op_registry.h 7.0 KB
Newer Older
S
superjomn 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
// Copyright (c) 2019 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.

#include <memory>
#include <string>
#include <unordered_map>
S
superjomn 已提交
18 19 20
#include "paddle/fluid/lite/core/kernel.h"
#include "paddle/fluid/lite/core/op_lite.h"
#include "paddle/fluid/lite/core/target_wrapper.h"
S
superjomn 已提交
21 22 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 49 50 51 52
#include "paddle/fluid/lite/utils/all.h"

namespace paddle {
namespace lite {

using KernelFunc = std::function<void()>;
using KernelFuncCreator = std::function<std::unique_ptr<KernelFunc>()>;

class LiteOpRegistry final : public Factory<OpLite> {
 public:
  static LiteOpRegistry &Global() {
    static auto *x = new LiteOpRegistry;
    return *x;
  }

 private:
  LiteOpRegistry() = default;
};

template <typename OpClass>
class OpLiteRegistor : public Registor<OpClass> {
 public:
  OpLiteRegistor(const std::string &op_type)
      : Registor<OpClass>([&] {
          LiteOpRegistry::Global().Register(
              op_type, []() -> std::unique_ptr<OpLite> {
                return std::unique_ptr<OpLite>(new OpClass);
              });
        }) {}
};

template <TargetType Target, PrecisionType Precision>
S
superjomn 已提交
53
using KernelRegistryForTarget = Factory<OpKernel<Target, Precision>>;
S
superjomn 已提交
54 55 56

class KernelRegistry final {
 public:
S
superjomn 已提交
57 58 59 60 61 62 63
  using any_kernel_registor_t =
      variant<KernelRegistryForTarget<TARGET(kCUDA), PRECISION(kFloat)> *,  //
              KernelRegistryForTarget<TARGET(kCUDA), PRECISION(kInt8)> *,   //
              KernelRegistryForTarget<TARGET(kX86), PRECISION(kFloat)> *,   //
              KernelRegistryForTarget<TARGET(kX86), PRECISION(kInt8)> *,    //
              KernelRegistryForTarget<TARGET(kHost), PRECISION(kFloat)> *   //
              >;
S
superjomn 已提交
64

S
superjomn 已提交
65
  KernelRegistry() {
S
superjomn 已提交
66 67 68 69 70 71 72 73 74 75 76 77
/*
using kernel_target_t =
    KernelRegistryForTarget<TARGET(kCUDA), PRECISION(kFloat)>;
registries_[0].set<kernel_target_t *>(
    &KernelRegistryForTarget<TARGET(kCUDA), PRECISION(kFloat)>::Global());
    */
#define INIT_FOR(target__, precision__)                                      \
  registries_[KernelRegistry::GetKernelOffset<TARGET(target__),              \
                                              PRECISION(precision__)>()]     \
      .set<KernelRegistryForTarget<TARGET(target__), PRECISION(precision__)> \
               *>(&KernelRegistryForTarget<TARGET(target__),                 \
                                           PRECISION(precision__)>::Global());
S
superjomn 已提交
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92
    // Currently, just register 2 kernel targets.
    INIT_FOR(kHost, kFloat);
#undef INIT_FOR
  }

  static KernelRegistry &Global() {
    static auto *x = new KernelRegistry;
    return *x;
  }

  template <TargetType Target, PrecisionType Precision>
  void Register(const std::string &name,
                typename KernelRegistryForTarget<Target, Precision>::creator_t
                    &&creator) {
    using kernel_registor_t = KernelRegistryForTarget<Target, Precision>;
S
superjomn 已提交
93 94
    registries_[GetKernelOffset<Target, Precision>()]
        .template get<kernel_registor_t *>()
S
superjomn 已提交
95 96 97
        ->Register(name, std::move(creator));
  }

S
update  
superjomn 已提交
98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
  template <TargetType Target, PrecisionType Precision>
  std::unique_ptr<KernelBase> Create(const std::string &op_type) {
    using kernel_registor_t = KernelRegistryForTarget<Target, Precision>;
    return registries_[GetKernelOffset<Target, Precision>()]
        .template get<kernel_registor_t *>()
        ->Create(op_type);
  }

  std::unique_ptr<KernelBase> Create(const std::string &op_type,
                                     TargetType target,
                                     PrecisionType precision) {
#define CREATE_KERNEL(target__)                                    \
  switch (precision) {                                             \
    case PRECISION(kFloat):                                        \
      return Create<TARGET(target__), PRECISION(kFloat)>(op_type); \
    default:                                                       \
      CHECK(false) << "not supported kernel place yet";            \
  }

    switch (target) {
      case TARGET(kHost): {
        CREATE_KERNEL(kHost);
      } break;
      case TARGET(kX86): {
        CREATE_KERNEL(kX86);
      } break;
      case TARGET(kCUDA): {
        CREATE_KERNEL(kCUDA);
      } break;
      default:
        CHECK(false) << "not supported kernel place";
    }

#undef CREATE_KERNEL
  }

S
superjomn 已提交
134 135 136 137 138 139 140
  // Get a kernel registry offset in all the registries.
  template <TargetType Target, PrecisionType Precision>
  static constexpr int GetKernelOffset() {
    return kNumTargets * static_cast<int>(Target) + static_cast<int>(Precision);
  }

 private:
S
superjomn 已提交
141
  std::array<any_kernel_registor_t, kNumTargets * kNumPrecisions> registries_;
S
superjomn 已提交
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186
};

template <TargetType target, PrecisionType precision, typename KernelType>
class KernelRegistor : public lite::Registor<KernelType> {
 public:
  KernelRegistor(const std::string op_type)
      : Registor<KernelType>([&] {
          KernelRegistry::Global().Register<target, precision>(
              op_type, [&]() -> std::unique_ptr<KernelType> {
                return std::unique_ptr<KernelType>(new KernelType);
              });
        }) {}
};

}  // namespace lite
}  // namespace paddle

// Operator registry
#define LITE_OP_REGISTER_INSTANCE(op_type__) op_type__##__registry__instance__
#define LITE_OP_REGISTER_FAKE(op_type__) op_type__##__registry__
#define REGISTER_LITE_OP(op_type__, OpClass)                              \
  static paddle::lite::OpLiteRegistor<OpClass> LITE_OP_REGISTER_INSTANCE( \
      op_type__)(#op_type__);

#define USE_LITE_OP(op_type__)                     \
  int LITE_OP_REGISTER_FAKE(op_type__)((unused)) = \
      LITE_OP_REGISTER_INSTANCE(op_type__).Touch();

// Kernel registry
#define LITE_KERNEL_REGISTER(op_type__, target__, precision__) \
  op_type__##target__##precision__##__registor__
#define LITE_KERNEL_REGISTER_INSTANCE(op_type__, target__, precision__) \
  op_type__##target__##precision__##__registor__instance__
#define LITE_KERNEL_REGISTER_FAKE(op_type__, target__, precision__) \
  LITE_KERNEL_REGISTER_INSTANCE(op_type__, target__, precision__)##__fake__

#define REGISTER_LITE_KERNEL(op_type__, target__, precision__, KernelClass) \
  static paddle::lite::KernelRegistor<TARGET(target__),                     \
                                      PRECISION(precision__), KernelClass>  \
      LITE_KERNEL_REGISTER_INSTANCE(op_type__, target__,                    \
                                    precision__)(#op_type__);

#define USE_LITE_KERNEL(op_type__, target__, precision__)                     \
  int LITE_KERNEL_REGISTER_FAKE(op_type__, target__, precision__)((unused)) = \
      LITE_KERNEL_REGISTER(op_type__, target__, precision__).Touch();