matrix_bit_code.h 4.0 KB
Newer Older
Y
Yancey1989 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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. */

#pragma once
W
weixing02 已提交
16 17 18
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/device_context.h"
Y
Yancey1989 已提交
19 20 21 22

namespace paddle {
namespace operators {
namespace math {
W
weixing02 已提交
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
/**
 * SimpleCodeTable class should support 3 functions:
 *
 * size_t size()
 *   return the number of ids
 *
 * int get_max_code_length()
 *   return the maximal code length
 *
 * SimpleCode operator()(size_t i)
 *   return the i-th code. Code class is descriebed below.
 *
 * SimpleCode class should support 3 functions:
 *
 * int get_length()
 *   return the length of the code
 *
 * size_t cal_index(int bit)
 *   bit ranges from 0 to get_length() - 1
 *   return the index for the (1+bit) level parent
 *
 * bool calc_bit(int bit)
 *   return true if the bit level parent is the right child of (1+bit) level
 *   parent
 *
 */
Y
Yancey1989 已提交
49 50 51 52 53 54

/**
 * return the 1-based index of the highest bit set
 *
 * for x > 0:
 * \f[
W
weixing02 已提交
55
 *    FindLastSet(x) = 1 + \floor*{\log_{2}x}
Y
Yancey1989 已提交
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
 * \f]
 */
inline constexpr size_t FindLastSet(size_t x) {
  return std::is_same<size_t, unsigned int>::value
             ? (x ? 8 * sizeof(x) - __builtin_clz(x) : 0)
             : (std::is_same<size_t, unsigned long>::value  // NOLINT
                    ? (x ? 8 * sizeof(x) - __builtin_clzl(x) : 0)
                    : (x ? 8 * sizeof(x) - __builtin_clzll(x) : 0));
}

struct SimpleCode {
  SimpleCode(size_t code, size_t num_classes) : c_(code + num_classes) {}
  inline size_t calc_index(int bit) const { return (c_ >> (bit + 1)) - 1; }
  inline bool calc_bit(int bit) const { return c_ & (1 << bit); }
  inline int get_length() const { return FindLastSet(c_) - 1; }

 private:
  size_t c_;
};

struct SimpleCodeTable {
  explicit SimpleCodeTable(size_t num_classes) : num_classes_(num_classes) {}
  SimpleCode operator()(size_t code) const {
    return SimpleCode(code, num_classes_);
  }
  size_t size() const { return num_classes_; }
  int get_max_code_length() const { return FindLastSet(num_classes_ - 1); }

 private:
  size_t num_classes_;
  int max_code_length_;
};

Y
Yancey1989 已提交
89
template <typename T>
Y
Yancey1989 已提交
90 91
class MatrixBitCodeFunctor {
 public:
Y
Yancey1989 已提交
92 93
  explicit MatrixBitCodeFunctor(size_t num_classes, const int64_t* ids)
      : num_classes_(num_classes), ids_(ids) {}
Y
Yancey1989 已提交
94 95 96
  /* For j < code_length
       tmat(i, j) += vec(0, index(i, j))
  */
W
weixing02 已提交
97
  void Add(framework::Tensor* tmat, const framework::Tensor& vec);
Y
Yancey1989 已提交
98

Y
Yancey1989 已提交
99 100 101
  /* For j < code_length
       vec(0, index(i, j)) += tmat(i, j)
  */
W
weixing02 已提交
102
  void AddGrad(const framework::Tensor& tmat, framework::Tensor* vec);
Y
Yancey1989 已提交
103 104

  /* For j < code_length
Y
Yancey1989 已提交
105
    sum(i, 0) = \sum_j bit(i, j) * tmat(i, j)
Y
Yancey1989 已提交
106
  */
W
weixing02 已提交
107
  void Sum(const framework::Tensor& tmat, framework::Tensor* sum, T scale_sum);
Y
Yancey1989 已提交
108

Y
Yancey1989 已提交
109 110 111
  /* For j < code_length
       tmat(i, j) -= bit(i, j)
  */
W
weixing02 已提交
112
  void Sub(framework::Tensor* tmat);
Y
Yancey1989 已提交
113 114 115
  /* For j < code_length
       input.row(i) += tmat(i, j) * weight.row(index(i, j))
  */
W
weixing02 已提交
116
  void Mul(framework::Tensor* tmat, const framework::Tensor& weight,
Y
Yancey1989 已提交
117
           const framework::Tensor& input);
Y
Yancey1989 已提交
118

Y
Yancey1989 已提交
119 120 121
  /* For index(i, j) >= 0:
      weight.row(index(i, j)) += tmat(i, j) * input.row(i)
  */
W
weixing02 已提交
122
  void MulGradWeight(const framework::Tensor& tmat, framework::Tensor* weight,
Y
Yancey1989 已提交
123 124 125 126
                     const framework::Tensor& input);
  /* For j < code_length
    input.row(i) += tmat(i, j) * weight.row(index(i, j))
  */
Y
Yancey1989 已提交
127
  void MulGradError(const framework::Tensor& tmat,
W
weixing02 已提交
128 129
                    const framework::Tensor& weight, framework::Tensor* input);

Y
Yancey1989 已提交
130 131
  size_t num_classes_;
  const int64_t* ids_;
Y
Yancey1989 已提交
132
};
Y
Yancey1989 已提交
133 134 135
}  // namespace math
}  // namespace operators
}  // namespace paddle