engine.h 22.7 KB
Newer Older
Y
Yan Chunwei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
/* 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 <NvInfer.h>
18
#include <map>
Y
Yan Chunwei 已提交
19
#include <memory>
20
#include <mutex>  // NOLINT
21
#include <string>
Y
Yan Chunwei 已提交
22
#include <unordered_map>
23
#include <unordered_set>
24
#include <utility>
25
#include <vector>
W
wanghuancoder 已提交
26

N
nhzlx 已提交
27
#include "paddle/fluid/framework/tensor.h"
28
#include "paddle/fluid/framework/tensor_util.h"
Z
Zhaolong Xing 已提交
29
#include "paddle/fluid/inference/api/paddle_analysis_config.h"
Y
Yan Chunwei 已提交
30 31
#include "paddle/fluid/inference/engine.h"
#include "paddle/fluid/inference/tensorrt/helper.h"
32
#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
N
nhzlx 已提交
33
#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
34
#include "paddle/fluid/inference/utils/singleton.h"
35
#include "paddle/fluid/platform/enforce.h"
36
#include "paddle/utils/any.h"
Y
Yan Chunwei 已提交
37

W
wanghuancoder 已提交
38 39 40 41 42 43
namespace paddle {
namespace framework {
class Tensor;
}  // namespace framework
}  // namespace paddle

Y
Yan Chunwei 已提交
44 45 46 47
namespace paddle {
namespace inference {
namespace tensorrt {

W
wanghuancoder 已提交
48 49 50 51
namespace plugin {
class PluginTensorRT;
}  // namespace plugin

52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
using FluidDT = framework::proto::VarType_Type;
using TRT_DT = nvinfer1::DataType;

namespace {  // NOLINT

TRT_DT FluidDataType2TRT(FluidDT type) {
  switch (type) {
    case FluidDT::VarType_Type_FP32:
      return TRT_DT::kFLOAT;
    case FluidDT::VarType_Type_INT32:
      return TRT_DT::kINT32;
    default:
      return TRT_DT::kINT32;
  }
  PADDLE_THROW(platform::errors::InvalidArgument(
      "unknown fluid datatype in TRT op converter"));
  return TRT_DT::kINT32;
}

// The T can be int32 or int64 type.
template <typename T>
nvinfer1::Dims Vec2TRT_Dims(const std::vector<T>& shape, std::string input,
                            bool with_dynamic_shape = false) {
75
  PADDLE_ENFORCE_GT(shape.size(), 0UL,
76
                    platform::errors::InvalidArgument(
77
                        "TensorRT's tensor input requires at least 1 "
78 79
                        "dimensions, but input %s has %d dims.",
                        input, shape.size()));
W
wenbin 已提交
80

81 82 83 84 85 86 87 88 89 90 91 92 93
  auto ShapeStr = [](const std::vector<T>& shape) {
    std::ostringstream os;
    os << "[";
    for (size_t i = 0; i < shape.size(); ++i) {
      if (i == shape.size() - 1) {
        os << shape[i];
      } else {
        os << shape[i] << ",";
      }
    }
    os << "]";
    return os.str();
  };
94 95
  if (!with_dynamic_shape) {
    if (shape.size() == 4UL) {
96 97 98 99 100 101
      if (shape[2] == -1 || shape[3] == -1) {
        PADDLE_THROW(platform::errors::InvalidArgument(
            "The input [%s] shape of trt subgraph is %s, please enable "
            "trt dynamic_shape mode by SetTRTDynamicShapeInfo.",
            input, ShapeStr(shape)));
      }
102
      return nvinfer1::Dims3(shape[1], shape[2], shape[3]);
W
wenbin 已提交
103 104 105 106 107 108 109 110
    } else if (shape.size() == 5UL) {
      if (shape[2] == -1 || shape[3] == -1 || shape[4] == -1) {
        PADDLE_THROW(platform::errors::InvalidArgument(
            "The input [%s] shape of trt subgraph is %s, please enable "
            "trt dynamic_shape mode by SetTRTDynamicShapeInfo.",
            input, ShapeStr(shape)));
      }
      return nvinfer1::Dims4(shape[1], shape[2], shape[3], shape[4]);
111
    } else if (shape.size() == 3UL) {
112 113 114 115 116 117
      if (shape[1] == -1 || shape[2] == -1) {
        PADDLE_THROW(platform::errors::InvalidArgument(
            "The input [%s] shape of trt subgraph is %s, please enable "
            "trt dynamic_shape mode by SetTRTDynamicShapeInfo.",
            input, ShapeStr(shape)));
      }
118 119
      return nvinfer1::Dims2(shape[1], shape[2]);
    }
120
    return nvinfer1::Dims3(shape[1], 1, 1);
121 122
  } else {
    if (shape.size() == 4UL) {
123
      return nvinfer1::Dims4(shape[0], shape[1], shape[2], shape[3]);
124 125 126
    } else if (shape.size() == 3UL) {
      return nvinfer1::Dims3(shape[0], shape[1], shape[2]);
    }
127 128 129 130 131 132
    nvinfer1::Dims dims;
    dims.nbDims = shape.size();
    for (size_t i = 0; i < shape.size(); i++) {
      dims.d[i] = shape[i];
    }
    return dims;
133 134 135 136
  }
}
}  // NOLINT

N
nhzlx 已提交
137
class TRTInt8Calibrator;
W
wanghuancoder 已提交
138

Y
Yan Chunwei 已提交
139 140 141 142
/*
 * TensorRT Engine.
 *
 * There are two alternative ways to use it, one is  to build from a paddle
143
 * protobuf model, another way is to manually construct the network.
Y
Yan Chunwei 已提交
144
 */
145 146
class TensorRTEngine {
  using DescType = ::paddle::framework::proto::BlockDesc;
147
  using ShapeMapType = std::map<std::string, std::vector<int>>;
148

Y
Yan Chunwei 已提交
149 150 151 152
 public:
  // Weight is model parameter.
  class Weight {
   public:
153
    Weight() = default;
154
    Weight(nvinfer1::DataType dtype, void* value, size_t num_elem) {
Y
Yan Chunwei 已提交
155 156 157 158
      w_.type = dtype;
      w_.values = value;
      w_.count = num_elem;
    }
159
    const nvinfer1::Weights& get() { return w_; }
Y
Yan Chunwei 已提交
160

161 162
    std::vector<int64_t> dims;

Y
Yan Chunwei 已提交
163 164 165 166
   private:
    nvinfer1::Weights w_;
  };

Z
Zhaolong Xing 已提交
167 168 169 170
  TensorRTEngine(
      int max_batch, int max_workspace,
      AnalysisConfig::Precision precision = AnalysisConfig::Precision::kFloat32,
      TRTInt8Calibrator* calibrator = nullptr, int device_id = 0,
171 172 173
      const ShapeMapType min_input_shape = {},
      const ShapeMapType max_input_shape = {},
      const ShapeMapType optim_input_shape = {},
174
      bool disable_trt_plugin_fp16 = false,
Z
Zhaolong Xing 已提交
175
      nvinfer1::ILogger& logger = NaiveLogger::Global())
Y
Yan Chunwei 已提交
176 177
      : max_batch_(max_batch),
        max_workspace_(max_workspace),
Z
Zhaolong Xing 已提交
178
        precision_(precision),
N
nhzlx 已提交
179
        calibrator_(calibrator),
N
nhzlx 已提交
180
        device_id_(device_id),
181 182 183
        min_input_shape_(min_input_shape),
        max_input_shape_(max_input_shape),
        optim_input_shape_(optim_input_shape),
184
        disable_trt_plugin_fp16_(disable_trt_plugin_fp16),
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
        logger_(logger) {
    if (min_input_shape_.size() != 0 && max_input_shape_.size() != 0 &&
        optim_input_shape_.size() != 0) {
      PADDLE_ENFORCE_EQ(
          min_input_shape_.size(), max_input_shape_.size(),
          platform::errors::InvalidArgument(
              "The min_input_shape_'s size(%d) should be equal to the "
              "size(%d) of max_input_shape_",
              min_input_shape_.size(), max_input_shape_.size()));
      PADDLE_ENFORCE_EQ(
          min_input_shape_.size(), optim_input_shape_.size(),
          platform::errors::InvalidArgument(
              "The min_input_shape_'s size(%d) should be equal to the "
              "size(%d) of optim_input_shape_",
              min_input_shape_.size(), optim_input_shape_.size()));
#if IS_TRT_VERSION_GE(6000)
      with_dynamic_shape_ = true;
#else
      LOG(WARNING) << "Using dynamic shape of TRT need ensure that the TRT "
                      "version should be at least 6.";
#endif
    }
207
    dy::initLibNvInferPlugins(&logger, "");
208
  }
Y
Yan Chunwei 已提交
209

210 211 212 213 214 215 216 217 218
  ~TensorRTEngine() {
    for (auto& attr : attrs_) {
      if (attr_dels_.find(attr.first) != attr_dels_.end()) {
        attr_dels_[attr.first]();
      }
    }
    attrs_.clear();
    attr_dels_.clear();
  }
Y
Yan Chunwei 已提交
219

220
  // Add an input and set its name, data type and dimension.
Y
Yan Chunwei 已提交
221 222 223 224 225 226 227
  nvinfer1::ITensor* DeclareInput(const std::string& name,
                                  nvinfer1::DataType dtype,
                                  const nvinfer1::Dims& dim);
  // Set the offset-th output from a layer as the network's output, and set its
  // name.
  void DeclareOutput(const nvinfer1::ILayer* layer, int offset,
                     const std::string& name);
L
Luo Tao 已提交
228 229
  // Set the itensor_map_[name] as the network's output, and set its name.
  void DeclareOutput(const std::string& name);
230
  void ClearTensorMap() { itensor_map_.clear(); }
Y
Yan Chunwei 已提交
231

L
Luo Tao 已提交
232 233 234
  void SetITensor(const std::string& name, nvinfer1::ITensor* tensor);
  // Get an ITensor called name.
  nvinfer1::ITensor* GetITensor(const std::string& name);
Y
Yan Chunwei 已提交
235 236

  nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); }
237 238 239 240 241 242 243 244 245 246 247 248
  nvinfer1::IExecutionContext* context() {
    std::unique_lock<std::mutex> lock(mutex_);
    const std::thread::id tid = std::this_thread::get_id();
    if (infer_context_.find(tid) == infer_context_.end()) {
      PADDLE_ENFORCE_NOT_NULL(
          infer_engine_,
          platform::errors::InvalidArgument(
              "You should build engine first and then set the context."));
      infer_context_[tid].reset(infer_engine_->createExecutionContext());
    }
    return infer_context_[tid].get();
  }
249 250 251 252 253 254 255 256 257 258
  void ResetContext() {
    std::unique_lock<std::mutex> lock(mutex_);
    const std::thread::id tid = std::this_thread::get_id();
    PADDLE_ENFORCE_NOT_NULL(
        infer_engine_,
        platform::errors::InvalidArgument(
            "You should build engine first and then set the context."));
    infer_context_[tid].reset(nullptr);
    infer_context_.erase(tid);
  }
N
nhzlx 已提交
259 260

  nvinfer1::IHostMemory* Serialize() {
261 262 263 264
    PADDLE_ENFORCE_NOT_NULL(
        infer_engine_,
        platform::errors::InvalidArgument(
            "The TensorRT engine must be built first before serialization"));
N
nhzlx 已提交
265 266 267 268 269
    ihost_memory_.reset(infer_engine_->serialize());
    return ihost_memory_.get();
  }

  void Deserialize(const std::string& engine_serialized_data) {
N
nhzlx 已提交
270
    freshDeviceId();
N
nhzlx 已提交
271
    infer_ptr<nvinfer1::IRuntime> runtime(createInferRuntime(&logger_));
272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294

    if (use_dla_) {
      if (precision_ != AnalysisConfig::Precision::kInt8 &&
          precision_ != AnalysisConfig::Precision::kHalf) {
        LOG(WARNING) << "TensorRT DLA must be used with int8 or fp16, but you "
                        "set float32, so DLA is not used.";
      } else if (runtime->getNbDLACores() == 0) {
        LOG(WARNING)
            << "TensorRT DLA is set by config, but your device does not have "
               "DLA, so DLA is not used.";
      } else {
        if (dla_core_ < 0 || dla_core_ >= runtime->getNbDLACores()) {
          dla_core_ = 0;
          LOG(WARNING) << "Invalid DLACore, must be 0 < DLACore < "
                       << runtime->getNbDLACores() << ", but got " << dla_core_
                       << ", so use use 0 as default.";
        }
        runtime->setDLACore(dla_core_);
        LOG(INFO) << "TensorRT DLA enabled in Deserialize(), DLACore "
                  << dla_core_;
      }
    }

295 296
    infer_engine_.reset(runtime->deserializeCudaEngine(
        engine_serialized_data.c_str(), engine_serialized_data.size()));
297

298 299 300 301 302 303 304 305
    PADDLE_ENFORCE_NOT_NULL(
        infer_engine_,
        platform::errors::Fatal(
            "Building TRT cuda engine failed when deserializing engine info. "
            "Please check:\n1. Your TRT serialization is generated and loaded "
            "on the same GPU architecture;\n2. The Paddle Inference version of "
            "generating serialization file and doing inference are "
            "consistent."));
N
nhzlx 已提交
306 307
  }

308 309
  void SetRuntimeBatch(size_t batch_size);
  int GetRuntimeBatch();
310 311 312 313 314 315 316

  bool WithFp16() {
    bool enable_fp16 = (precision_ == AnalysisConfig::Precision::kHalf);
    bool support_fp16 = infer_builder_->platformHasFastFp16();
    return enable_fp16 && support_fp16;
  }

N
nhzlx 已提交
317
  int GetDeviceId() { return device_id_; }
318

319 320
  nvinfer1::IPluginV2Layer* AddPlugin(nvinfer1::ITensor* const* inputs,
                                      int num_inputs, plugin::PluginTensorRT*);
321 322 323 324 325

  nvinfer1::IPluginV2Layer* AddPluginV2Ext(nvinfer1::ITensor* const* inputs,
                                           int num_inputs,
                                           plugin::PluginTensorRTV2Ext* plugin);

326 327 328 329 330 331 332
  void SetTensorDynamicRange(nvinfer1::ITensor* tensor, float range) {
    quant_dynamic_range_[tensor] = range;
  }

  float* GetWeightCPUData(const std::string& name,
                          framework::Tensor* weight_tensor, bool enable_int8,
                          const std::vector<float>& scale = {});
N
nhzlx 已提交
333 334 335 336 337 338 339 340

  // A pointer to CPU memory is needed of the TRT weight.
  // Before TRT runs, fluid loads weight into GPU storage.
  // so we need to copy the weights from GPU to CPU in our op converter.
  // We use a map to store these weights for the weight memory is not released
  // in advance, which affecting the construction of TRT Op.
  std::unordered_map<std::string /*name*/, std::unique_ptr<framework::Tensor>>
      weight_map;
Y
Yan Chunwei 已提交
341

342 343 344 345 346 347
  // When setting weight_map, a self-increasing suffix is needed for the names
  // so as to avoid repeatedly setting weights with the same name.
  void SetWeights(std::string w_name,
                  std::unique_ptr<framework::Tensor> w_tensor) {
    static int suffix_counter = 0;
    std::string suffix = std::to_string(suffix_counter);
P
Pei Yang 已提交
348 349
    std::string splitter = "__";
    weight_map[w_name + splitter + suffix] = std::move(w_tensor);
350 351 352
    suffix_counter += 1;
  }

353
  void SetUseOSS(bool use_oss) { use_oss_ = use_oss; }
354 355
  void SetUseDLA(bool use_dla) { use_dla_ = use_dla; }
  void SetDLACore(int dla_core) { dla_core_ = dla_core; }
356 357
  void SetWithErnie(bool with_ernie) { with_ernie_ = with_ernie; }

358 359 360 361 362 363
  void ClearWeights() {
    for (auto& weight_pair : weight_map) {
      weight_pair.second.reset(nullptr);
    }
  }

364 365 366 367 368 369 370 371 372 373
  // NOTE: The func bellow was modified to adapt the dynamic shape.
  // Initialize the inference network, so that TensorRT layers can add to this
  // network.
  void InitNetwork();
  // After finishing adding ops, freeze this network and creates the execution
  // environment.
  void FreezeNetwork();
  void Execute(int batch_size, std::vector<void*>* buffers,
               cudaStream_t stream = nullptr);

374
  nvinfer1::INetworkDefinition* network() { return infer_network_.get(); }
375 376 377 378

  ShapeMapType min_input_shape() { return min_input_shape_; }
  ShapeMapType max_input_shape() { return max_input_shape_; }
  ShapeMapType optim_input_shape() { return optim_input_shape_; }
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427

  bool AdjustDynamicShapeRange(const ShapeMapType& runtime_input_shape,
                               std::vector<std::string>* changed) {
    bool ret = false;
    changed->clear();
    for (const auto& it : runtime_input_shape) {
      auto name = it.first;
      auto input_shape = it.second;
      PADDLE_ENFORCE_EQ(
          min_input_shape_.count(name), true,
          platform::errors::InvalidArgument(
              "TRT dynamic_shape min_input_shape %s not found.", name));
      PADDLE_ENFORCE_EQ(min_input_shape_[name].size(), input_shape.size(),
                        platform::errors::InvalidArgument(
                            "TRT dynamic_shape min_input_shape %s size not "
                            "equal, the min_input_shape[%s].size()=%d"
                            ", but the runtime_input_shape[%s].size()=%d.",
                            name, name, min_input_shape_[name].size(), name,
                            input_shape.size()));
      auto bak_min_shape = min_input_shape_[name];
      auto bak_max_shape = max_input_shape_[name];
      bool min_change = false;
      bool max_change = false;
      for (size_t d = 0; d < input_shape.size(); ++d) {
        if (input_shape[d] < min_input_shape_[name][d]) {
          ret = true;
          min_change = true;
          min_input_shape_[name][d] = input_shape[d];
        }
        if (input_shape[d] > max_input_shape_[name][d]) {
          ret = true;
          max_change = true;
          max_input_shape_[name][d] = input_shape[d];
        }
      }

      if (min_change)
        LOG(INFO) << "refactor shape range: " << name << ", min_shape from "
                  << Vec2Str(bak_min_shape) << " to "
                  << Vec2Str(min_input_shape_[name]);
      if (max_change)
        LOG(INFO) << "refactor shape range: " << name << ", max_shape from "
                  << Vec2Str(bak_max_shape) << " to "
                  << Vec2Str(max_input_shape_[name]);
      if (min_change || max_change) changed->push_back(name);
    }
    return ret;
  }

428 429
  bool use_oss() { return use_oss_; }
  bool with_ernie() { return with_ernie_; }
430
  bool disable_trt_plugin_fp16() { return disable_trt_plugin_fp16_; }
431 432
  bool with_dynamic_shape() { return with_dynamic_shape_; }

433
#if IS_TRT_VERSION_GE(6000)
434 435 436
  nvinfer1::IPluginV2Layer* AddDynamicPlugin(
      nvinfer1::ITensor* const* inputs, int num_inputs,
      plugin::DynamicPluginTensorRT* plugin) {
437 438 439 440 441
    owned_pluginv2_.emplace_back(plugin);
    return network()->addPluginV2(inputs, num_inputs, *plugin);
  }
#endif

442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493
  bool Has(const std::string& attr_name) const {
    return attrs_.count(attr_name) > 0;
  }

  void Erase(const std::string& attr_name) {
    if (!Has(attr_name)) {
      return;
    }
    if (attr_dels_.find(attr_name) != attr_dels_.end()) {
      attr_dels_[attr_name]();
      attr_dels_.erase(attr_name);
    }
    attrs_.erase(attr_name);
  }

  // Set a pointer to the attribute. Engine takes ownership of the attribute.
  template <typename AttrType>
  void Set(const std::string& attr_name, AttrType* attr) {
    if (attrs_.count(attr_name) == 0) {
      PADDLE_ENFORCE_EQ(
          attrs_.count(attr_name), 0,
          platform::errors::AlreadyExists(
              "Attribute %s already set in trt engine.", attr_name));
    } else {
      VLOG(3) << "Setting the attribute " << attr_name << " for trt engine "
              << this;
    }
    attrs_[attr_name] = attr;
    attr_dels_[attr_name] = [attr, attr_name]() {
      VLOG(3) << "deleting " << attr_name;
      delete attr;
    };
  }

  // Set a pointer to the attribute. Engine doesn't take ownership. Caller
  // should delete the attribute.
  template <typename AttrType>
  void SetNotOwned(const std::string& attr_name, AttrType* attr) {
    PADDLE_ENFORCE_EQ(
        attrs_.count(attr_name), 0,
        platform::errors::AlreadyExists(
            "Attribute %s already set in trt engine.", attr_name));
    attrs_[attr_name] = attr;
  }

  // Get a reference to the attributed previously set.
  template <typename AttrType>
  AttrType& Get(const std::string& attr_name) const {
    PADDLE_ENFORCE_NE(attrs_.find(attr_name), attrs_.end(),
                      platform::errors::InvalidArgument(
                          "Attribute %s not found in trt engine.", attr_name));
    try {
494 495
      return *paddle::any_cast<AttrType*>(attrs_.at(attr_name));
    } catch (paddle::bad_any_cast&) {
496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517
      auto TypeToString = [](const std::type_info& info) -> std::string {
        if (std::type_index(info) == std::type_index(typeid(bool*))) {
          return "bool";
        } else if (std::type_index(info) == std::type_index(typeid(int*))) {
          return "int";
        } else if (std::type_index(info) ==
                   std::type_index(typeid(const int*))) {
          return "const int";
        } else if (std::type_index(info) ==
                   std::type_index(typeid(std::string*))) {
          return "std::string";
        }
        return info.name();
      };

      PADDLE_THROW(platform::errors::InvalidArgument(
          "Invalid type for attritube %s, expected: %s, actual: %s.", attr_name,
          TypeToString(typeid(AttrType*)),
          TypeToString(attrs_.at(attr_name).type())));
    }
  }

Y
Yan Chunwei 已提交
518
 private:
N
nhzlx 已提交
519 520 521 522 523
  // Each ICudaEngine object is bound to a specific GPU when it is instantiated,
  // ensure that the thread is associated with the correct device by calling
  // freshDeviceId().
  void freshDeviceId();

Y
Yan Chunwei 已提交
524 525
  // the max batch size
  int max_batch_;
526 527
  // the runtime batch size
  static int runtime_batch_;
Y
Yan Chunwei 已提交
528 529
  // the max memory size the engine uses
  int max_workspace_;
530

Z
Zhaolong Xing 已提交
531
  AnalysisConfig::Precision precision_;
N
nhzlx 已提交
532 533 534
  TRTInt8Calibrator* calibrator_;
  // batch size of the current data, will be updated each Executation.
  int batch_size_{-1};
N
nhzlx 已提交
535

N
nhzlx 已提交
536
  int device_id_;
537 538 539
  ShapeMapType min_input_shape_;
  ShapeMapType max_input_shape_;
  ShapeMapType optim_input_shape_;
540
  bool disable_trt_plugin_fp16_{false};
541
  bool use_oss_{false};
542 543
  bool use_dla_{false};
  int dla_core_{0};
544
  bool with_ernie_{false};
Y
Yan Chunwei 已提交
545 546 547
  nvinfer1::ILogger& logger_;

  // max data size for the buffers.
L
Luo Tao 已提交
548 549
  std::unordered_map<std::string /*name*/, nvinfer1::ITensor* /*ITensor*/>
      itensor_map_;
550

551
  std::vector<std::unique_ptr<plugin::PluginTensorRT>> owned_plugin_;
552
  std::vector<std::unique_ptr<plugin::PluginTensorRTV2Ext>> owned_plugin_v2ext_;
Y
Yan Chunwei 已提交
553 554 555 556

  // TensorRT related internal members
  template <typename T>
  struct Destroyer {
557 558 559 560 561
    void operator()(T* x) {
      if (x) {
        x->destroy();
      }
    }
Y
Yan Chunwei 已提交
562 563 564 565 566 567
  };
  template <typename T>
  using infer_ptr = std::unique_ptr<T, Destroyer<T>>;
  infer_ptr<nvinfer1::IBuilder> infer_builder_;
  infer_ptr<nvinfer1::INetworkDefinition> infer_network_;
  infer_ptr<nvinfer1::ICudaEngine> infer_engine_;
568 569
  std::unordered_map<std::thread::id, infer_ptr<nvinfer1::IExecutionContext>>
      infer_context_;
N
nhzlx 已提交
570
  infer_ptr<nvinfer1::IHostMemory> ihost_memory_;
571
  std::unordered_map<nvinfer1::ITensor*, float> quant_dynamic_range_;
572

573
  std::unordered_map<std::string, paddle::any> attrs_;
574 575
  std::unordered_map<std::string, std::function<void(void)>> attr_dels_;

576 577 578 579
  // For dynamic shape
  bool with_dynamic_shape_{false};
#if IS_TRT_VERSION_GE(6000)
  infer_ptr<nvinfer1::IBuilderConfig> infer_builder_config_;
580
  nvinfer1::IOptimizationProfile* optim_profile_;
581
  std::vector<std::unique_ptr<plugin::DynamicPluginTensorRT>> owned_pluginv2_;
582
#endif
583
  std::mutex mutex_;
Y
Yan Chunwei 已提交
584 585
};  // class TensorRTEngine

586
// Add a layer__ into engine__ with args ARGS.
Y
Yan Chunwei 已提交
587 588 589 590 591 592 593 594 595
// For example:
//
// Reference
// https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#charRNN_define_network
//
// will add a fully connected layer into the engine.
// TensorRT has too many layers, so that is not wise to add member functions for
// them, and an macro like this is more extensible when underlying TensorRT
// library add new layer supports.
596 597
#define TRT_ENGINE_ADD_LAYER(engine__, layer__, ...) \
  engine__->network()->add##layer__(__VA_ARGS__);
Y
Yan Chunwei 已提交
598

599 600 601 602 603 604 605 606 607 608 609 610
class TRTEngineManager {
 public:
  bool Empty() const { return engines_.size() == 0; }
  bool Has(const std::string& name) const {
    if (engines_.count(name) == 0) return false;
    return engines_.at(name).get() != nullptr;
  }

  TensorRTEngine* Get(const std::string& name) const {
    return engines_.at(name).get();
  }

Z
Zhaolong Xing 已提交
611 612 613 614
  TensorRTEngine* Create(
      std::string name, int max_batch, int max_workspace,
      AnalysisConfig::Precision precision = AnalysisConfig::Precision::kFloat32,
      TRTInt8Calibrator* calibrator = nullptr, int device_id = 0,
615 616 617
      const std::map<std::string, std::vector<int>> min_input_shape = {},
      const std::map<std::string, std::vector<int>> max_input_shape = {},
      const std::map<std::string, std::vector<int>> optim_input_shape = {},
618
      bool disable_trt_plugin_fp16 = false,
Z
Zhaolong Xing 已提交
619
      nvinfer1::ILogger& logger = NaiveLogger::Global()) {
620 621 622 623
    auto* p =
        new TensorRTEngine(max_batch, max_workspace, precision, calibrator,
                           device_id, min_input_shape, max_input_shape,
                           optim_input_shape, disable_trt_plugin_fp16, logger);
624 625 626 627 628 629 630 631 632 633 634 635 636 637
    engines_[name].reset(p);
    return p;
  }

  void DeleteAll() {
    for (auto& item : engines_) {
      item.second.reset(nullptr);
    }
  }

 private:
  std::unordered_map<std::string, std::unique_ptr<TensorRTEngine>> engines_;
};

Y
Yan Chunwei 已提交
638 639 640
}  // namespace tensorrt
}  // namespace inference
}  // namespace paddle