Matrix.h 49.2 KB
Newer Older
Z
zhangjinchao01 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

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

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */


#pragma once

#include <memory>
#include <thread>
#include <stdint.h>

#include "paddle/utils/Logging.h"
#include "paddle/utils/ThreadLocal.h"

#include <hl_gpu.h>

#include "MemoryHandle.h"
#include "paddle/utils/TypeDefs.h"
#include "Vector.h"
#include "paddle/utils/ThreadLocal.h"
#include "BaseMatrix.h"

namespace paddle {

enum SparseValueType { NO_VALUE = 0, FLOAT_VALUE = 1 };

/**
 * @brief  matrix sparse_format .
 *
 * nnz represents nonzero number in sparse matrix.
 *
 * SPARSE_CSR: row major matrix. length of row is height_ + 1, each element
 * represents row start index in Matrix. length of col and value are nnz.
 *
 * SPARSE_CSC: col major matrix. length of col is width_ + 1, each element
 * represents col start index in Matrix. length of col and value are nnz.
 *
 * @code
 * for example: [0, 1, 0, 2, 0;
 *               1, 0, 0, 0, 0;
 *               0, 0, 0, 2, 5];
 * SPARSE_CSR row   [0, 2, 3, 5];
 *            col   [1, 3, 0, 3, 4];
 *            value [1, 2, 1, 2, 5]
 * SPARSE_CSC col   [0, 1, 2, 2, 4, 5];
 *            row   [1, 0, 0, 2, 2];
 *            value [1, 1, 2, 2, 5]
 * @endcode
 */
enum SparseFormat { SPARSE_CSR = 0, SPARSE_CSC = 1 };

class Matrix;
class GpuMatrix;
class CpuMatrix;
class CpuSparseMatrix;
class GpuSparseMatrix;
typedef std::shared_ptr<Matrix> MatrixPtr;
typedef std::shared_ptr<GpuMatrix> GpuMatrixPtr;
typedef std::shared_ptr<CpuMatrix> CpuMatrixPtr;
typedef std::shared_ptr<GpuSparseMatrix> GpuSparseMatrixPtr;
typedef std::shared_ptr<CpuSparseMatrix> CpuSparseMatrixPtr;

/**
 * Copy or assignemnt constructor will share the data as opposed to making a
 * copy of the original data. To make a copy of the orinal data, use copyFrom()
 * instead.
 */
class Matrix : public BaseMatrix {
protected:
  Matrix(MemoryHandlePtr memHandle, size_t height, size_t width, bool trans,
         bool use_gpu);

  Matrix(real* data, size_t height, size_t width, bool trans, bool use_gpu);

  Matrix(real* data, size_t height, size_t width, size_t stride, bool trans,
         bool use_gpu);

  static ThreadLocal<MatrixPtr> tmpMat_;

public:
  size_t elementCnt_;  // maximal number of elements which can be held in data_
  MemoryHandlePtr memoryHandle_;

public:
  virtual ~Matrix() {}

  static MatrixPtr create(MemoryHandlePtr memHandle, size_t height,
                          size_t width, bool trans = false);
  static MatrixPtr create(size_t height, size_t width, bool trans = false,
                          bool useGpu = false);
  static MatrixPtr create(real* data, size_t height, size_t width,
                          bool trans = false, bool useGpu = false);
  static MatrixPtr create(real* data, size_t height, size_t width,
                          size_t stride, bool trans = false,
                          bool useGpu = false);

  static MatrixPtr createSparseMatrix(size_t height, size_t width, size_t nnz,
                                      SparseValueType valueType = FLOAT_VALUE,
                                      bool trans = false, bool useGpu = false);
  static MatrixPtr createSparseMatrix(size_t height, size_t width, size_t nnz,
                                      SparseValueType valueType = FLOAT_VALUE,
                                      SparseFormat foramt = SPARSE_CSR,
                                      bool trans = false, bool useGpu = false);

  static MatrixPtr createSparseMatrix(real* data, int* row, int* col,
                                      size_t height, size_t width,
                                      size_t nnz, /* used to allocate space */
                                      SparseValueType valueType, /*value type*/
                                      SparseFormat format, bool trans,
                                      bool useGpu);

  static void resizeOrCreateSparseMatrix(
      MatrixPtr& matrix, size_t height, size_t width, size_t nnz,
      SparseValueType valueType = FLOAT_VALUE, SparseFormat foramt = SPARSE_CSR,
      bool trans = false, bool useGpu = false);

  static void resizeOrCreate(MatrixPtr& a, size_t height, size_t width,
                             bool trans = false, bool useGpu = false);

  /**
   * @brief  set the data buffer used to hold the matrix data.
   *
   * caller should make sure that the size of data is at least
   * sizeof(real)*height*width.
   */
  void setData(real* data) {
    BaseMatrix::setData(data);
    memoryHandle_.reset();
  }

  /// the data should be contiguous
  void setData(real* data, size_t newHeight, size_t newWidth) {
    setData(data);
    height_ = newHeight;
    width_ = newWidth;
    elementCnt_ = newHeight * newWidth;
    stride_ = width_;
  }

  size_t getWidth() const { return width_; }
  size_t getHeight() const { return height_; }
  size_t getStride() const { return stride_; }
  size_t getElementCnt() const { return elementCnt_; }
  virtual real* getData() { return data_; }
  virtual const real* getData() const { return data_; }
  bool isTransposed() const { return trans_; }
  bool isContiguous() const { return stride_ == width_ || height_ == 1; }

  // If sparse matrix, need to dynamic_cast to CpuSparseMatrix/GpuSparseMatrix
  // befor call the following functions.
  // Declare these functions in the base class just easy to call them.
  // And these declarations should be moved to base class of sparse matrix
  // if refactor sparse matrix
  virtual int* getRows() const {
    LOG(FATAL) << "Not implemented";
    return nullptr;   //! suppress warning for no return value.
  }

  virtual int* getCols() const {
    LOG(FATAL) << "Not implemented";
    return nullptr;   //! suppress warning for no return value.
  }

  virtual SparseFormat getFormat() const {
    LOG(FATAL) << "Not implemented";
    return SPARSE_CSR;  //! suppress warning for no return value.
  }

  virtual SparseValueType getValueType() const {
    LOG(FATAL) << "Not implemented";
    return NO_VALUE;    //! suppress warning for no return value.
  }

  /**
   * @brief matrix elment-wise add
   *
   * Named add3 just because add/add2 has been used in BaseMatrix.cu
   * and they are not virtual function.
   */
  virtual void add3(MatrixPtr b) { LOG(FATAL) << "Not implemented"; }

  MemoryHandlePtr getMemoryHandle() const { return memoryHandle_; }

  virtual void zeroMem() { LOG(FATAL) << "Not implemented"; }

  virtual void resetOne() { LOG(FATAL) << "Not implemented"; }

  virtual void copyFrom(const Matrix& src) { LOG(FATAL) << "Not implemented"; }

  virtual void trimFrom(const CpuSparseMatrix& src) {
    LOG(FATAL) << "Not implemented";
  }

  // asynchronous copy
  virtual void copyFrom(const Matrix& src, hl_stream_t stream) {
    LOG(FATAL) << "Not implemented";
  }

  MatrixPtr subMatrix(size_t startRow, size_t endRow, size_t startCol,
                      size_t endCol);

  MatrixPtr subRowMatrix(size_t startRow, size_t endRow) {
    return subMatrix(startRow, endRow, 0, getWidth());
  }

  MatrixPtr subColMatrix(size_t startCol, size_t endCol) {
    return subMatrix(0, getHeight(), startCol, endCol);
  }

  virtual MatrixPtr subMatrix(size_t startRow, size_t numRows) {
    CHECK_LE(startRow + numRows, getHeight());
    return Matrix::create(getData() + startRow * getWidth(), numRows,
                          getWidth(), trans_, useGpu_);
  }
  virtual MatrixPtr subMatrix(size_t startRow, size_t numRows, MatrixPtr dest) {
    CHECK_LE(startRow + numRows, getHeight());
    CHECK_EQ(useGpu_, dest->useGpu_);
    dest->setData(this->rowBuf(startRow), numRows, getWidth());
    return dest;
  }

  /**
   * If this is GpuMatrix, src is assumed to be CPU memory
   *
   * If this is CpuMatrix, src is assumed to be CPU memory
   */
  virtual void copyFrom(const real* src, size_t size) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void copyFrom(const real* src, const int64_t* seq) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @brief convert a int vector to a real matrix.
   *
   * (1) source and dest are both in CPU.
   *
   * (2) sizes are exactly match.
   */
  virtual void copyFrom(const IVector& src) {
    LOG(FATAL) << "copy data from int vector only available on CpuMatrix.";
  }

256
  virtual void copyByRowIndex(Matrix& b, const IVector& rowIndex) {
Z
zhangjinchao01 已提交
257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @brief Create a matrix with the same type (GpuMatrix, CpuMatrix,
   *        NonValueSparseMatrix, etc.) as this.
   *
   * If height and width is zero, the new matrix will have the same size
   * as this, otherwise the new matrix will have the specified size.
   *
   */
  virtual MatrixPtr clone(size_t height = 0, size_t width = 0,
                          bool useGpu = false) {
    LOG(FATAL) << "Not implemented";
    return nullptr;
  }

  virtual real* getRowBuf(size_t row) {
    LOG(FATAL) << "Not implemented";
    return nullptr;
  }

  virtual real getElement(size_t x, size_t y) const {
    LOG(FATAL) << "Not implemented";
    return 0;
  }

  virtual real getSum() {
    LOG(FATAL) << "Not implemented";
    return 0;
  }

  virtual void accumulateColSum(Matrix& src) {
    LOG(FATAL) << "Not implemented";
  }

  virtual real getAbsSum() {
    LOG(FATAL) << "Not implemented";
    return 0;
  }

  /**
   * @note Original data may not be preserved after resize().
   */
  virtual void resize(size_t newHeight, size_t newWidth) = 0;

  /**
   * @note This should only be used for sparse matrix.
   */
  virtual void resize(size_t newHeight, size_t newWidth,
                      size_t newNnz, /* total item used to allocate space */
                      SparseValueType valueType, SparseFormat format) = 0;

  /**
   * @brief This should only be used for sparse matrix.
   *
   * Currently must be called for each row in order.
   * The matrix is not valid until setRow is called for the last row.
   */
  virtual void setRow(size_t row, size_t colNum, const unsigned int* cols,
                      const real* values) = 0;

  virtual MatrixPtr getTranspose() = 0;

  /**
   * @brief  hard transpose.
   *
   * allocate matTrans' memory outside, then set memAlloc as false;
   * else set as true.
   */
  virtual void transpose(MatrixPtr matTrans, bool memAlloc) {
    LOG(FATAL) << "Not implemented";
  }

public:
  /// Only set all variables to 0 or NULL but not free them.
  virtual void clear() {
    height_ = 0;
    width_ = 0;
    data_ = NULL;
  }

  void reshape(size_t height, size_t width);

  /// add b to each sample of this.
  virtual void addBias(Matrix& b, real scale) {
    LOG(FATAL) << "Not implemented";
  }

  /// add each sample from a to this.
  virtual void collectBias(Matrix& a, real scale) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void sequenceAvgForward(Matrix& a, const IVector& startsPos,
    int mode) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @code
   * this = scaleAB*(a*b) + scaleT*this
   * @endcode
   */
  virtual void mul(const MatrixPtr a, const MatrixPtr b, real scaleAB,
                   real scaleT) {
    LOG(FATAL) << "Not implemented";
  }

  /// Add a vector (column) b to matrix a, column by column.
  virtual void addColumnVector(const Matrix& b) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @code
   * For j < codeLength:
   *   this(i, j) += vec(index(i, j), 0)
   * where index(i, j) = ((codes(i) + numClasses) >> (j + 1)) - 1
   * @endcode
   */
  virtual void addByBitCode(size_t numClasses, const IVector& codes,
                            const Matrix& vec) {
    (void)numClasses;
    (void)codes;
    (void)vec;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   * For j < codeLength:
   *   vec(index(i, j), 0) += this(i, j)
   * where index is same as the index for addByBitCode
   * @endcode
   */
  virtual void addByBitCodeBackward(size_t numClasses, const IVector& codes,
                                    Matrix& vec) {
    (void)numClasses;
    (void)codes;
    (void)vec;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   * For j < codeLength:
   *   this(i, j) += <mat.row(index(i, j)), input.row(i)>
   * where index is same as the index for addByBitCode
   * @endcode
   */
  virtual void mulByBitCode(size_t numClasses, const IVector& codes,
                            const Matrix& mat, const Matrix& input) {
    (void)numClasses;
    (void)codes;
    (void)mat;
    (void)input;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   * For j < codeLength:
   *   mat.row(index(i, j)) += this(i, j) * input.row(i)
   * where index is same as the index for addByBitCode
   * @endcode
   */
  virtual void mulByBitCodeBackwardWeight(size_t numClasses,
                                          const IVector& codes, Matrix& mat,
                                          const Matrix& input) {
    (void)numClasses;
    (void)codes;
    (void)mat;
    (void)input;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   * For j < codeLength:
   *   input.row(i) += this(i, j) * mat.row(index(i, j))
   * where index is same as the index for addByBitCode
   * @endcode
   */
  virtual void mulByBitCodeBackwardError(size_t numClasses,
                                         const IVector& codes,
                                         const Matrix& mat, Matrix& input) {
    (void)numClasses;
    (void)codes;
    (void)mat;
    (void)input;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   * For j < codeLength
   *   sum(i, 0) = scaleSum * \sum_j  bit(i, j) * this(i, j)
   * where bit(i, j) = ((codes(i) + numClasses) & 2^j) ? 1 : 0
   * @endcode
   */
  virtual void sumByBitCode(size_t numClasses, IVector& codes, Matrix& sum,
                            real scaleSum) {
    (void)numClasses;
    (void)codes;
    (void)sum;
    (void)scaleSum;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   * For j < codeLength
   *  this(i, j) -= bit(i, j)
   * where bit(i, j) is same as that for sumByBitCode
   * @endcode
   */
  virtual void subByBitCode(size_t numClasses_, IVector& codes) {
    (void)numClasses_;
    (void)codes;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * add the sum of each row of this to mat
   */
  virtual void rowSum(Matrix& sum) {
    (void)sum;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * set the max of each row of this to mat
   */
  virtual void rowMax(Matrix& max) {
    (void)max;
    LOG(FATAL) << "Not implemeted";
  }

496 497 498
  /**
   * set the max of each column of this to mat
   */
Z
zhangjinchao01 已提交
499 500
  virtual void colMax(Matrix& max) { LOG(FATAL) << "not implemented"; }

501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521
  /**
   * @brief Get the top k elements of each column of this matrix.
   *
   * The row ids and values of these elements are stored in
   * maxIds and max respectively. where k is the size of maxIds.
   * And note that the top k elements are not sorted.
   */
  virtual void colMax(IVector& maxIds, Matrix& maxVal) {
    LOG(FATAL) << "not implemented";
  }

  virtual void maxoutForward(Matrix& a, IVector& id, size_t channels,
                             size_t groups) {
    LOG(FATAL) << "not implemented";
  }

  virtual void maxoutBackward(Matrix& a, IVector& id, size_t channels,
                              size_t groups) {
    LOG(FATAL) << "not implemented";
  }

Z
zhangjinchao01 已提交
522 523 524 525 526 527
  virtual void rowMaxId(IVector& maxIds) { LOG(FATAL) << "Not implemented"; }

  /**
   * @brief Get the top k elements of each row of this matrix.
   *
   * The column ids and values of these elements are stored in
528 529
   * maxIds and max respectively. where k is the size of maxIds.
   * And note that the top k elements are not sorted.
Z
zhangjinchao01 已提交
530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611
   */
  virtual void rowMax(IVector& maxIds, Matrix& max) {
    LOG(FATAL) << "Not implemented";
  }

  /// normalize each row so that the sum of each row is 1.
  virtual void rowNormalizeL1(Matrix& out) {
    (void)out;
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   *  this = a*b
   * @endcode
   */
  virtual void mul(const MatrixPtr a, const MatrixPtr b) {
    LOG(FATAL) << "Not implemented";
  }

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

  /**
   * @code
   * this = this* b
   * @endcode
   */
  virtual void rightMul(Matrix& b) { LOG(FATAL) << "Not implemented"; }

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

  /**
   * @code
   * this = a*this)
   * @endcode
   */
  virtual void leftMul(Matrix& a) { LOG(FATAL) << "Not implemented"; }

  /// merge the element for each col.
  virtual void colMerge(Matrix& src) { LOG(FATAL) << "Not implemented"; }

  /// copy -log(output[label]) to this->data[i].
  virtual void oneHotCrossEntropy(Matrix& output, IVector& label) {
    LOG(FATAL) << "Not implemented";
  }

  /// calculate the error of outputV according to label.
  virtual void oneHotCrossEntropyBp(Matrix& outputV, IVector& label) {
    LOG(FATAL) << "Not implemented";
  }

  /// copy -log(output[label]) to this->data[i].
  virtual void oneHotCrossEntropyWithSelfNorm(Matrix& output, IVector& label,
                                              real alpha) {
    LOG(FATAL) << "Not implemented";
  }

  /// calculate the error of outputV according to label.
  virtual void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV,
                                                IVector& label,
                                                real alpha) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * \f[
   *  a[i] = \sum_{j=-(N-1)/2}^{(N-1)/2} b_{i+j} * c_{j}
   * \f]
612
   *
Z
zhangjinchao01 已提交
613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768
   * b contains M elements,
   * c contains N elements (N is odd),
   * b's index arithmetic is computed modulo M,
   * c's index arithmetic is computed modulo N.
   */
  virtual void circularConv(Matrix& b, Matrix& c) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void circularConvDerivative(Matrix& output, Matrix& prevOut1,
                                      Matrix& prevOut2, Matrix& prevGrad1,
                                      Matrix& prevGrad2) {
    LOG(FATAL) << "Not implemented";
  }


  /* output_ij = exp(this_{ij}) / (sum_j exp(this_ij)) */
  virtual void softmax(Matrix& output) {
    (void)output;
    LOG(FATAL) << "Not implemeted";
  }
  virtual void sequenceSoftmax(Matrix& output, const IVector& index) {
    (void)output;
    LOG(FATAL) << "Not implemeted";
  }

  virtual void softmaxBackward(Matrix& outputV) {
    (void)outputV;
    LOG(FATAL) << "Not implemeted";
  }

  /*
    sum_i = sum_j this_ij * output_ij
    this_ij = output_ij* (this_ij - sum_i)
  */
  virtual void softmaxDerivative(Matrix& output, Matrix& sftmaxSum) {
    LOG(FATAL) << "Not implemented";
  }

  /// calculate the sum of squares diff cost.
  virtual void sumOfSquares(Matrix& output, Matrix& label) {
    LOG(FATAL) << "Not implemented";
  }

  /// gradient of sumOfSquares.
  virtual void sumOfSquaresBp(Matrix& outputV, Matrix& label) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void tanh(Matrix& output) { LOG(FATAL) << "Not implemented"; }

  virtual void tanhDerivative(Matrix& output) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void softrelu(Matrix& output) { LOG(FATAL) << "Not implemented"; }

  virtual void softreluDerivative(Matrix& output) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void scaledTanh(Matrix& output, real p1, real p2) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * cosine similarity, for each row i,
   *   this[i] = cos(output1[i], output2[i])
   *
   * output2 can only have one row, then for each row i,
   *   this[i] = cos(output1[i], output2[0])
   */
  virtual void cosSim(Matrix& output1, Matrix& output2, real scale = 1.0f) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void cosSimDerivative(Matrix& output, Matrix& prevOut1,
                                Matrix& prevOut2, Matrix& prevGrad1,
                                Matrix& prevGrad2, real scale = 1.0f) {
    LOG(FATAL) << "Not implemented";
  }

  /// print out the values of elements to os
  virtual void print(std::ostream& os) const {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * print a part of the matrix
   * from the (top,left) value to the (height, width) value (not included)
   */
  virtual void print(std::ostream& os, size_t height, size_t width) const {
    LOG(FATAL) << "Not implemented";
  }

  /// print one row to os
  virtual void printOneRow(std::ostream& os, size_t idx) const {
    LOG(FATAL) << "Not implemented";
  }

  virtual void check(std::ostream& os, Matrix& refMat, bool printDiff = true) {}

  virtual real getMin() {
    LOG(FATAL) << "Not implemented";
    return 0;
  }
  virtual real getMax() {
    LOG(FATAL) << "Not implemented";
    return 0;
  }

  virtual void randomizeUniform() { LOG(FATAL) << "Not implemented"; }

  /**
   * @brief  calulate the error of classification
   *
   * output[i] = 1 if row i is an error.
   *
   * output[i] = 0 if row i is correct.
   */
  virtual void classificationError(MatrixPtr output, IVectorPtr label) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * This function is used to calculate the convolution:
   *
   * It will expand a feature matrix according to the
   * convolution filters
   */
  virtual void convExpand(Matrix& feature, int feaImgHeight, int feaImgWidth,
                          int channels, int blockH, int blockW, int strideH,
                          int strideW, int paddingH, int paddingW,
                          int outputH, int outputW) {
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * This function is the reverse implementation of convExpand:
   *
   * Its function is to restore a expanded-matrix into a feature matrix
   */
  virtual void convShrink(Matrix& expandColMat, int thisImgHeight,
                          int thisImgWidth, int channels, int blockH,
                          int blockW, int strideH, int strideW, int paddingH,
                          int paddingW, int outputH, int outputW,
                          real alpha = 1.0f, real beta = 0.0f) {
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * Pooling forward operation, pick out the largest element
   * in the sizeX of value
   */
  virtual void maxPoolForward(Matrix& inputMat, size_t imgSizeH,
                              size_t imgSizeW, size_t channels, size_t sizeX,
769 770 771
                              size_t sizeY, size_t strideH, size_t strideW,
                              size_t outputH, size_t outputW,
                              size_t paddingH, size_t paddingW) {
Z
zhangjinchao01 已提交
772 773 774 775 776 777
    LOG(FATAL) << "Not implemeted";
  }

  /// Pooling backward operation.
  virtual void maxPoolBackward(Matrix& image, size_t imgSizeH, size_t imgSizeW,
                               Matrix& outGrad, Matrix& outV, size_t sizeX,
778 779 780 781
                               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
782 783 784 785 786
    LOG(FATAL) << "Not implemeted";
  }

  /// Pooling forward operation, caculate the average of sizeX elements.
  virtual void avgPoolForward(Matrix& input, size_t imgSizeH, size_t imgSizeW,
787 788 789 790
                              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) {
Z
zhangjinchao01 已提交
791 792 793 794
    LOG(FATAL) << "Not implemeted";
  }

  virtual void avgPoolBackward(Matrix& input, size_t imgSizeH, size_t imgSizeW,
795 796
                               size_t sizeX, size_t sizeY,
                               size_t strideH, size_t strideW,
Z
zhangjinchao01 已提交
797
                               size_t outputH, size_t outputW,
798 799
                               real scaleTargets, real scaleOutput,
                               size_t paddingH, size_t paddingW) {
Z
zhangjinchao01 已提交
800 801 802 803 804 805 806
    LOG(FATAL) << "Not implemeted";
  }

  /// normalize-operation.
  virtual void crossMapNormalFwd(Matrix& input, size_t imgSizeH,
                                 size_t imgSizeW, Matrix& denoms,
                                 size_t channels, size_t sizeX, float scale,
807
                                 float pow) {
Z
zhangjinchao01 已提交
808 809 810 811 812 813 814
    LOG(FATAL) << "Not implemeted";
  }

  virtual void crossMapNormalBwd(Matrix& localGrad, Matrix& denoms,
                                 Matrix& preOutV, Matrix& localOutV,
                                 size_t channels, size_t imgSizeH,
                                 size_t imgSizeW, size_t size, float scale,
815
                                 float pow) {
Z
zhangjinchao01 已提交
816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * 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_input[i].
   */
  virtual void maxSequenceForward(Matrix& input, const IVector& sequence,
                                  IVector& index) {
    LOG(FATAL) << "Not implemeted";
  }

  virtual void maxSequenceBackward(Matrix& outputGrad, const IVector& sequence,
                                   IVector& index) {
    LOG(FATAL) << "Not implemeted";
  }

  virtual void contextProjectionForward(MatrixPtr input, MatrixPtr weight,
                                        const IVector& sequence,
                                        int contextLength,
                                        int contextStart, size_t beginPad,
                                        bool isPadding) {
    LOG(FATAL) << "Not implemeted";
  }

  virtual void contextProjectionBackward(MatrixPtr inputGrad,
                                         MatrixPtr weightGrad,
                                         const IVector& sequence,
                                         int contextLength,
                                         int contextStart, size_t beginPad,
                                         bool isPadding) {
    LOG(FATAL) << "Not implemeted";
  }

  virtual void contextProjectionBackwardData(MatrixPtr inputGrad,
                                             const IVector& sequence,
                                             int contextLength,
                                             int contextStart) {
    LOG(FATAL) << "Not implemeted";
  }

  virtual void contextProjectionBackwardWeight(MatrixPtr weightGrad,
                                               const IVector& sequence,
                                               int contextLength,
                                               int contextStart, int totalPad,
                                               size_t beginPad) {
    LOG(FATAL) << "Not implemeted";
  }

  /**
   * @code
   * this.row[i] += table.row[ids[i]]
   * if ids[i] == -1, it will be ignored
   * @endcode
   */
  virtual void selectRows(Matrix& table, IVector& ids) {
    (void)table;
    (void)ids;
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @code
   * this[i] = table[i, id[i]]
   * @endcode
   */
  virtual void selectElements(Matrix& table, IVector& ids) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @code
   * table.row[ids[i]] += this.row[i]
   * if ids[i] == -1, it will be ignored
   * @endcode
   */
  virtual void addToRows(Matrix& table, IVector& ids) {
    (void)table;
    (void)ids;
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @code
   * table[i, id[i]] += this[i]
   * @endcode
   */
  virtual void addElements(Matrix& table, IVector& ids) {
    LOG(FATAL) << "Not implemented";
  }
  /**
   * @brief  cross entropy for multi binary labels
   *
   * @code
   * this[i] = -sum(label[i][j]*log(output[i][j])
   *           + (1-label[i][j])*log(1-output[i][j]))
916
   * @endcode
Z
zhangjinchao01 已提交
917 918 919 920 921 922 923 924 925 926 927
   */
  virtual void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @brief  The gradient of cross entropy for multi binary labels on output
   *
   * @code
   * this[i][j] = -label[i][j]/output[i][j]
   *              + (1-label[i][j])/(1-output[i][j])
928
   * @endcode
Z
zhangjinchao01 已提交
929 930 931 932 933 934 935
   */
  virtual void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label) {
    LOG(FATAL) << "Not implemented";
  }

  /**
   * @brief  Calculate the classification error for multi binary labels
936
   *
Z
zhangjinchao01 已提交
937 938 939 940
   * @code
   * this[i] = sum((output[i][j] >= threshold && label[i][j] == 0)
   *            || (output[i][j] < threshold && label[i][j] == 1))
   *            / output->getWidth()
941
   * @endcode
Z
zhangjinchao01 已提交
942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005
   */
  virtual void classificationErrorMulti(Matrix& output, Matrix& label,
                                        real threshold) {
    LOG(FATAL) << "Not implemented";
  }

  virtual void paramReluForward(Matrix& data, Matrix& W) {
    LOG(FATAL) << "Not implemented";
  }
  virtual void paramReluBackwardW(Matrix& oGrad, Matrix& data) {
    LOG(FATAL) << "Not implemented";
  }
  virtual void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) {
    LOG(FATAL) << "Not implemented";
  }
};

inline std::ostream& operator<<(std::ostream& os, const Matrix& mat) {
  mat.print(os);
  return os;
}

class GpuMatrix : public Matrix {
public:
  GpuMatrix();

  GpuMatrix(size_t height, size_t width, bool trans = false);
  GpuMatrix(real* data, size_t height, size_t width, bool trans = false)
      : Matrix(data, height, width, trans, true) {}
  GpuMatrix(real* data, size_t height, size_t width, size_t stride,
            bool trans = false)
      : Matrix(data, height, width, stride, trans, true) {}
  GpuMatrix(GpuMemHandlePtr dataHandle, size_t height, size_t width,
            bool trans = false)
      : Matrix(dataHandle, height, width, trans, true) {}
  ~GpuMatrix();

  void zeroMem();
  void resetOne();

  void resize(size_t newHeight, size_t newWidth);
  void resize(size_t newHeight, size_t newWidth,
              size_t newNnz, /* used to allocate space */
              SparseValueType valueType, SparseFormat format) {
    LOG(FATAL) << "Only Support Sparse Matrix";
  }
  void setRow(size_t row, size_t colNum, const unsigned int* cols,
              const real* values) {
    LOG(FATAL) << "Only Support Sparse Matrix";
  }

  /**
   * Copy the data from cpu_memory buffer
   */
  void copyFrom(const real* hostSrc, size_t size);

  void copyFrom(const real* hostSrc, const int64_t* seq);

  void copyFrom(const Matrix& src, hl_stream_t stream);

  void copyFrom(const Matrix& src);

  void copyFrom(const IVector& src);

1006
  void copyByRowIndex(Matrix& b, const IVector& rowIndex);
Z
zhangjinchao01 已提交
1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111

  MatrixPtr clone(size_t height, size_t width, bool useGpu = false);

  real getElement(size_t x, size_t y) const;

  real* getRow(size_t row) { return BaseMatrix::rowBuf(row); }
  virtual real* getRowBuf(size_t row) { return getRow(row); }

  real getSum();
  void accumulateColSum(Matrix& src);
  real getAbsSum();

  MatrixPtr getTranspose();
  void transpose(MatrixPtr matTrans, bool memAlloc);

  /// add b to each sample of this.
  void addBias(Matrix& b, real scale);

  /**
   * @code
   * add each sample from a to this.
   * @endcode
   */
  void collectBias(Matrix& a, real scale);

  void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode);

  /**
   * @code
   * this.row[i] += table.row[ids[i]]
   * @endcode
   */
  virtual void selectRows(Matrix& table, IVector& ids);

  /**
   * @code
   * this[i] = table[i, id[i]]
   * @endcode
   */
  virtual void selectElements(Matrix& table, IVector& ids);

  /**
   * @code
   * table.row[ids[i]] += this.row[i]
   * @endcode
   */
  virtual void addToRows(Matrix& table, IVector& ids);

  void addColumnVector(const Matrix& b);

  /**
   * @code
   * this = scaleAB*(a*b) + scaleT*this
   * @endcode
   */
  void mul(const MatrixPtr a, const MatrixPtr b, real scaleAB, real scaleT);

  /**
   * @code
   * this = a*b
   * @endcode
   */
  void mul(const MatrixPtr a, const MatrixPtr b);

  void mul(const GpuMatrix& a, const GpuMatrix& b, real scaleAB, real scaleT);

  void mul(const GpuSparseMatrix& a, const GpuMatrix& b, real scaleAB,
           real scaleT);

  void mul(const GpuMatrix& a, const GpuSparseMatrix& b, real scaleAB,
           real scaleT);

  /**
   * @code
   * this = scaleAB*(this*b) +  scaleT*this
   * @endcode
   */
  void rightMul(Matrix& b, real scaleAB, real scaleT);

  /**
   * @code
   * this = this* b
   * @endcode
   */
  void rightMul(Matrix& b);

  /**
   * @code
   * this = scaleAB*(a*this) +  scaleT*this
   * @endcode
   */
  void leftMul(Matrix& a, real scaleAB, real scaleT);

  /**
   * @code
   * this = a*this
   * @endcode
   */
  void leftMul(Matrix& a);

  void colMerge(Matrix& src);
  void rowSum(Matrix& sum);
  void rowMax(Matrix& max);
  void rowMax(IVector& maxIds, Matrix& max);
  void colMax(Matrix& max);
1112 1113 1114
  void colMax(IVector& maxIds, Matrix& max);
  void maxoutForward(Matrix& a, IVector& id, size_t channels, size_t groups);
  void maxoutBackward(Matrix& a, IVector& id, size_t channels, size_t groups);
Z
zhangjinchao01 已提交
1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166

  void oneHotCrossEntropy(Matrix& output, IVector& label);
  void oneHotCrossEntropyBp(Matrix& outputV, IVector& label);
  void oneHotCrossEntropyWithSelfNorm(Matrix& output, IVector& label,
                                      real alpha);
  void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, IVector& label,
                                        real alpha);

  void softmax(Matrix& output);
  void sequenceSoftmax(Matrix& output, const IVector& index);
  void softmaxBackward(Matrix& outputV);
  void softmaxDerivative(Matrix& output, Matrix& sftmaxSum);

  /// calculate the sum of squares diff cost.
  void sumOfSquares(Matrix& output, Matrix& label);

  /// gradient of sumOfSquares.
  void sumOfSquaresBp(Matrix& outputV, Matrix& label);
  void tanh(Matrix& output);
  void tanhDerivative(Matrix& output);
  void softrelu(Matrix& output);
  void softreluDerivative(Matrix& output);
  void scaledTanh(Matrix& output, real p1, real p2);

  void cosSim(Matrix& output1, Matrix& output2, real scale);
  void cosSimDerivative(Matrix& output, Matrix& prevOut1, Matrix& prevOut2,
                        Matrix& prevGrad1, Matrix& prevGrad2, real scale);

  virtual void print(std::ostream& os) const;
  virtual void print(std::ostream& os, size_t height, size_t width) const;

  void paramReluForward(Matrix& data, Matrix& W);
  void paramReluBackwardW(Matrix& oGrad, Matrix& data);
  void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W);

  void check(std::ostream& os, Matrix& refMat, bool printDiff = true);
  void randomizeUniform();

  void classificationError(MatrixPtr output, IVectorPtr label);

  void convExpand(Matrix& feature, int feaImgHeight, int feaImgWidth,
                  int channels, int blockH, int blockW, int strideH,
                  int strideW, int paddingH, int paddingW,
                  int outputH, int outputW);

  void convShrink(Matrix& expandColMat, int thisImgHeight, int thisImgWidth,
                  int channels, int blockH, int blochW, int strideH,
                  int strideW, int paddingH, int paddingWreal,
                  int outputH, int outputW,
                  real alpha = 1.0f, real beta = 0.0f);

  void maxPoolForward(Matrix& inputMat, size_t imgSizeH, size_t imgSizeW,
1167 1168 1169 1170
                      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);
Z
zhangjinchao01 已提交
1171 1172

  void maxPoolBackward(Matrix& image, size_t imgSizeH, size_t imgSizeW,
1173 1174 1175 1176 1177
                       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
1178 1179

  void avgPoolForward(Matrix& input, size_t imgSizeH, size_t imgSizeW,
1180 1181 1182 1183
                      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);
Z
zhangjinchao01 已提交
1184 1185

  void avgPoolBackward(Matrix& input, size_t imgSizeH, size_t imgSizeW,
1186 1187 1188 1189 1190
                       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
1191 1192 1193

  void crossMapNormalFwd(Matrix& input, size_t imgSizeH, size_t imgSizeW,
                         Matrix& denoms, size_t channels, size_t sizeX,
1194
                         float scale, float pow);
Z
zhangjinchao01 已提交
1195 1196 1197

  void crossMapNormalBwd(Matrix& localGrad, Matrix& denoms, Matrix& preOutV,
                         Matrix& localOutV, size_t channels, size_t imgSizeH,
1198 1199
                         size_t imgSizeW, size_t sizeX,
                         float scale, float pow);
Z
zhangjinchao01 已提交
1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270

  void maxSequenceForward(Matrix& input, const IVector& sequence,
                          IVector& index);

  void maxSequenceBackward(Matrix& outputGrad, const IVector& sequence,
                           IVector& index);

  void contextProjectionForward(MatrixPtr input, MatrixPtr weight,
                                const IVector& sequence, int contextLength,
                                int contextStart, size_t beginPad,
                                bool isPadding);

  void contextProjectionBackwardData(MatrixPtr inputGrad,
                                     const IVector& sequence,
                                     int contextLength, int contextStart);

  void contextProjectionBackwardWeight(MatrixPtr weightGrad,
                                       const IVector& sequence,
                                       int contextLength,
                                       int contextStart, int totalPad,
                                       size_t beginPad);
};

class CpuMatrix : public Matrix {
public:
  CpuMatrix(size_t height, size_t width, bool trans = false);
  CpuMatrix(real* data, size_t height, size_t width, bool trans = false)
      : Matrix(data, height, width, trans, false) {}
  CpuMatrix(real* data, size_t height, size_t width, size_t stride,
            bool trans = false)
      : Matrix(data, height, width, stride, trans, false) {}

  CpuMatrix(CpuMemHandlePtr dataHandle, size_t height, size_t width,
            bool trans = false)
      : Matrix(dataHandle, height, width, trans, false) {}

  ~CpuMatrix();

  void zeroMem();
  void resetOne();
  void resize(size_t newHeight, size_t newWidth);
  void resize(size_t newHeight, size_t newWidth,
              size_t newNnz, /* used to allocate space */
              SparseValueType valueType, SparseFormat format) {
    LOG(FATAL) << "Only Support Sparse Matrix";
  }
  void setRow(size_t row, size_t colNum, const unsigned int* cols,
              const real* values) {
    LOG(FATAL) << "Only Support Sparse Matrix";
  }

  real getElement(size_t x, size_t y) const;
  real getSum();
  void accumulateColSum(Matrix& src);
  real getAbsSum();

  MatrixPtr getTranspose();
  void transpose(MatrixPtr matTrans, bool memAlloc);

  void copyFrom(const Matrix& src);

  void copyFrom(const Matrix& src, hl_stream_t stream);

  void copyFrom(const real* cpuSrc, size_t size);

  void copyFrom(const real* cpuSrc, const int64_t* seq);

  void copyFrom(const IVector& src);

  void copyFrom(CpuSparseMatrix& src);

1271
  void copyByRowIndex(Matrix& b, const IVector& rowIndex);
Z
zhangjinchao01 已提交
1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286

  MatrixPtr clone(size_t height, size_t width, bool useGpu = false);

  void convExpand(Matrix& feature, int feaImgHeight, int feaImgWidth,
                  int channels, int blcokH, int blockW, int strideH,
                  int strideW, int paddingH, int paddingW,
                  int outputH, int outputW);

  void convShrink(Matrix& expandFeat, int thisImgHeight, int thisImgWidth,
                  int channels, int blockH, int blockW, int strideH,
                  int strideW, int paddingH, int paddingW,
                  int outputH, int outputW,
                  real alpha = 1.0f, real beta = 0.0f);

  void maxPoolForward(Matrix& inputMat, size_t imgSizeH, size_t imgSizeW,
1287 1288 1289 1290
                      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);
Z
zhangjinchao01 已提交
1291 1292

  void maxPoolBackward(Matrix& image, size_t imgSizeH, size_t imgSizeW,
1293 1294 1295 1296 1297 1298
                       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
1299 1300

  void avgPoolForward(Matrix& input, size_t imgSizeH, size_t imgSizeW,
1301 1302 1303 1304
                      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);
Z
zhangjinchao01 已提交
1305 1306

  void avgPoolBackward(Matrix& input, size_t imgSizeH, size_t imgSizeW,
1307 1308 1309 1310 1311
                       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
1312 1313 1314

  void crossMapNormalFwd(Matrix& input, size_t imgSizeH, size_t imgSizeW,
                         Matrix& denoms, size_t channels, size_t sizeX,
1315
                         float scale, float pow);
Z
zhangjinchao01 已提交
1316 1317 1318

  void crossMapNormalBwd(Matrix& localGrad, Matrix& denoms, Matrix& preOutV,
                         Matrix& localOutV, size_t channels, size_t imgSizeH,
1319 1320
                         size_t imgSizeW, size_t sizeX,
                         float scale, float pow);
Z
zhangjinchao01 已提交
1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361

  void maxSequenceForward(Matrix& input, const IVector& sequence,
                          IVector& index);

  void maxSequenceBackward(Matrix& outputGrad, const IVector& sequence,
                           IVector& index);

  void contextProjectionForward(MatrixPtr input, MatrixPtr weight,
                                const IVector& sequence, int contextLength,
                                int contextStart, size_t beginPad,
                                bool isPadding);

  void contextProjectionBackward(MatrixPtr inputGrad, MatrixPtr weightGrad,
                                 const IVector& sequence, int contextLength,
                                 int contextStart, size_t beginPad,
                                 bool isPadding);

  real* getRow(size_t row) { return BaseMatrix::rowBuf(row); }
  virtual real* getRowBuf(size_t row) { return getRow(row); }

public:
  /// add b to each sample of this.
  void addBias(Matrix& b, real scale);

  /// add each sample of a to this.
  void collectBias(Matrix& a, real scale);

  void sequenceAvgForward(Matrix& a, const IVector& startsPos, int mode);


  /**
   * @code
   * this.row[i] += table.row[ids[i]]
   * @endcode
   */
  virtual void selectRows(Matrix& table, IVector& ids);

  /**
   * @code
   * table.row[ids[i]] += this.row[i]
   * @endcode
1362
   */
Z
zhangjinchao01 已提交
1363 1364 1365 1366 1367 1368
  virtual void addToRows(Matrix& table, IVector& ids);

  /**
   * @code
   * this[i] = table[i, id[i]]
   * @endcode
1369
   */
Z
zhangjinchao01 已提交
1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424
  virtual void selectElements(Matrix& table, IVector& ids);

  /**
   * @code
   * table[i, id[i]] += this[i]
   * @endcode
   */
  virtual void addElements(Matrix& table, IVector& ids);

  /**
   * use abstract getRow() to get row from table.
   *
   * Define table as template instead of virtual class for performance sake.
   * internal used by above two virtual funcs.
   */
  template <typename TableMatType>
  void selectRowsImp(TableMatType& table, IVector& ids);
  template <typename TableMatType>
  void addToRowsImp(TableMatType& table, IVector& ids);

  void addColumnVector(const Matrix& b);

  void mul(const MatrixPtr a, const MatrixPtr b, real scaleAB, real scaleT);
  void mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT);

  void mul(CpuMatrix* a, CpuSparseMatrix* b, real scaleAB, real scaleT);

  static void mul(CpuMatrix* a, CpuMatrix* b, CpuSparseMatrix* c, real scaleAB,
                  real scaleT);

  /**
   * c = a * b
   *
   * use abstract getRow() to get row from B,C.
   * Define B,C as template instead of virtual class for performance sake.
   */
  template <typename MatBType, typename MatCType>
  static void mul(CpuSparseMatrix* a, MatBType* b, MatCType* c, real scaleAB,
                  real scaleT);

  virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT);

  void mul(const MatrixPtr a, const MatrixPtr b);

  void rightMul(Matrix& b, real scaleAB, real scaleT);
  void rightMul(Matrix& b);

  void leftMul(Matrix& a, real scaleAB, real scaleT);
  void leftMul(Matrix& a);
  void colMerge(Matrix& src);
  void rowSum(Matrix& sum);
  void rowMaxId(IVector& maxIds);
  void rowMax(Matrix& max);
  void rowMax(IVector& maxIds, Matrix& maxVal);
  void colMax(Matrix& max);
1425 1426 1427
  void colMax(IVector& maxIds, Matrix& maxVal);
  void maxoutForward(Matrix& a, IVector& id, size_t channels, size_t groups);
  void maxoutBackward(Matrix& a, IVector& id, size_t channels, size_t groups);
Z
zhangjinchao01 已提交
1428 1429 1430 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 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555
  void rowNormalizeL1(Matrix& out);

  void oneHotCrossEntropy(Matrix& output, IVector& label);
  void oneHotCrossEntropyBp(Matrix& outputV, IVector& label);
  void oneHotCrossEntropyWithSelfNorm(Matrix& output, IVector& label,
                                      real alpha);
  void oneHotCrossEntropyWithSelfNormBp(Matrix& outputV, IVector& label,
                                        real alpha);

  void circularConv(Matrix& b, Matrix& c);
  void circularConvDerivative(Matrix& output, Matrix& prevOut1,
                              Matrix& prevOut2, Matrix& prevGrad1,
                              Matrix& prevGrad2);

  void softmax(Matrix& output);
  void sequenceSoftmax(Matrix& output, const IVector& index);
  void softmaxDerivative(Matrix& output, Matrix& sftmaxSum);

  /// calculate the sum of squares diff cost.
  void sumOfSquares(Matrix& output, Matrix& label);

  /// gradient of sumOfSquares.
  void sumOfSquaresBp(Matrix& outputV, Matrix& label);

  void tanh(Matrix& output);
  void tanhDerivative(Matrix& output);

  void softrelu(Matrix& output);
  void softreluDerivative(Matrix& output);
  void scaledTanh(Matrix& output, real p1, real p2);

  void cosSim(Matrix& output1, Matrix& output2, real scale);
  void cosSimDerivative(Matrix& output, Matrix& prevOut1, Matrix& prevOut2,
                        Matrix& prevGrad1, Matrix& prevGrad2, real scale);

  void print(std::ostream& os) const;
  void print(std::ostream& os, size_t height, size_t width) const;
  void printOneRow(std::ostream& os, size_t idx) const;

  void paramReluForward(Matrix& data, Matrix& W);
  void paramReluBackwardW(Matrix& oGrad, Matrix& data);
  void paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W);

  void check(std::ostream& os, Matrix& refMat, bool printDiff = true);

  real getMin();
  real getMax();

  void randomizeUniform();

  void classificationError(MatrixPtr output, IVectorPtr label);

  void addByBitCode(size_t numClasses, const IVector& codes, const Matrix& vec);

  void addByBitCodeBackward(size_t numClasses, const IVector& codes,
                            Matrix& vec);

  void mulByBitCode(size_t numClasses, const IVector& codes, const Matrix& mat,
                    const Matrix& input);

  void mulByBitCodeBackwardWeight(size_t numClasses, const IVector& codes,
                                  Matrix& mat, const Matrix& input);

  void mulByBitCodeBackwardError(size_t numClasses, const IVector& codes,
                                 const Matrix& mat, Matrix& input);

  void sumByBitCode(size_t numClasses, IVector& codes, Matrix& sum,
                    real scaleSum);

  void subByBitCode(size_t numClasses_, IVector& codes);

  void multiBinaryLabelCrossEntropy(Matrix& output, Matrix& label);
  void multiBinaryLabelCrossEntropyBp(Matrix& output, Matrix& label);
  void classificationErrorMulti(Matrix& output, Matrix& label, real threshold);
};

class SharedCpuMatrix : public CpuMatrix {
public:
  /* blockNum is number of partitions of the matrix  */
  SharedCpuMatrix(int blockNum, size_t height, size_t width, bool trans = false)
      : CpuMatrix(height, width, trans) {
    initShared(blockNum);
  }
  SharedCpuMatrix(int blockNum, real* data, size_t height, size_t width,
                  bool trans = false)
      : CpuMatrix(data, height, width, trans) {
    initShared(blockNum);
  }

  SharedCpuMatrix(int blockNum, CpuMemHandlePtr dataHandle, size_t height,
                  size_t width, bool trans = false)
      : CpuMatrix(dataHandle, height, width, trans) {
    initShared(blockNum);
  }

  SharedCpuMatrix(CpuMemHandlePtr dataHandle, size_t height,
                  size_t width, bool trans = false)
      : CpuMatrix(dataHandle, height, width, trans) {
    initBlock(1);
  }

  ~SharedCpuMatrix() {}

public:
  virtual void mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB, real scaleT);
  void add(Matrix& b, real p1, real p2);
  void add(real p1, real p2);

private:
  void initShared(int blockNum);
  void initBlock(int blockNum);

  int blockNum_;
  std::vector<std::unique_ptr<std::mutex>> blockLocks_;
  ThreadLocal<CpuMatrixPtr> localBuf_;
  ThreadLocal<std::vector<int>> localBufRows_;
  ThreadLocal<std::vector<int>> blockSeq_;
};

typedef struct { unsigned int col; } sparse_non_value_t;

typedef struct {
  unsigned int col;
  float value;
} sparse_float_value_t;

}  // namespace paddle
#include "ExecViaCpu.h"