kernel_base.h 4.7 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 *
 * 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
16
#include "paddle/fluid/operators/jit/macro.h"
T
tensor-tang 已提交
17 18 19 20
#include "paddle/fluid/platform/macros.h"

namespace paddle {
namespace operators {
T
tensor-tang 已提交
21
namespace jit {
T
tensor-tang 已提交
22

23
typedef enum {
T
tensor-tang 已提交
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
  kNone = 0,
  kVMul = 1,
  kVAdd = 2,
  kVAddRelu,
  kVSub,
  kVScal,
  kVAddBias,
  kVRelu,
  kVIdentity,
  kVExp,
  kVSigmoid,
  kVTanh,
  kLSTMCtHt,
  kLSTMC1H1,
  kGRUH1,
  kGRUHtPart1,
  kGRUHtPart2,
  kCRFDecoding,
  kLayerNorm,
  kNCHW16CMulNC,
T
tensor-tang 已提交
44
  kSeqPool,
45
} KernelType;
T
tensor-tang 已提交
46

47 48 49 50 51 52 53
typedef enum {
  kNonePoolType = 0,
  kSum,
  kAvg,
  kSqrt,
} SeqPoolType;

T
tensor-tang 已提交
54
template <typename T>
55
struct XYZNTuples {
T
tensor-tang 已提交
56 57 58 59 60
  typedef T data_type;
  typedef int attr_type;
  typedef void (*func_type)(const T*, const T*, T*, int);
};

61 62 63
template <typename T>
struct AXYNTuples : public XYZNTuples<T> {};

64 65 66 67 68 69 70
template <typename T>
struct XYNTuples {
  typedef T data_type;
  typedef int attr_type;
  typedef void (*func_type)(const T*, T*, int);
};

T
tensor-tang 已提交
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
typedef struct {
  void* gates;  // gates: x_ch, x_ih, x_fh, x_oh
  const void* ct_1;
  void* ct;
  void* ht;
  /* weight_peephole and checked data are only used in peephole*/
  const void* wp{nullptr};  //  W_ic, W_fc, W_oc
  void* checked{nullptr};   // size: 2 * d
} lstm_t;

typedef struct {
  void* gates;  // gates: {x_update, x_reset; x_state}
  const void* ht_1;
  void* ht;
} gru_t;

struct rnn_attr_s {
  int d;
  KernelType act_gate, act_cand;
  rnn_attr_s() = default;
T
tensor-tang 已提交
91
  explicit rnn_attr_s(int _d, KernelType _act_gate, KernelType _act_cand)
T
tensor-tang 已提交
92 93 94 95 96 97 98
      : d(_d), act_gate(_act_gate), act_cand(_act_cand) {}
};

struct lstm_attr_s : public rnn_attr_s {
  bool use_peephole;
  KernelType act_cell;
  lstm_attr_s() = default;
T
tensor-tang 已提交
99 100
  explicit lstm_attr_s(int _d, KernelType _act_gate, KernelType _act_cand,
                       KernelType _act_cell, bool _use_peephole = false)
T
tensor-tang 已提交
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
      : rnn_attr_s(_d, _act_gate, _act_cand),
        use_peephole(_use_peephole),
        act_cell(_act_cell) {}
};

typedef struct rnn_attr_s gru_attr_t;
typedef struct lstm_attr_s lstm_attr_t;

template <typename T>
struct LSTMTuples {
  typedef T data_type;
  typedef lstm_attr_t attr_type;
  typedef void (*func_type)(lstm_t*, const lstm_attr_t*);
};

116 117 118 119 120 121 122
template <typename T>
struct GRUTuples {
  typedef T data_type;
  typedef gru_attr_t attr_type;
  typedef void (*func_type)(gru_t*, const gru_attr_t*);
};

123
typedef struct seq_pool_attr_s {
T
tensor-tang 已提交
124 125
  int h, w;
  SeqPoolType type;
126 127 128
  seq_pool_attr_s() = default;
  explicit seq_pool_attr_s(int height, int width, SeqPoolType pool_type)
      : h(height), w(width), type(pool_type) {}
T
tensor-tang 已提交
129 130 131 132 133 134 135 136 137
} seq_pool_attr_t;

template <typename T>
struct SeqPoolTuples {
  typedef T data_type;
  typedef seq_pool_attr_t attr_type;
  typedef void (*func_type)(const T*, T*, const seq_pool_attr_t*);
};

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
template <typename T>
struct CRFDecodingTuples {
  typedef T data_type;
  typedef int attr_type;
  typedef void (*func_type)(const int, const T*, const T*, T*, int*, int);
};

template <typename T>
struct LayerNormTuples {
  typedef T data_type;
  typedef int attr_type;
  typedef void (*func_type)(T*, T*, T*, T*, const T*, const T*, int,
                            const float, int);
};

T
tensor-tang 已提交
153 154 155 156 157 158 159 160
// nChw16c = nChw16c .* NC
template <typename T>
struct NCHW16CMulNCTuples {
  typedef T data_type;
  typedef int attr_type;
  typedef void (*func_type)(const T*, const T*, T*, int, int);
};

T
tensor-tang 已提交
161 162 163 164
// Just for adding to kernel pool without template
class Kernel {
 public:
  Kernel() = default;
T
tensor-tang 已提交
165
  virtual ~Kernel() = default;
T
tensor-tang 已提交
166 167 168
  DISABLE_COPY_AND_ASSIGN(Kernel);
};

T
tensor-tang 已提交
169
template <typename KernelTuples>
T
tensor-tang 已提交
170
class KernelMore : public Kernel {
171
 public:
T
tensor-tang 已提交
172 173 174
  using T = typename KernelTuples::data_type;
  using Func = typename KernelTuples::func_type;
  using Attr = typename KernelTuples::attr_type;
T
tensor-tang 已提交
175
  virtual Func GetFunc() const { return func; }
T
tensor-tang 已提交
176 177
  virtual bool UseMe(const Attr& attr) const = 0;
  virtual const char* ImplType() const = 0;
T
tensor-tang 已提交
178 179 180 181 182

 protected:
  Func func{nullptr};
};

T
tensor-tang 已提交
183
template <typename KernelTuples>
T
tensor-tang 已提交
184
class ReferKernel : public KernelMore<KernelTuples> {
T
tensor-tang 已提交
185 186
 public:
  // Refer code can always be used
T
tensor-tang 已提交
187
  bool UseMe(const typename KernelTuples::attr_type& attr) const override {
T
tensor-tang 已提交
188 189
    return true;
  }
T
tensor-tang 已提交
190
  const char* ImplType() const override { return "Refer"; }
T
tensor-tang 已提交
191 192
};

T
tensor-tang 已提交
193
}  // namespace jit
T
tensor-tang 已提交
194 195
}  // namespace operators
}  // namespace paddle