matrix_bit_code.cc 12.5 KB
Newer Older
Y
Yancey1989 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2017 PaddlePaddle Authors. 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. */

W
weixing02 已提交
15
#include "paddle/fluid/operators/math/matrix_bit_code.h"
Y
Yu Yang 已提交
16

Y
Yancey1989 已提交
17 18 19 20
namespace paddle {
namespace operators {
namespace math {

Y
Yancey1989 已提交
21
template <typename T>
Y
Yu Yang 已提交
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
struct MatrixBitCodeFunctorAdd : public boost::static_visitor<void> {
  const framework::Tensor &vec_;
  framework::Tensor *tmat_;

  MatrixBitCodeFunctorAdd(const framework::Tensor &vec, framework::Tensor *tmat)
      : vec_(vec), tmat_(tmat) {}

  template <typename CodeTable>
  void operator()(const CodeTable &code_table) {
    size_t batch_size = tmat_->dims()[0];
    size_t width = tmat_->dims()[1];
    auto *tmat_data = tmat_->data<T>();
    auto *vec_data = vec_.data<T>();
    for (size_t i = 0; i < batch_size; ++i) {
      auto code = code_table.get_code(i);
      int code_length = code.get_length();
      for (int j = 0; j < code_length; ++j) {
        size_t index = code.calc_index(j);
        tmat_data[i * width + j] += vec_data[index];
      }
Y
Yancey1989 已提交
42 43
    }
  }
Y
Yu Yang 已提交
44 45 46 47 48 49 50
};

template <typename T>
void MatrixBitCodeFunctor<T>::Add(const framework::Tensor &vec,
                                  framework::Tensor *tmat) {
  MatrixBitCodeFunctorAdd<T> func(vec, tmat);
  code_table_.apply_visitor(func);
Y
Yancey1989 已提交
51 52
}

Y
Yancey1989 已提交
53
template <typename T>
Y
Yu Yang 已提交
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73
struct MatrixBitCodeFunctorAddGrad : public boost::static_visitor<void> {
  const framework::Tensor &tmat_;
  framework::Tensor *vec_;
  MatrixBitCodeFunctorAddGrad(const framework::Tensor &tmat,
                              framework::Tensor *vec)
      : tmat_(tmat), vec_(vec) {}

  template <typename CodeTable>
  void operator()(const CodeTable &table) {
    size_t batch_size = tmat_.dims()[0];
    size_t width = tmat_.dims()[1];
    auto *vec_data = vec_->data<T>();
    auto *tmat_data = tmat_.data<T>();
    for (size_t i = 0; i < batch_size; ++i) {
      auto code = table.get_code(i);
      int code_length = code.get_length();
      for (int j = 0; j < code_length; ++j) {
        size_t index = code.calc_index(j);
        vec_data[index] += tmat_data[i * width + j];
      }
Y
Yancey1989 已提交
74 75
    }
  }
Y
Yu Yang 已提交
76 77 78 79 80 81 82
};

template <typename T>
void MatrixBitCodeFunctor<T>::AddGrad(const framework::Tensor &tmat,
                                      framework::Tensor *vec) {
  MatrixBitCodeFunctorAddGrad<T> func(tmat, vec);
  code_table_.apply_visitor(func);
Y
Yancey1989 已提交
83 84
}

Y
Yancey1989 已提交
85
template <typename T>
Y
Yu Yang 已提交
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
struct MatrixBitCodeFunctorSum : public boost::static_visitor<void> {
  const framework::Tensor &tmat_;
  framework::Tensor *sum_;
  T scale_sum_;

  MatrixBitCodeFunctorSum(const framework::Tensor &tmat, framework::Tensor *sum,
                          T scale_sum)
      : tmat_(tmat), sum_(sum), scale_sum_(scale_sum) {}

  template <typename CodeTable>
  void operator()(const CodeTable &code_table) {
    size_t num_samples = tmat_.dims()[0];
    size_t o_width = tmat_.dims()[1];
    auto *tmat_data = tmat_.data<T>();
    auto *sum_data = sum_->data<T>();
    for (size_t i = 0; i < num_samples; ++i) {
      T sm = static_cast<T>(0.0);
      auto code = code_table.get_code(i);
      int code_length = code.get_length();
      for (int j = 0; j < code_length; ++j) {
        if (code.calc_bit(j)) {
          // calc_bit starts from right most bit, while data in tmat[i] is in
          // the
          // reverse order.
          sm += tmat_data[i * o_width + j];
        }
Y
Yancey1989 已提交
112
      }
Y
Yu Yang 已提交
113
      sum_data[i] = scale_sum_ * sm;
Y
Yancey1989 已提交
114 115
    }
  }
Y
Yu Yang 已提交
116 117 118 119 120 121 122
};

template <typename T>
void MatrixBitCodeFunctor<T>::Sum(const framework::Tensor &tmat,
                                  framework::Tensor *sum, T scale_sum) {
  MatrixBitCodeFunctorSum<T> func(tmat, sum, scale_sum);
  code_table_.apply_visitor(func);
Y
Yancey1989 已提交
123
}
Y
Yancey1989 已提交
124

Y
Yancey1989 已提交
125
template <typename T>
Y
Yu Yang 已提交
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
struct MatrixBitCodeFunctorMul : public boost::static_visitor<void> {
  framework::Tensor *tmat_;
  const framework::Tensor &weight_;
  const framework::Tensor &input_;

  MatrixBitCodeFunctorMul(framework::Tensor *tmat,
                          const framework::Tensor &weight,
                          const framework::Tensor &input)
      : tmat_(tmat), weight_(weight), input_(input) {}

  template <typename CodeTable>
  void operator()(const CodeTable &code_table) {
    auto blas =
        GetBlas<platform::CPUDeviceContext, T>(platform::CPUDeviceContext());
    size_t num_samples = tmat_->dims()[0];
    size_t tmat_width = tmat_->dims()[1];
    size_t input_width = input_.dims()[1];
    size_t weight_width = weight_.dims()[1];
    auto tmat_value = tmat_->data<T>();
    auto weight_value = weight_.data<T>();
    auto input_value = input_.data<T>();
    for (size_t i = 0; i < num_samples; ++i) {
      auto code = code_table.get_code(i);
      int code_length = code.get_length();
      const T *input_row = input_value + input_width * i;
      for (int j = 0; j < code_length; ++j) {
        size_t index = code.calc_index(j);
        const T *weight_row = weight_value + weight_width * index;
        T sum = blas.DOT(input_width, weight_row, input_row);
        tmat_value[i * tmat_width + j] += sum;
      }
Y
Yancey1989 已提交
157 158
    }
  }
Y
Yu Yang 已提交
159 160 161 162 163 164 165 166
};

template <typename T>
void MatrixBitCodeFunctor<T>::Mul(framework::Tensor *tmat,
                                  const framework::Tensor &weight,
                                  const framework::Tensor &input) {
  MatrixBitCodeFunctorMul<T> func(tmat, weight, input);
  code_table_.apply_visitor(func);
Y
Yancey1989 已提交
167 168
}

Y
Yu Yang 已提交
169 170 171 172 173 174
template <typename T, size_t N>
class ReservedVector : public std::vector<T> {
 public:
  ReservedVector() { this->reserve(N); }
};

Y
Yancey1989 已提交
175
template <typename T>
Y
Yu Yang 已提交
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
struct MatrixBitCodeFunctorMulGradWeight : public boost::static_visitor<void> {
  const framework::Tensor &tmat_;
  framework::Tensor *weight_;
  const framework::Tensor &input_;
  MatrixBitCodeFunctorMulGradWeight(const framework::Tensor &tmat,
                                    framework::Tensor *weight,
                                    const framework::Tensor &input)
      : tmat_(tmat), weight_(weight), input_(input) {}
  template <typename CodeTable>
  void operator()(const CodeTable &code_table) {
    auto blas =
        GetBlas<platform::CPUDeviceContext, T>(platform::CPUDeviceContext());
    size_t num_samples = tmat_.dims()[0];
    size_t input_width = input_.dims()[1];
    size_t tmat_width = tmat_.dims()[1];
    size_t weight_width = weight_->dims()[1];
    auto tmat_value = tmat_.data<T>();
    auto weight_value = weight_->data<T>();
    auto input_value = input_.data<T>();

    std::map<int, ReservedVector<std::pair<T, const T *>, 8u>> ops;
    for (size_t i = 0; i < num_samples; ++i) {
      auto code = code_table.get_code(i);
      int code_length = code.get_length();
      const T *input_value_row = input_value + input_width * i;
      const T *tmat_row = tmat_value + i * tmat_width;
      for (int j = 0; j < code_length; ++j) {
        ops[code.calc_index(j)].emplace_back(tmat_row[j], input_value_row);
      }
J
JiabinYang 已提交
205
    }
Y
Yu Yang 已提交
206 207 208 209 210 211 212 213
    for (auto &op : ops) {
      auto &op_in_row = op.second;
      for (auto &pair : op_in_row) {
        auto &scale = pair.first;
        auto *input_row = pair.second;
        T *weight_row = weight_value + op.first * weight_width;
        blas.AXPY(input_width, scale, input_row, weight_row);
      }
Y
Yancey1989 已提交
214
    }
Y
Yancey1989 已提交
215
  }
Y
Yu Yang 已提交
216 217 218 219 220 221 222 223
};

template <typename T>
void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor &tmat,
                                            framework::Tensor *weight,
                                            const framework::Tensor &input) {
  MatrixBitCodeFunctorMulGradWeight<T> func(tmat, weight, input);
  code_table_.apply_visitor(func);
Y
Yancey1989 已提交
224
}
Y
Yancey1989 已提交
225

J
JiabinYang 已提交
226
template <typename T>
Y
Yu Yang 已提交
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
struct MatrixBitCodeFunctorMulGradWeightSR
    : public boost::static_visitor<void> {
  const framework::Tensor &tmat_;
  framework::SelectedRows *weight_;
  const framework::Tensor &input_;

  MatrixBitCodeFunctorMulGradWeightSR(const framework::Tensor &tmat,
                                      framework::SelectedRows *weight,
                                      const framework::Tensor &input)
      : tmat_(tmat), weight_(weight), input_(input) {}

  template <typename CodeTable>
  void operator()(const CodeTable &code_table) {
    auto blas =
        GetBlas<platform::CPUDeviceContext, T>(platform::CPUDeviceContext());
    size_t num_samples = tmat_.dims()[0];
    size_t input_width = input_.dims()[1];
    size_t tmat_width = tmat_.dims()[1];
    size_t weight_width = weight_->value().dims()[1];
    auto tmat_value = tmat_.data<T>();
    auto weight_value = weight_->mutable_value()->data<T>();
    auto input_value = input_.data<T>();

    std::unordered_map<int, std::vector<std::pair<T, const T *>>> ops;
    ops.reserve(weight_->rows().size());

    for (size_t i = 0; i < num_samples; ++i) {
      auto code = code_table.get_code(i);
      int code_length = code.get_length();
      const T *input_value_row = input_value + input_width * i;
      const T *tmat_row = tmat_value + i * tmat_width;
      for (int j = 0; j < code_length; ++j) {
        ops[code.calc_index(j)].emplace_back(tmat_row[j], input_value_row);
      }
J
JiabinYang 已提交
261 262
    }

Y
Yu Yang 已提交
263 264 265 266 267 268 269 270
    for (auto &row : weight_->rows()) {
      auto &op_in_row = ops[row];
      for (auto &pair : op_in_row) {
        auto &scale = pair.first;
        auto *input_row = pair.second;
        blas.AXPY(input_width, scale, input_row, weight_value);
      }
      weight_value += weight_width;
J
JiabinYang 已提交
271 272
    }
  }
Y
Yu Yang 已提交
273 274 275 276 277 278 279 280
};

template <typename T>
void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor &tmat,
                                            framework::SelectedRows *weight,
                                            const framework::Tensor &input) {
  MatrixBitCodeFunctorMulGradWeightSR<T> func(tmat, weight, input);
  code_table_.apply_visitor(func);
J
JiabinYang 已提交
281
}
J
JiabinYang 已提交
282

Y
Yancey1989 已提交
283
template <typename T>
Y
Yu Yang 已提交
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
struct MatrixBitCodeFunctorMulGradError : public boost::static_visitor<void> {
  const framework::Tensor &tmat_;
  const framework::Tensor &weight_;
  framework::Tensor *input_;

  MatrixBitCodeFunctorMulGradError(const framework::Tensor &tmat,
                                   const framework::Tensor &weight,
                                   framework::Tensor *input)
      : tmat_(tmat), weight_(weight), input_(input) {}
  template <typename CodeTable>
  void operator()(const CodeTable &code_table) {
    size_t num_samples = tmat_.dims()[0];
    size_t tmat_width = tmat_.dims()[1];
    size_t input_width = input_->dims()[1];
    size_t weight_width = weight_.dims()[1];
    auto tmat_value = tmat_.data<T>();
    auto weight_value = weight_.data<T>();
    auto input_value = input_->data<T>();

    for (size_t i = 0; i < num_samples; ++i) {
      auto code = code_table.get_code(i);
      int code_length = code.get_length();
      for (int j = 0; j < code_length; ++j) {
        size_t index = code.calc_index(j);

        for (size_t k = 0; k < input_width; ++k) {
          input_value[input_width * i + k] +=
              tmat_value[i * tmat_width + j] *
              weight_value[weight_width * index + k];
        }
Y
Yancey1989 已提交
314 315 316
      }
    }
  }
Y
Yu Yang 已提交
317 318 319 320 321 322 323 324
};

template <typename T>
void MatrixBitCodeFunctor<T>::MulGradError(const framework::Tensor &tmat,
                                           const framework::Tensor &weight,
                                           framework::Tensor *input) {
  MatrixBitCodeFunctorMulGradError<T> func(tmat, weight, input);
  code_table_.apply_visitor(func);
Y
Yancey1989 已提交
325 326 327
}

template <typename T>
Y
Yu Yang 已提交
328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344
struct MatrixBitCodeFunctorSub : public boost::static_visitor<void> {
  framework::Tensor *tmat_;

  explicit MatrixBitCodeFunctorSub(framework::Tensor *tmat) : tmat_(tmat) {}

  template <typename CodeTable>
  void operator()(const CodeTable &code_table) {
    size_t num_samples = tmat_->dims()[0];
    size_t o_width = tmat_->dims()[1];
    auto *tmat_data = tmat_->data<T>();
    for (size_t i = 0; i < num_samples; ++i) {
      auto code = code_table.get_code(i);
      int code_length = code.get_length();
      for (int j = 0; j < code_length; ++j) {
        if (code.calc_bit(j)) {
          tmat_data[i * o_width + j] -= 1;
        }
Y
Yancey1989 已提交
345 346 347
      }
    }
  }
Y
Yu Yang 已提交
348 349 350 351 352 353
};

template <typename T>
void MatrixBitCodeFunctor<T>::Sub(framework::Tensor *tmat) {
  MatrixBitCodeFunctorSub<T> func(tmat);
  code_table_.apply_visitor(func);
Y
Yancey1989 已提交
354 355
}

Y
Yancey1989 已提交
356 357 358
template class MatrixBitCodeFunctor<float>;
template class MatrixBitCodeFunctor<double>;

Y
Yancey1989 已提交
359 360 361
}  // namespace math
}  // namespace operators
}  // namespace paddle