“f32ca6369099f5d3776ae87d431b9b39ea8eba3e”上不存在“paddle/fluid/operators/reader/ctr_reader.cc”
analysis_predictor.cc 87.2 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>
18

19
#include <algorithm>
N
nhzlx 已提交
20
#include <fstream>
21
#include <memory>
22
#include <set>
23
#include <string>
24
#include <utility>
25
#include <vector>
26

W
Wilber 已提交
27
#include "paddle/fluid//platform/device/gpu/gpu_types.h"
28
#include "paddle/fluid/framework/feed_fetch_method.h"
29
#include "paddle/fluid/framework/feed_fetch_type.h"
30
#include "paddle/fluid/framework/generator.h"
Y
Yan Chunwei 已提交
31
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
32
#include "paddle/fluid/framework/ir/pass.h"
33
#include "paddle/fluid/framework/naive_executor.h"
34
#include "paddle/fluid/framework/op_proto_maker.h"
35
#include "paddle/fluid/framework/scope.h"
J
JingZhuangzhuang 已提交
36
#include "paddle/fluid/framework/transfer_scope_cache.h"
Y
Yan Chunwei 已提交
37
#include "paddle/fluid/framework/var_type_traits.h"
38
#include "paddle/fluid/framework/version.h"
39
#include "paddle/fluid/inference/analysis/helper.h"
40
#include "paddle/fluid/inference/analysis/passes/convert_to_mixed_precision.h"
Y
Yan Chunwei 已提交
41
#include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
42
#include "paddle/fluid/inference/api/helper.h"
43
#include "paddle/fluid/inference/api/infer_context.h"
44
#include "paddle/fluid/inference/api/paddle_analysis_config.h"
45
#include "paddle/fluid/inference/api/paddle_inference_api.h"
L
luotao1 已提交
46
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
W
Wilber 已提交
47
#include "paddle/fluid/inference/api/resource_manager.h"
48
#include "paddle/fluid/inference/utils/io_utils.h"
49
#include "paddle/fluid/inference/utils/model_utils.h"
50
#include "paddle/fluid/inference/utils/singleton.h"
51
#include "paddle/fluid/memory/memcpy.h"
52
#include "paddle/fluid/platform/cpu_helper.h"
53
#include "paddle/fluid/platform/device/gpu/gpu_info.h"
54
#include "paddle/fluid/platform/device_context.h"
55
#include "paddle/fluid/platform/place.h"
T
tensor-tang 已提交
56
#include "paddle/fluid/platform/profiler.h"
57
#include "paddle/phi/api/ext/op_meta_info.h"
58 59
#include "paddle/phi/common/backend.h"
#include "paddle/phi/common/data_type.h"
W
Wilber 已提交
60
#include "paddle/phi/common/place.h"
W
Wilber 已提交
61
#include "paddle/phi/core/enforce.h"
62 63
#include "paddle/utils/string/split.h"

64
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
65 66 67 68
#include "paddle/fluid/distributed/fleet_executor/fleet_executor.h"
#include "paddle/fluid/distributed/fleet_executor/fleet_executor_desc.pb.h"
#include "paddle/fluid/distributed/fleet_executor/task_node.h"
#endif
T
tensor-tang 已提交
69

70 71 72 73
#ifdef PADDLE_WITH_MKLML
#include "paddle/fluid/platform/dynload/mklml.h"
#endif

74 75 76 77
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/inference/api/mkldnn_quantizer.h"
#endif

78 79 80 81
#ifdef PADDLE_WITH_ONNXRUNTIME
#include "paddle/fluid/inference/api/onnxruntime_predictor.h"
#endif

Y
Yan Chunwei 已提交
82 83
#if PADDLE_WITH_TENSORRT
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
84
#include "paddle/fluid/inference/tensorrt/helper.h"
85
#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
Y
Yan Chunwei 已提交
86 87
#endif

88 89 90 91
#ifdef PADDLE_WITH_IPU
#include "paddle/fluid/platform/device/ipu/paddle_ipu_handler.h"
#endif

92 93
namespace paddle {

N
nhzlx 已提交
94
using inference::Singleton;
N
nhzlx 已提交
95
#if PADDLE_WITH_TENSORRT
N
nhzlx 已提交
96 97
using inference::tensorrt::TRTCalibratorEngine;
using inference::tensorrt::TRTCalibratorEngineManager;
98
using inference::tensorrt::TRTInt8Calibrator;
N
nhzlx 已提交
99
#endif
100

101 102
int AnalysisPredictor::clone_num_ = 1;

103 104 105 106
namespace {
bool IsPersistable(const framework::VarDesc *var) {
  if (var->Persistable() &&
      var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
107 108
      var->GetType() != framework::proto::VarType::FETCH_LIST &&
      var->GetType() != framework::proto::VarType::RAW) {
109 110 111 112
    return true;
  }
  return false;
}
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149

phi::DataType ConvertPrecision(AnalysisConfig::Precision precision) {
  switch (precision) {
    case AnalysisConfig::Precision::kFloat32:
      return phi::DataType::FLOAT32;
    case AnalysisConfig::Precision::kHalf:
      return phi::DataType::FLOAT16;
    case AnalysisConfig::Precision::kBf16:
      return phi::DataType::BFLOAT16;
    case AnalysisConfig::Precision::kInt8:
      return phi::DataType::INT8;
    default:
      PADDLE_THROW(paddle::platform::errors::InvalidArgument(
          "Paddle Inference not support precision. We now only support "
          "Float32, Half, Bfloat16 and Int8"));
      return phi::DataType::FLOAT32;
  }
}

phi::Backend ConvertBackend(AnalysisConfig::Backend backend) {
  switch (backend) {
    case AnalysisConfig::Backend::kGPU:
      // NOTE: phi also support phi::Backend::GPUDNN.
      return phi::Backend::GPU;
    case AnalysisConfig::Backend::kNPU:
      return phi::Backend::NPU;
    case AnalysisConfig::Backend::kXPU:
      return phi::Backend::XPU;
    case AnalysisConfig::Backend::kCPU:
      return phi::Backend::CPU;
    default:
      PADDLE_THROW(paddle::platform::errors::InvalidArgument(
          "Paddle Inference not support backend, we now only support GPU, XPU, "
          "NPU and CPU."));
      return phi::Backend::CPU;
  }
}
150 151
}  // namespace

C
ccrrong 已提交
152 153
bool PaddleTensorToLoDTensor(const PaddleTensor &pt,
                             framework::LoDTensor *t,
154
                             const platform::Place &place) {
155
  framework::DDim ddim = phi::make_ddim(pt.shape);
156 157 158 159 160 161 162
  void *input_ptr;
  if (pt.dtype == PaddleDType::INT64) {
    input_ptr = t->mutable_data<int64_t>(ddim, place);
  } else if (pt.dtype == PaddleDType::FLOAT32) {
    input_ptr = t->mutable_data<float>(ddim, place);
  } else if (pt.dtype == PaddleDType::INT32) {
    input_ptr = t->mutable_data<int32_t>(ddim, place);
163 164
  } else if (pt.dtype == PaddleDType::FLOAT16) {
    input_ptr = t->mutable_data<float16>(ddim, place);
165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180
  } else {
    LOG(ERROR) << "unsupported feed type " << pt.dtype;
    return false;
  }

  PADDLE_ENFORCE_NOT_NULL(
      input_ptr,
      paddle::platform::errors::Fatal(
          "Cannot convert to LoDTensor because LoDTensor creation failed."));
  PADDLE_ENFORCE_NOT_NULL(
      pt.data.data(),
      paddle::platform::errors::InvalidArgument(
          "The data contained in the input PaddleTensor is illegal."));

  if (platform::is_cpu_place(place)) {
    // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
C
ccrrong 已提交
181 182
    std::memcpy(
        static_cast<void *>(input_ptr), pt.data.data(), pt.data.length());
J
jianghaicheng 已提交
183 184
  } else if (platform::is_ipu_place(place)) {
#ifdef PADDLE_WITH_IPU
C
ccrrong 已提交
185 186
    std::memcpy(
        static_cast<void *>(input_ptr), pt.data.data(), pt.data.length());
J
jianghaicheng 已提交
187 188 189 190
#else
    PADDLE_THROW(paddle::platform::errors::Fatal(
        "Not compile with WITH_IPU, should not reach here."));
#endif
191
  } else if (platform::is_gpu_place(place)) {
C
ccrrong 已提交
192 193
    PADDLE_ENFORCE_EQ(platform::is_xpu_place(place),
                      false,
194 195
                      platform::errors::InvalidArgument(
                          "Only one choice can be made between CPU and XPU."));
196
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
197
    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
L
Leo Chen 已提交
198
    auto *dev_ctx = static_cast<const phi::GPUContext *>(pool.Get(place));
199
    auto dst_gpu_place = place;
C
ccrrong 已提交
200 201 202 203 204
    memory::Copy(dst_gpu_place,
                 static_cast<void *>(input_ptr),
                 platform::CPUPlace(),
                 pt.data.data(),
                 pt.data.length(),
205 206 207 208 209
                 dev_ctx->stream());
#else
    PADDLE_THROW(paddle::platform::errors::Fatal(
        "Not compile with CUDA, should not reach here."));
#endif
210 211
  } else if (platform::is_xpu_place(place)) {
#ifdef PADDLE_WITH_XPU
212
    auto dst_xpu_place = place;
C
ccrrong 已提交
213 214 215 216 217
    memory::Copy(dst_xpu_place,
                 static_cast<void *>(input_ptr),
                 platform::CPUPlace(),
                 pt.data.data(),
                 pt.data.length());
218 219 220 221 222 223 224
#else
    PADDLE_THROW(paddle::platform::errors::Fatal(
        "Not compile with XPU, should not reach here."));
#endif
  } else {
    PADDLE_THROW(paddle::platform::errors::InvalidArgument(
        "The analysis predictor supports CPU, GPU and XPU now."));
225 226 227 228 229 230 231 232 233 234
  }
  // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
  framework::LoD lod;
  for (auto &level : pt.lod) {
    lod.emplace_back(level);
  }
  t->set_lod(lod);
  return true;
}

Y
Yan Chunwei 已提交
235
bool AnalysisPredictor::Init(
236 237
    const std::shared_ptr<framework::Scope> &parent_scope,
    const std::shared_ptr<framework::ProgramDesc> &program) {
M
minqiyang 已提交
238
  VLOG(3) << "Predictor::init()";
239 240
  if (config_.with_profile_) {
    LOG(WARNING) << "Profiler is activated, which might affect the performance";
241 242
    auto tracking_device = config_.use_gpu() ? platform::ProfilerState::kAll
                                             : platform::ProfilerState::kCPU;
T
tensor-tang 已提交
243
    platform::EnableProfiler(tracking_device);
244
  } else {
245 246
    VLOG(2) << "Profiler is deactivated, and no profiling report will be "
               "generated.";
T
tensor-tang 已提交
247 248
  }

249
  // no matter with or without MKLDNN
L
luotao1 已提交
250
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
251

252 253 254
  if (!PrepareScope(parent_scope)) {
    return false;
  }
255 256 257

  InitPlace();

258 259 260 261 262 263 264
  if (!CreateExecutor()) {
    return false;
  }
  if (!PrepareProgram(program)) {
    return false;
  }

265 266 267
  // Get the feed_target_names and fetch_target_names
  PrepareFeedFetch();

268 269 270
  // Prepare executor, create local variables.
  if (!PrepareExecutor()) {
    return true;
Y
Yan Chunwei 已提交
271
  }
272

273 274 275 276 277 278 279 280 281 282 283 284 285
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  // TODO(inference): Now only gpu with external stream support private
  // device_context.
  if (config_.use_gpu_ && config_.use_external_stream_) {
    private_context_ = true;
  }
  if (private_context_) {
    if (!status_is_cloned_) {
      predictor_stream_ = config_.GetExecStream();
    }
    // NOTE: If the external_stream equals to global_device_contexts's stream,
    // then fallback.
    auto global_stream =
L
Leo Chen 已提交
286
        static_cast<phi::GPUContext *>(
287 288 289 290 291 292
            platform::DeviceContextPool::Instance().Get(place_))
            ->stream();
    if (predictor_stream_ != global_stream) {
      InitResourceManager(predictor_stream_);
      InitDeviceContexts();
    }
Y
Yan Chunwei 已提交
293
  }
294
#endif
295 296
  return true;
}
297

298
void AnalysisPredictor::InitPlace() {
299
  if (config_.use_gpu()) {
C
ccrrong 已提交
300 301
    PADDLE_ENFORCE_EQ(config_.use_xpu(),
                      false,
302 303
                      platform::errors::InvalidArgument(
                          "Only one choice can be made between CPU and XPU."));
304
    place_ = paddle::platform::CUDAPlace(config_.gpu_device_id());
305
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
306
    if (config_.thread_local_stream_enabled()) {
W
Wilber 已提交
307 308
      LOG_FIRST_N(WARNING, 1) << "We will remove this interface in the future. "
                                 "Please use config.SetExecStream instead.";
309 310
    }
#endif
311
  } else if (config_.use_xpu()) {
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334
    if (config_.lite_engine_enabled()) {
#ifdef LITE_SUBGRAPH_WITH_XPU
      // Currently, Paddle-Lite's XPU user interface only supports the transfer
      // of Host data pointers. If it is currently used as a subgraph, execution
      // efficiency will be sacrificed, so it is temporarily set to cpu place.
      // And, the current lite engine of xpu must execute all parts of the
      // model.
      place_ = paddle::platform::CPUPlace();
#else
      PADDLE_THROW(platform::errors::Unavailable(
          "You tried to use an XPU lite engine, but Paddle was not compiled "
          "with it."));
#endif  // LITE_SUBGRAPH_WITH_XPU
    } else {
#ifdef PADDLE_WITH_XPU
      place_ = paddle::platform::XPUPlace(config_.xpu_device_id());
#else
      PADDLE_THROW(platform::errors::Unavailable(
          "You tried to use XPU forward propagation (inference without lite "
          "engine), but Paddle was not compiled "
          "with WITH_XPU."));
#endif  // PADDLE_WITH_XPU
    }
W
Wilber 已提交
335 336 337 338 339 340 341 342
  } else if (config_.use_npu()) {
#ifdef PADDLE_WITH_ASCEND_CL
    place_ = paddle::platform::NPUPlace(config_.npu_device_id());
#else
    PADDLE_THROW(platform::errors::Unavailable(
        "You tried to use NPU forward propagation, but Paddle was not compiled "
        "with WITH_ASCEND_CL."));
#endif
343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
  } else if (config_.NNAdapter().use_nnadapter) {
    if (config_.lite_engine_enabled()) {
      place_ = paddle::platform::CPUPlace();
#ifndef LITE_SUBGRAPH_WITH_NNADAPTER
      PADDLE_THROW(
          platform::errors::Unavailable("You tried to use an NNAdapter lite "
                                        "engine, but Paddle was not compiled "
                                        "with it."));
#endif  // LITE_SUBGRAPH_WITH_NNADAPTER
    } else {
      PADDLE_THROW(
          platform::errors::Unavailable("You tried to use NNadapter forward "
                                        "propagation (inference without lite "
                                        "engine), but Paddle was not compiled "
                                        "with LITE_WITH_NNADAPTER."));
    }
J
jianghaicheng 已提交
359 360 361 362 363 364 365
  } else if (config_.use_ipu()) {
#ifdef PADDLE_WITH_IPU
    place_ = paddle::platform::IPUPlace();
#else
    PADDLE_THROW(platform::errors::Unavailable(
        "You tried to use IPU forward propagation, but Paddle was not compiled "
        "with WITH_IPU."));
366 367 368 369 370 371 372 373 374
#endif
  } else if (config_.use_custom_device()) {
#ifdef PADDLE_WITH_CUSTOM_DEVICE
    place_ = paddle::platform::CustomPlace(config_.custom_device_type());
#else
    PADDLE_THROW(platform::errors::Unavailable(
        "You tried to use CustomDevice forward propagation, but Paddle was not "
        "compiled "
        "with WITH_CUSTOM_DEVICE."));
J
jianghaicheng 已提交
375
#endif
376 377 378
  } else {
    place_ = paddle::platform::CPUPlace();
  }
379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395
}

void AnalysisPredictor::InitResourceManager(void *stream) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  predictor_stream_ =
      ResourceManager::Instance().InitGPUResource(place_, stream);
#endif
}

void AnalysisPredictor::InitDeviceContexts() {
// Init GPUContext.
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  if (place_.GetType() == phi::AllocationType::GPU) {
    device_contexts_.emplace(
        place_, std::async(std::launch::deferred, [=] {
          auto *gpu_resource =
              ResourceManager::Instance().GetGPUResource(predictor_stream_);
W
Wilber 已提交
396
          auto *gpu_context = new InferGPUContext(place_);
397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417
          gpu_context->SetAllocator(
              memory::allocation::AllocatorFacade::Instance()
                  .GetAllocator(place_, gpu_resource->GetStream())
                  .get());
          gpu_context->SetPinnedAllocator(
              memory::allocation::AllocatorFacade::Instance()
                  .GetAllocator(paddle::platform::CUDAPinnedPlace())
                  .get());
          gpu_context->SetHostAllocator(
              memory::allocation::AllocatorFacade::Instance()
                  .GetAllocator(platform::CPUPlace())
                  .get());
          gpu_context->SetZeroAllocator(
              memory::allocation::AllocatorFacade::Instance()
                  .GetZeroAllocator(place_)
                  .get());
          gpu_context->SetGenerator(
              framework::DefaultCUDAGenerator(place_.GetDeviceId()).get());
          gpu_context->SetHostGenerator(framework::DefaultCPUGenerator().get());

          gpu_context->SetStream(gpu_resource->GetStream());
418
          gpu_context->SetBlasHandle(gpu_resource->GetBlasHandleCreator());
419
          gpu_context->SetBlasTensorCoreHandle(
420 421 422 423 424 425 426 427
              gpu_resource->GetBlasTensorCoreHandleCreator());
          gpu_context->SetBlasTF32Handle(
              gpu_resource->GetBlasTF32TensorCoreHandleCreator());
          gpu_context->SetDnnHandle(gpu_resource->GetDnnHandleCreator());
          gpu_context->SetSolverHandle(
              gpu_resource->GetSolverDnHandleCreator());
          gpu_context->SetSparseHandle(gpu_resource->GetSparseHandleCreator());
          gpu_context->SetEigenDevice(gpu_resource->GetGpuEigenDeviceCreator());
428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521
          gpu_context->SetComputeCapability(
              gpu_resource->GetGpuComputeCapability());
          gpu_context->SetMaxThreadsPerBlock(
              gpu_resource->GetGpuMaxThreadsPerBlock());
          gpu_context->SetMaxThreadsPerMultiProcessor(
              gpu_resource->GetGpuMaxThreadsPerMp());
          gpu_context->SetMaxGridDimSize(gpu_resource->GetGpuMaxGridDimSize());
          gpu_context->SetMultiProcessors(
              gpu_resource->GetGPUMultiProcessors());
          gpu_context->SetDriverVersion(gpu_resource->GetGpuDriverVersion());
          gpu_context->SetRuntimeVersion(gpu_resource->GetGpuRuntimeVersion());
          VLOG(1) << "thread id is " << std::this_thread::get_id()
                  << ", stream id is "
                  << reinterpret_cast<void *>(gpu_resource->GetStream())
                  << ", allotor ptr is "
                  << reinterpret_cast<void *>(
                         memory::allocation::AllocatorFacade::Instance()
                             .GetAllocator(place_, gpu_resource->GetStream())
                             .get());
          return std::unique_ptr<phi::DeviceContext>(gpu_context);
        }));
  }
#endif
  // TODO(Inference): Support other backends.
}

void *AnalysisPredictor::GetExecStream() const {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  if (place_.GetType() == phi::AllocationType::GPU) {
    if (private_context_) {
      return predictor_stream_;
    } else {
      paddle::platform::DeviceContextPool &pool =
          paddle::platform::DeviceContextPool::Instance();
      return reinterpret_cast<const phi::GPUContext *>(pool.Get(place_))
          ->stream();
    }
  } else {
    return nullptr;
  }
  return nullptr;
#else
  // TODO(inference): Support other backends.
  return nullptr;
#endif
}

const void *AnalysisPredictor::GetDeviceContexts() const {
  if (private_context_) {
    return &device_contexts_;
  } else {
    paddle::platform::DeviceContextPool &pool =
        paddle::platform::DeviceContextPool::Instance();
    const auto &dev_ctxs = pool.device_contexts();
    return &dev_ctxs;
  }
}

bool AnalysisPredictor::PrepareScope(
    const std::shared_ptr<framework::Scope> &parent_scope) {
  if (parent_scope) {
    PADDLE_ENFORCE_NOT_NULL(
        parent_scope,
        platform::errors::PreconditionNotMet(
            "Both program and parent_scope should be set in Clone mode."));
    scope_ = parent_scope;
    status_is_cloned_ = true;
  } else {
    paddle::framework::InitDevices();
    paddle::framework::InitDefaultKernelSignatureMap();
    // TODO(wilber): we need to release memory occupied by weights.
    scope_.reset(new paddle::framework::Scope());
    status_is_cloned_ = false;
  }
  sub_scope_ = &scope_->NewScope();
  return true;
}

bool AnalysisPredictor::PrepareProgram(
    const std::shared_ptr<framework::ProgramDesc> &program) {
  if (!program) {
    if (!LoadProgramDesc()) return false;
    // If not cloned, the parameters should be loaded.
    // If config_.ir_optim() is True, parameters is loaded in
    // OptimizeInferenceProgram(), but other persistable variables
    // (like RAW type var) are not created in scope.
    // If config_.ir_optim() is False, parameters is loaded in LoadParameters(),
    // still need to create other persistable variables.
    // So in both case, create persistable variables at first.
    executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);

    // if enable_ir_optim_ is false,
    // the analysis pass(op fuse, graph analysis, trt subgraph, mkldnn etc) will
    // not be executed.
522 523
    model_precision_ =
        paddle::inference::GetModelPrecision(*inference_program_);
524 525 526 527 528 529 530 531 532 533 534 535 536
    OptimizeInferenceProgram();
  } else {
    // If the program is passed from external, no need to optimize it, this
    // logic is used in the clone scenario.
    inference_program_ = program;
  }

  executor_->CreateVariables(*inference_program_, 0, false, sub_scope_);

  return true;
}

bool AnalysisPredictor::CreateExecutor() {
537 538 539
  executor_.reset(new paddle::framework::NaiveExecutor(place_));
  return true;
}
W
wenbin 已提交
540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558

static bool IsPrepareDataOptTargetOp(framework::OpDesc *op) {
  // here is prepare data optimization related bad cases:
  // let's assume an op behind conditional_block and if conditional_block
  // chooses branch 1, the op need to call prepare data. else the op don't need
  // to call prepare data. In running, if predictor chooses branch 2, then
  // optimization takes effect, later issue is followed if predictor chooses
  // branch 1, because the op lost chance to prepare data.
  std::vector<std::string> op_type = {"conditional_block_infer",
                                      "select_input"};
  for (const auto &type : op_type) {
    if (op->Type() == type) {
      return true;
    }
  }
  return false;
}

static void DisablePrepareDataOpt(
C
ccrrong 已提交
559 560
    std::shared_ptr<framework::ProgramDesc> inference_program,
    int block,
W
wenbin 已提交
561 562 563 564 565 566 567 568 569
    bool pre_disable_opt) {
  bool disable_opt = false;
  auto &infer_block = inference_program->Block(block);
  for (auto *op : infer_block.AllOps()) {
    if (disable_opt || pre_disable_opt) {
      op->SetAttr("inference_force_prepare_data", true);
    }
    if (op->HasAttr("sub_block")) {
      int blockID = op->GetBlockAttrId("sub_block");
C
ccrrong 已提交
570 571
      DisablePrepareDataOpt(
          inference_program, blockID, disable_opt || pre_disable_opt);
W
wenbin 已提交
572 573
    }
    // disable prepare data if unfriendly op is found
W
wenbin 已提交
574 575 576
    if (!disable_opt) {
      disable_opt = IsPrepareDataOptTargetOp(op);
    }
W
wenbin 已提交
577 578 579
  }
}

580
bool AnalysisPredictor::PrepareExecutor() {
581
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
582 583 584 585 586
  if (config_.dist_config().use_dist_model()) {
    VLOG(3) << "use_dist_model is enabled, will init FleetExecutor.";
    return PrepareFleetExecutor();
  }
#endif
W
wenbin 已提交
587 588
  DisablePrepareDataOpt(inference_program_, 0, false);

C
ccrrong 已提交
589 590
  executor_->Prepare(
      sub_scope_, *inference_program_, 0, config_.use_feed_fetch_ops_);
591

592 593 594
  PADDLE_ENFORCE_NOT_NULL(sub_scope_,
                          platform::errors::PreconditionNotMet(
                              "The sub_scope should not be nullptr."));
Y
Yan Chunwei 已提交
595

596 597 598
  return true;
}

599
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635
bool AnalysisPredictor::PrepareFleetExecutor() {
  VLOG(3) << "AnalysisPredictor::PrepareFleetExecutor()";
  if (config_.dist_config().nranks() > 1 && !CommInit()) {
    return false;
  }
  task_node_.reset(new distributed::TaskNode(inference_program_.get(),
                                             config_.dist_config().rank()));
  // With auto cut, there is no concept of pp, no need to add dependency.
  task_node_->SetType("Compute");
  task_node_->Init(config_.use_feed_fetch_ops_enabled());
  executor_desc_ = distributed::FleetExecutorDesc();
  executor_desc_.set_cur_rank(config_.dist_config().rank());
  std::unordered_map<int64_t, int64_t> id_to_rank;
  for (int i = 0; i < config_.dist_config().nranks(); ++i) {
    distributed::RankInfo *rank_info = executor_desc_.add_cluster_info();
    rank_info->set_rank(i);
    rank_info->set_ip_port(config_.dist_config().trainer_endpoints()[i]);
    id_to_rank.insert({i, i});
  }
  fleet_exe_.reset(new distributed::FleetExecutor(executor_desc_));
  // NOTE: Vars of feed fetch ops are not persistable,
  // which will result in that those vars will be created in
  // the subscope (microscope) in fleet executor. This will
  // cause that the GetInputTensor/GetOutputTensor funct
  // in analysis predictor cannot find those vars in the scope
  // returned by the DistModel, since DistModel only return the
  // root scope. So, those vars must  to be created in the root
  // scope instead of in the microscope
  std::vector<std::string> feed_fetch_vars;
  for (auto pair : idx2feeds_) {
    feed_fetch_vars.emplace_back(pair.second);
  }
  for (auto pair : idx2fetches_) {
    feed_fetch_vars.emplace_back(pair.second);
  }
  fleet_exe_->Init(config_.dist_config().carrier_id(),
C
ccrrong 已提交
636 637 638 639 640 641 642
                   *(inference_program_.get()),
                   scope_.get(),
                   place_,
                   1,
                   {task_node_.get()},
                   id_to_rank,
                   feed_fetch_vars);
643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678
  return true;
}

bool AnalysisPredictor::CommInit() {
  std::map<int64_t, std::vector<int64_t>> ring_id_to_ranks{};
  std::map<int64_t, std::vector<int64_t>> rank_to_ring_ids{};
  if (!LoadConverterConfig(&ring_id_to_ranks, &rank_to_ring_ids)) {
    VLOG(3) << "Load converter config failed, DistModel init failed.";
    return false;
  }
  std::unique_ptr<framework::ProgramDesc> comm_init_program(
      new framework::ProgramDesc());
  framework::BlockDesc *comm_init_block = comm_init_program->MutableBlock(0);
  std::vector<int64_t> &ring_ids =
      rank_to_ring_ids[config_.dist_config().rank()];
  int64_t order = 0;
  std::string var_name_base = "comm_init_";
  for (int64_t ring_id : ring_ids) {
    VLOG(3) << "Init comm for ring id: " << ring_id;
    int64_t ranks_in_group = ring_id_to_ranks[ring_id].size();
    int64_t rank_in_group = 0;
    std::vector<int64_t> &ranks = ring_id_to_ranks[ring_id];
    for (int64_t rank : ranks) {
      if (config_.dist_config().rank() == rank) {
        break;
      }
      rank_in_group += 1;
    }
    std::vector<std::string> peer_endpoints;
    for (int64_t rank : ranks) {
      if (config_.dist_config().rank() == rank) {
        continue;
      }
      peer_endpoints.emplace_back(
          config_.dist_config().trainer_endpoints()[rank]);
    }
C
ccrrong 已提交
679 680 681 682 683 684
    InsertCommOp(var_name_base + std::to_string(order),
                 ranks_in_group,
                 rank_in_group,
                 peer_endpoints,
                 comm_init_block,
                 ring_id);
685 686 687 688 689 690 691 692 693 694 695
    order += 1;
  }
  framework::NaiveExecutor e(place_);
  e.CreateVariables(*comm_init_program, 0, true, scope_.get());
  e.Prepare(scope_.get(), *comm_init_program, 0, false);
  e.Run();
  VLOG(3) << "Comm init successful.";
  return true;
}

void AnalysisPredictor::InsertCommOp(
C
ccrrong 已提交
696 697 698 699 700
    std::string tmp_var_name,
    int nranks,
    int rank,
    const std::vector<std::string> &peer_endpoints,
    framework::BlockDesc *block,
701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756
    int ring_id) {
  /*
   * tmp_var_name: the var name for var comm_id
   * nranks: number of total ranks
   * rank: the rank of local rank in the comm group
   * peer_endpoints: peer's endpoints
   * block: the block where to insert the comm ops
   * ring_id: the ring_id to be inited
   */
  const std::string &endpoint = config_.dist_config().current_endpoint();
  std::stringstream ss;
  ss << "Init comm with tmp var: " << tmp_var_name
     << ". The ring id is: " << ring_id << ". The group has: " << nranks
     << " ranks. Current rank in the group is: " << rank
     << ". The endpoint is: " << endpoint << ". Peer endpoints are: ";
  for (auto ep : peer_endpoints) {
    ss << ep << ", ";
  }
  VLOG(3) << ss.str();
  if (config_.use_gpu()) {
    framework::VarDesc *new_var = block->Var(tmp_var_name);
    new_var->SetType(framework::proto::VarType::RAW);
    new_var->SetPersistable(true);
    framework::OpDesc *gen_nccl_id_op = block->AppendOp();
    gen_nccl_id_op->SetType("c_gen_nccl_id");
    gen_nccl_id_op->SetOutput("Out", {tmp_var_name});
    gen_nccl_id_op->SetAttr("rank", rank);
    gen_nccl_id_op->SetAttr("endpoint",
                            config_.dist_config().current_endpoint());
    gen_nccl_id_op->SetAttr("other_endpoints", peer_endpoints);
    gen_nccl_id_op->SetAttr("ring_id", ring_id);
    gen_nccl_id_op->SetAttr("op_role",
                            static_cast<int>(framework::OpRole::kForward));
    gen_nccl_id_op->CheckAttrs();
    framework::OpDesc *comm_init_op = block->AppendOp();
    comm_init_op->SetType("c_comm_init");
    comm_init_op->SetInput("X", {tmp_var_name});
    comm_init_op->SetAttr("rank", rank);
    comm_init_op->SetAttr("nranks", nranks);
    comm_init_op->SetAttr("ring_id", ring_id);
    comm_init_op->SetAttr("op_role",
                          static_cast<int>(framework::OpRole::kForward));
    comm_init_op->CheckAttrs();
  } else {
    LOG(WARNING) << "DistModelInf doesn't init comm.";
    // TODO(fleet exe dev): comm init for more devices
  }
}

bool AnalysisPredictor::LoadConverterConfig(
    std::map<int64_t, std::vector<int64_t>> *ring_id_to_ranks,
    std::map<int64_t, std::vector<int64_t>> *rank_to_ring_ids) {
  VLOG(3) << "Going to load converter config from: "
          << config_.dist_config().comm_init_config() << "\n";
  std::ifstream fin(config_.dist_config().comm_init_config(), std::ios::in);
  PADDLE_ENFORCE_EQ(
C
ccrrong 已提交
757 758
      static_cast<bool>(fin.is_open()),
      true,
759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830
      platform::errors::NotFound(
          "Cannot open file %s, please confirm whether the file is normal.",
          config_.dist_config().comm_init_config()));
  std::string line;
  bool ring_to_rank{true};
  // Reading config from file, the config file should like these format
  //  [ring_id -> ranks]
  //  0,0,1,2,3
  //  1,0,1
  //  2,2,3
  //  21,0,1
  //  22,1,2
  //  23,2,3
  //  [rank -> ring_ids]
  //  0,0,1,21
  //  1,0,1,21,22
  //  2,0,2,22,23
  //  3,0,2,23
  while (std::getline(fin, line)) {
    std::vector<std::string> one_line = paddle::string::Split(line, ',');
    if (one_line.size() == 1) {
      // start a new section of the config
      if (line == "[ring_id -> ranks]") {
        ring_to_rank = true;
      } else if (line == "[rank -> ring_ids]") {
        ring_to_rank = false;
      }
    } else {
      // parse key - values pairs in one section
      int64_t key = std::stoll(one_line[0]);
      for (size_t i = 1; i < one_line.size(); ++i) {
        int64_t val = std::stoll(one_line[i]);
        if (ring_to_rank) {
          if (ring_id_to_ranks->find(key) == ring_id_to_ranks->end()) {
            ring_id_to_ranks->insert({key, std::vector<int64_t>()});
          }
          ring_id_to_ranks->at(key).emplace_back(val);
        } else {
          if (rank_to_ring_ids->find(key) == rank_to_ring_ids->end()) {
            rank_to_ring_ids->insert({key, std::vector<int64_t>()});
          }
          rank_to_ring_ids->at(key).emplace_back(val);
        }
        // NOTE: add more configuration sections here
      }
    }
  }
  std::stringstream ss;
  ss << "Loaded the following converter config:\n";
  ss << "ring_id_to_ranks:\n";
  for (auto pair : *ring_id_to_ranks) {
    int64_t key = pair.first;
    ss << "\t" << key << "\t->\t";
    for (auto value : pair.second) {
      ss << value << "\t";
    }
    ss << "\n";
  }
  ss << "rank_to_ring_ids:\n";
  for (auto pair : *rank_to_ring_ids) {
    int64_t key = pair.first;
    ss << "\t" << key << "\t->\t";
    for (auto value : pair.second) {
      ss << value << "\t";
    }
    ss << "\n";
  }
  VLOG(3) << ss.str();
  return true;
}
#endif

831 832
void AnalysisPredictor::MkldnnPreSet(const std::vector<PaddleTensor> &inputs) {
#ifdef PADDLE_WITH_MKLDNN
W
Wilber 已提交
833 834 835 836 837 838 839 840 841 842 843 844
  std::vector<std::vector<int>> inputs_shape;
  for (size_t i = 0; i < inputs.size(); ++i) {
    inputs_shape.emplace_back(inputs[i].shape);
  }
  MkldnnPreSet(inputs_shape);
#endif
}

void AnalysisPredictor::MkldnnPreSet(
    const std::vector<std::vector<int>> &inputs_shape) {
#ifdef PADDLE_WITH_MKLDNN
  VLOG(2) << "AnalysisPredictor::ZeroCopyRun get_cur_mkldnn_session_id="
845
          << platform::MKLDNNDeviceContext::tls().get_cur_mkldnn_session_id();
846 847 848
  // In cache clearing mode.
  if (config_.mkldnn_cache_capacity_ > 0) {
    VLOG(2) << "In mkldnn cache clear mode.";
849 850 851
    platform::MKLDNNDeviceContext::tls().set_cur_mkldnn_session_id(
        platform::MKLDNNDeviceContextThreadLocals::
            kMKLDNNSessionID_CacheClearing);
852 853
    // Set current_input_shape for caching dynamic shape.
    std::stringstream ss;
W
Wilber 已提交
854 855 856
    for (size_t i = 0; i < inputs_shape.size(); ++i) {
      for (size_t j = 0; j < inputs_shape[i].size(); ++j) {
        ss << inputs_shape[i][j] << "-";
857 858 859
      }
    }
    VLOG(2) << "Set input shape=" << ss.str();
860
    platform::MKLDNNDeviceContext::tls().set_cur_input_shape_str(ss.str());
861
  }
862 863 864
  platform::MKLDNNDeviceContext::tls().set_cur_input_shape_cache_capacity(
      config_.mkldnn_cache_capacity_);

865 866 867 868 869 870
#endif
}

void AnalysisPredictor::MkldnnPostReset() {
#ifdef PADDLE_WITH_MKLDNN
  // In cache clearing mode.
871 872 873 874
  if (config_.mkldnn_cache_capacity_ > 0 &&
      static_cast<platform::MKLDNNDeviceContext *>(
          (&platform::DeviceContextPool::Instance())->Get(platform::CPUPlace()))
              ->GetCachedObjectsNumber() > 0) {
875 876 877 878 879 880 881 882
    if (VLOG_IS_ON(2)) {
      auto shape_blob_size = static_cast<platform::MKLDNNDeviceContext *>(
                                 (&platform::DeviceContextPool::Instance())
                                     ->Get(platform::CPUPlace()))
                                 ->GetShapeBlobSize();
      CHECK_LE(shape_blob_size,
               static_cast<size_t>(config_.mkldnn_cache_capacity_));
    }
883 884 885
    // We cannot reset to the default cache settings
    // as there maybe CopyToCPU method used and oneDNN
    // primitives are used there so cache would grow
886 887 888 889
  }
#endif
}

890 891 892
bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
                            std::vector<PaddleTensor> *output_data,
                            int batch_size) {
893
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
894 895 896
#ifdef PADDLE_WITH_MKLDNN
  if (config_.use_mkldnn_) MkldnnPreSet(inputs);
#endif
M
minqiyang 已提交
897
  VLOG(3) << "Predictor::predict";
898 899 900 901
  inference::Timer timer;
  timer.tic();
  // set feed variable
  framework::Scope *scope = sub_scope_ ? sub_scope_ : scope_.get();
C
ccrrong 已提交
902 903 904
  PADDLE_ENFORCE_NOT_NULL(
      scope,
      platform::errors::PreconditionNotMet("The scope should not be nullptr."));
905 906
  if (!SetFeed(inputs, scope)) {
    LOG(ERROR) << "fail to set feed";
Y
Yan Chunwei 已提交
907
    return false;
908
  }
M
Michal Gallus 已提交
909

910 911 912 913 914 915 916 917 918
#ifdef PADDLE_WITH_TENSORRT
  if (config_.tensorrt_engine_enabled()) {
    inference::tensorrt::TensorRTEngine::predictor_id_per_thread =
        predictor_id_;
    VLOG(3) << "thread_local var predictor_id in TensorRTEngine is set to: "
            << inference::tensorrt::TensorRTEngine::predictor_id_per_thread;
  }
#endif

919 920 921
  // Run the inference program
  // if share variables, we need not create variables
  executor_->Run();
922

923 924 925 926
  // get fetch variable
  if (!GetFetch(output_data, scope)) {
    LOG(ERROR) << "fail to get fetches";
    return false;
T
tensor-tang 已提交
927
  }
Y
Yan Chunwei 已提交
928

M
minqiyang 已提交
929
  VLOG(3) << "predict cost: " << timer.toc() << "ms";
Y
Yan Chunwei 已提交
930

Y
Yan Chunwei 已提交
931 932 933 934 935
  // 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.
936 937 938
  if (sub_scope_) {
    tensor_array_batch_cleaner_.CollectNoTensorVars(sub_scope_);
  }
Y
Yan Chunwei 已提交
939
  tensor_array_batch_cleaner_.ResetNoTensorVars();
940 941 942 943

  // recover the cpu_math_library_num_threads to 1, in order to avoid thread
  // conflict when integrating it into deployment service.
  paddle::platform::SetNumThreads(1);
944 945
#ifdef PADDLE_WITH_MKLDNN
  if (config_.use_mkldnn_) MkldnnPostReset();
T
Tao Luo 已提交
946
#endif
947
#if defined(PADDLE_WITH_MKLML)
T
Tao Luo 已提交
948 949 950 951
  // Frees unused memory allocated by the Intel® MKL Memory Allocator to
  // avoid memory leak. See:
  // https://software.intel.com/en-us/mkl-developer-reference-c-mkl-free-buffers
  platform::dynload::MKL_Free_Buffers();
952
#endif
953 954
  return true;
}
955

956 957
bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                                framework::Scope *scope) {
M
minqiyang 已提交
958
  VLOG(3) << "Predictor::set_feed";
959 960 961 962 963 964 965 966 967 968
  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) {
969 970
    framework::LoDTensor *input = &feed_tensors_[i];
    if (!PaddleTensorToLoDTensor(inputs[i], input, place_)) {
971 972 973
      return false;
    }
    int idx = -1;
974
    if (config_.specify_input_name_) {
T
tensor-tang 已提交
975 976
      auto name = inputs[i].name;
      if (feed_names_.find(name) == feed_names_.end()) {
T
tensor-tang 已提交
977 978
        LOG(ERROR) << "feed names from program do not have name: [" << name
                   << "] from specified input";
T
tensor-tang 已提交
979 980
      }
      idx = feed_names_[name];
981
    } else {
R
Ruibiao Chen 已提交
982
      idx = PADDLE_GET_CONST(int, feeds_[i]->GetAttr("col"));
983
    }
984
    framework::SetFeedVariable(scope, *input, "feed", idx);
985 986 987 988 989 990 991 992
  }
  return true;
}

template <typename T>
void AnalysisPredictor::GetFetchOne(const framework::LoDTensor &fetch,
                                    PaddleTensor *output) {
  // set shape.
993
  auto shape = phi::vectorize(fetch.dims());
994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010
  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
1011
  VLOG(3) << "Predictor::get_fetch";
Y
Yan Chunwei 已提交
1012 1013
  outputs->resize(fetches_.size());
  for (size_t i = 0; i < fetches_.size(); ++i) {
R
Ruibiao Chen 已提交
1014
    int idx = PADDLE_GET_CONST(int, fetches_[i]->GetAttr("col"));
1015
    PADDLE_ENFORCE_EQ(
C
ccrrong 已提交
1016 1017
        static_cast<size_t>(idx),
        i,
1018
        platform::errors::InvalidArgument(
C
ccrrong 已提交
1019 1020
            "Fetch op's col attr(%d) should be equal to the index(%d)",
            idx,
1021
            i));
1022
    framework::FetchType &fetch_var =
1023
        framework::GetFetchVariable(*scope, "fetch", idx);
R
Ruibiao Chen 已提交
1024
    auto &fetch = PADDLE_GET(framework::LoDTensor, fetch_var);
1025
    auto type = framework::TransToProtoVarType(fetch.dtype());
1026
    auto output = &(outputs->at(i));
Y
Yan Chunwei 已提交
1027
    output->name = fetches_[idx]->Input("X")[0];
Y
Yu Yang 已提交
1028
    if (type == framework::proto::VarType::FP32) {
1029 1030
      GetFetchOne<float>(fetch, output);
      output->dtype = PaddleDType::FLOAT32;
Y
Yu Yang 已提交
1031
    } else if (type == framework::proto::VarType::INT64) {
1032 1033
      GetFetchOne<int64_t>(fetch, output);
      output->dtype = PaddleDType::INT64;
1034 1035 1036
    } else if (type == framework::proto::VarType::INT32) {
      GetFetchOne<int32_t>(fetch, output);
      output->dtype = PaddleDType::INT32;
1037 1038 1039
    } else if (type == framework::proto::VarType::FP16) {
      GetFetchOne<float16>(fetch, output);
      output->dtype = PaddleDType::FLOAT16;
1040
    } else {
1041 1042
      LOG(ERROR) << "unknown type, only support float32, float16, int64 and "
                    "int32 now.";
1043 1044
    }
  }
Y
Yan Chunwei 已提交
1045 1046
  return true;
}
1047

1048
void AnalysisPredictor::PrepareArgument() {
1049
  argument_.SetUseGPU(config_.use_gpu());
1050
  argument_.SetUseFcPadding(config_.use_fc_padding());
1051
  argument_.SetGPUDeviceId(config_.gpu_device_id());
1052
  argument_.SetEnableAnalysisOptim(config_.enable_ir_optim_);
1053 1054 1055 1056 1057 1058 1059 1060
  if (model_precision_ == phi::DataType::FLOAT32) {
    argument_.SetEnableMemoryOptim(config_.enable_memory_optim());
  } else {
    // TODO(inference): mixed precision temporarily not support memory_optim
    LOG_FIRST_N(WARNING, 1) << "mixed precision model temporarily not support "
                               "memory optim, so we just turn off that.";
    argument_.SetEnableMemoryOptim(false);
  }
T
Tao Luo 已提交
1061
  argument_.SetModelFromMemory(config_.model_from_memory_);
Y
Yan Chunwei 已提交
1062
  // Analyze inference_program
1063
  argument_.SetPredictorID(predictor_id_);
1064
  argument_.SetOptimCacheDir(config_.opt_cache_dir_);
1065 1066
  if (!config_.model_dir().empty()) {
    argument_.SetModelDir(config_.model_dir());
T
Tao Luo 已提交
1067
  } else {
C
ccrrong 已提交
1068 1069
    PADDLE_ENFORCE_EQ(config_.prog_file().empty(),
                      false,
1070 1071
                      platform::errors::PreconditionNotMet(
                          "Either model_dir or prog_file should be set."));
N
nhzlx 已提交
1072
    std::string dir = inference::analysis::GetDirRoot(config_.prog_file());
N
nhzlx 已提交
1073

1074 1075
    argument_.SetModelProgramPath(config_.prog_file());
    argument_.SetModelParamsPath(config_.params_file());
Y
Yan Chunwei 已提交
1076
  }
1077

1078
  argument_.SetTensorRtPrecisionMode(config_.tensorrt_precision_mode_);
1079
  argument_.SetTensorRtUseOSS(config_.trt_use_varseqlen_);
1080
  argument_.SetTensorRtWithInterleaved(config_.trt_with_interleaved_);
1081 1082
  argument_.SetTensorRtTransformerPosid(config_.tensorrt_transformer_posid_);
  argument_.SetTensorRtTransformerMaskid(config_.tensorrt_transformer_maskid_);
1083 1084 1085 1086 1087
  argument_.SetMinInputShape(config_.min_input_shape_);
  argument_.SetMaxInputShape(config_.max_input_shape_);
  argument_.SetOptimInputShape(config_.optim_input_shape_);
  argument_.SetTensorRtTunedDynamicShape(
      config_.tuned_tensorrt_dynamic_shape());
1088
  if (config_.use_gpu() && config_.tensorrt_engine_enabled()) {
Y
Yan Chunwei 已提交
1089
    LOG(INFO) << "TensorRT subgraph engine is enabled";
1090 1091 1092
    argument_.SetUseTensorRT(true);
    argument_.SetTensorRtWorkspaceSize(config_.tensorrt_workspace_size_);
    argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
1093
    argument_.SetTensorRtMinSubgraphSize(config_.tensorrt_min_subgraph_size_);
1094
    argument_.SetTensorRtDisabledOPs(config_.trt_disabled_ops_);
1095 1096
    argument_.SetTensorRtUseDLA(config_.trt_use_dla_);
    argument_.SetTensorRtDLACore(config_.trt_dla_core_);
N
nhzlx 已提交
1097
    argument_.SetTensorRtUseStaticEngine(config_.trt_use_static_engine_);
1098
    argument_.SetTensorRtUseCalibMode(config_.trt_use_calib_mode_);
1099
    argument_.SetCloseTrtPluginFp16(config_.disable_trt_plugin_fp16_);
1100 1101 1102
    argument_.SetTensorRtShapeRangeInfoPath(config_.shape_range_info_path());
    argument_.SetTensorRtAllowBuildAtRuntime(
        config_.trt_allow_build_at_runtime());
1103
    argument_.SetTensorRtUseInspector(config_.trt_use_inspector_);
W
Wojciech Uss 已提交
1104
  }
1105

D
denglin-github 已提交
1106 1107 1108 1109
  if (config_.dlnne_enabled()) {
    LOG(INFO) << "Dlnne subgraph is enabled";
    argument_.SetUseDlnne(true);
    argument_.SetDlnneMinSubgraphSize(config_.dlnne_min_subgraph_size_);
D
denglin-github 已提交
1110 1111 1112 1113 1114 1115 1116 1117
    argument_.SetDlnneMaxBatchSize(config_.dlnne_max_batchsize_);
    argument_.SetDlnneUseStaticBatch(config_.dlnne_use_static_batch_);
    argument_.SetDlnneWeightShareMode(config_.dlnne_weight_share_mode_);
    argument_.SetDlnneDisableNodesByOutputs(
        config_.dlnne_disable_nodes_by_outputs_);
    argument_.SetDlnneInputShapeDict(config_.dlnne_input_shape_dict_);
    argument_.SetDlnneUseCalibMode(config_.dlnne_use_calib_mode_);
    argument_.SetDlnnePrecisionMode(config_.dlnne_precision_mode_);
D
denglin-github 已提交
1118 1119
  }

石晓伟 已提交
1120
  if (config_.lite_engine_enabled()) {
W
Wilber 已提交
1121 1122
    argument_.SetCpuMathLibraryNumThreads(
        config_.cpu_math_library_num_threads());
石晓伟 已提交
1123 1124 1125
    argument_.SetLitePrecisionMode(config_.lite_precision_mode_);
    argument_.SetLitePassesFilter(config_.lite_passes_filter_);
    argument_.SetLiteOpsFilter(config_.lite_ops_filter_);
1126 1127 1128
    argument_.SetLiteZeroCopy(config_.lite_zero_copy_);
    argument_.SetUseXpu(config_.use_xpu_);
    argument_.SetXpuL3WorkspaceSize(config_.xpu_l3_workspace_size_);
W
Wilber 已提交
1129 1130 1131 1132 1133
    argument_.SetXpuLocked(config_.xpu_locked_);
    argument_.SetXpuAutotune(config_.xpu_autotune_);
    argument_.SetXpuAutotuneFile(config_.xpu_autotune_file_);
    argument_.SetXpuPrecision(config_.xpu_precision_);
    argument_.SetXpuAdaptiveSeqlen(config_.xpu_adaptive_seqlen_);
1134
    argument_.SetXpuDeviceId(config_.xpu_device_id_);
1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154
    // NNAdapter related
    argument_.SetUseNNAdapter(config_.NNAdapter().use_nnadapter);
    argument_.SetNNAdapterDeviceNames(
        config_.NNAdapter().nnadapter_device_names);
    argument_.SetNNAdapterContextProperties(
        config_.NNAdapter().nnadapter_context_properties);
    argument_.SetNNAdapterModelCacheDir(
        config_.NNAdapter().nnadapter_model_cache_dir);
    argument_.SetNNAdapterSubgraphPartitionConfigBuffer(
        config_.NNAdapter().nnadapter_subgraph_partition_config_buffer);
    argument_.SetNNAdapterSubgraphPartitionConfigPath(
        config_.NNAdapter().nnadapter_subgraph_partition_config_path);
    std::vector<std::string> buffer_keys;
    std::vector<std::vector<char>> buffer_vals;
    for (auto it : config_.NNAdapter().nnadapter_model_cache_buffers) {
      buffer_keys.emplace_back(it.first);
      buffer_vals.emplace_back(it.second);
    }
    argument_.SetNNAdapterModelCacheToken(buffer_keys);
    argument_.SetNNAdapterModelCacheBuffer(buffer_vals);
石晓伟 已提交
1155 1156 1157
    LOG(INFO) << "Lite subgraph engine is enabled";
  }

1158
#ifdef PADDLE_WITH_IPU
J
jianghaicheng 已提交
1159 1160
  argument_.SetUseIpu(config_.use_ipu_);
  argument_.SetIpuDeviceNum(config_.ipu_device_num());
1161
  argument_.SetIpuMicroBatchSize(config_.ipu_micro_batch_size_);
J
jianghaicheng 已提交
1162 1163
  argument_.SetIpuEnablePipelining(config_.ipu_enable_pipelining_);
  argument_.SetIpuBatchesPerStep(config_.ipu_batches_per_step_);
1164 1165 1166 1167 1168 1169
  argument_.SetIpuEnableFp16(config_.ipu_enable_fp16_);
  argument_.SetIpuReplicaNum(config_.ipu_replica_num_);
  argument_.SetIpuAvailableMemoryProportion(
      config_.ipu_available_memory_proportion_);
  argument_.SetIpuEnableHalfPartial(config_.ipu_enable_half_partial_);
#endif
J
jianghaicheng 已提交
1170

1171 1172 1173
  argument_.SetUseNpu(config_.use_npu_);
  argument_.SetNPUDeviceId(config_.npu_device_id());

1174
  if (config_.use_mkldnn_) {
Y
Yan Chunwei 已提交
1175
    LOG(INFO) << "MKLDNN is enabled";
1176 1177 1178
    argument_.SetMKLDNNEnabledOpTypes(config_.mkldnn_enabled_op_types_);
  }

1179 1180 1181 1182 1183 1184 1185 1186
#ifdef PADDLE_WITH_MKLDNN
  if (config_.mkldnn_quantizer_enabled()) {
    LOG(INFO) << "Quantization is enabled";
    argument_.SetQuantizeEnabledOpTypes(
        config_.mkldnn_quantizer_config()->enabled_op_types());
    argument_.SetQuantizeExcludedOpIds(
        config_.mkldnn_quantizer_config()->excluded_op_ids());
  }
1187 1188 1189 1190
  if (config_.use_mkldnn_bfloat16_) {
    LOG(INFO) << "Bfloat16 is enabled";
    argument_.SetBfloat16EnabledOpTypes(config_.bfloat16_enabled_op_types_);
  }
B
baoachun 已提交
1191 1192 1193 1194 1195 1196

  if (config_.use_mkldnn_int8_) {
    LOG(INFO) << "Int8 is enabled";
    argument_.SetQuantizeEnabledOpTypes(config_.quantize_enabled_op_types_);
    argument_.SetQuantizeExcludedOpIds(config_.quantize_excluded_op_ids_);
    argument_.SetQuantVarScales({});
1197
    argument_.SetCalibrationFilePath(config_.calibration_file_path_);
B
baoachun 已提交
1198
  }
1199 1200
#endif

1201
  auto passes = config_.pass_builder()->AllPasses();
1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235
  if (model_precision_ != phi::DataType::FLOAT32) {
    LOG(INFO) << "Model is mixed precision type with " << model_precision_
              << ", we will use a new PassStrategy. Note that only the GPU "
                 "backend is supported for now.";
    passes.clear();
    if (config_.tensorrt_engine_enabled()) {
      for (const auto &pass : kTrtLowerPrecisionPasses) {
        passes.push_back(pass);
      }
    } else if (config_.use_gpu()) {
      for (const auto &pass : kGpuLowerPrecisionPasses) {
        passes.push_back(pass);
      }
    }

    const auto &deleted_passes = config_.pass_builder()->GetAllDeletedPasses();
    for (const auto &it : deleted_passes) {
      auto iterator = std::find(passes.begin(), passes.end(), it);
      if (iterator != passes.end()) {
        passes.erase(iterator);
      }
    }

    if (config_.ir_debug_) {
      auto it = std::begin(passes);
      while (it != std::end(passes)) {
        if (*it != "graph_viz_pass") {
          it = passes.insert(it + 1, "graph_viz_pass");
        } else {
          ++it;
        }
      }
    }
  }
Y
Yan Chunwei 已提交
1236 1237 1238 1239
  if (!config_.ir_optim()) {
    passes.clear();
    LOG(INFO) << "ir_optim is turned off, no IR pass will be executed";
  }
1240
  argument_.SetDisableLogs(config_.glog_info_disabled());
1241
  argument_.SetIrAnalysisPasses(passes);
Y
Yan Chunwei 已提交
1242
  argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses());
1243
  argument_.SetScopeNotOwned(scope_.get());
1244

1245
  // mixed precison.
1246
  argument_.SetModelPrecision(static_cast<int>(model_precision_));
1247
  argument_.SetMixedBlackList(config_.mixed_black_list_);
1248 1249 1250 1251 1252
}

// NOTE All the members in AnalysisConfig should be copied to Argument.
void AnalysisPredictor::OptimizeInferenceProgram() {
  PrepareArgument();
1253 1254
  Analyzer().Run(&argument_);

1255
  PADDLE_ENFORCE_EQ(
C
ccrrong 已提交
1256 1257
      argument_.scope_valid(),
      true,
1258
      platform::errors::InvalidArgument("The argument scope should be valid."));
1259 1260
  VLOG(5) << "to prepare executor";
  ARGUMENT_CHECK_FIELD((&argument_), ir_analyzed_program);
Y
Yan Chunwei 已提交
1261
  inference_program_.reset(
1262 1263 1264 1265 1266
      new framework::ProgramDesc(argument_.ir_analyzed_program()),
      [](framework::ProgramDesc *prog) {
// Note, please do NOT use any member variables, because member variables may
// have been destructed in multiple threads.
#if PADDLE_WITH_TENSORRT
W
Wilber 已提交
1267 1268 1269 1270
        auto &block = prog->Block(0);
        for (auto &op_desc : block.AllOps()) {
          if (op_desc->Type() == "tensorrt_engine") {
            std::string engine_key =
R
Ruibiao Chen 已提交
1271
                PADDLE_GET_CONST(std::string, op_desc->GetAttr("engine_key"));
W
Wilber 已提交
1272
            int engine_predictor_id =
R
Ruibiao Chen 已提交
1273
                PADDLE_GET_CONST(int, op_desc->GetAttr("predictor_id"));
W
Wilber 已提交
1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284
            std::string engine_name =
                engine_key + std::to_string(engine_predictor_id);
            if (paddle::inference::Singleton<
                    inference::tensorrt::TRTEngineManager>::Global()
                    .Has(engine_name)) {
              paddle::inference::Singleton<
                  inference::tensorrt::TRTEngineManager>::Global()
                  .DeleteKey(engine_name);
            }
          }
        }
1285 1286 1287
#endif
        delete prog;
      });
1288 1289 1290 1291
  // The config and argument take a lot of storage,
  // when the predictor settings are complete, we release these stores.
  argument_.PartiallyRelease();
  config_.PartiallyRelease();
1292
  LOG(INFO) << "======= optimize end =======";
Y
Yan Chunwei 已提交
1293
}
1294 1295

template <>
1296 1297 1298
std::unique_ptr<PaddlePredictor>
CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
    const AnalysisConfig &config) {
W
Wilber 已提交
1299 1300
  // TODO(NHZlX): Should add the link to the doc of
  // paddle_infer::CreatePredictor<paddle_infer::Config>
P
Pei Yang 已提交
1301 1302 1303 1304
  if (config.glog_info_disabled()) {
    FLAGS_logtostderr = 1;
    FLAGS_minloglevel = 2;  // GLOG_ERROR
  }
M
minqiyang 已提交
1305
  VLOG(3) << "create AnalysisConfig";
1306
  PADDLE_ENFORCE_EQ(
C
ccrrong 已提交
1307 1308
      config.is_valid(),
      true,
1309 1310
      platform::errors::InvalidArgument(
          "Note: Each config can only be used for one predictor."));
1311

1312 1313 1314 1315
  // Register custom operators compiled by the user.
  // This function can only be executed once per process.
  static std::once_flag custom_operators_registered;
  std::call_once(custom_operators_registered,
1316
                 []() { inference::RegisterAllCustomOperator(); });
1317

1318
  if (config.use_gpu()) {
1319 1320 1321 1322 1323 1324
    static std::once_flag gflags_initialized;
    static bool process_level_allocator_enabled;

    std::call_once(gflags_initialized, [&]() {
      std::vector<std::string> gflags;
      PADDLE_ENFORCE_GE(
C
ccrrong 已提交
1325 1326
          config.memory_pool_init_size_mb(),
          0.f,
1327 1328 1329
          platform::errors::InvalidArgument(
              "The size of memory pool should be greater than 0."));
      PADDLE_ENFORCE_GE(
C
ccrrong 已提交
1330 1331
          config.gpu_device_id(),
          0,
1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344
          platform::errors::InvalidArgument(
              "Invalid device id (%d). The device id should be greater than 0.",
              config.gpu_device_id()));
      gflags.push_back("dummy");

      float fraction_of_gpu_memory = config.fraction_of_gpu_memory_for_pool();
      if (fraction_of_gpu_memory > 0.95f) {
        LOG(ERROR)
            << "Allocate too much memory for the GPU memory pool, assigned "
            << config.memory_pool_init_size_mb() << " MB";
        LOG(ERROR) << "Try to shink the value by setting "
                      "AnalysisConfig::EnableGpu(...)";
      }
1345

1346 1347 1348 1349 1350 1351 1352
      if (fraction_of_gpu_memory >= 0.0f || fraction_of_gpu_memory <= 0.95f) {
        std::string flag = "--fraction_of_gpu_memory_to_use=" +
                           std::to_string(fraction_of_gpu_memory);
        VLOG(3) << "set flag: " << flag;
        gflags.push_back(flag);
      }

1353 1354 1355 1356 1357 1358 1359 1360 1361
      // TODO(Shixiaowei02): Add a mandatory scheme to use the thread local
      // allocator when multi-stream is enabled.
      if (config.thread_local_stream_enabled()) {
        gflags.push_back("--allocator_strategy=thread_local");
        process_level_allocator_enabled = false;
      } else {
        process_level_allocator_enabled = true;
      }

1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376
      if (framework::InitGflags(gflags)) {
        VLOG(3) << "The following gpu analysis configurations only take effect "
                   "for the first predictor: ";
        for (size_t i = 1; i < gflags.size(); ++i) {
          VLOG(3) << gflags[i];
        }
      } else {
        LOG(WARNING) << "The one-time configuration of analysis predictor "
                        "failed, which may be due to native predictor called "
                        "first and its configurations taken effect.";
      }
    });

    if (config.thread_local_stream_enabled() &&
        process_level_allocator_enabled) {
1377 1378 1379 1380 1381 1382
      PADDLE_THROW(platform::errors::Fatal(
          "When binding threads and streams, the use of "
          "process-level allocators will result in undefined result "
          "errors due to memory asynchronous operations."
          "The thread and stream binding configuration of all "
          "predictors should be the same in a single process."));
1383 1384 1385 1386
    }
  }

  std::unique_ptr<PaddlePredictor> predictor(new AnalysisPredictor(config));
1387 1388
  // Each config can only be used for one predictor.
  config.SetInValid();
1389 1390
  auto predictor_p = dynamic_cast<AnalysisPredictor *>(predictor.get());

1391 1392 1393 1394
#ifdef PADDLE_WITH_TENSORRT
  paddle::framework::ir::patterns::KeyCounter::Instance().CleanCounter();
#endif

1395 1396 1397 1398 1399
  if (!predictor_p->Init(nullptr)) {
    return nullptr;
  }

  if (config.mkldnn_quantizer_enabled() && !predictor_p->MkldnnQuantize()) {
1400 1401
    return nullptr;
  }
1402

G
Gabor Buella 已提交
1403
  return predictor;
1404 1405
}

1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417
bool AnalysisPredictor::MkldnnQuantize() {
#if PADDLE_WITH_MKLDNN
  if (!mkldnn_quantizer_)
    mkldnn_quantizer_ = new AnalysisPredictor::MkldnnQuantizer(
        *this, config_.mkldnn_quantizer_config());
  return mkldnn_quantizer_->Quantize();
#else
  LOG(ERROR) << "Please compile with MKLDNN first to use MkldnnQuantizer";
  return false;
#endif
}

1418
void AnalysisPredictor::PrepareFeedFetch() {
1419 1420 1421
  PADDLE_ENFORCE_NOT_NULL(sub_scope_,
                          platform::errors::InvalidArgument(
                              "The sub_scope should not be nullptr."));
1422
  CreateFeedFetchVar(sub_scope_);
1423 1424
  for (auto *op : inference_program_->Block(0).AllOps()) {
    if (op->Type() == "feed") {
R
Ruibiao Chen 已提交
1425
      int idx = PADDLE_GET_CONST(int, op->GetAttr("col"));
1426 1427 1428 1429 1430
      if (feeds_.size() <= static_cast<size_t>(idx)) {
        feeds_.resize(idx + 1);
      }
      feeds_[idx] = op;
      feed_names_[op->Output("Out")[0]] = idx;
N
nhzlx 已提交
1431
      idx2feeds_[idx] = op->Output("Out")[0];
1432
    } else if (op->Type() == "fetch") {
R
Ruibiao Chen 已提交
1433
      int idx = PADDLE_GET_CONST(int, op->GetAttr("col"));
Y
Yan Chunwei 已提交
1434 1435
      if (fetches_.size() <= static_cast<size_t>(idx)) {
        fetches_.resize(idx + 1);
1436
      }
Y
Yan Chunwei 已提交
1437
      fetches_[idx] = op;
N
nhzlx 已提交
1438
      idx2fetches_[idx] = op->Input("X")[0];
1439 1440 1441 1442
    }
  }
}

1443
void AnalysisPredictor::CreateFeedFetchVar(framework::Scope *scope) {
C
ccrrong 已提交
1444 1445 1446
  PADDLE_ENFORCE_NOT_NULL(
      scope,
      platform::errors::InvalidArgument("The scope should not be nullptr."));
1447
  auto *var = scope->Var("feed");
1448
  var->GetMutable<framework::FeedList>();
1449
  var = scope->Var("fetch");
1450
  var->GetMutable<framework::FetchList>();
1451 1452
}

N
nhzlx 已提交
1453 1454 1455 1456 1457 1458 1459 1460
std::vector<std::string> AnalysisPredictor::GetInputNames() {
  std::vector<std::string> input_names;
  for (auto &item : idx2feeds_) {
    input_names.push_back(item.second);
  }
  return input_names;
}

1461 1462 1463 1464 1465 1466
std::map<std::string, std::vector<int64_t>>
AnalysisPredictor::GetInputTensorShape() {
  std::map<std::string, std::vector<int64_t>> input_shapes;
  std::vector<std::string> names = GetInputNames();
  for (std::string name : names) {
    auto *var = inference_program_->Block(0).FindVar(name);
C
ccrrong 已提交
1467 1468 1469
    PADDLE_ENFORCE_NOT_NULL(
        var,
        platform::errors::PreconditionNotMet("Input %s does not exist.", name));
1470 1471 1472 1473 1474
    input_shapes[name] = var->GetShape();
  }
  return input_shapes;
}

1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505
std::map<std::string, paddle_infer::DataType>
AnalysisPredictor::GetInputTypes() {
  std::map<std::string, paddle_infer::DataType> input_type;
  std::vector<std::string> names = GetInputNames();
  for (const auto &name : names) {
    auto *var = inference_program_->Block(0).FindVar(name);
    PADDLE_ENFORCE_NOT_NULL(
        var,
        platform::errors::PreconditionNotMet(
            "Input %s does not exist inference_program_.", name));
    auto dtype = var->GetDataType();
    if (dtype == paddle::framework::proto::VarType::FP32) {
      input_type[name] = paddle_infer::DataType::FLOAT32;
    } else if (dtype == paddle::framework::proto::VarType::FP16) {
      input_type[name] = paddle_infer::DataType::FLOAT16;
    } else if (dtype == paddle::framework::proto::VarType::INT64) {
      input_type[name] = paddle_infer::DataType::INT64;
    } else if (dtype == paddle::framework::proto::VarType::INT32) {
      input_type[name] = paddle_infer::DataType::INT32;
    } else if (dtype == paddle::framework::proto::VarType::UINT8) {
      input_type[name] = paddle_infer::DataType::UINT8;
    } else if (dtype == paddle::framework::proto::VarType::INT8) {
      input_type[name] = paddle_infer::DataType::INT8;
    } else {
      PADDLE_THROW(paddle::platform::errors::Unimplemented(
          "Unsupported data type `%s` when get input dtype ", dtype));
    }
  }
  return input_type;
}

N
nhzlx 已提交
1506 1507 1508 1509 1510 1511 1512 1513
std::vector<std::string> AnalysisPredictor::GetOutputNames() {
  std::vector<std::string> output_names;
  for (auto &item : idx2fetches_) {
    output_names.push_back(item.second);
  }
  return output_names;
}

1514 1515
std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
    const std::string &name) {
1516
  framework::Scope *scope;
1517
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
1518 1519 1520 1521 1522 1523 1524 1525
  if (config_.dist_config().use_dist_model()) {
    scope = scope_.get();
  } else {
    scope = executor_->scope();
  }
#else
  scope = executor_->scope();
#endif
1526
  PADDLE_ENFORCE_NOT_NULL(
1527
      scope->FindVar(name),
1528
      platform::errors::PreconditionNotMet(
1529
          "The variable named %s is not found in the scope of the executor.",
1530
          name));
1531 1532
  std::unique_ptr<ZeroCopyTensor> res(new ZeroCopyTensor(
      static_cast<void *>(scope), this->GetDeviceContexts()));
1533 1534
  res->input_or_output_ = true;
  res->SetName(name);
N
nhzlx 已提交
1535 1536
  if (platform::is_cpu_place(place_)) {
    res->SetPlace(PaddlePlace::kCPU);
J
jianghaicheng 已提交
1537 1538 1539 1540
  } else if (platform::is_ipu_place(place_)) {
    // Currently, IPUPlace's tensor copy between cpu and ipu has been set in
    // IpuBackend.
    res->SetPlace(PaddlePlace::kCPU);
1541
  } else if (platform::is_xpu_place(place_)) {
1542 1543 1544 1545 1546 1547 1548 1549
    if (config_.lite_engine_enabled()) {
      // Currently, Paddle-Lite's XPU user interface only supports the transfer
      // of host data pointers. If it is currently used as a subgraph, execution
      // efficiency will be sacrificed, so it is temporarily set to cpu place.
      // And, the current lite engine of xpu must execute all parts of the
      // model.
      res->SetPlace(PaddlePlace::kCPU);
    } else {
1550
      auto xpu_place = place_;
1551 1552
      res->SetPlace(PaddlePlace::kXPU, xpu_place.GetDeviceId());
    }
W
Wilber 已提交
1553
  } else if (platform::is_npu_place(place_)) {
1554
    auto npu_place = place_;
W
Wilber 已提交
1555
    res->SetPlace(PaddlePlace::kNPU, npu_place.GetDeviceId());
1556 1557 1558 1559 1560 1561
  } else if (platform::is_custom_place(place_)) {
    auto custom_place = place_;
    auto paddleplace = static_cast<PaddlePlace>(
        static_cast<size_t>(PaddlePlace::kCUSTOM) +
        phi::GetOrRegisterGlobalDeviceTypeId(place_.GetDeviceType()));
    res->SetPlace(paddleplace, custom_place.GetDeviceId());
N
nhzlx 已提交
1562
  } else {
1563
    auto gpu_place = place_;
N
nhzlx 已提交
1564 1565
    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
  }
1566 1567 1568 1569 1570
  return res;
}

std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
    const std::string &name) {
1571
  framework::Scope *scope;
1572
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
1573 1574 1575 1576 1577 1578 1579 1580
  if (config_.dist_config().use_dist_model()) {
    scope = scope_.get();
  } else {
    scope = executor_->scope();
  }
#else
  scope = executor_->scope();
#endif
1581
  PADDLE_ENFORCE_NOT_NULL(
1582
      scope->FindVar(name),
1583
      platform::errors::PreconditionNotMet(
1584
          "The variable named %s is not found in the scope of the executor.",
1585
          name));
1586 1587
  std::unique_ptr<ZeroCopyTensor> res(new ZeroCopyTensor(
      static_cast<void *>(scope), this->GetDeviceContexts()));
1588 1589
  res->input_or_output_ = false;
  res->SetName(name);
N
nhzlx 已提交
1590 1591
  if (platform::is_cpu_place(place_)) {
    res->SetPlace(PaddlePlace::kCPU);
J
jianghaicheng 已提交
1592 1593 1594 1595
  } else if (platform::is_ipu_place(place_)) {
    // Currently, IPUPlace's tensor copy between cpu and ipu has been set in
    // IpuBackend.
    res->SetPlace(PaddlePlace::kCPU);
1596
  } else if (platform::is_xpu_place(place_)) {
1597 1598 1599 1600 1601 1602 1603 1604
    if (config_.lite_engine_enabled()) {
      // Currently, Paddle-Lite's XPU user interface only supports the transfer
      // of host data pointers. If it is currently used as a subgraph, execution
      // efficiency will be sacrificed, so it is temporarily set to cpu place.
      // And, the current lite engine of xpu must execute all parts of the
      // model.
      res->SetPlace(PaddlePlace::kCPU);
    } else {
1605
      auto xpu_place = place_;
1606 1607
      res->SetPlace(PaddlePlace::kXPU, xpu_place.GetDeviceId());
    }
W
Wilber 已提交
1608
  } else if (platform::is_npu_place(place_)) {
1609
    auto npu_place = place_;
W
Wilber 已提交
1610
    res->SetPlace(PaddlePlace::kNPU, npu_place.GetDeviceId());
1611 1612 1613 1614 1615 1616
  } else if (platform::is_custom_place(place_)) {
    auto custom_place = place_;
    auto paddleplace = static_cast<PaddlePlace>(
        static_cast<size_t>(PaddlePlace::kCUSTOM) +
        phi::GetOrRegisterGlobalDeviceTypeId(place_.GetDeviceType()));
    res->SetPlace(paddleplace, custom_place.GetDeviceId());
N
nhzlx 已提交
1617
  } else {
1618
    auto gpu_place = place_;
N
nhzlx 已提交
1619 1620
    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
  }
1621 1622 1623 1624
  return res;
}

bool AnalysisPredictor::ZeroCopyRun() {
1625
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
1626 1627 1628 1629 1630 1631 1632 1633 1634 1635
  if (config_.dist_config().use_dist_model()) {
    VLOG(3) << "ZeroCopyRun will use the fleet executor.";
    inference::Timer timer;
    timer.tic();
    fleet_exe_->Run(config_.dist_config().carrier_id());
    VLOG(3) << "Fleet executor inf runs once use: "
            << std::to_string(timer.toc()) << "ms";
    return true;
  }
#endif
1636 1637 1638
  if (private_context_) {
    paddle::platform::DeviceContextPool::SetDeviceContexts(&device_contexts_);
  }
1639
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
W
Wilber 已提交
1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650
#ifdef PADDLE_WITH_MKLDNN
  if (config_.use_mkldnn_) {
    std::vector<std::vector<int>> shape_vector;
    auto names = GetInputNames();
    for (size_t i = 0; i < names.size(); ++i) {
      auto in_tensor = GetInputTensor(names[i]);
      shape_vector.emplace_back(in_tensor->shape());
    }
    MkldnnPreSet(shape_vector);
  }
#endif
1651 1652 1653 1654 1655 1656 1657 1658 1659 1660

#ifdef PADDLE_WITH_TENSORRT
  if (config_.tensorrt_engine_enabled()) {
    inference::tensorrt::TensorRTEngine::predictor_id_per_thread =
        predictor_id_;
    VLOG(3) << "thread_local var predictor_id in TensorRTEngine is set to: "
            << inference::tensorrt::TensorRTEngine::predictor_id_per_thread;
  }
#endif

1661
  executor_->Run();
1662 1663 1664 1665 1666

  if (config_.shape_range_info_collected()) {
    CollectShapeRangeInfo();
  }

Y
Yan Chunwei 已提交
1667
  // Fix TensorArray reuse not cleaned bug.
Y
Yan Chunwei 已提交
1668
  tensor_array_batch_cleaner_.CollectTensorArrays(sub_scope_);
Y
Yan Chunwei 已提交
1669
  tensor_array_batch_cleaner_.ResetTensorArray();
1670 1671 1672 1673

  // recover the cpu_math_library_num_threads to 1, in order to avoid thread
  // conflict when integrating it into deployment service.
  paddle::platform::SetNumThreads(1);
1674 1675 1676
  if (private_context_) {
    paddle::platform::DeviceContextPool::SetDeviceContexts(nullptr);
  }
W
Wilber 已提交
1677 1678 1679
#ifdef PADDLE_WITH_MKLDNN
  if (config_.use_mkldnn_) MkldnnPostReset();
#endif
1680
#if defined(PADDLE_WITH_MKLML)
T
Tao Luo 已提交
1681 1682 1683 1684 1685
  // Frees unused memory allocated by the Intel® MKL Memory Allocator to
  // avoid memory leak. See:
  // https://software.intel.com/en-us/mkl-developer-reference-c-mkl-free-buffers
  platform::dynload::MKL_Free_Buffers();
#endif
1686 1687 1688
  return true;
}

W
Wilber 已提交
1689 1690
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
bool AnalysisPredictor::ExpRunWithExternalStream(const gpuStream_t stream) {
W
Wilber 已提交
1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715
  if (!private_context_) {
    PADDLE_THROW(platform::errors::Fatal(
        "Please use config.SetExecStream to init gpu resources, and then we "
        "will bind gpu resources to execution stream."));
  }

  if (stream != predictor_stream_) {
#ifdef PADDLE_WITH_HIP
    hipStreamSynchronize(static_cast<gpuStream_t>(predictor_stream_));
#else
    cudaStreamSynchronize(static_cast<gpuStream_t>(predictor_stream_));
#endif
    ResourceManager::Instance().GpuResourceReBindStream(predictor_stream_,
                                                        stream);
    predictor_stream_ = stream;

    auto *dev_ctxs = reinterpret_cast<const std::map<
        phi::Place,
        std::shared_future<std::unique_ptr<phi::DeviceContext>>> *>(
        this->GetDeviceContexts());
    auto *dev_ctx =
        static_cast<InferGPUContext *>(dev_ctxs->at(place_).get().get());
    dev_ctx->SetStream(stream);
  }

W
Wilber 已提交
1716 1717 1718 1719
  return ZeroCopyRun();
}
#endif

1720 1721 1722 1723 1724 1725
void AnalysisPredictor::CollectShapeRangeInfo() {
  // if use gpu, sync first.
  if (config_.use_gpu()) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
    paddle::platform::DeviceContextPool &pool =
        paddle::platform::DeviceContextPool::Instance();
1726
    auto gpu_place = place_;
L
Leo Chen 已提交
1727
    auto *dev_ctx = static_cast<const phi::GPUContext *>(pool.Get(gpu_place));
1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764
#ifdef PADDLE_WITH_HIP
    hipStreamSynchronize(dev_ctx->stream());
#else
    cudaStreamSynchronize(dev_ctx->stream());
#endif
#endif
  }

  std::vector<std::string> var_names = sub_scope_->LocalVarNames();
  for (const auto &name : var_names) {
    auto *var = sub_scope_->GetVar(name);
    if (!var->IsType<framework::LoDTensor>()) {
      continue;
    }
    framework::DDim dim = var->Get<framework::LoDTensor>().dims();
    std::vector<int32_t> shape(dim.size());
    for (size_t i = 0; i < shape.size(); ++i) shape[i] = dim[i];
    shape_info_[name].emplace_back(shape);
  }
}

void AnalysisPredictor::StatisticShapeRangeInfo() {
  std::map<std::string, std::vector<int32_t>> min_shapes;
  std::map<std::string, std::vector<int32_t>> max_shapes;
  std::map<std::string, std::vector<int32_t>> opt_shapes;
  for (auto it : shape_info_) {
    auto name = it.first;
    auto shapes = it.second;

    std::vector<int32_t> min_shape(shapes[0].begin(), shapes[0].end());
    std::vector<int32_t> max_shape(shapes[0].begin(), shapes[0].end());
    std::vector<int32_t> opt_shape(shapes[0].begin(), shapes[0].end());

    auto ShapeMaxFreq = [](const std::map<int32_t, int32_t> &m) -> int32_t {
      std::vector<std::pair<int32_t, int32_t>> counter;
      for (auto &it : m) counter.push_back(it);
      std::sort(
C
ccrrong 已提交
1765 1766
          counter.begin(),
          counter.end(),
1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787
          [](std::pair<int32_t, int32_t> &a, std::pair<int32_t, int32_t> &b) {
            return a.second > b.second;
          });
      return counter[0].first;
    };

    for (size_t d = 0; d < shapes[0].size(); ++d) {
      std::map<int32_t, int32_t> counter;
      for (size_t i = 0; i < shapes.size(); ++i) {
        counter[shapes[i][d]] += 1;
        if (shapes[i][d] < min_shape[d]) min_shape[d] = shapes[i][d];
        if (shapes[i][d] > max_shape[d]) max_shape[d] = shapes[i][d];
      }
      opt_shape[d] = ShapeMaxFreq(counter);
    }

    min_shapes[name] = min_shape;
    max_shapes[name] = max_shape;
    opt_shapes[name] = opt_shape;
  }

C
ccrrong 已提交
1788 1789
  inference::SerializeShapeRangeInfo(
      config_.shape_range_info_path(), min_shapes, max_shapes, opt_shapes);
1790 1791
}

1792 1793
bool AnalysisPredictor::LoadProgramDesc() {
  // Initialize the inference program
1794
  std::string filename;
1795 1796
  if (!config_.model_dir().empty()) {
    filename = config_.model_dir() + "/__model__";
1797
  } else if (!config_.prog_file().empty()) {
1798 1799 1800
    // 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`.
1801
    filename = config_.prog_file();
1802
  } else {
1803
    if (config_.model_dir().empty() && config_.prog_file().empty()) {
1804 1805 1806 1807
      LOG(ERROR)
          << "Either model_dir or (prog_file, param_file) should be set.";
      return false;
    }
1808
    LOG(ERROR) << string::Sprintf(
C
ccrrong 已提交
1809 1810
        "not valid model path '%s' or program path '%s'.",
        config_.model_dir(),
1811
        config_.params_file());
1812 1813
    return false;
  }
1814 1815 1816

  // Create ProgramDesc
  framework::proto::ProgramDesc proto;
T
Tao Luo 已提交
1817
  if (!config_.model_from_memory()) {
T
Tao Luo 已提交
1818 1819 1820
    std::string pb_content;
    // Read binary
    std::ifstream fin(filename, std::ios::in | std::ios::binary);
1821
    PADDLE_ENFORCE_EQ(
C
ccrrong 已提交
1822 1823
        static_cast<bool>(fin.is_open()),
        true,
1824 1825 1826
        platform::errors::NotFound(
            "Cannot open file %s, please confirm whether the file is normal.",
            filename));
T
Tao Luo 已提交
1827 1828 1829 1830 1831 1832 1833 1834
    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 {
1835
    proto.ParseFromString(config_.prog_file());
T
Tao Luo 已提交
1836
  }
1837 1838 1839 1840 1841 1842
  inference_program_.reset(new framework::ProgramDesc(proto));
  return true;
}

bool AnalysisPredictor::LoadParameters() {
  PADDLE_ENFORCE_NOT_NULL(inference_program_.get(),
1843 1844
                          platform::errors::PreconditionNotMet(
                              "The inference program should be loaded first."));
T
Tao Luo 已提交
1845

1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865
  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);

1866
      if (!config_.params_file().empty()) {
1867 1868 1869 1870 1871 1872
        params.push_back(new_var->Name());
      } else {
        // append_op
        framework::OpDesc *op = load_block->AppendOp();
        op->SetType("load");
        op->SetOutput("Out", {new_var->Name()});
1873
        op->SetAttr("file_path", {config_.model_dir() + "/" + new_var->Name()});
1874 1875 1876 1877 1878
        op->CheckAttrs();
      }
    }
  }

1879
  if (!config_.params_file().empty()) {
1880 1881 1882 1883 1884 1885
    // 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);
1886
    op->SetAttr("file_path", {config_.params_file()});
1887 1888 1889 1890
    op->CheckAttrs();
  }

  // Use NaiveExecutor to Load parameters.
S
superjomn 已提交
1891
  framework::NaiveExecutor e(place_);
1892 1893 1894 1895
  e.Prepare(scope_.get(), *load_program, 0, false);
  e.Run();
  VLOG(3) << "get " << scope_->LocalVarNames().size() << " vars after load";

1896 1897
  return true;
}
1898

1899 1900 1901 1902 1903
uint64_t AnalysisPredictor::TryShrinkMemory() {
  ClearIntermediateTensor();
  return paddle::memory::Release(place_);
}

1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922
void AnalysisPredictor::ClearIntermediateTensor() {
  PADDLE_ENFORCE_NOT_NULL(inference_program_.get(),
                          platform::errors::PreconditionNotMet(
                              "The inference program should be loaded first."));
  const auto &global_block = inference_program_->MutableBlock(0);
  for (auto *var : global_block->AllVars()) {
    if (!IsPersistable(var)) {
      const std::string name = var->Name();
      auto *variable = executor_->scope()->FindVar(name);
      if (variable != nullptr && variable->IsType<framework::LoDTensor>() &&
          name != "feed" && name != "fetch") {
        VLOG(3) << "Clear Intermediate Tensor: " << name;
        auto *t = variable->GetMutable<framework::LoDTensor>();
        t->clear();
      }
    }
  }
}

N
nhzlx 已提交
1923
#if PADDLE_WITH_TENSORRT
N
nhzlx 已提交
1924
bool AnalysisPredictor::SaveTrtCalibToDisk() {
C
ccrrong 已提交
1925 1926
  PADDLE_ENFORCE_EQ(config_.tensorrt_engine_enabled(),
                    true,
1927 1928
                    platform::errors::PreconditionNotMet(
                        "This func can be invoked only in trt mode"));
N
nhzlx 已提交
1929 1930 1931
  auto &block = inference_program_->Block(0);
  for (auto &op_desc : block.AllOps()) {
    if (op_desc->Type() == "tensorrt_engine") {
R
Ruibiao Chen 已提交
1932
      std::string engine_name = PADDLE_GET_CONST(
1933
          std::string, op_desc->GetAttr("calibration_engine_key"));
N
nhzlx 已提交
1934
      if (!Singleton<TRTCalibratorEngineManager>::Global().Has(engine_name)) {
N
nhzlx 已提交
1935 1936 1937 1938
        LOG(ERROR) << "You should run the predictor(with trt) on the real data "
                      "to generate calibration info";
        return false;
      }
N
nhzlx 已提交
1939 1940
      TRTCalibratorEngine *calib_engine =
          Singleton<TRTCalibratorEngineManager>::Global().Get(engine_name);
N
nhzlx 已提交
1941
      LOG(INFO) << "Wait for calib threads done.";
N
nhzlx 已提交
1942
      calib_engine->calib_->waitAndSetDone();
N
nhzlx 已提交
1943 1944
      LOG(INFO) << "Generating TRT Calibration table data, this may cost a lot "
                   "of time...";
N
nhzlx 已提交
1945 1946 1947
      calib_engine->thr_->join();
      std::string calibration_table_data =
          calib_engine->calib_->getCalibrationTableAsString();
N
nhzlx 已提交
1948

N
nhzlx 已提交
1949
      if (calibration_table_data.empty()) {
N
nhzlx 已提交
1950 1951 1952
        LOG(ERROR) << "the calibration table is empty.";
        return false;
      }
N
nhzlx 已提交
1953

N
nhzlx 已提交
1954 1955 1956 1957 1958
      std::string model_opt_cache_dir =
          argument_.Has("model_dir")
              ? argument_.model_dir()
              : inference::analysis::GetDirRoot(argument_.model_program_path());

N
nhzlx 已提交
1959
      std::string calibration_table_data_path =
N
nhzlx 已提交
1960 1961 1962 1963
          inference::analysis::GetTrtCalibPath(
              inference::analysis::GetOrCreateModelOptCacheDir(
                  model_opt_cache_dir),
              engine_name);
N
nhzlx 已提交
1964 1965 1966 1967 1968

      std::ofstream ofile(calibration_table_data_path, std::ios::out);
      LOG(INFO) << "Write Paddle-TRT INT8 calibration table data to file "
                << calibration_table_data_path;
      ofile << calibration_table_data;
N
nhzlx 已提交
1969 1970 1971 1972
      ofile.close();
    }
  }
  // Free all calibrator resources.
N
nhzlx 已提交
1973
  Singleton<TRTCalibratorEngineManager>::Global().DeleteALL();
N
nhzlx 已提交
1974 1975
  return true;
}
N
nhzlx 已提交
1976
#endif
N
nhzlx 已提交
1977

1978
AnalysisPredictor::~AnalysisPredictor() {
N
nhzlx 已提交
1979
#if PADDLE_WITH_TENSORRT
N
nhzlx 已提交
1980
  if (config_.tensorrt_engine_enabled() &&
N
nhzlx 已提交
1981 1982
      config_.tensorrt_precision_mode_ == AnalysisConfig::Precision::kInt8 &&
      Singleton<TRTCalibratorEngineManager>::Global().Has()) {
N
nhzlx 已提交
1983 1984
    SaveTrtCalibToDisk();
  }
N
nhzlx 已提交
1985
#endif
1986
  if (config_.with_profile_) {
1987 1988 1989 1990
    platform::DisableProfiler(platform::EventSortingKey::kTotal,
                              "./profile.log");
  }
  if (sub_scope_) {
J
JingZhuangzhuang 已提交
1991 1992 1993 1994 1995 1996 1997 1998 1999
    if (framework::global_transfer_scope_key().find(sub_scope_) !=
        framework::global_transfer_scope_key().end()) {
      auto scope_key_set = framework::global_transfer_scope_key()[sub_scope_];
      for (auto iter = scope_key_set.begin(); iter != scope_key_set.end();
           iter++) {
        framework::global_transfer_data_cache().erase(*iter);
      }
      framework::global_transfer_scope_key().erase(sub_scope_);
    }
2000 2001
    scope_->DeleteScope(sub_scope_);
  }
Y
Yan Chunwei 已提交
2002

2003 2004 2005 2006 2007 2008
#if PADDLE_WITH_MKLDNN
  if (mkldnn_quantizer_) {
    delete mkldnn_quantizer_;
    mkldnn_quantizer_ = nullptr;
  }
#endif
2009

2010 2011 2012
  if (config_.shape_range_info_collected()) {
    StatisticShapeRangeInfo();
  }
2013 2014 2015 2016 2017
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  if (predictor_stream_ != nullptr) {
    ResourceManager::Instance().DestroyGPUResource(predictor_stream_);
  }
#endif
W
Wilber 已提交
2018 2019 2020
  if (place_.GetType() != phi::AllocationType::UNDEFINED) {
    memory::Release(place_);
  }
2021
  device_contexts_.clear();
2022 2023
}

2024
std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone(void *stream) {
Y
Yan Chunwei 已提交
2025
  std::lock_guard<std::mutex> lk(clone_mutex_);
2026
  auto *x = new AnalysisPredictor(config_);
2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037
  x->status_is_cloned_ = true;
  if (config_.use_external_stream_ && stream == nullptr) {
    PADDLE_THROW(platform::errors::InvalidArgument(
        "config has been configured to use external stream, but the Clone "
        "function has not received a valid stream parameter."));
  } else if (!config_.use_external_stream_ && stream != nullptr) {
    PADDLE_THROW(platform::errors::InvalidArgument(
        "config has not been configured to use external stream, but the Clone "
        "function has received a stream parameter."));
  }
  x->predictor_stream_ = stream;
2038
  x->Init(scope_, inference_program_);
2039
  x->executor_->ResetTrtOps(++AnalysisPredictor::clone_num_);
2040 2041 2042
  return std::unique_ptr<PaddlePredictor>(x);
}

2043
std::string AnalysisPredictor::GetSerializedProgram() const {
Y
Yan Chunwei 已提交
2044 2045 2046
  return inference_program_->Proto()->SerializeAsString();
}

2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085
// Add SaveOptimModel
void AnalysisPredictor::SaveOptimModel(const std::string &dir) {
  // save model
  std::string model_name = dir + "/model";
  std::ofstream outfile;
  outfile.open(model_name, std::ios::out | std::ios::binary);
  std::string inference_prog_desc = GetSerializedProgram();
  outfile << inference_prog_desc;
  // save params
  framework::ProgramDesc save_program;
  auto *save_block = save_program.MutableBlock(0);

  const framework::ProgramDesc &main_program = program();
  const framework::BlockDesc &global_block = main_program.Block(0);
  std::vector<std::string> save_var_list;
  for (framework::VarDesc *var : global_block.AllVars()) {
    if (IsPersistable(var)) {
      framework::VarDesc *new_var = save_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);

      save_var_list.push_back(new_var->Name());
    }
  }
  std::sort(save_var_list.begin(), save_var_list.end());
  auto *op = save_block->AppendOp();
  op->SetType("save_combine");
  op->SetInput("X", save_var_list);
  op->SetAttr("file_path", dir + "/params");
  op->CheckAttrs();

  platform::CPUPlace place;
  framework::Executor exe(place);
  exe.Run(save_program, scope(), 0, true, true);
}

Y
Yan Chunwei 已提交
2086
template <>
2087 2088
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<AnalysisConfig>(
    const AnalysisConfig &config) {
W
Wilber 已提交
2089
  LOG(WARNING) << "Deprecated. Please use CreatePredictor instead.";
2090 2091
  return CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
      config);
Y
Yan Chunwei 已提交
2092 2093
}

2094
}  // namespace paddle
2095 2096 2097

#if PADDLE_WITH_TENSORRT
USE_TRT_CONVERTER(elementwise_add_weight);
S
shentanyue 已提交
2098 2099 2100 2101
USE_TRT_CONVERTER(elementwise_sub_weight);
USE_TRT_CONVERTER(elementwise_mul_weight);
USE_TRT_CONVERTER(elementwise_div_weight);
USE_TRT_CONVERTER(elementwise_pow_weight);
2102 2103 2104 2105 2106 2107 2108
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);
2109
USE_TRT_CONVERTER(transpose);
2110
USE_TRT_CONVERTER(transpose2);
2111
USE_TRT_CONVERTER(flatten);
2112
USE_TRT_CONVERTER(flatten_contiguous_range);
2113
USE_TRT_CONVERTER(matmul);
2114 2115
USE_TRT_CONVERTER(conv2d);
USE_TRT_CONVERTER(relu);
Z
zhupengyang 已提交
2116 2117
USE_TRT_CONVERTER(exp);
USE_TRT_CONVERTER(log);
2118 2119 2120 2121 2122 2123 2124 2125 2126
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);
2127 2128
USE_TRT_CONVERTER(hard_sigmoid);
USE_TRT_CONVERTER(hard_swish);
2129
USE_TRT_CONVERTER(split);
2130 2131
USE_TRT_CONVERTER(prelu);
USE_TRT_CONVERTER(conv2d_transpose);
H
hjchen2 已提交
2132
USE_TRT_CONVERTER(leaky_relu);
2133 2134
USE_TRT_CONVERTER(shuffle_channel);
USE_TRT_CONVERTER(swish);
L
LielinJiang 已提交
2135
USE_TRT_CONVERTER(silu);
2136
USE_TRT_CONVERTER(group_norm);
2137
USE_TRT_CONVERTER(instance_norm);
P
Pei Yang 已提交
2138 2139 2140
USE_TRT_CONVERTER(layer_norm);
USE_TRT_CONVERTER(gelu);
USE_TRT_CONVERTER(multihead_matmul);
2141 2142
USE_TRT_CONVERTER(fused_embedding_eltwise_layernorm);
USE_TRT_CONVERTER(skip_layernorm);
2143
USE_TRT_CONVERTER(slice);
2144
USE_TRT_CONVERTER(scale);
2145
USE_TRT_CONVERTER(stack);
P
Pei Yang 已提交
2146
USE_TRT_CONVERTER(clip);
2147
USE_TRT_CONVERTER(gather);
2148
USE_TRT_CONVERTER(anchor_generator);
Z
zlsh80826 已提交
2149
USE_TRT_CONVERTER(yolo_box);
2150
USE_TRT_CONVERTER(yolo_box_head);
2151
USE_TRT_CONVERTER(arg_max);
2152
USE_TRT_CONVERTER(roi_align);
2153
USE_TRT_CONVERTER(affine_channel);
Z
zlsh80826 已提交
2154
USE_TRT_CONVERTER(multiclass_nms);
2155
USE_TRT_CONVERTER(multiclass_nms3);
2156
USE_TRT_CONVERTER(nearest_interp);
2157
USE_TRT_CONVERTER(nearest_interp_v2);
2158
USE_TRT_CONVERTER(bilinear_interp_v2);
W
Wangzheee 已提交
2159
USE_TRT_CONVERTER(reshape);
2160
USE_TRT_CONVERTER(reshape2);
2161 2162
USE_TRT_CONVERTER(reduce_sum);
USE_TRT_CONVERTER(gather_nd);
W
wenbin 已提交
2163
USE_TRT_CONVERTER(reduce_mean);
W
wenbin 已提交
2164
USE_TRT_CONVERTER(tile);
W
wenbin 已提交
2165 2166
USE_TRT_CONVERTER(conv3d);
USE_TRT_CONVERTER(conv3d_transpose);
W
wangxinxin08 已提交
2167
USE_TRT_CONVERTER(mish);
W
wangxinxin08 已提交
2168
USE_TRT_CONVERTER(deformable_conv);
F
feng_shuai 已提交
2169
USE_TRT_CONVERTER(pool3d)
2170 2171
USE_TRT_CONVERTER(fused_preln_embedding_eltwise_layernorm)
USE_TRT_CONVERTER(preln_skip_layernorm)
2172 2173
USE_TRT_CONVERTER(preln_residual_bias)
USE_TRT_CONVERTER(c_allreduce_sum)
F
feng_shuai 已提交
2174
USE_TRT_CONVERTER(roll)
F
feng_shuai 已提交
2175
USE_TRT_CONVERTER(strided_slice)
Z
zhoutianzi666 已提交
2176 2177
USE_TRT_CONVERTER(rnn)
USE_TRT_CONVERTER(fill_constant_batch_size_like)
2178
USE_TRT_CONVERTER(transformer_input_convert)
C
ccrrong 已提交
2179
USE_TRT_CONVERTER(cast)
2180 2181
USE_TRT_CONVERTER(recover_padding)
USE_TRT_CONVERTER(remove_padding)
C
ccrrong 已提交
2182
USE_TRT_CONVERTER(equal);
2183 2184
USE_TRT_CONVERTER(top_k)
USE_TRT_CONVERTER(top_k_v2)
2185 2186
USE_TRT_CONVERTER(squeeze2)
USE_TRT_CONVERTER(unsqueeze2)
2187 2188
USE_TRT_CONVERTER(sum)
USE_TRT_CONVERTER(shape)
2189
USE_TRT_CONVERTER(fill_constant)
2190
USE_TRT_CONVERTER(fused_token_prune)
2191 2192 2193 2194
#if PADDLE_WITH_CUSPARSELT && IS_TRT_VERSION_GE(8000)
USE_TRT_CONVERTER(sparse_fc)
USE_TRT_CONVERTER(sparse_multihead_matmul)
#endif
2195
#endif
W
Wilber 已提交
2196 2197 2198 2199 2200 2201

namespace paddle_infer {

Predictor::Predictor(const Config &config) {
  const_cast<Config *>(&config)->SwitchUseFeedFetchOps(false);
  // The second parameter indicates that the discard log is not printed
2202 2203 2204 2205 2206 2207 2208 2209 2210 2211
  if (config.use_onnxruntime()) {
#ifdef PADDLE_WITH_ONNXRUNTIME
    if (config.use_gpu()) {
      LOG(WARNING) << "The current ONNXRuntime backend doesn't support GPU,"
                      "and it falls back to use Paddle Inference.";
    } else if (!paddle::CheckConvertToONNX(config)) {
      LOG(WARNING)
          << "Paddle2ONNX do't support convert the Model, fall back to using "
             "Paddle Inference.";
    } else {
C
ccrrong 已提交
2212 2213 2214 2215
      predictor_ =
          paddle::CreatePaddlePredictor<Config,
                                        paddle::PaddleEngineKind::kONNXRuntime>(
              config);
2216 2217 2218 2219 2220 2221 2222 2223 2224
      return;
    }
#else
    LOG(WARNING)
        << "The onnxruntime backend isn't enabled,"
           " and please re-compile Paddle with WITH_ONNXRUNTIME option,"
           "fall back to using Paddle Inference.";
#endif
  }
C
ccrrong 已提交
2225 2226 2227 2228
  predictor_ =
      paddle::CreatePaddlePredictor<Config,
                                    paddle::PaddleEngineKind::kAnalysis>(
          config);
W
Wilber 已提交
2229 2230 2231 2232 2233
}

std::vector<std::string> Predictor::GetInputNames() {
  return predictor_->GetInputNames();
}
2234 2235 2236 2237

std::map<std::string, DataType> Predictor::GetInputTypes() {
  return predictor_->GetInputTypes();
}
W
Wilber 已提交
2238 2239

std::unique_ptr<Tensor> Predictor::GetInputHandle(const std::string &name) {
2240
  return predictor_->GetInputTensor(name);
W
Wilber 已提交
2241 2242 2243 2244 2245 2246 2247
}

std::vector<std::string> Predictor::GetOutputNames() {
  return predictor_->GetOutputNames();
}

std::unique_ptr<Tensor> Predictor::GetOutputHandle(const std::string &name) {
2248
  return predictor_->GetOutputTensor(name);
W
Wilber 已提交
2249 2250 2251 2252
}

bool Predictor::Run() { return predictor_->ZeroCopyRun(); }

2253 2254
std::unique_ptr<Predictor> Predictor::Clone(void *stream) {
  auto analysis_pred = predictor_->Clone(stream);
W
Wilber 已提交
2255 2256 2257 2258 2259 2260 2261 2262
  std::unique_ptr<Predictor> pred(new Predictor(std::move(analysis_pred)));
  return pred;
}

void Predictor::ClearIntermediateTensor() {
  predictor_->ClearIntermediateTensor();
}

2263 2264
uint64_t Predictor::TryShrinkMemory() { return predictor_->TryShrinkMemory(); }

2265 2266
void *Predictor::GetExecStream() const { return predictor_->GetExecStream(); }

W
Wilber 已提交
2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284
int GetNumBytesOfDataType(DataType dtype) {
  switch (dtype) {
    case DataType::FLOAT32:
      return sizeof(float);
    case DataType::INT64:
      return sizeof(int64_t);
    case DataType::INT32:
      return sizeof(int32_t);
    case DataType::UINT8:
      return sizeof(uint8_t);
    default:
      assert(false);
      return -1;
  }
}

std::string GetVersion() { return paddle::get_version(); }

2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300
std::tuple<int, int, int> GetTrtCompileVersion() {
#ifdef PADDLE_WITH_TENSORRT
  return paddle::inference::tensorrt::GetTrtCompileVersion();
#else
  return std::tuple<int, int, int>{0, 0, 0};
#endif
}

std::tuple<int, int, int> GetTrtRuntimeVersion() {
#ifdef PADDLE_WITH_TENSORRT
  return paddle::inference::tensorrt::GetTrtRuntimeVersion();
#else
  return std::tuple<int, int, int>{0, 0, 0};
#endif
}

W
Wilber 已提交
2301 2302 2303 2304
std::string UpdateDllFlag(const char *name, const char *value) {
  return paddle::UpdateDllFlag(name, value);
}

2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324
void ConvertToMixedPrecision(const std::string &model_file,
                             const std::string &params_file,
                             const std::string &mixed_model_file,
                             const std::string &mixed_params_file,
                             PrecisionType mixed_precision,
                             BackendType backend,
                             bool keep_io_types,
                             std::unordered_set<std::string> black_list) {
  auto phi_backend = paddle::ConvertBackend(backend);
  auto phi_precision = paddle::ConvertPrecision(mixed_precision);
  paddle::inference::analysis::ConvertToMixedPrecision(model_file,
                                                       params_file,
                                                       mixed_model_file,
                                                       mixed_params_file,
                                                       phi_precision,
                                                       phi_backend,
                                                       keep_io_types,
                                                       black_list);
}

W
Wilber 已提交
2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335
}  // namespace paddle_infer

namespace paddle_infer {
std::shared_ptr<Predictor> CreatePredictor(const Config &config) {  // NOLINT
  std::shared_ptr<Predictor> predictor(new Predictor(config));
  return predictor;
}

namespace services {
PredictorPool::PredictorPool(const Config &config, size_t size) {
  PADDLE_ENFORCE_GE(
C
ccrrong 已提交
2336 2337
      size,
      1UL,
W
Wilber 已提交
2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355
      paddle::platform::errors::InvalidArgument(
          "The predictor pool size should be greater than 1, but it's (%d)",
          size));
  Config copy_config(config);
  main_pred_.reset(new Predictor(config));
  for (size_t i = 0; i < size - 1; i++) {
    if (config.tensorrt_engine_enabled()) {
      Config config_tmp(copy_config);
      preds_.push_back(
          std::move(std::unique_ptr<Predictor>(new Predictor(config_tmp))));
    } else {
      preds_.push_back(std::move(main_pred_->Clone()));
    }
  }
}

Predictor *PredictorPool::Retrive(size_t idx) {
  PADDLE_ENFORCE_LT(
C
ccrrong 已提交
2356 2357
      idx,
      preds_.size() + 1,
W
Wilber 已提交
2358
      paddle::platform::errors::InvalidArgument(
C
ccrrong 已提交
2359 2360
          "There are (%d) predictors in the pool, but the idx is (%d)",
          idx,
W
Wilber 已提交
2361 2362 2363 2364 2365 2366 2367
          preds_.size() + 1));
  if (idx == 0) {
    return main_pred_.get();
  }
  return preds_[idx - 1].get();
}
}  // namespace services
W
Wilber 已提交
2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387

namespace experimental {

// Note: Can only be used under thread_local semantics.
bool InternalUtils::RunWithExternalStream(paddle_infer::Predictor *p,
                                          cudaStream_t stream) {
#ifdef PADDLE_WITH_CUDA
  auto pred = dynamic_cast<paddle::AnalysisPredictor *>(p->predictor_.get());
  return pred->ExpRunWithExternalStream(stream);
#endif
  return false;
}
bool InternalUtils::RunWithExternalStream(paddle_infer::Predictor *p,
                                          hipStream_t stream) {
#ifdef PADDLE_WITH_HIP
  auto pred = dynamic_cast<paddle::AnalysisPredictor *>(p->predictor_.get());
  return pred->ExpRunWithExternalStream(stream);
#endif
  return false;
}
W
Wilber 已提交
2388

2389 2390 2391 2392 2393 2394
void InternalUtils::UpdateConfigInterleaved(paddle_infer::Config *c,
                                            bool with_interleaved) {
#ifdef PADDLE_WITH_CUDA
  c->trt_with_interleaved_ = with_interleaved;
#endif
}
W
Wilber 已提交
2395

2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409
void InternalUtils::SetTransformerPosid(
    paddle_infer::Config *c, const std::string &tensorrt_transformer_posid) {
#ifdef PADDLE_WITH_CUDA
  c->tensorrt_transformer_posid_ = tensorrt_transformer_posid;
#endif
}

void InternalUtils::SetTransformerMaskid(
    paddle_infer::Config *c, const std::string &tensorrt_transformer_maskid) {
#ifdef PADDLE_WITH_CUDA
  c->tensorrt_transformer_maskid_ = tensorrt_transformer_maskid;
#endif
}

W
Wilber 已提交
2410 2411 2412 2413 2414
void InternalUtils::SyncStream(paddle_infer::Predictor *p) {
#ifdef PADDLE_WITH_CUDA
  auto *pred = dynamic_cast<paddle::AnalysisPredictor *>(p->predictor_.get());
  paddle::platform::DeviceContextPool &pool =
      paddle::platform::DeviceContextPool::Instance();
L
Leo Chen 已提交
2415
  auto *dev_ctx = reinterpret_cast<phi::GPUContext *>(pool.Get(pred->place_));
W
Wilber 已提交
2416 2417 2418 2419 2420 2421 2422 2423 2424
  cudaStreamSynchronize(dev_ctx->stream());
#endif
}
void InternalUtils::SyncStream(cudaStream_t stream) {
#ifdef PADDLE_WITH_CUDA
  cudaStreamSynchronize(stream);
#endif
}

W
Wilber 已提交
2425
}  // namespace experimental
W
Wilber 已提交
2426
}  // namespace paddle_infer