program.h 5.2 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 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
// 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.

#pragma once
#include <list>
#include <memory>
#include <string>
#include <utility>
#include <vector>
#include "paddle/fluid/lite/core/kernel.h"
#include "paddle/fluid/lite/core/mir/node.h"
#include "paddle/fluid/lite/core/op_lite.h"
#include "paddle/fluid/lite/core/op_registry.h"
#include "paddle/fluid/lite/model_parser/compatible_pb.h"
#ifdef LITE_WITH_PROFILE
#include "paddle/fluid/lite/core/profile/basic_profiler.h"
#endif  // LITE_WITH_PROFILE

namespace paddle {
namespace lite {

static const char kKernelTypeAttr[] = "__@kernel_type_attr@__";

// A program is used to represent a code program, in Paddle, a code program
// contains:
// - main block, which is a list of OpLite
// - scope: which contains all the weights
struct Program {
40 41
 public:
  explicit Program(const std::shared_ptr<Scope>& root) { scope_ = root; }
T
tensor-tang 已提交
42 43 44
  Program(const framework::proto::ProgramDesc& desc,
          const std::shared_ptr<Scope>& root,
          const std::vector<Place>& valid_places)
45 46
      : scope_(root), valid_places_(valid_places), desc_(desc) {
    CHECK(scope_) << "scope should be init first";
T
tensor-tang 已提交
47 48 49 50 51
    PrepareWorkspace(desc);
    Build(desc);
  }

  std::unique_ptr<Program> Clone() const {
52
    std::unique_ptr<Program> res(new Program(desc_, scope_, valid_places_));
T
tensor-tang 已提交
53 54 55
    return res;
  }

56 57 58 59 60 61 62 63 64 65 66
  const std::list<std::string>& weights() const { return weights_; }
  const std::list<std::string>& tmp_vars() const { return tmp_vars_; }
  std::list<std::string>* mutable_weights() { return &weights_; }
  std::list<std::string>* mutable_tmp_vars() { return &tmp_vars_; }

  const std::list<std::shared_ptr<OpLite>>& ops() const { return ops_; }
  std::list<std::shared_ptr<OpLite>>* mutable_ops() { return &ops_; }

  lite::Scope* exec_scope() { return exec_scope_; }
  lite::Scope* scope() { return scope_.get(); }

T
tensor-tang 已提交
67 68
 private:
  // Build from a program and scope.
69
  void Build(const framework::proto::ProgramDesc& program);
T
tensor-tang 已提交
70
  // Create temporary variables.
71 72 73 74 75 76 77 78 79 80 81 82
  void PrepareWorkspace(const framework::proto::ProgramDesc& program);

 private:
  std::list<std::string> tmp_vars_;
  std::list<std::string> weights_;
  std::list<std::shared_ptr<OpLite>> ops_;
  // the scope to run the kernels, NOTE this is the execution scope.
  std::shared_ptr<lite::Scope> scope_;
  std::vector<Place> valid_places_;
  // Runtime scope.
  lite::Scope* exec_scope_{};
  const framework::proto::ProgramDesc desc_;
T
tensor-tang 已提交
83 84
};

85 86 87
struct Instruction {
  Instruction(const std::shared_ptr<OpLite>& op,
              std::unique_ptr<KernelBase>&& kernel)
T
tensor-tang 已提交
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106
      : op_(op), kernel_(std::move(kernel)) {
#ifdef LITE_WITH_PROFILE
    profile_id_ = profile::BasicProfiler<profile::BasicTimer>::Global()
                      .NewRcd(kernel_->SerializedKernelType())
                      .id();
#endif  // LITE_WITH_PROFILE
  }

  void Run() {
#ifdef LITE_WITH_PROFILE
    profile::ProfileBlock x(profile_id_);
#endif  // LITE_WITH_PROFILE
    CHECK(op_);
    CHECK(kernel_);
    if (first_epoch_) {
      first_epoch_ = false;
      CHECK(op_->CheckShape());
    }
    op_->InferShape();
T
tensor-tang 已提交
107
    kernel_->Launch();
T
tensor-tang 已提交
108 109
  }

110
  friend std::ostream& operator<<(std::ostream& os, const Instruction& other) {
T
tensor-tang 已提交
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
    os << other.kernel_->summary() << "\t(" << other.kernel_->doc() << ")";
    return os;
  }

  const OpLite* op() const { return op_.get(); }
  const KernelBase* kernel() const { return kernel_.get(); }

 private:
  std::shared_ptr<OpLite> op_;
  std::unique_ptr<KernelBase> kernel_;
  bool first_epoch_{true};

#ifdef LITE_WITH_PROFILE
  // for profiler
  int profile_id_{-1};
#endif  // LITE_WITH_PROFILE
};

/*
 * A program contains kernels for runtime.
 */
class RuntimeProgram {
 public:
134
  explicit RuntimeProgram(std::vector<Instruction>&& insts)
T
tensor-tang 已提交
135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163
      : instructions_(std::move(insts)) {
    if (instructions_.empty()) {
      LOG(FATAL) << "no instructions";
    }
  }

  void Run() {
    for (auto& inst : instructions_) {
      LOG(INFO) << ">> Running kernel: " << inst;
      inst.Run();
    }
  }

  // Serialize the graph and save to the disk.
  void PersistModel(const std::string& dir,
                    const framework::proto::ProgramDesc& desc);

  void set_exec_scope(lite::Scope* x) { exec_scope_ = x; }
  lite::Scope* exec_scope() { return exec_scope_; }

  size_t num_instructions() const { return instructions_.size(); }

 protected:
  std::string SerializeProgram(const framework::proto::ProgramDesc& desc);
  void SaveParams(const std::string& dir,
                  const framework::proto::ProgramDesc& desc);

 private:
  RuntimeProgram(const RuntimeProgram&) = delete;
164
  std::vector<Instruction> instructions_;
T
tensor-tang 已提交
165 166 167 168 169
  lite::Scope* exec_scope_{};
};

}  // namespace lite
}  // namespace paddle