executor.cc 12.7 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Q
qijun 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

Y
Yi Wang 已提交
15
#include "paddle/fluid/framework/executor.h"
Y
Yang Yang 已提交
16

Y
Yang Yang 已提交
17
#include <set>
Y
Yang Yang 已提交
18

Y
Yang Yu 已提交
19
#include "gflags/gflags.h"
20
#include "paddle/fluid/framework/channel.h"
Y
Yi Wang 已提交
21 22 23 24 25 26 27
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/platform/place.h"
Y
Yang Yu 已提交
28

D
dzhwinter 已提交
29
DECLARE_bool(benchmark);
Y
Yang Yu 已提交
30 31 32
DEFINE_bool(check_nan_inf, false,
            "Checking whether operator produce NAN/INF or not. It will be "
            "extremely slow so please use this flag wisely.");
Q
qijun 已提交
33 34 35 36

namespace paddle {
namespace framework {

Y
Yu Yang 已提交
37 38 39 40 41 42 43 44 45
struct ExecutorPrepareContext {
  ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id)
      : prog_(prog), block_id_(block_id) {}

  framework::ProgramDesc prog_;
  size_t block_id_;
  std::vector<std::unique_ptr<OperatorBase>> ops_;
};

D
dzhwinter 已提交
46
Executor::Executor(const platform::Place& place) : place_(place) {}
Q
qijun 已提交
47

48 49
static void CreateTensor(Variable* var, proto::VarType::Type var_type) {
  if (var_type == proto::VarType::LOD_TENSOR) {
Q
QI JUN 已提交
50
    var->GetMutable<LoDTensor>();
51
  } else if (var_type == proto::VarType::SELECTED_ROWS) {
Q
QI JUN 已提交
52
    var->GetMutable<SelectedRows>();
53
  } else if (var_type == proto::VarType::FEED_MINIBATCH) {
Q
QI JUN 已提交
54
    var->GetMutable<FeedFetchList>();
55
  } else if (var_type == proto::VarType::FETCH_LIST) {
Q
QI JUN 已提交
56
    var->GetMutable<FeedFetchList>();
57
  } else if (var_type == proto::VarType::STEP_SCOPES) {
Y
Yu Yang 已提交
58
    var->GetMutable<std::vector<framework::Scope>>();
59
  } else if (var_type == proto::VarType::LOD_RANK_TABLE) {
Y
Yu Yang 已提交
60
    var->GetMutable<LoDRankTable>();
61
  } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
Y
Yu Yang 已提交
62
    var->GetMutable<LoDTensorArray>();
63
  } else if (var_type == proto::VarType::PLACE_LIST) {
Y
Yang Yu 已提交
64
    var->GetMutable<platform::PlaceList>();
65
  } else if (var_type == proto::VarType::READER) {
F
fengjiayi 已提交
66
    var->GetMutable<ReaderHolder>();
67 68
  } else if (var_type == proto::VarType::CHANNEL) {
    var->GetMutable<ChannelHolder>();
T
typhoonzero 已提交
69 70
  } else if (var_type == proto::VarType::RAW) {
    // GetMutable will be called in operator
Q
QI JUN 已提交
71 72
  } else {
    PADDLE_THROW(
Y
Yu Yang 已提交
73
        "Variable type %d is not in "
F
fengjiayi 已提交
74
        "[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
T
typhoonzero 已提交
75
        "LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, RAW]",
Y
Yu Yang 已提交
76
        var_type);
Q
QI JUN 已提交
77 78 79
  }
}

Y
Yang Yu 已提交
80 81
static void CheckTensorNANOrInf(const std::string& name,
                                const framework::Tensor& tensor) {
Y
Yang Yu 已提交
82
  if (tensor.memory_size() == 0) {
Y
Yang Yu 已提交
83 84
    return;
  }
Y
Yang Yu 已提交
85 86
  if (tensor.type().hash_code() != typeid(float).hash_code() &&
      tensor.type().hash_code() != typeid(double).hash_code()) {
Y
Yang Yu 已提交
87 88
    return;
  }
Y
Yi Wang 已提交
89 90 91 92
  PADDLE_ENFORCE(!framework::TensorContainsInf(tensor),
                 "Tensor %s contains Inf", name);
  PADDLE_ENFORCE(!framework::TensorContainsNAN(tensor),
                 "Tensor %s contains NAN", name);
Y
Yang Yu 已提交
93 94
}

Y
Yu Yang 已提交
95
void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
T
typhoonzero 已提交
96
                   bool create_local_scope, bool create_vars) {
Y
Yu Yang 已提交
97 98 99
  auto* ctx = Prepare(pdesc, block_id);
  RunPreparedContext(ctx, scope, create_local_scope, create_vars);
  delete ctx;
Q
qijun 已提交
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
// Check whether the block already has feed operators and feed_holder.
// Return false if the block does not have any feed operators.
// If some feed operators have been prepended to the block, check that
// the info contained in these feed operators matches the feed_targets
// and feed_holder_name. Raise exception when any mismatch is found.
// Return true if the block has feed operators and holder of matching info.
static bool has_feed_operators(
    BlockDesc* block, std::map<std::string, const LoDTensor*>& feed_targets,
    const std::string& feed_holder_name) {
  size_t feed_count = 0;
  for (auto* op : block->AllOps()) {
    if (op->Type() == kFeedOpType) {
      feed_count++;
      PADDLE_ENFORCE_EQ(op->Input("X")[0], feed_holder_name,
                        "Input to feed op should be '%s'", feed_holder_name);
      std::string feed_target_name = op->Output("Out")[0];
      PADDLE_ENFORCE(
          feed_targets.find(feed_target_name) != feed_targets.end(),
          "Feed operator output name '%s' cannot be found in 'feed_targets'",
          feed_target_name);
    }
  }

  if (feed_count > 0) {
    PADDLE_ENFORCE_EQ(
        feed_count, feed_targets.size(),
        "The number of feed operators should match 'feed_targets'");

    // When feed operator are present, so should be feed_holder
    auto var = block->FindVar(feed_holder_name);
    PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable",
                            feed_holder_name);
134
    PADDLE_ENFORCE_EQ(var->GetType(), proto::VarType::FEED_MINIBATCH,
135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173
                      "'%s' variable should be 'FEED_MINIBATCH' type",
                      feed_holder_name);
  }

  return feed_count > 0;
}

// Check whether the block already has fetch operators and fetch_holder.
// Return false if the block does not have any fetch operators.
// If some fetch operators have been appended to the block, check that
// the info contained in these fetch operators matches the fetch_targets
// and fetch_holder_name. Raise exception when any mismatch is found.
// Return true if the block has fetch operators and holder of matching info.
static bool has_fetch_operators(
    BlockDesc* block, std::map<std::string, LoDTensor*>& fetch_targets,
    const std::string& fetch_holder_name) {
  size_t fetch_count = 0;
  for (auto* op : block->AllOps()) {
    if (op->Type() == kFetchOpType) {
      fetch_count++;
      PADDLE_ENFORCE_EQ(op->Output("Out")[0], fetch_holder_name,
                        "Output of fetch op should be '%s'", fetch_holder_name);
      std::string fetch_target_name = op->Input("X")[0];
      PADDLE_ENFORCE(
          fetch_targets.find(fetch_target_name) != fetch_targets.end(),
          "Fetch operator input name '%s' cannot be found in 'fetch_targets'",
          fetch_target_name);
    }
  }

  if (fetch_count > 0) {
    PADDLE_ENFORCE_EQ(
        fetch_count, fetch_targets.size(),
        "The number of fetch operators should match 'fetch_targets'");

    // When fetch operator are present, so should be fetch_holder
    auto var = block->FindVar(fetch_holder_name);
    PADDLE_ENFORCE_NOT_NULL(var, "Block should already have a '%s' variable",
                            fetch_holder_name);
174
    PADDLE_ENFORCE_EQ(var->GetType(), proto::VarType::FETCH_LIST,
175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192
                      "'%s' variable should be 'FETCH_LIST' type",
                      fetch_holder_name);
  }

  return fetch_count > 0;
}

void Executor::Run(const ProgramDesc& program, Scope* scope,
                   std::map<std::string, const LoDTensor*>& feed_targets,
                   std::map<std::string, LoDTensor*>& fetch_targets,
                   const std::string& feed_holder_name,
                   const std::string& fetch_holder_name) {
  auto* copy_program = new ProgramDesc(program);
  auto* global_block = copy_program->MutableBlock(0);

  if (!has_feed_operators(global_block, feed_targets, feed_holder_name)) {
    // create feed_holder variable
    auto* feed_holder = global_block->Var(feed_holder_name);
193
    feed_holder->SetType(proto::VarType::FEED_MINIBATCH);
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
    feed_holder->SetPersistable(true);

    int i = 0;
    for (auto& feed_target : feed_targets) {
      std::string var_name = feed_target.first;
      VLOG(3) << "feed target's name: " << var_name;

      // prepend feed op
      auto* op = global_block->PrependOp();
      op->SetType(kFeedOpType);
      op->SetInput("X", {feed_holder_name});
      op->SetOutput("Out", {var_name});
      op->SetAttr("col", {static_cast<int>(i)});
      op->CheckAttrs();

      i++;
    }
  }

  // map the data of feed_targets to feed_holder
  for (auto* op : global_block->AllOps()) {
    if (op->Type() == kFeedOpType) {
      std::string feed_target_name = op->Output("Out")[0];
      int idx = boost::get<int>(op->GetAttr("col"));
      SetFeedVariable(scope, *feed_targets[feed_target_name], feed_holder_name,
                      idx);
    }
  }

  if (!has_fetch_operators(global_block, fetch_targets, fetch_holder_name)) {
    // create fetch_holder variable
    auto* fetch_holder = global_block->Var(fetch_holder_name);
226
    fetch_holder->SetType(proto::VarType::FETCH_LIST);
227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
    fetch_holder->SetPersistable(true);

    int i = 0;
    for (auto& fetch_target : fetch_targets) {
      std::string var_name = fetch_target.first;
      VLOG(3) << "fetch target's name: " << var_name;

      // append fetch op
      auto* op = global_block->AppendOp();
      op->SetType(kFetchOpType);
      op->SetInput("X", {var_name});
      op->SetOutput("Out", {fetch_holder_name});
      op->SetAttr("col", {static_cast<int>(i)});
      op->CheckAttrs();

      i++;
    }
  }

  Run(*copy_program, scope, 0, true, true);

  // obtain the data of fetch_targets from fetch_holder
  for (auto* op : global_block->AllOps()) {
    if (op->Type() == kFetchOpType) {
      std::string fetch_target_name = op->Input("X")[0];
      int idx = boost::get<int>(op->GetAttr("col"));
      *fetch_targets[fetch_target_name] =
          GetFetchVariable(*scope, fetch_holder_name, idx);
    }
  }

  delete copy_program;
}

Y
Yu Yang 已提交
261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307
ExecutorPrepareContext* Executor::Prepare(const ProgramDesc& program,
                                          int block_id) {
  auto* ctx = new ExecutorPrepareContext(program, block_id);
  PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
  auto& block = program.Block(block_id);
  for (auto& op_desc : block.AllOps()) {
    ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
  }
  return ctx;
}

void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                                  bool create_local_scope, bool create_vars) {
  auto& block = ctx->prog_.Block(ctx->block_id_);

  Scope* local_scope = scope;
  if (create_vars) {
    if (create_local_scope) {
      local_scope = &scope->NewScope();
      for (auto& var : block.AllVars()) {
        if (var->Name() == framework::kEmptyVarName) {
          continue;
        }

        if (var->Persistable()) {
          auto* ptr = scope->Var(var->Name());
          CreateTensor(ptr, var->GetType());
          VLOG(3) << "Create Variable " << var->Name()
                  << " global, which pointer is " << ptr;
        } else {
          auto* ptr = local_scope->Var(var->Name());
          CreateTensor(ptr, var->GetType());
          VLOG(3) << "Create Variable " << var->Name()
                  << " locally, which pointer is " << ptr;
        }
      }
    } else {
      for (auto& var : block.AllVars()) {
        auto* ptr = local_scope->Var(var->Name());
        CreateTensor(ptr, var->GetType());
        VLOG(3) << "Create variable " << var->Name() << ", which pointer is "
                << ptr;
      }
    }  // if (create_local_scope)
  }    // if (create_vars)

  for (auto& op : ctx->ops_) {
Y
Yang Yang 已提交
308 309 310 311 312 313
    // TODO(ty):
    // e.g. sgd should wait for allreduce to be finished
    // if op's input is params' grad:
    //     sync with allreduce stream
    // SyncMultipleStreams(op);

Y
Yu Yang 已提交
314 315 316 317
    VLOG(4) << place_ << " " << op->DebugStringEx(local_scope);
    op->Run(*local_scope, place_);
    VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);

Y
Yang Yang 已提交
318 319 320 321 322 323 324
    // TODO(ty):
    // e.g. allreduce shoudl wait for fc_grad to be finished.
    // if op's output is params' grad:
    //     sync with computation stream
    //     apply allreduce on allreduce stream
    // SyncMultipleStreams(op);

Y
Yu Yang 已提交
325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349
    if (FLAGS_benchmark) {
      VLOG(2) << "Memory used after operator " + op->Type() + " running: "
              << memory::memory_usage(place_);
    }
    if (FLAGS_check_nan_inf) {
      for (auto& vname : op->OutputVars(true)) {
        auto* var = local_scope->FindVar(vname);
        if (var == nullptr) continue;
        if (var->IsType<framework::LoDTensor>()) {
          CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
        }
      }
    }
  }
  if (create_vars && create_local_scope) {
    scope->DeleteScope(local_scope);
  }
  if (FLAGS_benchmark) {
    VLOG(2) << "-------------------------------------------------------";
    VLOG(2) << "Memory used after deleting local scope: "
            << memory::memory_usage(place_);
    VLOG(2) << "-------------------------------------------------------";
  }
}

Q
qijun 已提交
350 351
}  // namespace framework
}  // namespace paddle