SparseMatrix.h 6.8 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

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 <cstddef>
#include "CpuSparseMatrix.h"
H
hedaoyuan 已提交
18
#include "Matrix.h"
Z
zhangjinchao01 已提交
19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36

namespace paddle {

typedef std::shared_ptr<_hl_sparse_matrix_s> hl_sparse_matrix_s_ptr;

class GpuSparseMatrix : public Matrix {
public:
  MemoryHandlePtr sMemoryHandle_;
  int* rows_;
  int* cols_;
  real* value_;
  const char* end_; /* point to the end of sMemoryHandle_ */

  hl_sparse_matrix_s_ptr sMatrix_;
  SparseValueType valueType_;
  SparseFormat format_;

public:
37 38
  GpuSparseMatrix(size_t height,
                  size_t width,
Z
zhangjinchao01 已提交
39 40
                  size_t nnz, /* used to allocate space */
                  SparseValueType valueType = FLOAT_VALUE,
41 42
                  SparseFormat format_ = SPARSE_CSR,
                  bool trans = false);
Z
zhangjinchao01 已提交
43

44 45 46 47
  GpuSparseMatrix(GpuMemHandlePtr dataHandle,
                  hl_sparse_matrix_s_ptr sMatrix,
                  size_t height,
                  size_t width,
Z
zhangjinchao01 已提交
48 49
                  size_t nnz, /* used to allocate space */
                  SparseValueType valueType = FLOAT_VALUE,
50 51
                  SparseFormat format_ = SPARSE_CSR,
                  bool trans = false,
Z
zhangjinchao01 已提交
52 53
                  MemoryHandlePtr sMemoryHandle = NULL);

54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
  GpuSparseMatrix(real* value,
                  int* rows,
                  int* cols,
                  size_t height,
                  size_t width,
                  size_t nnz,
                  SparseValueType valueType,
                  SparseFormat format,
                  bool trans);

  GpuSparseMatrix(hl_sparse_matrix_s_ptr sMatrix,
                  size_t height,
                  size_t width,
                  size_t nnz,
                  SparseValueType valueType,
                  SparseFormat format,
                  bool trans,
                  MemoryHandlePtr sMemoryHandle);
Z
zhangjinchao01 已提交
72 73 74 75 76 77 78 79 80 81 82 83 84

protected:
  struct Element {
    int row;
    int col;
    real val;
    Element(int rowIn, int colIn, real valIn)
        : row(rowIn), col(colIn), val(valIn) {}
  };

public:
  ~GpuSparseMatrix() {}

85 86
  void resize(size_t newHeight,
              size_t newWidth,
Z
zhangjinchao01 已提交
87
              size_t newNnz, /* used to allocate space */
88 89
              SparseValueType valueType,
              SparseFormat format);
Z
zhangjinchao01 已提交
90 91 92 93 94 95 96

  void resize(size_t newHeight, size_t newWidth);

  void sparseResizeCSR();

  void sparseResizeCSC();

97 98 99
  void resizeCSR(size_t newHeight,
                 size_t newWidth,
                 size_t newNnz,
Z
zhangjinchao01 已提交
100 101
                 SparseValueType valueType);

102 103 104
  void resizeCSC(size_t newHeight,
                 size_t newWidth,
                 size_t newNnz,
Z
zhangjinchao01 已提交
105 106
                 SparseValueType valueType);

107
  void mul(const GpuMatrix& a, const GpuMatrix& b, real scaleAB, real scaleT);
Z
zhangjinchao01 已提交
108 109 110 111
  /// B = A , B.trans = !A.trans
  MatrixPtr getTranspose();

  /// B = A'
112
  void transpose(MatrixPtr& matTrans, bool memAlloc);
Z
zhangjinchao01 已提交
113 114 115 116 117 118 119 120 121 122 123 124 125 126

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

  void copyFrom(const IVector& src) { LOG(FATAL) << "not implemented"; }
  void copyFrom(const IVector& src, hl_stream_t stream) {
    LOG(FATAL) << "not implemented";
  }

  template <class T>
  void copyFrom(int64_t* ids, int64_t* indices, T* data, hl_stream_t stream);

127 128 129
  void setRow(size_t row,
              size_t colNum,
              const unsigned int* cols,
Z
zhangjinchao01 已提交
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
              const real* values);
  SparseValueType getValueType() const;
  SparseFormat getFormat() const { return format_; }

  const int* getRowCols(size_t x) const { return cols_ + rows_[x]; }
  const real* getRowValues(size_t x) const { return value_ + rows_[x]; }
  size_t getColNum(size_t x) const { return rows_[x + 1] - rows_[x]; }
  void print(std::ostream& os) const;

  /**
   * @brief only set value_ of FLOAT_VALUE sparse matrix to zero
   */
  void zeroMem();

  /**
   * @brief sparseMatrix += denseMatrix
   *
   * Named add3 just because add/add2 has been used in BaseMatrix.cu
   * and they are not virtual function.
   *
   * Only add value of same (row, col) index in dense matrix
   * and do not use others values.
   *
   * @param[in]  b   dense matrix
   */
  void add3(GpuMatrix* b);
  void add3(MatrixPtr b);

  /**
   * @brief sparseMatrix[i,j] += bias[j], (j is the col index of sparse matrix)
   *
   * @param[in]  b      bias, dense matrix and height = 1
   * @param[in]  scale  scale of b
   */
  void addBias(Matrix& b, real scale);

  /**
   * @brief return rows, which is gpu address
   */
  int* getRows() const {
    CHECK(sMatrix_.get()) << "sMatrix_ is NULL";
    return hl_sparse_matrix_get_rows(sMatrix_.get());
  }

  /**
   * @brief return cols, which is gpu address
   */
  int* getCols() const {
    CHECK(sMatrix_.get()) << "sMatrix_ is NULL";
    return hl_sparse_matrix_get_cols(sMatrix_.get());
  }

  /**
   * @brief return value, which is gpu address
   */
  real* getValue() const {
    CHECK(sMatrix_.get()) << "sMatrix_ is NULL";
    return hl_sparse_matrix_get_value(sMatrix_.get());
  }

  /**
   * @brief return value_ of sparse matrix
   *
   * Some times CpuSparseMatrix maybe Matrix,
   * if getValue, must dynamic_cast to CpuSparseMatrix,
   * getData is convenient to get value
   */
  real* getData() { return getValue(); }
198
  const real* getData() const { return getValue(); }
Z
zhangjinchao01 已提交
199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217

  /**
   * @brief  Get top k value of each row in sparse matrix.
   *
   * Store the value in maxVal and theirs index in maxIds.
   * k = maxVal.width
   *
   * @param[out]  maxIds    index of top k
   * @param[out]  maxVal    value of top k
   */
  void rowMax(IVector& maxIds, Matrix& maxVal);

protected:
  void sparseResize();

  void copyRow(int offsets, size_t colNum, const sparse_non_value_t* row);
  void copyRow(int offsets, size_t colNum, const sparse_float_value_t* row);

public:
218
  void mul(const Matrix& a, const Matrix& b, real scaleAB, real scaleT);
Z
zhangjinchao01 已提交
219 220 221 222 223 224 225 226 227 228

  void copyFrom(CpuSparseMatrix& src, hl_stream_t stream);
  void copyFrom(GpuSparseMatrix& src, hl_stream_t stream);

  void trimFrom(const CpuSparseMatrix& src);
  void trimFromCSR(const CpuSparseMatrix& src);
  void trimFromCSC(const CpuSparseMatrix& src);

  // BaseMatrixT interface
public:
229
  bool isSparse() const { return true; }
Z
zhangjinchao01 已提交
230 231 232 233

private:
  using Matrix::mul;
  using Matrix::copyFrom;
H
hedaoyuan 已提交
234 235 236
  using Matrix::rowMax;
  using Matrix::print;
  using Matrix::subMatrix;
Z
zhangjinchao01 已提交
237 238 239
};

}  // namespace paddle