CpuSparseMatrix.cpp 23.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

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 "hl_gpu.h"
#include "CpuSparseMatrix.h"
#include "SparseMatrix.h"
#include "paddle/math/MathUtils.h"
#include "paddle/utils/Util.h"
#include "float.h"

namespace paddle {

const size_t CpuSparseMatrix::DEFAULT_AVG_WIDTH;

26 27 28 29 30
CpuSparseMatrix::CpuSparseMatrix(size_t height,
                                 size_t width,
                                 size_t nnz,
                                 SparseValueType valueType,
                                 SparseFormat format,
Z
zhangjinchao01 已提交
31 32 33 34 35
                                 bool trans)
    : Matrix(NULL, height, width, trans, false) {
  resize(height, width, nnz, valueType, format);
}

36 37 38 39 40 41
CpuSparseMatrix::CpuSparseMatrix(CpuMemHandlePtr dataHandle,
                                 size_t height,
                                 size_t width,
                                 size_t nnz,
                                 SparseValueType valueType,
                                 SparseFormat format,
Z
zhangjinchao01 已提交
42 43 44 45 46
                                 bool trans)
    : Matrix(dataHandle, height, width, trans, false) {
  resize(height, width, nnz, valueType, format);
}

47 48 49 50 51 52 53 54
CpuSparseMatrix::CpuSparseMatrix(real* data,
                                 int* rows,
                                 int* cols,
                                 size_t height,
                                 size_t width,
                                 size_t nnz,
                                 SparseValueType valueType,
                                 SparseFormat format,
Z
zhangjinchao01 已提交
55 56 57 58 59 60 61 62 63 64 65 66
                                 bool trans)
    : Matrix(NULL, height, width, trans, false) {
  cols_ = cols;
  rows_ = rows;
  value_ = data;
  height_ = height;
  width_ = width;
  elementCnt_ = nnz;
  valueType_ = valueType;
  format_ = format;
}

67 68 69 70 71
void CpuSparseMatrix::resize(size_t newHeight,
                             size_t newWidth,
                             size_t newNnz,
                             SparseValueType valueType,
                             SparseFormat format) {
Z
zhangjinchao01 已提交
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
  CHECK_LE(newNnz, newHeight * newWidth);
  size_t newSize = 0;
  if (format == SPARSE_CSR) {
    newSize = (newHeight + 1) * sizeof(int) + newNnz * sizeof(int);
  } else {
    newSize = (newWidth + 1) * sizeof(int) + newNnz * sizeof(int);
  }

  if (NO_VALUE != valueType) {
    newSize += newNnz * sizeof(real);
  }

  if (NULL == memoryHandle_.get() || newSize > memoryHandle_->getSize()) {
    memoryHandle_ = std::make_shared<CpuMemoryHandle>(newSize);
  }

  height_ = newHeight;
  width_ = newWidth;
  elementCnt_ = newNnz;
  valueType_ = valueType;
  format_ = format;
  sparseResize();
}
void CpuSparseMatrix::sparseResize() {
  if (format_ == SPARSE_CSR) {
    rows_ = reinterpret_cast<int*>(
        reinterpret_cast<char*>(memoryHandle_->getBuf()));
    cols_ = reinterpret_cast<int*>(
        reinterpret_cast<char*>(memoryHandle_->getBuf()) +
        (height_ + 1) * sizeof(int));
    if (NO_VALUE != valueType_) {
      value_ = reinterpret_cast<real*>(
          reinterpret_cast<char*>(memoryHandle_->getBuf()) +
          (height_ + 1) * sizeof(int) + elementCnt_ * sizeof(int));
    } else {
      value_ = NULL;
    }
  } else {
    cols_ = reinterpret_cast<int*>(
        reinterpret_cast<char*>(memoryHandle_->getBuf()));
    rows_ = reinterpret_cast<int*>(
        reinterpret_cast<char*>(memoryHandle_->getBuf()) +
        (width_ + 1) * sizeof(int));
    if (NO_VALUE != valueType_) {
      value_ = reinterpret_cast<real*>(
          reinterpret_cast<char*>(memoryHandle_->getBuf()) +
          (width_ + 1) * sizeof(int) + elementCnt_ * sizeof(int));
    } else {
      value_ = NULL;
    }
  }
}

void CpuSparseMatrix::resize(size_t newHeight, size_t newWidth) {
126 127 128 129 130
  resize(newHeight,
         newWidth,
         newHeight * std::min(DEFAULT_AVG_WIDTH, newWidth),
         valueType_,
         format_);
Z
zhangjinchao01 已提交
131 132 133 134
}

MatrixPtr CpuSparseMatrix::getTranspose() {
  if (!memoryHandle_ && !value_) {
135 136
    MatrixPtr dest(new CpuSparseMatrix(
        height_, width_, elementCnt_, valueType_, format_, true));
Z
zhangjinchao01 已提交
137 138 139
    return dest;
  } else if (memoryHandle_) {
    MatrixPtr dest(new CpuSparseMatrix(
140 141 142 143 144 145 146
        std::dynamic_pointer_cast<CpuMemoryHandle>(memoryHandle_),
        height_,
        width_,
        elementCnt_,
        valueType_,
        format_,
        true));
Z
zhangjinchao01 已提交
147 148
    return dest;
  } else if (value_) {
149 150 151 152 153 154 155 156 157
    MatrixPtr dest(new CpuSparseMatrix(value_,
                                       rows_,
                                       cols_,
                                       height_,
                                       width_,
                                       elementCnt_,
                                       valueType_,
                                       format_,
                                       true));
Z
zhangjinchao01 已提交
158 159 160 161 162 163 164 165 166 167 168 169 170
    return dest;
  } else {
    return NULL;
  }
}

SparseValueType CpuSparseMatrix::getValueType() { return valueType_; }

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

  if (dynamic_cast<CpuMatrix*>(a.get()) && dynamic_cast<CpuMatrix*>(b.get())) {
    CpuMatrix::mul(dynamic_cast<CpuMatrix*>(a.get()),
171 172 173 174
                   dynamic_cast<CpuMatrix*>(b.get()),
                   this,
                   scaleAB,
                   scaleT);
Z
zhangjinchao01 已提交
175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276
  } else {
    LOG(FATAL) << "not supported";
  }
}

void CpuSparseMatrix::add3(CpuMatrix* b) {
  CHECK(getFormat() != SPARSE_CSC) << "Not supported";
  CHECK(height_ == b->getHeight());
  CHECK(width_ == b->getWidth());
  real* A = getValue();
  real* B = b->getData();
  int* cols = getCols();
  for (size_t i = 0; i < height_; i++) {
    size_t start = getRowStartIdx(i);
    size_t end = getRowStartIdx(i + 1);
    for (size_t j = start; j < end; j++) {
      A[j] = B[i * width_ + cols[j]];
    }
  }
}

void CpuSparseMatrix::add3(MatrixPtr b) {
  if (dynamic_cast<CpuMatrix*>(b.get())) {
    add3(dynamic_cast<CpuMatrix*>(b.get()));
  } else {
    LOG(FATAL) << "not supported";
  }
}

void CpuSparseMatrix::addBias(Matrix& b, real scale) {
  CHECK_EQ(b.getHeight(), (size_t)1);
  CHECK_EQ(width_, b.getWidth());
  real* A = getValue();
  real* B = b.getData();
  int* cols = getCols();
  size_t nnz = getElementCnt();
  for (size_t i = 0; i < nnz; i++) {
    A[i] += scale * B[cols[i]];
  }
}

template <class T>
void printBuf(std::ostream& os, T* a, size_t len, const char* name) {
  os << "\n: " << name << " [";
  for (size_t i = 0; i < len; i++) {
    os << a[i] << " ";
  }
  os << "]\n";
}

void CpuSparseMatrix::print(std::ostream& os) const {
  size_t rowSize = format_ == SPARSE_CSC ? elementCnt_ : height_ + 1;
  size_t colSize = format_ == SPARSE_CSC ? width_ + 1 : elementCnt_;
  printBuf(os, rows_, rowSize, "row");
  printBuf(os, cols_, colSize, "col");
  if (valueType_ == FLOAT_VALUE) {
    printBuf(os, value_, elementCnt_, "value");
  }
  return;
}

void CpuSparseMatrix::printOneRow(std::ostream& os, size_t idx) const {
  CHECK_LT(idx, height_);
  if (format_ == SPARSE_CSC) {
    LOG(FATAL) << "SPARSE_CSC not supported";
    return;
  }

  const int* col = getRowCols(idx);
  size_t num = getColNum(idx);
  if (num > 0) {
    if (valueType_ == FLOAT_VALUE) {
      const real* data = getRowValues(idx);
      os << col[0] << ":" << data[0];
      for (size_t i = 1; i < num; ++i) {
        os << " " << col[i] << ":" << data[i];
      }
    } else {
      os << col[0];
      for (size_t i = 1; i < num; ++i) {
        os << " " << col[i];
      }
    }
  }
  os << ";";
}

void CpuSparseMatrix::randomizeUniform() {
  CHECK_LE(elementCnt_, height_ * width_);
  if (valueType_ == FLOAT_VALUE) {
    real* data = getValue();
    for (size_t i = 0; i < elementCnt_; ++i) {
      *data++ = rand() / static_cast<real>(RAND_MAX);  // NOLINT
    }
  }
  if (format_ == SPARSE_CSR) {
    sparseRand(rows_, cols_, elementCnt_, height_ + 1, width_, false);
  } else {
    sparseRand(cols_, rows_, elementCnt_, width_ + 1, height_, false);
  }
}

277 278
void CpuSparseMatrix::copyFrom(std::vector<int>& rows,
                               std::vector<int>& cols,
Z
zhangjinchao01 已提交
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
                               std::vector<real>& values) {
  size_t size = format_ == SPARSE_CSR ? cols.size() : rows.size();
  resize(height_, width_, size, valueType_, format_);
  if (valueType_ == FLOAT_VALUE) {
    memcpy(&value_[0], &values[0], sizeof(real) * values.size());
  }
  memcpy(&cols_[0], &cols[0], sizeof(int) * cols.size());
  memcpy(&rows_[0], &rows[0], sizeof(int) * rows.size());
}

// Copy from a CpuMatrix, only supported in sparse_float_value_t
// SparseMatrix.
void CpuSparseMatrix::copyFrom(const CpuMatrix& src) {
  CHECK_EQ(getHeight(), src.getHeight());
  CHECK_EQ(getWidth(), src.getWidth());
  CHECK(!src.trans_ && !trans_);
  if (format_ == SPARSE_CSR) {
    std::vector<int> rows(getHeight() + 1);
    std::vector<int> cols;
    std::vector<real> values;
    rows[0] = 0;
    for (size_t r = 0; r < getHeight(); ++r) {
      for (size_t c = 0; c < getWidth(); ++c) {
        real v = src.getElement(r, c);
        if (fabs(v) > FLT_EPSILON) {
          cols.push_back(c);
          values.push_back(v);
        }
      }
      rows[r + 1] = values.size();
    }
    copyFrom(rows, cols, values);
  } else {
    std::vector<int> cols(getWidth() + 1);
    std::vector<int> rows;
    std::vector<real> values;
    cols[0] = 0;
    for (size_t r = 0; r < getWidth(); ++r) {
      for (size_t c = 0; c < getHeight(); ++c) {
        real v = src.getElement(c, r);
        if (fabs(v) > FLT_EPSILON) {
          rows.push_back(c);
          values.push_back(v);
        }
      }
      cols[r + 1] = values.size();
    }
    copyFrom(rows, cols, values);
  }
}

MatrixPtr CpuSparseMatrix::clone(size_t height, size_t width, bool useGpu) {
  if (height == 0 && width == 0) {
    height = height_;
    width = width_;
  }
  CHECK(width && height);
  if (!useGpu) {
337 338
    return std::make_shared<CpuSparseMatrix>(
        height, width, 0, valueType_, format_);
Z
zhangjinchao01 已提交
339
  } else {
340 341
    return std::make_shared<GpuSparseMatrix>(
        height, width, elementCnt_, valueType_, format_);
Z
zhangjinchao01 已提交
342 343 344 345 346 347 348 349
  }
}

MatrixPtr CpuSparseMatrix::subMatrix(size_t startRow, size_t numRows) {
  CHECK_LE(startRow + numRows, height_);
  CHECK_EQ(format_, SPARSE_CSR);
  if (valueType_ == NO_VALUE) {
    return std::make_shared<CpuSparseMatrix>(
350 351 352 353 354 355 356 357
        nullptr,
        rows_ + startRow,
        cols_,
        numRows,
        width_,
        rows_[startRow + numRows] - rows_[startRow],
        valueType_,
        format_,
Z
zhangjinchao01 已提交
358 359 360
        trans_);
  } else {
    return std::make_shared<CpuSparseMatrix>(
361 362 363 364 365 366 367 368
        value_,
        rows_ + startRow,
        cols_,
        numRows,
        width_,
        rows_[startRow + numRows] - rows_[startRow],
        valueType_,
        format_,
Z
zhangjinchao01 已提交
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
        trans_);
  }
}

/* mem MUST be alloced outside (memAlloc=false) */
void CpuSparseMatrix::transpose(MatrixPtr matTrans, bool memAlloc) {
  CHECK(!memAlloc);
  CpuSparseMatrix* mat = dynamic_cast<CpuSparseMatrix*>(matTrans.get());
  if (format_ == SPARSE_CSR) {
    /*statistic element number in each col*/
    int* colCounters = mat->getRows() + 1;
    memset(colCounters, 0, sizeof(int) * width_);
    for (size_t i = 0; i < elementCnt_; ++i) {
      int col = cols_[i];
      colCounters[col]++;
    }
    /*fill mat rows */
    mat->getRows()[0] = 0;
    for (size_t i = 1; i < width_ + 1; i++) {
      mat->getRows()[i] = mat->getRows()[i - 1] + mat->getRows()[i];
    }
    /*fill mat values and cols*/
    std::vector<int> colNumVec(width_, 0);
    if (valueType_ == FLOAT_VALUE) {
      for (size_t i = 0; i < height_; i++) {
        for (int j = rows_[i]; j < rows_[i + 1]; j++) {
          int colIdx = cols_[j];
          int index = mat->getRows()[colIdx] + colNumVec[colIdx];
          mat->getCols()[index] = i;
          mat->getValue()[index] = value_[j];
          colNumVec[colIdx]++;
        }
      }
    } else {
      for (size_t i = 0; i < height_; i++) {
        for (int j = rows_[i]; j < rows_[i + 1]; j++) {
          int colIdx = cols_[j];
          int index = mat->getRows()[colIdx] + colNumVec[colIdx];
          mat->getCols()[index] = i;
          colNumVec[colIdx]++;
        }
      }
    }
  } else {
    /*statistic element number in each row*/
    int* rowCounters = mat->getCols() + 1;
    memset(rowCounters, 0, sizeof(int) * height_);
    for (size_t i = 0; i < elementCnt_; ++i) {
      int row = rows_[i];
      rowCounters[row]++;
    }

    /*fill mat cols */
    mat->getCols()[0] = 0;
    for (size_t i = 1; i < height_ + 1; i++) {
      mat->getCols()[i] = mat->getCols()[i - 1] + mat->getCols()[i];
    }
    /*fill mat values and rows*/
    std::vector<int> rowNumVec(height_, 0);
    if (valueType_ == FLOAT_VALUE) {
      for (size_t i = 0; i < width_; i++) {
        for (int j = cols_[i]; j < cols_[i + 1]; j++) {
          int rowIdx = rows_[j];
          int index = mat->getCols()[rowIdx] + rowNumVec[rowIdx];
          mat->getRows()[index] = i;
          mat->getValue()[index] = value_[j];
          rowNumVec[rowIdx]++;
        }
      }
    } else {
      for (size_t i = 0; i < width_; i++) {
        for (int j = cols_[i]; j < cols_[i + 1]; j++) {
          int rowIdx = rows_[j];
          int index = mat->getCols()[rowIdx] + rowNumVec[rowIdx];
          mat->getRows()[index] = i;
          rowNumVec[rowIdx]++;
        }
      }
    }
  }
}

451 452 453 454
void CpuSparseMatrix::setRow(size_t row,
                             size_t colNum,
                             const unsigned int* cols,
                             const real* values) {
Z
zhangjinchao01 已提交
455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542
  if (format_ == SPARSE_CSR) {
    CHECK_LT(row, height_);
    CHECK(NULL != cols);
    if (0 == row) {
      rows_[row] = 0;
    }
    rows_[row + 1] = rows_[row] + colNum;
    for (size_t i = 0; i < colNum; ++i) {
      cols_[rows_[row] + i] = cols[i];
    }
    if (valueType_ == NO_VALUE) {
      CHECK(!values);
    } else {
      for (size_t i = 0; i < colNum; ++i) {
        value_[rows_[row] + i] = values[i];
      }
    }
  } else {
    LOG(FATAL) << "not supported";
  }
}

void CpuSparseMatrix::fillRowIndices(IVectorPtr& outVec) const {
  if (format_ == SPARSE_CSR) {
    auto nnz = getElementCnt();
    IVector::resizeOrCreate(outVec, nnz, false);
    auto out = outVec->getData();
    int* rows = getRows();
    for (size_t i = 0; i < height_; i++) {
      for (int j = rows[i]; j < rows[i + 1]; j++) {
        out[j] = i;
      }
    }
  } else {
    LOG(FATAL) << "SPARSE_CSC not supported";
  }
}

ThreadLocal<std::vector<CpuSparseMatrixPtr>> CpuSparseMatrix::cpuLocalMats_;

CpuSparseMatrixPtr CpuSparseMatrix::getTmpSparseMatrix(size_t height,
                                                       size_t width) {
  std::vector<CpuSparseMatrixPtr>* localMats = cpuLocalMats_.get();
  auto it = localMats->begin();
  while (it != localMats->end()) {
    if (it->unique()) {
      (*it)->resize(height, width, elementCnt_, valueType_, format_);
      return *it;
    }
  }
  localMats->emplace_back(std::make_shared<CpuSparseMatrix>(
      height, width, elementCnt_, valueType_, format_, false));
  return localMats->back();
}

void CpuSparseMatrix::copyFrom(const Matrix& src, hl_stream_t stream) {
  if (dynamic_cast<const GpuSparseMatrix*>(&src)) {
    auto tmpSrc = dynamic_cast<const GpuSparseMatrix*>(&src);
    copyFrom(*tmpSrc, stream);
  } else if (dynamic_cast<const CpuSparseMatrix*>(&src)) {
    auto tmpSrc = dynamic_cast<const CpuSparseMatrix*>(&src);
    copyFrom(*tmpSrc);
  } else if (dynamic_cast<const CpuMatrix*>(&src)) {
    auto tmpSrc = dynamic_cast<const CpuMatrix*>(&src);
    copyFrom(*tmpSrc);
  } else {
    LOG(FATAL) << "not implemented";
  }
}

void CpuSparseMatrix::copyFrom(const Matrix& src) {
  if (dynamic_cast<const CpuSparseMatrix*>(&src)) {
    auto tmpSrc = dynamic_cast<const CpuSparseMatrix*>(&src);
    copyFrom(*tmpSrc);
  } else if (dynamic_cast<const CpuMatrix*>(&src)) {
    auto tmpSrc = dynamic_cast<const CpuMatrix*>(&src);
    copyFrom(*tmpSrc);
  } else {
    LOG(FATAL) << "not implemented";
  }
}

void CpuSparseMatrix::copyFrom(const GpuSparseMatrix& src, hl_stream_t stream) {
  CHECK_EQ(height_, src.getHeight());
  CHECK_EQ(width_, src.getWidth());
  CHECK_EQ(size_t(elementCnt_), src.getElementCnt());
  size_t valSize = valueType_ == NO_VALUE ? 0 : elementCnt_;
  if (format_ == SPARSE_CSC)
543 544 545 546 547 548 549 550
    hl_memcpy_from_csc_matrix(value_,
                              valSize,
                              rows_,
                              elementCnt_,
                              cols_,
                              width_ + 1,
                              src.sMatrix_.get(),
                              stream);
Z
zhangjinchao01 已提交
551
  else
552 553 554 555 556 557 558 559
    hl_memcpy_from_csr_matrix(value_,
                              valSize,
                              rows_,
                              height_ + 1,
                              cols_,
                              elementCnt_,
                              src.sMatrix_.get(),
                              stream);
Z
zhangjinchao01 已提交
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
}

void CpuSparseMatrix::copyFrom(const CpuSparseMatrix& src) {
  CHECK_EQ(height_, src.getHeight());
  CHECK_EQ(width_, src.getWidth());
  CHECK_EQ(format_, src.getFormat());
  int start = format_ == SPARSE_CSR ? src.getRows()[0] : src.getCols()[0];
  if (format_ == SPARSE_CSR) {
    size_t totalColNum = 0;
    for (size_t i = 0; i < height_; ++i) {
      totalColNum += src.getColNum(i);
    }
    resize(height_, width_, totalColNum, valueType_, format_);
    rows_[0] = 0;
    for (size_t i = 0; i < height_; ++i) {
      rows_[i + 1] = rows_[i] + src.getColNum(i);
    }
    memcpy(cols_, src.getCols() + start, totalColNum * sizeof(int));
  } else {
    size_t totalColNum = 0;
    for (size_t i = 0; i < width_; ++i) {
      totalColNum += src.getRowNum(i);
    }
    resize(height_, width_, totalColNum, valueType_, format_);
    cols_[0] = 0;
    for (size_t i = 0; i < width_; ++i) {
      cols_[i + 1] = cols_[i] + src.getRowNum(i);
    }
    memcpy(rows_, src.getRows() + start, totalColNum * sizeof(int));
  }

  // if have different value type, only copy rows and cols
  if (valueType_ == FLOAT_VALUE && src.getValueType() == FLOAT_VALUE) {
    memcpy(value_, src.getValue() + start, elementCnt_ * sizeof(real));
  }
}

597 598
void CpuSparseMatrix::copyRow(int offsets,
                              size_t colNum,
Z
zhangjinchao01 已提交
599 600 601 602 603 604
                              const sparse_non_value_t* row) {
  for (size_t j = 0; j < colNum; j++) {
    cols_[offsets + j] = row[j].col;
  }
}

605 606
void CpuSparseMatrix::copyRow(int offsets,
                              size_t colNum,
Z
zhangjinchao01 已提交
607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658
                              const sparse_float_value_t* row) {
  for (size_t j = 0; j < colNum; j++) {
    cols_[offsets + j] = row[j].col;
    value_[offsets + j] = row[j].value;
  }
}

template <class T>
void CpuSparseMatrix::copyFrom(int64_t* ids, int64_t* indices, T* data) {
  size_t totalColNum = 0;
  for (size_t i = 0; i < height_; ++i) {
    int64_t id = ids[i];
    totalColNum += indices[id + 1] - indices[id];
  }
  valueType_ = typeid(T) == typeid(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE;

  resize(height_, width_, totalColNum, valueType_, format_);

  rows_[0] = 0;
  for (size_t i = 0; i < height_; ++i) {
    int64_t id = ids[i];
    T* row = data + indices[id];
    size_t colNum = indices[id + 1] - indices[id];
    rows_[i + 1] = rows_[i] + colNum;
    copyRow(rows_[i], colNum, row);
  }
}

template <class T>
void CpuSparseMatrix::copyFrom(int64_t* indices, T* data) {
  CHECK(format_ == SPARSE_CSR);
  size_t totalColNum = indices[height_] - indices[0];
  valueType_ = typeid(T) == typeid(sparse_non_value_t) ? NO_VALUE : FLOAT_VALUE;
  resize(height_, width_, totalColNum, valueType_, format_);

  rows_[0] = 0;
  for (size_t i = 0; i < height_; ++i) {
    T* row = data + indices[i];
    size_t colNum = indices[i + 1] - indices[i];
    rows_[i + 1] = rows_[i] + colNum;
    copyRow(rows_[i], colNum, row);
  }
}

void CpuSparseMatrix::trimFrom(const CpuSparseMatrix& src) {
  CHECK_EQ(height_, src.getHeight());
  CHECK_LE(width_, src.getWidth());
  CHECK_EQ(format_, src.getFormat());
  CHECK_EQ(valueType_, src.getValueType());
  if (format_ == SPARSE_CSR) {
    int* srcCols = src.getCols();
    size_t numLessWidth =
659 660
        std::count_if(srcCols,
                      srcCols + src.getElementCnt(),
Z
zhangjinchao01 已提交
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
                      [this](size_t n) { return n < this->width_; });
    resize(height_, width_, numLessWidth, valueType_, format_);
    rows_[0] = 0;
    size_t index = 0;
    for (size_t r = 0; r < height_; ++r) {
      for (int i = src.getRows()[r]; i < src.getRows()[r + 1]; ++i) {
        if (srcCols[i] < static_cast<int>(width_)) {
          cols_[index] = srcCols[i];
          if (valueType_ == FLOAT_VALUE) {
            value_[index] = src.getValue()[i];
          }
          ++index;
        }
      }
      rows_[r + 1] = index;
    }
    CHECK_EQ(index, numLessWidth);
  } else {
    size_t numLessWidth = src.getCols()[width_] - src.getCols()[0];
    resize(height_, width_, numLessWidth, valueType_, format_);
    cols_[0] = 0;
    size_t index = 0;
    // note: c < width_, not src.getWidth();
    for (size_t c = 0; c < width_; ++c) {
      for (int i = src.getCols()[c]; i < src.getCols()[c + 1]; ++i) {
        rows_[index] = src.getRows()[i];
        if (valueType_ == FLOAT_VALUE) {
          value_[index] = src.getValue()[i];
        }
        ++index;
      }
      cols_[c + 1] = index;
    }
    CHECK_EQ(index, numLessWidth);
  }
}

void CpuSparseMatrix::zeroMem() {
  CHECK(valueType_ == FLOAT_VALUE);
700
  memset(value_, 0, elementCnt_ * sizeof(real));
Z
zhangjinchao01 已提交
701 702
}

703 704
template void CpuSparseMatrix::copyFrom(int64_t* ids,
                                        int64_t* indices,
Z
zhangjinchao01 已提交
705 706
                                        sparse_non_value_t* data);

707 708
template void CpuSparseMatrix::copyFrom(int64_t* ids,
                                        int64_t* indices,
Z
zhangjinchao01 已提交
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
                                        sparse_float_value_t* data);

template void CpuSparseMatrix::copyFrom(int64_t* indices,
                                        sparse_non_value_t* data);

template void CpuSparseMatrix::copyFrom(int64_t* indices,
                                        sparse_float_value_t* data);

void CpuSparseMatrix::rowMax(IVector& maxIds, Matrix& maxVal) {
  size_t numSamples = getHeight();
  size_t beam = maxVal.getWidth();
  CHECK_EQ(maxIds.getSize(), numSamples * beam);
  CHECK_EQ(maxVal.getHeight(), numSamples);
  maxVal.zeroMem();
  int* outids = maxIds.getData();
  real* outvalues = maxVal.getData();

  typedef std::pair<real, size_t> valuepair;
  std::vector<valuepair> vec;
  for (size_t i = 0; i < numSamples; i++) {
    vec.clear();

    auto num = getColNum(i);
    auto ids = getRowCols(i);
    auto values = getRowValues(i);
    for (size_t j = 0; j < num; j++) {
      vec.push_back(std::make_pair(values[j], ids[j]));
    }

    size_t outsize = std::min(num, beam);
739 740 741
    std::partial_sort(vec.begin(),
                      vec.begin() + outsize,
                      vec.end(),
Z
zhangjinchao01 已提交
742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757
                      [](const valuepair& a, const valuepair& b) {
                        return a.first > b.first;
                      });
    for (size_t j = 0; j < outsize; j++) {
      outids[i * beam + j] = vec[j].second;
      outvalues[i * beam + j] = vec[j].first;
    }
    if (outsize < beam) {
      // if the number of values to sort are less than the output size,
      // use -1 to indicate the end of valid sorted values.
      outids[i * beam + outsize] = -1;
    }
  }
}

}  // namespace paddle