jit_kernel.h 2.5 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 31 32 33 34 35 36 37 38
#define SIGMOID_THRESHOLD_MIN -40.0
#define SIGMOID_THRESHOLD_MAX 13.0

#define AVX_FLOAT_BLOCK 8
#define AVX_DOUBLE_BLOCK 4
#define AVX2_FLOAT_BLOCK 8
#define AVX2_DOUBLE_BLOCK 4
#define AVX512_FLOAT_BLOCK 16
#define AVX512_DOUBLE_BLOCK 8

T
tensor-tang 已提交
39
typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block;
T
tensor-tang 已提交
40

T
tensor-tang 已提交
41
class Kernel {
T
tensor-tang 已提交
42 43 44 45 46
 public:
  Kernel() {}
  virtual ~Kernel() = default;

 private:
T
tensor-tang 已提交
47 48 49 50 51
  DISABLE_COPY_AND_ASSIGN(Kernel);
};

class KernelPool {
 public:
T
tensor-tang 已提交
52
  static KernelPool &Instance();
T
tensor-tang 已提交
53 54 55 56

  template <typename Ker, typename... ARGS>
  const std::shared_ptr<Ker> Get(ARGS... args);

T
tensor-tang 已提交
57 58
  const std::shared_ptr<Kernel> Get(const std::string &key) const;

T
tensor-tang 已提交
59 60
 private:
  KernelPool() = default;
T
tensor-tang 已提交
61
  std::unordered_map<std::string, std::shared_ptr<Kernel>> kers_;
T
tensor-tang 已提交
62 63 64 65

  DISABLE_COPY_AND_ASSIGN(KernelPool);
};

T
tensor-tang 已提交
66 67 68 69 70 71 72
template <typename T>
class VMulKernel : public Kernel {
 public:
  explicit VMulKernel(int n);
  void (*Compute)(const int n, const T *, const T *, T *);
};

T
tensor-tang 已提交
73 74 75 76 77 78 79
template <typename T>
class VAddKernel : public Kernel {
 public:
  explicit VAddKernel(int n);
  void (*Compute)(const int n, const T *, const T *, T *);
};

T
tensor-tang 已提交
80 81 82
template <typename T>
class LSTMKernel : public Kernel {
 public:
T
tensor-tang 已提交
83 84
  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
85

T
tensor-tang 已提交
86 87
  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
88 89

 private:
T
tensor-tang 已提交
90
  int d_, d2_, d3_;
T
tensor-tang 已提交
91 92 93 94
  std::function<void(const int, const T *, T *)> act_gate_, act_cell_,
      act_cand_;
};

T
tensor-tang 已提交
95 96 97 98
}  // namespace jitkernel
}  // namespace math
}  // namespace operators
}  // namespace paddle