program.h 5.9 KB
Newer Older
S
superjomn 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// 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>
S
superjomn 已提交
17
#include <memory>
S
superjomn 已提交
18
#include <string>
S
superjomn 已提交
19
#include <utility>
S
superjomn 已提交
20 21
#include <vector>
#include "paddle/fluid/lite/core/kernel.h"
S
Superjomn 已提交
22
#include "paddle/fluid/lite/core/mir/node.h"
S
superjomn 已提交
23 24
#include "paddle/fluid/lite/core/op_lite.h"
#include "paddle/fluid/lite/core/op_registry.h"
Y
Yan Chunwei 已提交
25 26 27
#ifdef LITE_WITH_PROFILE
#include "paddle/fluid/lite/core/profile/basic_profiler.h"
#endif  // LITE_WITH_PROFILE
S
superjomn 已提交
28 29 30 31

namespace paddle {
namespace lite {

S
superjomn 已提交
32
static const char kKernelTypeAttr[] = "__@kernel_type_attr@__";
S
Superjomn 已提交
33

S
superjomn 已提交
34 35 36 37 38 39 40 41
// 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 {
  std::list<std::string> tmp_vars;
  std::list<std::string> weights;
  std::list<std::shared_ptr<OpLite>> ops;
S
superjomn 已提交
42
  // the scope to run the kernels, NOTE this is the execution scope.
S
superjomn 已提交
43
  std::shared_ptr<lite::Scope> scope;
44
  std::vector<Place> valid_places;
S
superjomn 已提交
45 46
  // Runtime scope.
  lite::Scope* exec_scope{};
47
  const framework::proto::ProgramDesc desc;
S
superjomn 已提交
48 49

  explicit Program(const std::shared_ptr<Scope>& root) { scope = root; }
50
  Program(const framework::proto::ProgramDesc& desc,
S
superjomn 已提交
51
          const std::shared_ptr<Scope>& root,
52 53
          const std::vector<Place>& valid_places)
      : scope(root), valid_places(valid_places), desc(desc) {
S
Superjomn 已提交
54
    CHECK(scope) << "scope should be init first";
S
superjomn 已提交
55
    PrepareWorkspace(desc);
S
Superjomn 已提交
56
    Build(desc);
S
superjomn 已提交
57 58 59
  }

  std::unique_ptr<Program> Clone() const {
60
    std::unique_ptr<Program> res(new Program(desc, scope, valid_places));
S
superjomn 已提交
61 62 63 64 65
    return res;
  }

 private:
  // Build from a program and scope.
S
Superjomn 已提交
66
  void Build(const framework::proto::ProgramDesc& program) {
S
superjomn 已提交
67 68
    CHECK(ops.empty()) << "Executor duplicate Build found";
    // Create operators.
69 70 71
    for (const auto& proto_op_desc : program.blocks(0).ops()) {
      lite::OpDesc op_desc(proto_op_desc);
      auto op_type = op_desc.Type();
72
      // if (op_type == "feed" || op_type == "fetch") continue;
S
Superjomn 已提交
73
      VLOG(4) << "create Op [" << op_type << "]";
Y
Yan Chunwei 已提交
74
      LOG(INFO) << "create Op [" << op_type << "]";
S
Superjomn 已提交
75 76
      auto op = LiteOpRegistry::Global().Create(op_type);
      CHECK(op) << "no Op found for " << op_type;
S
superjomn 已提交
77
      ops.emplace_back(std::move(op));
78
      ops.back()->Attach(op_desc, exec_scope);
S
superjomn 已提交
79 80 81 82
    }
  }

  // Create temporary variables.
83
  void PrepareWorkspace(const framework::proto::ProgramDesc& program) {
S
superjomn 已提交
84 85
    CHECK(!exec_scope) << "Duplicate PrepareWorkspace found";
    exec_scope = &scope->NewScope();
86
    // Create Feed and Fetch var.
87 88
    scope->Var("feed")->GetMutable<std::vector<lite::Tensor>>();
    scope->Var("fetch")->GetMutable<std::vector<lite::Tensor>>();
S
superjomn 已提交
89

90 91
    tmp_vars.push_back("feed");
    tmp_vars.push_back("fetch");
Y
Yan Chunwei 已提交
92
    CHECK(!program.blocks().empty());
93 94 95 96 97
    for (auto proto_var_desc : program.blocks(0).vars()) {
      lite::VarDesc var_desc(proto_var_desc);
      if (!var_desc.Persistable()) {
        tmp_vars.push_back(var_desc.Name());
        exec_scope->Var(var_desc.Name());
98
      } else {
99 100
        if (var_desc.Name() == "feed" || var_desc.Name() == "fetch") continue;
        weights.push_back(var_desc.Name());
S
superjomn 已提交
101 102 103 104 105
      }
    }
  }
};

S
Superjomn 已提交
106 107 108
struct Instruct {
  Instruct(const std::shared_ptr<OpLite>& op,
           std::unique_ptr<KernelBase>&& kernel)
Y
Yan Chunwei 已提交
109 110 111 112 113 114 115
      : 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
  }
S
Superjomn 已提交
116 117

  void Run() {
Y
Yan Chunwei 已提交
118 119 120
#ifdef LITE_WITH_PROFILE
    profile::ProfileBlock x(profile_id_);
#endif  // LITE_WITH_PROFILE
S
Superjomn 已提交
121 122
    CHECK(op_);
    CHECK(kernel_);
S
Superjomn 已提交
123
    if (first_epoch_) {
124
      first_epoch_ = false;
S
superjomn 已提交
125
      CHECK(op_->CheckShape());
126
    }
S
Superjomn 已提交
127 128 129 130
    op_->InferShape();
    kernel_->Run();
  }

S
Superjomn 已提交
131
  friend std::ostream& operator<<(std::ostream& os, const Instruct& other) {
S
superjomn 已提交
132 133 134 135
    os << other.kernel_->summary() << "\t(" << other.kernel_->doc() << ")";
    return os;
  }

S
Superjomn 已提交
136 137 138
  const OpLite* op() const { return op_.get(); }
  const KernelBase* kernel() const { return kernel_.get(); }

S
Superjomn 已提交
139 140 141
 private:
  std::shared_ptr<OpLite> op_;
  std::unique_ptr<KernelBase> kernel_;
142
  bool first_epoch_{true};
Y
Yan Chunwei 已提交
143 144 145 146 147

#ifdef LITE_WITH_PROFILE
  // for profiler
  int profile_id_{-1};
#endif  // LITE_WITH_PROFILE
S
Superjomn 已提交
148 149 150 151 152 153 154
};

/*
 * A program contains kernels for runtime.
 */
class RuntimeProgram {
 public:
S
Superjomn 已提交
155
  explicit RuntimeProgram(std::vector<Instruct>&& insts)
S
superjomn 已提交
156
      : instructions_(std::move(insts)) {
S
Superjomn 已提交
157 158
    if (instructions_.empty()) {
      LOG(FATAL) << "no instructions";
S
superjomn 已提交
159 160
    }
  }
S
Superjomn 已提交
161 162 163

  void Run() {
    for (auto& inst : instructions_) {
S
superjomn 已提交
164
      LOG(INFO) << ">> Running kernel: " << inst;
S
Superjomn 已提交
165 166 167 168
      inst.Run();
    }
  }

S
Superjomn 已提交
169
  // Serialize the graph and save to the disk.
S
Superjomn 已提交
170
  void PersistModel(const std::string& dir,
S
Superjomn 已提交
171 172
                    const framework::proto::ProgramDesc& desc);

173 174 175
  void set_exec_scope(lite::Scope* x) { exec_scope_ = x; }
  lite::Scope* exec_scope() { return exec_scope_; }

S
Superjomn 已提交
176 177
  size_t num_instructions() const { return instructions_.size(); }

S
Superjomn 已提交
178
 protected:
S
Superjomn 已提交
179 180 181
  std::string SerializeProgram(const framework::proto::ProgramDesc& desc);
  void SaveParams(const std::string& dir,
                  const framework::proto::ProgramDesc& desc);
S
Superjomn 已提交
182

S
Superjomn 已提交
183 184
 private:
  RuntimeProgram(const RuntimeProgram&) = delete;
S
Superjomn 已提交
185
  std::vector<Instruct> instructions_;
186
  lite::Scope* exec_scope_{};
S
Superjomn 已提交
187 188
};

S
superjomn 已提交
189 190
}  // namespace lite
}  // namespace paddle