dataMtx, /* scaleSum= */ 1, /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2713 2714
    } else if (mode == 2) {
      // divide by square root of sequenceLength
2715 2716 2717
      outMtx->sumCols(*dataMtx,
                      (real)1 / std::sqrt(sequenceLength),
                      /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
2718 2719 2720 2721 2722 2723
    } else {
      LOG(FATAL) << "should not reach here";
    }
  }
}

L
Luo Tao 已提交
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 2750 2751 2752 2753 2754 2755 2756 2757 2758
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
2759
/* this = scaleAB*(a*b) + scaleT*this*/
2760 2761
void CpuMatrix::mul(const Matrix& a,
                    const Matrix& b,
2762
                    real scaleAB,
Z
zhangjinchao01 已提交
2763 2764
                    real scaleT) {
  CHECK(!isTransposed()) << "Not supported";
2765 2766 2767 2768
  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
2769

2770 2771 2772 2773 2774 2775
  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
2776 2777 2778 2779 2780
  } else {
    LOG(FATAL) << "Not supported";
  }
}

2781 2782 2783
void CpuMatrix::mul(CpuSparseMatrix* a,
                    CpuMatrix* b,
                    real scaleAB,
Z
zhangjinchao01 已提交
2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797
                    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;
2798
  bool a_trans, b_trans;
Z
zhangjinchao01 已提交
2799 2800 2801
  if (!a->isTransposed()) {
    a_col = a->getWidth();
    a_row = a->getHeight();
2802
    a_trans = false;
Z
zhangjinchao01 已提交
2803 2804 2805
  } else {
    a_col = a->getHeight();
    a_row = a->getWidth();
2806
    a_trans = true;
Z
zhangjinchao01 已提交
2807 2808 2809 2810
  }
  if (!b->isTransposed()) {
    b_col = b->getWidth();
    b_row = b->getHeight();
2811
    b_trans = false;
Z
zhangjinchao01 已提交
2812 2813 2814
  } else {
    b_col = b->getHeight();
    b_row = b->getWidth();
2815
    b_trans = true;
Z
zhangjinchao01 已提交
2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831
  }

  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();
2832
  BlasGemm<DEVICE_TYPE_CPU, real>::compute(
L
Liu Yiqun 已提交
2833
      a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc);
Z
zhangjinchao01 已提交
2834 2835
}

2836 2837
void CpuMatrix::mul(
    CpuMatrix* a, CpuMatrix* b, CpuSparseMatrix* c, real scaleAB, real scaleT) {
Z
zhangjinchao01 已提交
2838 2839 2840 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 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943
  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";
  }
}

2944 2945 2946
void CpuMatrix::mul(CpuMatrix* a,
                    CpuSparseMatrix* b,
                    real scaleAB,
Z
zhangjinchao01 已提交
2947 2948 2949 2950 2951 2952 2953 2954 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
                    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) {
2984 2985
            colVecAddTo(
                C + j, A + rows[i], B[i], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006
          }
        }
      }
    } 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) {
3007 3008
            colVecAddTo(
                C + rows[j], A + i, B[j], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032
          }
        }
      }
    }
  } 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) {
3033 3034
            colVecAddTo(
                C + cols[i], A + j, B[i], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055
          }
        }
      }
    } 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) {
3056 3057
            colVecAddTo(
                C + i, A + cols[j], B[j], height_, width_, a->getWidth());
Z
zhangjinchao01 已提交
3058 3059 3060 3061 3062 3063 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 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155
          }
        }
      }
    }
  }
}

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>
3156 3157
void CpuMatrix::mul(
    CpuSparseMatrix* a, MatBType* b, MatCType* c, real scaleAB, real scaleT) {
Z
zhangjinchao01 已提交
3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 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
  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>(
3260 3261 3262 3263
    CpuSparseMatrix* a,
    CpuMatrix* b,
    SparseRowCpuMatrix* c,
    real scaleAB,
Z
zhangjinchao01 已提交
3264 3265
    real scaleT);
template void CpuMatrix::mul<CpuMatrix, SparseAutoGrowRowCpuMatrix>(
3266 3267 3268 3269 3270
    CpuSparseMatrix* a,
    CpuMatrix* b,
    SparseAutoGrowRowCpuMatrix* c,
    real scaleAB,
    real scaleT);
Z
zhangjinchao01 已提交
3271 3272 3273 3274 3275 3276
template void CpuMatrix::mul<CpuMatrix, CacheRowCpuMatrix>(CpuSparseMatrix* a,
                                                           CpuMatrix* b,
                                                           CacheRowCpuMatrix* c,
                                                           real scaleAB,
                                                           real scaleT);

3277
#ifndef PADDLE_MOBILE_INFERENCE
3278 3279 3280
void SharedCpuMatrix::mul(CpuSparseMatrix* a,
                          CpuMatrix* b,
                          real scaleAB,
Z
zhangjinchao01 已提交
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
                          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);
    }
3320 3321
    std::shuffle(
        blockSeq.begin(), blockSeq.end(), ThreadLocalRandomEngine::get());
Z
zhangjinchao01 已提交
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
  }
  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) {
3359 3360
          vecAddTo(
              localC + bufPos * width, B + cols[j] * width, value[j], width);
Z
zhangjinchao01 已提交
3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405
        }
      }
    }

    {
      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);
  }
}

3406
#endif
Z
zhangjinchao01 已提交
3407 3408 3409 3410 3411 3412
/* 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 */
3413
void CpuMatrix::mul(const Matrix& a, const Matrix& b) {
Z
zhangjinchao01 已提交
3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444
  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);

3445
  sum.sumRows(*this, /* scaleSum= */ 1, /* scaleDest= */ 0);
Z
zhangjinchao01 已提交
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 3476
}

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
3477
/* Get the top k elements of each row of this matrix */
Z
zhangjinchao01 已提交
3478 3479 3480 3481 3482 3483 3484
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
3485
  CHECK_EQ(maxVal.getWidth(), beam);
Z
zhangjinchao01 已提交
3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497

  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(
3498 3499 3500
        vec.begin(),
        vec.begin() + beam,
        vec.end(),
Z
zhangjinchao01 已提交
3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516
        [](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);
}

3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535
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(
3536 3537 3538
        vec.begin(),
        vec.begin() + beam,
        vec.end(),
3539 3540 3541 3542 3543 3544 3545 3546 3547 3548
        [](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;
    }
  }
}

3549 3550 3551
void CpuMatrix::maxoutForward(Matrix& a,
                              IVector& id,
                              size_t channels,
3552 3553 3554 3555 3556 3557 3558 3559
                              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
3560
  const real* input = a.getData();
3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583
  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);
  }
}

3584 3585 3586
void CpuMatrix::maxoutBackward(Matrix& a,
                               IVector& id,
                               size_t channels,
3587 3588 3589 3590 3591 3592 3593 3594 3595
                               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
3596 3597
  real* inputG = getData();
  const real* outG = a.getData();
3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611
  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
3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636
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 */
3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650
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
3651

3652 3653
  // top k matrix classification
  cpuOutput->rowMax(*cpuTopIds, *cpuTopVal);
Z
zhangjinchao01 已提交
3654

3655 3656 3657 3658
  size_t dim = cpuOutput->getWidth();
  real* result = this->getData();
  int* ids = cpuTopIds->getData();
  int* lbl = cpuLabel->getData();
Z
zhangjinchao01 已提交
3659 3660 3661
  for (size_t i = 0; i < numSamples; ++i) {
    CHECK_GE(lbl[i], 0);
    CHECK_LT((size_t)lbl[i], dim);
3662 3663 3664 3665 3666

    for (size_t j = 0; j < topkSize; ++j) {
      if (ids[j + i * topkSize] == lbl[i]) {
        result[i] = 0;
        break;
Z
zhangjinchao01 已提交
3667
      }
3668
      result[i] = 1.0f;
Z
zhangjinchao01 已提交
3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713
    }
  }
}

/* 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
*/
3714 3715
void CpuMatrix::oneHotCrossEntropyWithSelfNorm(Matrix& output,
                                               IVector& label,
Z
zhangjinchao01 已提交
3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745
                                               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
*/
3746 3747
void CpuMatrix::oneHotCrossEntropyWithSelfNormBp(Matrix& output,
                                                 IVector& label,
Z
zhangjinchao01 已提交
3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 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
                                                 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());

3828 3829 3830 3831 3832 3833 3834 3835 3836 3837
  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
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 3863 3864 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
  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);
3890 3891
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906
            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);
3907 3908
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929
            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;
    }
  }

3930 3931 3932 3933
  BaseMatrix::sumOfSquaredDiffs(output,
                                label,
                                /* scaleSum= */ 1,
                                /* scaleDest= */ 1);
Z
zhangjinchao01 已提交
3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962
}

/* 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);
3963 3964
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978
            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);
3979 3980
               j < labelptr->getRowStartIdx(i + 1);
               ++j) {
Z
zhangjinchao01 已提交
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 4007 4008 4009 4010 4011 4012 4013
            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;
    }
  }
}

4014
void CpuMatrix::smoothL1(Matrix& output, Matrix& label, real destScale) {
G
gaoyuan 已提交
4015 4016 4017 4018 4019 4020 4021 4022 4023
  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
4024

G
gaoyuan 已提交
4025
  real* cost = getData();
D
dangqingqing 已提交
4026
  real* out = output.getData();
G
gaoyuan 已提交
4027 4028
  real* lbl = label.getData();

D
dangqingqing 已提交
4029
  for (size_t i = 0; i < numSamples; ++i, out += dim, lbl += dim) {
G
gaoyuan 已提交
4030
    for (size_t j = 0; j < dim; ++j) {
D
dangqingqing 已提交
4031
      real absVal = std::fabs(out[j] - lbl[j]);
4032
      cost[i] *= destScale;
D
dangqingqing 已提交
4033 4034
      if (absVal < 1.0)
        cost[i] += 0.5 * absVal * absVal;
G
gaoyuan 已提交
4035
      else
D
dangqingqing 已提交
4036
        cost[i] += absVal - 0.5;
G
gaoyuan 已提交
4037 4038 4039 4040
    }
  }
}

4041
void CpuMatrix::smoothL1Bp(Matrix& output, Matrix& label, real destScale) {
G
gaoyuan 已提交
4042 4043 4044 4045 4046 4047 4048 4049
  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
4050 4051
  CHECK_EQ(getWidth(), dim);

G
gaoyuan 已提交
4052 4053
  real* out = output.getData();
  real* lbl = label.getData();
D
dangqingqing 已提交
4054
  real* grad = getData();
G
gaoyuan 已提交
4055

D
dangqingqing 已提交
4056
  for (size_t i = 0; i < numSamples; ++i, out += dim, grad += dim, lbl += dim) {
G
gaoyuan 已提交
4057
    for (size_t j = 0; j < dim; ++j) {
D
dangqingqing 已提交
4058
      real val = out[j] - lbl[j];
4059
      grad[j] *= destScale;
D
dangqingqing 已提交
4060 4061 4062 4063 4064
      if (std::fabs(val) < 1) {
        grad[j] += val;
      } else {
        grad[j] += (real(0) < val) - (val < real(0));
      }
G
gaoyuan 已提交
4065 4066 4067 4068
    }
  }
}

Z
zhangjinchao01 已提交
4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097 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 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168
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();
X
xzl 已提交
4169
  real* output = data_;
Z
zhangjinchao01 已提交
4170 4171
  size_t numElements = data.getWidth();
  size_t numSamples = data.getHeight();
H
hedaoyuan 已提交
4172 4173
  size_t paraSize = W.getHeight() * W.getWidth();
  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
X
xzl 已提交
4174

H
hedaoyuan 已提交
4175
  size_t partial_sum = numElements / paraSize;
X
xzl 已提交
4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192
  if (paraSize == numElements) {
    for (size_t n = 0; n < numSamples * numElements; ++n) {
      output[n] = input[n] > 0 ? input[n] : input[n] * w[n % numElements];
    }
    return;
  }

#if defined(__ARM_NEON__) || defined(__ARM_NEON)
  for (size_t n = 0; n < numSamples; ++n) {
    for (size_t i = 0; i < paraSize; i++) {
      neon::prelu(
          input + i * partial_sum, w[i], output + i * partial_sum, partial_sum);
    }
    input = input + numElements;
    output = output + numElements;
  }
#else
Z
zhangjinchao01 已提交
4193 4194
  for (size_t n = 0, k = 0; n < numSamples; ++n) {
    for (size_t i = 0; i < numElements; ++i, ++k) {
X
xzl 已提交
4195
      output[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
Z
zhangjinchao01 已提交
4196 4197
    }
  }
X
xzl 已提交
4198
#endif
Z
zhangjinchao01 已提交
4199 4200 4201 4202 4203 4204 4205 4206
}

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
4207 4208 4209
  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
4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223
  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
4224 4225 4226
  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
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 4288 4289 4290 4291 4292 4293 4294 4295 4296 4297 4298 4299 4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335
  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];
      }
    }
  }
}

4336 4337
void CpuMatrix::circularConvDerivative(
    Matrix& outG, Matrix& in0, Matrix& in1, Matrix& inG0, Matrix& inG1) {
Z
zhangjinchao01 已提交
4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356
  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;
4357 4358 4359 4360 4361 4362
  for (size_t x = 0; x < height; ++x,
              outGV += width0,
              inV0 += width0,
              inV1 += width1,
              inGV0 += width0,
              inGV1 += width1) {
Z
zhangjinchao01 已提交
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
    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 */
4431 4432
void CpuMatrix::classificationErrorMulti(Matrix& output,
                                         Matrix& label,
Z
zhangjinchao01 已提交
4433 4434 4435 4436 4437 4438 4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466
                                         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
4467 4468 4469 4470 4471
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
4472 4473 4474
                                const size_t numChannels,
                                const real ratioH,
                                const real ratioW) {
L
liaogang 已提交
4475 4476 4477
  CHECK(dynamic_cast<const CpuMatrix*>(&in));

  size_t outputW = getWidth();
L
liaogang 已提交
4478
  size_t batchSize = getHeight();
L
liaogang 已提交
4479 4480
  size_t inputW = in.getWidth();
  size_t inputH = in.getHeight();
L
liaogang 已提交
4481 4482
  size_t inPosOffset = inImgH * inImgW;
  size_t outPosOffset = outImgH * outImgW;
L
liaogang 已提交
4483
  (void)(inputH);
L
liaogang 已提交
4484 4485

  real* outData = getData();
4486
  const real* inData = in.getData();
L
liaogang 已提交
4487 4488 4489 4490

  if (inImgH == outImgH && inImgW == outImgW) {
    this->copyFrom(in);
  } else {
4491
    for (size_t k = 0; k < batchSize; ++k) {  // loop for batches
L
liaogang 已提交
4492 4493 4494
      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
4495 4496
        real h1lambda = ratioH * i - h;
        real h2lambda = 1 - h1lambda;
L
liaogang 已提交
4497

L
liaogang 已提交
4498 4499 4500
        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
4501 4502
          real w1lambda = ratioW * j - w;
          real w2lambda = 1 - w1lambda;
L
liaogang 已提交
4503 4504 4505
          // 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
4506
          for (size_t c = 0; c < numChannels; ++c) {  // loop for channels
L
liaogang 已提交
4507
            // bilinear interpolation
L
liaogang 已提交
4508
            outPos[0] =
4509 4510 4511
                h2lambda * (w2lambda * inPos[0] + w1lambda * inPos[wid]) +
                h1lambda * (w2lambda * inPos[hid * inImgW] +
                            w1lambda * inPos[hid * inImgW + wid]);
L
liaogang 已提交
4512 4513
            inPos += inPosOffset;
            outPos += outPosOffset;
L
liaogang 已提交
4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525
          }
        }
      }
    }
  }
}

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
4526 4527 4528
                                 const size_t numChannels,
                                 const real ratioH,
                                 const real ratioW) {
L
liaogang 已提交
4529 4530 4531 4532 4533
  CHECK(dynamic_cast<const CpuMatrix*>(&out));

  size_t inputW = getWidth();
  size_t inputH = getHeight();
  size_t outputW = out.getWidth();
L
liaogang 已提交
4534
  size_t batchSize = out.getHeight();
L
liaogang 已提交
4535 4536
  size_t inPosOffset = inImgH * inImgW;
  size_t outPosOffset = outImgH * outImgW;
L
liaogang 已提交
4537
  (void)(inputH);
L
liaogang 已提交
4538 4539 4540 4541 4542

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

  if (inImgH == outImgH && inImgW == outImgW) {
L
liaogang 已提交
4543
    this->add(const_cast<Matrix&>(out));
L
liaogang 已提交
4544
  } else {
4545
    for (size_t k = 0; k < batchSize; ++k) {  // loop for batches
L
liaogang 已提交
4546 4547 4548
      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
4549 4550
        real h1lambda = ratioH * i - h;
        real h2lambda = 1 - h1lambda;
L
liaogang 已提交
4551 4552 4553
        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
4554 4555
          real w1lambda = ratioW * j - w;
          real w2lambda = 1 - w1lambda;
L
liaogang 已提交
4556 4557 4558

          real* inPos = &inGrad[k * inputW + h * inImgW + w];
          const real* outPos = &outGrad[k * outputW + i * outImgW + j];
L
liaogang 已提交
4559
          for (size_t c = 0; c < numChannels; ++c) {  // loop for channels
L
liaogang 已提交
4560 4561 4562 4563
            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
4564 4565
            inPos += inPosOffset;
            outPos += outPosOffset;
L
liaogang 已提交
4566 4567 4568 4569 4570 4571 4572
          }
        }
      }
    }
  }
}

C
chengduoZH 已提交
4573 4574 4575 4576 4577 4578 4579 4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633
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
4634
  int outDepth = (depth + 2 * paddingD - filterD) / strideD + 1;
C
chengduoZH 已提交
4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661
  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
4662 4663 4664 4665 4666 4667 4668 4669
////////////////////////////////////////////////////////////////
//               functions executed via cpu                   //
////////////////////////////////////////////////////////////////

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