kernels.h 4.8 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 19 20 21 22 23
/* 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 <memory>  // for shared_ptr
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/operators/jitkernels/jitcode_base.h"
#include "paddle/fluid/operators/jitkernels/kernel_base.h"
#include "paddle/fluid/operators/jitkernels/kernel_key.h"
T
tensor-tang 已提交
24
#include "paddle/fluid/platform/place.h"
T
tensor-tang 已提交
25 26 27 28 29

namespace paddle {
namespace operators {
namespace jitkernels {

T
tensor-tang 已提交
30 31
// TODO(TJ): rename file to kernel_pool

T
tensor-tang 已提交
32 33 34
template <KernelType KT>
class JitCodePool {
 public:
T
tensor-tang 已提交
35
  JitCodePool() = default;
T
tensor-tang 已提交
36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
  static JitCodePool& Instance() {
    static thread_local JitCodePool<KT> g_jit_codes;
    return g_jit_codes;
  }

  std::shared_ptr<const JitBase> Get(size_t key) const {
    if (codes_.find(key) == codes_.end()) {
      return nullptr;
    }
    return codes_.at(key);
  }

  void Insert(size_t key, const std::shared_ptr<const JitBase>& value) {
    codes_.insert({key, value});
  }

 private:
  std::unordered_map<size_t, std::shared_ptr<const JitBase>> codes_;
  DISABLE_COPY_AND_ASSIGN(JitCodePool);
};

T
tensor-tang 已提交
57
// TODO(TJ): std::tuple<T, Func, Attr>
T
tensor-tang 已提交
58 59 60 61 62 63 64
template <typename T, typename Func, typename Attr>
struct KernelAttr {
  typedef T data_type;
  typedef Func return_type;
  typedef Attr attr_type;
};

T
tensor-tang 已提交
65 66 67 68
typedef std::unique_ptr<const Kernel> KernelPtr;
typedef std::unordered_map<KernelKey, std::vector<KernelPtr>, KernelKey::Hash>
    KernelMap;

T
tensor-tang 已提交
69 70 71
class KernelPool {
 public:
  static KernelPool& Instance();
T
tensor-tang 已提交
72
  KernelPool() = default;
T
tensor-tang 已提交
73 74 75 76 77 78 79 80 81 82 83 84 85
  KernelMap& AllKernels() { return pool_; }
  void Insert(const KernelKey& key, KernelPtr value) {
    if (pool_.find(key) == pool_.end()) {
      pool_.emplace(key, std::vector<KernelPtr>());
    }
    pool_.at(key).emplace_back(std::move(value));
  }

 private:
  KernelMap pool_;
  DISABLE_COPY_AND_ASSIGN(KernelPool);
};

T
tensor-tang 已提交
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121
// Every kernel should have refer code and it should be used in unit tests,
// so refer kernels should have it's independent kernel pool
class ReferKernelPool {
 public:
  static ReferKernelPool& Instance();
  ReferKernelPool() = default;
  KernelMap& AllKernels() { return pool_; }
  void Insert(const KernelKey& key, KernelPtr value) {
    if (pool_.find(key) == pool_.end()) {
      pool_.emplace(key, std::vector<KernelPtr>());
    }
    pool_.at(key).emplace_back(std::move(value));
  }

 private:
  KernelMap pool_;
  DISABLE_COPY_AND_ASSIGN(ReferKernelPool);
};

// Refer code do not related with attr, and always on CPUPlace
template <KernelType KT, typename T, typename Func, typename Attr>
inline Func GetRefer() {
  auto& ref_pool = ReferKernelPool().Instance().AllKernels();
  KernelKey kkey(KT, platform::CPUPlace());
  auto ref_iter = ref_pool.find(kkey);
  PADDLE_ENFORCE(ref_iter != ref_pool.end(),
                 "Every Kernel should have reference function.");
  auto& ref_impls = ref_iter->second;
  for (auto& impl : ref_impls) {
    auto i = dynamic_cast<const ReferKernel<T, Func, Attr>*>(impl.get());
    if (i) {
      return i->GetFunc();
    }
  }
  return nullptr;
}
T
tensor-tang 已提交
122 123 124 125 126

// TODO(TJ): make tuple? named KernelAttr
template <KernelType KT, typename T, typename Func, typename Attr,
          typename PlaceType = platform::CPUPlace>
Func Get(Attr attr) {
T
tensor-tang 已提交
127 128 129 130 131 132 133 134 135 136 137 138 139
  // size_t key = GetKey<Attr>(attr);
  // auto jitcode = JitCodePool<KT>().Instance().Get(key);
  // if (jitcode) {
  //   return jitcode->template getCode<Func>();
  // }

  if (std::is_same<PlaceType, platform::CPUPlace>::value &&
      std::is_same<T, float>::value) {  // TODO(TJ): float move to create
    // auto p = CreateJitCode<KT, Attr>(attr);
    // if (p) {
    //   JitCodePool<KT>().Instance().Insert(key, p);
    //   return p->template getCode<Func>();
    // }
T
tensor-tang 已提交
140 141
  }

T
tensor-tang 已提交
142
  // pool: (KernelKey(type, place), vector<Kernel>)
T
tensor-tang 已提交
143 144 145 146
  auto& pool = KernelPool().Instance().AllKernels();
  KernelKey kkey(KT, PlaceType());
  auto iter = pool.find(kkey);
  if (iter != pool.end()) {
T
tensor-tang 已提交
147 148 149
    auto& impls = iter->second;
    for (auto& impl : impls) {
      auto i = dynamic_cast<const KernelImpl<T, Func, Attr>*>(impl.get());
T
tensor-tang 已提交
150 151 152 153 154 155
      if (i && i->UseMe(attr)) {
        return i->GetFunc();
      }
    }
  }

T
tensor-tang 已提交
156 157
  // The last implementation should be reference function on CPUPlace.
  return GetRefer<KT, T, Func, Attr>();
T
tensor-tang 已提交
158 159 160 161 162
}

}  // namespace jitkernels
}  // namespace operators
}  // namespace paddle