jit_kernel.h 3.6 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
/* 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
#include <functional>
#include <memory>  // for shared_ptr
#include <string>
T
tensor-tang 已提交
19
#include <unordered_map>
T
tensor-tang 已提交
20
#include "paddle/fluid/platform/cpu_info.h"
T
tensor-tang 已提交
21 22 23 24 25 26 27 28
#include "paddle/fluid/platform/macros.h"

// Note: Only support on CPU yet.
namespace paddle {
namespace operators {
namespace math {
namespace jitkernel {

T
tensor-tang 已提交
29 30
#define SIGMOID_THRESHOLD_MIN -40.0
#define SIGMOID_THRESHOLD_MAX 13.0
T
tensor-tang 已提交
31
#define EXP_MAX_INPUT 40.0
T
tensor-tang 已提交
32 33 34 35
#define AVX_FLOAT_BLOCK 8
#define AVX2_FLOAT_BLOCK 8
#define AVX512_FLOAT_BLOCK 16

T
tensor-tang 已提交
36
typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
T
tensor-tang 已提交
37

T
tensor-tang 已提交
38
class Kernel {
T
tensor-tang 已提交
39
 public:
T
tensor-tang 已提交
40
  Kernel() = default;
T
tensor-tang 已提交
41
  virtual ~Kernel() = default;
T
tensor-tang 已提交
42 43 44
  int num_{0};
  int end_{0};
  int rest_{0};
T
tensor-tang 已提交
45 46 47 48 49
  DISABLE_COPY_AND_ASSIGN(Kernel);
};

class KernelPool {
 public:
T
tensor-tang 已提交
50
  static KernelPool &Instance();
T
tensor-tang 已提交
51 52

  template <typename Ker, typename... ARGS>
T
tensor-tang 已提交
53
  std::shared_ptr<const Ker> Get(ARGS... args);
T
tensor-tang 已提交
54

T
tensor-tang 已提交
55
  std::shared_ptr<const Kernel> Get(const std::string &key) const;
T
tensor-tang 已提交
56

T
tensor-tang 已提交
57 58
 private:
  KernelPool() = default;
T
tensor-tang 已提交
59
  std::unordered_map<std::string, std::shared_ptr<const Kernel>> kers_;
T
tensor-tang 已提交
60 61 62 63

  DISABLE_COPY_AND_ASSIGN(KernelPool);
};

T
tensor-tang 已提交
64 65 66
template <typename T>
class VMulKernel : public Kernel {
 public:
T
tensor-tang 已提交
67
  virtual void Compute(const T *x, const T *y, T *z) const = 0;
T
tensor-tang 已提交
68 69
};

T
tensor-tang 已提交
70 71 72
template <typename T>
class VAddKernel : public Kernel {
 public:
T
tensor-tang 已提交
73
  virtual void Compute(const T *x, const T *y, T *z) const = 0;
T
tensor-tang 已提交
74 75
};

T
tensor-tang 已提交
76 77 78
template <typename T>
class VScalKernel : public Kernel {
 public:
T
tensor-tang 已提交
79 80
  virtual void Compute(const T a, const T *x, T *y) const = 0;
  virtual void Compute(const T a, T *x) const = 0;
T
tensor-tang 已提交
81 82
};

T
tensor-tang 已提交
83 84 85
template <typename T>
class VAddBiasKernel : public Kernel {
 public:
T
tensor-tang 已提交
86
  virtual void Compute(const T a, const T *x, T *y) const = 0;
T
tensor-tang 已提交
87 88
};

T
tensor-tang 已提交
89
template <typename T>
T
tensor-tang 已提交
90
class VActKernel : public Kernel {
T
tensor-tang 已提交
91
 public:
T
tensor-tang 已提交
92
  virtual void Compute(const T *x, T *y) const = 0;
T
tensor-tang 已提交
93 94 95
};

template <typename T>
T
tensor-tang 已提交
96
class VReluKernel : public VActKernel<T> {
T
tensor-tang 已提交
97
 public:
T
tensor-tang 已提交
98
  virtual void Compute(const T *x, T *y) const = 0;
T
tensor-tang 已提交
99 100 101
};

template <typename T>
T
tensor-tang 已提交
102
class VIdentityKernel : public VActKernel<T> {
T
tensor-tang 已提交
103
 public:
T
tensor-tang 已提交
104
  virtual void Compute(const T *x, T *y) const = 0;
T
tensor-tang 已提交
105 106
};

T
tensor-tang 已提交
107
template <typename T>
T
tensor-tang 已提交
108
class VExpKernel : public VActKernel<T> {
T
tensor-tang 已提交
109
 public:
T
tensor-tang 已提交
110 111
  virtual void Compute(const T *x, T *y) const = 0;
};
T
tensor-tang 已提交
112

T
tensor-tang 已提交
113 114 115 116 117
template <typename T>
class VSigmoidKernel : public VActKernel<T> {
 public:
  virtual void Compute(const T *x, T *y) const = 0;
};
T
tensor-tang 已提交
118

T
tensor-tang 已提交
119 120 121 122 123 124 125 126 127
template <typename T>
class VTanhKernel : public VActKernel<T> {
 public:
  virtual void Compute(const T *x, T *y) const = 0;
};

template <typename T>
class LSTMKernel : public Kernel {
 public:
T
tensor-tang 已提交
128
  virtual void ComputeCtHt(T *gates, const T *ct_1, T *ct, T *ht,
129 130
                           /* below only used in peephole*/
                           const T *wp_data = nullptr,
T
tensor-tang 已提交
131
                           T *checked = nullptr) const = 0;
132 133 134 135 136

  // compute c1 and h1 without c0 or h0
  virtual void ComputeC1H1(T *gates, T *ct, T *ht,
                           /* below only used in peephole*/
                           const T *wp_data = nullptr) const = 0;
T
tensor-tang 已提交
137 138
};

T
tensor-tang 已提交
139 140 141 142
}  // namespace jitkernel
}  // namespace math
}  // namespace operators
}  // namespace paddle