jit_kernel.h 3.2 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 36

#define AVX_FLOAT_BLOCK 8
#define AVX2_FLOAT_BLOCK 8
#define AVX512_FLOAT_BLOCK 16

T
tensor-tang 已提交
37
typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
T
tensor-tang 已提交
38

T
tensor-tang 已提交
39
class Kernel {
T
tensor-tang 已提交
40
 public:
T
tensor-tang 已提交
41
  Kernel() = default;
T
tensor-tang 已提交
42 43 44
  virtual ~Kernel() = default;

 private:
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 int n, 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 int n, 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 int n, const T a, const T *x, T *y) const = 0;
  virtual void Compute(const int n, const T a, T *x) const = 0;
T
tensor-tang 已提交
81 82
};

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

T
tensor-tang 已提交
89 90 91
template <typename T>
class VExpKernel : public Kernel {
 public:
T
tensor-tang 已提交
92 93 94 95 96 97 98 99 100 101 102 103 104
  virtual void Compute(const int n, const T *x, T *y) const = 0;
};

template <typename T>
class VSigmoidKernel : public Kernel {
 public:
  virtual void Compute(const int n, const T *x, T *y) const = 0;
};

template <typename T>
class VTanhKernel : public Kernel {
 public:
  virtual void Compute(const int n, const T *x, T *y) const = 0;
T
tensor-tang 已提交
105 106
};

T
tensor-tang 已提交
107 108 109
template <typename T>
class LSTMKernel : public Kernel {
 public:
T
tensor-tang 已提交
110 111
  explicit LSTMKernel(int d, const std::string &act_gate,
                      const std::string &act_cand, const std::string &act_cell);
T
tensor-tang 已提交
112

T
tensor-tang 已提交
113 114
  void (*jit_ker)(T *, const T *, T *, T *);
  std::function<void(T *, const T *, T *, T *)> ComputeCtHt, ComputeCtHt_NoC0H0;
T
tensor-tang 已提交
115 116

 private:
T
tensor-tang 已提交
117
  int d_, d2_, d3_;
T
tensor-tang 已提交
118 119 120 121
  std::function<void(const int, const T *, T *)> act_gate_, act_cell_,
      act_cand_;
};

T
tensor-tang 已提交
122 123 124 125
}  // namespace jitkernel
}  // namespace math
}  // namespace operators
}  // namespace paddle