analysis_predictor.cc 19.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

Y
Yan Chunwei 已提交
15
#include "paddle/fluid/inference/api/analysis_predictor.h"
16 17
#include <glog/logging.h>
#include <algorithm>
18
#include <memory>
19 20
#include <string>
#include <vector>
21
#include "paddle/fluid/framework/feed_fetch_method.h"
22
#include "paddle/fluid/framework/feed_fetch_type.h"
Y
Yan Chunwei 已提交
23
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
24
#include "paddle/fluid/framework/ir/pass.h"
25
#include "paddle/fluid/framework/naive_executor.h"
26
#include "paddle/fluid/framework/scope.h"
27
#include "paddle/fluid/inference/api/helper.h"
28
#include "paddle/fluid/inference/api/paddle_inference_api.h"
L
luotao1 已提交
29
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
30 31 32
#if PADDLE_WITH_TENSORRT
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#endif
33
#include "paddle/fluid/inference/utils/singleton.h"
34
#include "paddle/fluid/memory/memcpy.h"
35
#include "paddle/fluid/platform/cpu_helper.h"
T
tensor-tang 已提交
36 37 38
#include "paddle/fluid/platform/profiler.h"

DECLARE_bool(profile);
39 40 41

namespace paddle {

42 43
using contrib::AnalysisConfig;

44 45 46 47 48 49 50 51 52 53 54
namespace {
bool IsPersistable(const framework::VarDesc *var) {
  if (var->Persistable() &&
      var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
      var->GetType() != framework::proto::VarType::FETCH_LIST) {
    return true;
  }
  return false;
}
}  // namespace

Y
Yan Chunwei 已提交
55
bool AnalysisPredictor::Init(
56 57
    const std::shared_ptr<framework::Scope> &parent_scope,
    const std::shared_ptr<framework::ProgramDesc> &program) {
M
minqiyang 已提交
58
  VLOG(3) << "Predictor::init()";
T
tensor-tang 已提交
59 60 61 62 63 64 65 66
  if (FLAGS_profile) {
    LOG(WARNING) << "Profiler is actived, might affect the performance";
    LOG(INFO) << "You can turn off by set gflags '-profile false'";
    auto tracking_device = config_.use_gpu ? platform::ProfilerState::kAll
                                           : platform::ProfilerState::kCPU;
    platform::EnableProfiler(tracking_device);
  }

67
  // no matter with or without MKLDNN
L
luotao1 已提交
68
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
69

70 71 72 73 74 75 76 77 78 79 80 81 82
  if (!PrepareScope(parent_scope)) {
    return false;
  }
  if (!CreateExecutor()) {
    return false;
  }
  if (!PrepareProgram(program)) {
    return false;
  }

  // Prepare executor, create local variables.
  if (!PrepareExecutor()) {
    return true;
Y
Yan Chunwei 已提交
83
  }
84 85 86 87 88 89 90 91 92

  // Get the feed_target_names and fetch_target_names
  PrepareFeedFetch();

  return true;
}

bool AnalysisPredictor::PrepareScope(
    const std::shared_ptr<framework::Scope> &parent_scope) {
Y
Yan Chunwei 已提交
93
  if (parent_scope) {
94 95 96
    PADDLE_ENFORCE_NOT_NULL(
        parent_scope,
        "Both program and parent_scope should be set in Clone mode.");
Y
Yan Chunwei 已提交
97
    scope_ = parent_scope;
98
    status_is_cloned_ = true;
Y
Yan Chunwei 已提交
99 100 101
  } else {
    paddle::framework::InitDevices(false);
    scope_.reset(new paddle::framework::Scope());
102
    status_is_cloned_ = false;
Y
Yan Chunwei 已提交
103
  }
104 105 106 107 108
  sub_scope_ = &scope_->NewScope();
  return true;
}
bool AnalysisPredictor::PrepareProgram(
    const std::shared_ptr<framework::ProgramDesc> &program) {
109 110
  if (!program) {
    if (!LoadProgramDesc()) return false;
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131

    // Optimize the program, and load parameters and modify them in the
    // scope_.
    // This will change the scope_ address.
    if (config_.enable_ir_optim) {
      status_ir_optim_enabled_ = true;
      OptimizeInferenceProgram();
    } else {
      // If the parent_scope is passed, we assert that the persistable variables
      // are already created, so just create the no persistable variables.

      // If not cloned, the parameters should be loaded
      // OptimizeInferenceProgram.
      // So in both cases, just the local variables are needed to load, not the
      // parematers.
      executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);

      // Load parameters
      LOG(INFO) << "load parameters ";
      LoadParameters();
    }
Y
Yan Chunwei 已提交
132
  } else {
133 134
    // If the program is passed from external, no need to optimize it, this
    // logic is used in the clone scenario.
135 136
    inference_program_ = program;
  }
M
Michal Gallus 已提交
137

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
  executor_->CreateVariables(*inference_program_, 0, false, sub_scope_);

  return true;
}
bool AnalysisPredictor::CreateExecutor() {
  if (config_.use_gpu) {
    status_use_gpu_ = true;
    place_ = paddle::platform::CUDAPlace(config_.device);
  } else {
    place_ = paddle::platform::CPUPlace();
  }
  executor_.reset(new paddle::framework::NaiveExecutor(place_));
  return true;
}
bool AnalysisPredictor::PrepareExecutor() {
  executor_->Prepare(sub_scope_, *inference_program_, 0,
154 155
                     config_.use_feed_fetch_ops);

156
  PADDLE_ENFORCE_NOT_NULL(sub_scope_);
Y
Yan Chunwei 已提交
157

158 159 160
  return true;
}

L
luotao1 已提交
161
void AnalysisPredictor::SetMkldnnThreadID(int tid) {
L
luotao1 已提交
162 163 164 165 166 167 168
#ifdef PADDLE_WITH_MKLDNN
  platform::set_cur_thread_id(tid);
#else
  LOG(ERROR) << "Please compile with MKLDNN first to use MKLDNN";
#endif
}

169 170 171
bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
                            std::vector<PaddleTensor> *output_data,
                            int batch_size) {
M
minqiyang 已提交
172
  VLOG(3) << "Predictor::predict";
173 174 175 176 177 178
  inference::Timer timer;
  timer.tic();
  // set feed variable
  framework::Scope *scope = sub_scope_ ? sub_scope_ : scope_.get();
  if (!SetFeed(inputs, scope)) {
    LOG(ERROR) << "fail to set feed";
Y
Yan Chunwei 已提交
179
    return false;
180
  }
M
Michal Gallus 已提交
181

182 183 184
  // Run the inference program
  // if share variables, we need not create variables
  executor_->Run();
185

186 187 188 189
  // get fetch variable
  if (!GetFetch(output_data, scope)) {
    LOG(ERROR) << "fail to get fetches";
    return false;
T
tensor-tang 已提交
190
  }
M
minqiyang 已提交
191
  VLOG(3) << "predict cost: " << timer.toc() << "ms";
Y
Yan Chunwei 已提交
192

Y
Yan Chunwei 已提交
193 194 195 196 197 198 199
  // All the containers in the scope will be hold in inference, but the
  // operators assume that the container will be reset after each batch.
  // Here is a bugfix, collect all the container variables, and reset then to a
  // bool; the next time, the operator will call MutableData and construct a new
  // container again, so that the container will be empty for each batch.
  tensor_array_batch_cleaner_.CollectNoTensorVars(sub_scope_);
  tensor_array_batch_cleaner_.ResetNoTensorVars();
200 201
  return true;
}
202

203 204
bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                                framework::Scope *scope) {
M
minqiyang 已提交
205
  VLOG(3) << "Predictor::set_feed";
206 207 208 209 210 211 212 213 214 215 216 217 218 219
  if (inputs.size() != feeds_.size()) {
    LOG(ERROR) << "wrong feed input size, need " << feeds_.size() << " but get "
               << inputs.size();
    return false;
  }

  // Cache the inputs memory for better concurrency performance.
  feed_tensors_.resize(inputs.size());

  for (size_t i = 0; i < inputs.size(); ++i) {
    auto &input = feed_tensors_[i];
    framework::DDim ddim = framework::make_ddim(inputs[i].shape);
    void *input_ptr;
    if (inputs[i].dtype == PaddleDType::INT64) {
220
      input_ptr = input.mutable_data<int64_t>(ddim, place_);
221
    } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
222
      input_ptr = input.mutable_data<float>(ddim, place_);
223 224 225 226 227
    } else {
      LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
      return false;
    }

228 229 230 231 232 233
    if (platform::is_cpu_place(place_)) {
      // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
      std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
                  inputs[i].data.length());
    } else {
#ifdef PADDLE_WITH_CUDA
Q
qingqing01 已提交
234 235 236 237
      platform::DeviceContextPool &pool =
          platform::DeviceContextPool::Instance();
      auto *dev_ctx =
          static_cast<const platform::CUDADeviceContext *>(pool.Get(place_));
238 239 240
      auto dst_gpu_place = boost::get<platform::CUDAPlace>(place_);
      memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
                   platform::CPUPlace(), inputs[i].data.data(),
Q
qingqing01 已提交
241
                   inputs[i].data.length(), dev_ctx->stream());
242 243 244 245
#else
      PADDLE_THROW("Not compile with CUDA, should not reach here.");
#endif
    }
246 247 248 249 250 251 252 253
    // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
    framework::LoD lod;
    for (auto &level : inputs[i].lod) {
      lod.emplace_back(level);
    }
    input.set_lod(lod);
    int idx = -1;
    if (config_.specify_input_name) {
T
tensor-tang 已提交
254 255
      auto name = inputs[i].name;
      if (feed_names_.find(name) == feed_names_.end()) {
T
tensor-tang 已提交
256 257
        LOG(ERROR) << "feed names from program do not have name: [" << name
                   << "] from specified input";
T
tensor-tang 已提交
258 259
      }
      idx = feed_names_[name];
260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
    } else {
      idx = boost::get<int>(feeds_[i]->GetAttr("col"));
    }
    framework::SetFeedVariable(scope, input, "feed", idx);
  }
  return true;
}

template <typename T>
void AnalysisPredictor::GetFetchOne(const framework::LoDTensor &fetch,
                                    PaddleTensor *output) {
  // set shape.
  auto shape = framework::vectorize(fetch.dims());
  output->shape.assign(shape.begin(), shape.end());
  // set data.
  const T *data = fetch.data<T>();
  int num_elems = inference::VecReduceToInt(shape);
  output->data.Resize(num_elems * sizeof(T));
  // The fetched tensor output by fetch op, should always in CPU memory, so just
  // copy.
  memcpy(output->data.data(), data, num_elems * sizeof(T));
  // set lod
  output->lod.clear();
  for (auto &level : fetch.lod()) {
    output->lod.emplace_back(level.begin(), level.end());
  }
}

bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
                                 framework::Scope *scope) {
M
minqiyang 已提交
290
  VLOG(3) << "Predictor::get_fetch";
291 292 293 294 295 296 297 298
  outputs->resize(fetchs_.size());
  for (size_t i = 0; i < fetchs_.size(); ++i) {
    int idx = boost::get<int>(fetchs_[i]->GetAttr("col"));
    PADDLE_ENFORCE((size_t)idx == i);
    framework::LoDTensor &fetch =
        framework::GetFetchVariable(*scope, "fetch", idx);
    auto type = fetch.type();
    auto output = &(outputs->at(i));
299
    output->name = fetchs_[idx]->Input("X")[0];
Y
Yu Yang 已提交
300
    if (type == framework::proto::VarType::FP32) {
301 302
      GetFetchOne<float>(fetch, output);
      output->dtype = PaddleDType::FLOAT32;
Y
Yu Yang 已提交
303
    } else if (type == framework::proto::VarType::INT64) {
304 305 306 307 308 309
      GetFetchOne<int64_t>(fetch, output);
      output->dtype = PaddleDType::INT64;
    } else {
      LOG(ERROR) << "unknown type, only support float32 and int64 now.";
    }
  }
Y
Yan Chunwei 已提交
310 311
  return true;
}
312

313
// NOTE All the members in AnalysisConfig should be copied to Argument.
Y
Yan Chunwei 已提交
314
void AnalysisPredictor::OptimizeInferenceProgram() {
315 316 317
  status_program_optimized_ = true;

  argument_.SetUseGPU(config_.use_gpu);
S
superjomn 已提交
318
  argument_.SetGPUDeviceId(config_.device);
T
Tao Luo 已提交
319
  argument_.SetModelFromMemory(config_.model_from_memory_);
Y
Yan Chunwei 已提交
320 321
  // Analyze inference_program
  if (!config_.model_dir.empty()) {
322
    argument_.SetModelDir(config_.model_dir);
T
Tao Luo 已提交
323 324 325 326 327
  } else {
    PADDLE_ENFORCE(
        !config_.param_file.empty(),
        "Either model_dir or (param_file, prog_file) should be set.");
    PADDLE_ENFORCE(!config_.prog_file.empty());
328 329
    argument_.SetModelProgramPath(config_.prog_file);
    argument_.SetModelParamsPath(config_.param_file);
Y
Yan Chunwei 已提交
330
  }
331

332 333 334 335
  if (config_.use_gpu && config_.use_tensorrt_) {
    argument_.SetUseTensorRT(true);
    argument_.SetTensorRtWorkspaceSize(config_.tensorrt_workspace_size_);
    argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
336
    argument_.SetTensorRtMinSubgraphSize(config_.tensorrt_min_subgraph_size_);
W
Wojciech Uss 已提交
337
  }
338

339 340 341 342
  if (config_.use_mkldnn_) {
    argument_.SetMKLDNNEnabledOpTypes(config_.mkldnn_enabled_op_types_);
  }

343 344 345 346 347 348 349 350 351
  auto passes = config_.pass_builder()->AllPasses();
  if (!config_.enable_ir_optim) passes.clear();
  argument_.SetIrAnalysisPasses(passes);
  argument_.SetScopeNotOwned(const_cast<framework::Scope *>(scope_.get()));
  Analyzer().Run(&argument_);

  PADDLE_ENFORCE(argument_.scope_valid());
  VLOG(5) << "to prepare executor";
  ARGUMENT_CHECK_FIELD((&argument_), ir_analyzed_program);
Y
Yan Chunwei 已提交
352
  inference_program_.reset(
353
      new framework::ProgramDesc(argument_.ir_analyzed_program()));
354
  LOG(INFO) << "== optimize end ==";
Y
Yan Chunwei 已提交
355
}
356 357

template <>
358 359
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
    AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) {
M
minqiyang 已提交
360
  VLOG(3) << "create AnalysisConfig";
361 362 363 364 365 366 367 368 369 370 371 372 373
  if (config.use_gpu) {
    // 1. GPU memeroy
    PADDLE_ENFORCE_GT(
        config.fraction_of_gpu_memory, 0.f,
        "fraction_of_gpu_memory in the config should be set to range (0., 1.]");
    PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device);
    std::vector<std::string> flags;
    if (config.fraction_of_gpu_memory >= 0.0f ||
        config.fraction_of_gpu_memory <= 0.95f) {
      flags.push_back("dummpy");
      std::string flag = "--fraction_of_gpu_memory_to_use=" +
                         std::to_string(config.fraction_of_gpu_memory);
      flags.push_back(flag);
M
minqiyang 已提交
374
      VLOG(3) << "set flag: " << flag;
375 376 377 378 379
      framework::InitGflags(flags);
    }
  }

  std::unique_ptr<PaddlePredictor> predictor(new AnalysisPredictor(config));
380
  if (!dynamic_cast<AnalysisPredictor *>(predictor.get())->Init(nullptr)) {
381 382
    return nullptr;
  }
383
  return std::move(predictor);
384 385
}

386
void AnalysisPredictor::PrepareFeedFetch() {
387 388
  PADDLE_ENFORCE_NOT_NULL(sub_scope_);
  CreateFeedFetchVar(sub_scope_);
389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406
  for (auto *op : inference_program_->Block(0).AllOps()) {
    if (op->Type() == "feed") {
      int idx = boost::get<int>(op->GetAttr("col"));
      if (feeds_.size() <= static_cast<size_t>(idx)) {
        feeds_.resize(idx + 1);
      }
      feeds_[idx] = op;
      feed_names_[op->Output("Out")[0]] = idx;
    } else if (op->Type() == "fetch") {
      int idx = boost::get<int>(op->GetAttr("col"));
      if (fetchs_.size() <= static_cast<size_t>(idx)) {
        fetchs_.resize(idx + 1);
      }
      fetchs_[idx] = op;
    }
  }
}

407 408 409 410 411 412 413 414
void AnalysisPredictor::CreateFeedFetchVar(framework::Scope *scope) {
  PADDLE_ENFORCE_NOT_NULL(scope);
  auto *var = scope->Var("feed");
  var->GetMutable<framework::FeedFetchList>();
  var = scope->Var("fetch");
  var->GetMutable<framework::FeedFetchList>();
}

415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436
std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
    const std::string &name) {
  PADDLE_ENFORCE(executor_->scope()->FindVar(name), "no name called %s", name);
  std::unique_ptr<ZeroCopyTensor> res(
      new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
  res->input_or_output_ = true;
  res->SetName(name);
  return res;
}

std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
    const std::string &name) {
  PADDLE_ENFORCE(executor_->scope()->FindVar(name), "no name called %s", name);
  std::unique_ptr<ZeroCopyTensor> res(
      new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
  res->input_or_output_ = false;
  res->SetName(name);
  return res;
}

bool AnalysisPredictor::ZeroCopyRun() {
  executor_->Run();
Y
Yan Chunwei 已提交
437
  // Fix TensorArray reuse not cleaned bug.
Y
Yan Chunwei 已提交
438
  tensor_array_batch_cleaner_.CollectTensorArrays(sub_scope_);
Y
Yan Chunwei 已提交
439
  tensor_array_batch_cleaner_.ResetTensorArray();
440 441 442 443 444
  return true;
}

bool AnalysisPredictor::LoadProgramDesc() {
  // Initialize the inference program
445
  std::string filename;
446
  if (!config_.model_dir.empty()) {
447
    filename = config_.model_dir + "/__model__";
448 449 450 451
  } else if (!config_.prog_file.empty() && !config_.param_file.empty()) {
    // All parameters are saved in a single file.
    // The file names should be consistent with that used
    // in Python API `fluid.io.save_inference_model`.
452
    filename = config_.prog_file;
453
  } else {
454 455 456 457 458
    if (config_.model_dir.empty() && config_.prog_file.empty()) {
      LOG(ERROR)
          << "Either model_dir or (prog_file, param_file) should be set.";
      return false;
    }
459 460 461 462 463
    LOG(ERROR) << string::Sprintf(
        "not valid model path '%s' or program path '%s'.", config_.model_dir,
        config_.param_file);
    return false;
  }
464 465 466

  // Create ProgramDesc
  framework::proto::ProgramDesc proto;
T
Tao Luo 已提交
467
  if (!config_.model_from_memory()) {
T
Tao Luo 已提交
468 469 470
    std::string pb_content;
    // Read binary
    std::ifstream fin(filename, std::ios::in | std::ios::binary);
T
Tao Luo 已提交
471 472
    PADDLE_ENFORCE(static_cast<bool>(fin.is_open()), "Cannot open file %s",
                   filename);
T
Tao Luo 已提交
473 474 475 476 477 478 479 480 481 482
    fin.seekg(0, std::ios::end);
    pb_content.resize(fin.tellg());
    fin.seekg(0, std::ios::beg);
    fin.read(&(pb_content.at(0)), pb_content.size());
    fin.close();

    proto.ParseFromString(pb_content);
  } else {
    proto.ParseFromString(config_.prog_file);
  }
483 484 485 486 487 488 489
  inference_program_.reset(new framework::ProgramDesc(proto));
  return true;
}

bool AnalysisPredictor::LoadParameters() {
  PADDLE_ENFORCE_NOT_NULL(inference_program_.get(),
                          "The inference program should be loaded first.");
T
Tao Luo 已提交
490

491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535
  const auto &global_block = inference_program_->MutableBlock(0);

  // create a temporary program to load parameters.

  std::unique_ptr<framework::ProgramDesc> load_program(
      new framework::ProgramDesc());
  framework::BlockDesc *load_block = load_program->MutableBlock(0);
  std::vector<std::string> params;

  for (auto *var : global_block->AllVars()) {
    if (IsPersistable(var)) {
      VLOG(3) << "persistable variable's name: " << var->Name();

      framework::VarDesc *new_var = load_block->Var(var->Name());
      new_var->SetShape(var->GetShape());
      new_var->SetDataType(var->GetDataType());
      new_var->SetType(var->GetType());
      new_var->SetLoDLevel(var->GetLoDLevel());
      new_var->SetPersistable(true);

      if (!config_.param_file.empty()) {
        params.push_back(new_var->Name());
      } else {
        // append_op
        framework::OpDesc *op = load_block->AppendOp();
        op->SetType("load");
        op->SetOutput("Out", {new_var->Name()});
        op->SetAttr("file_path", {config_.model_dir + "/" + new_var->Name()});
        op->CheckAttrs();
      }
    }
  }

  if (!config_.param_file.empty()) {
    // sort paramlist to have consistent ordering
    std::sort(params.begin(), params.end());
    // append just the load_combine op
    framework::OpDesc *op = load_block->AppendOp();
    op->SetType("load_combine");
    op->SetOutput("Out", params);
    op->SetAttr("file_path", {config_.param_file});
    op->CheckAttrs();
  }

  // Use NaiveExecutor to Load parameters.
S
superjomn 已提交
536
  framework::NaiveExecutor e(place_);
537 538 539 540
  e.Prepare(scope_.get(), *load_program, 0, false);
  e.Run();
  VLOG(3) << "get " << scope_->LocalVarNames().size() << " vars after load";

541 542
  return true;
}
543 544 545 546 547 548 549 550 551 552 553

AnalysisPredictor::~AnalysisPredictor() {
  if (FLAGS_profile) {
    platform::DisableProfiler(platform::EventSortingKey::kTotal,
                              "./profile.log");
  }
  if (sub_scope_) {
    scope_->DeleteScope(sub_scope_);
  }
}

554 555 556 557 558 559
std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
  auto *x = new AnalysisPredictor(config_);
  x->Init(scope_, inference_program_);
  return std::unique_ptr<PaddlePredictor>(x);
}

Y
Yan Chunwei 已提交
560 561
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<contrib::AnalysisConfig>(
562
    const contrib::AnalysisConfig &config) {
Y
Yan Chunwei 已提交
563 564 565 566
  return CreatePaddlePredictor<contrib::AnalysisConfig,
                               PaddleEngineKind::kAnalysis>(config);
}

567
}  // namespace paddle
568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589

#if PADDLE_WITH_TENSORRT
USE_TRT_CONVERTER(elementwise_add_weight);
USE_TRT_CONVERTER(elementwise_add_tensor);
USE_TRT_CONVERTER(elementwise_sub_tensor);
USE_TRT_CONVERTER(elementwise_div_tensor);
USE_TRT_CONVERTER(elementwise_mul_tensor);
USE_TRT_CONVERTER(elementwise_max_tensor);
USE_TRT_CONVERTER(elementwise_min_tensor);
USE_TRT_CONVERTER(elementwise_pow_tensor);
USE_TRT_CONVERTER(mul);
USE_TRT_CONVERTER(conv2d);
USE_TRT_CONVERTER(relu);
USE_TRT_CONVERTER(sigmoid);
USE_TRT_CONVERTER(tanh);
USE_TRT_CONVERTER(fc);
USE_TRT_CONVERTER(pool2d);
USE_TRT_CONVERTER(softmax);
USE_TRT_CONVERTER(batch_norm);
USE_TRT_CONVERTER(concat);
USE_TRT_CONVERTER(dropout);
USE_TRT_CONVERTER(pad);
590
USE_TRT_CONVERTER(split);
591 592
USE_TRT_CONVERTER(prelu);
USE_TRT_CONVERTER(conv2d_transpose);
H
hjchen2 已提交
593
USE_TRT_CONVERTER(leaky_relu);
594
#endif