analysis_predictor.cc 71.4 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"
Y
Yan Chunwei 已提交
30
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
31
#include "paddle/fluid/framework/ir/pass.h"
32
#include "paddle/fluid/framework/naive_executor.h"
33
#include "paddle/fluid/framework/op_proto_maker.h"
34
#include "paddle/fluid/framework/scope.h"
Y
Yan Chunwei 已提交
35
#include "paddle/fluid/framework/var_type_traits.h"
36
#include "paddle/fluid/framework/version.h"
37
#include "paddle/fluid/inference/analysis/helper.h"
Y
Yan Chunwei 已提交
38
#include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
39
#include "paddle/fluid/inference/api/helper.h"
40
#include "paddle/fluid/inference/api/paddle_inference_api.h"
L
luotao1 已提交
41
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
42
#include "paddle/fluid/inference/utils/io_utils.h"
43
#include "paddle/fluid/inference/utils/singleton.h"
44
#include "paddle/fluid/memory/memcpy.h"
45
#include "paddle/fluid/platform/cpu_helper.h"
46
#include "paddle/fluid/platform/device/gpu/gpu_info.h"
47
#include "paddle/fluid/platform/device_context.h"
48
#include "paddle/fluid/platform/place.h"
T
tensor-tang 已提交
49
#include "paddle/fluid/platform/profiler.h"
50
#include "paddle/phi/api/ext/op_meta_info.h"
W
Wilber 已提交
51
#include "paddle/phi/common/place.h"
52 53
#include "paddle/utils/string/split.h"

54
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
55 56 57 58
#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
59

60 61 62 63
#ifdef PADDLE_WITH_MKLML
#include "paddle/fluid/platform/dynload/mklml.h"
#endif

64 65 66 67
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/inference/api/mkldnn_quantizer.h"
#endif

68 69 70 71
#ifdef PADDLE_WITH_ONNXRUNTIME
#include "paddle/fluid/inference/api/onnxruntime_predictor.h"
#endif

Y
Yan Chunwei 已提交
72 73
#if PADDLE_WITH_TENSORRT
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
74
#include "paddle/fluid/inference/tensorrt/helper.h"
75
#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
Y
Yan Chunwei 已提交
76 77
#endif

78 79 80 81
#ifdef PADDLE_WITH_IPU
#include "paddle/fluid/platform/device/ipu/paddle_ipu_handler.h"
#endif

82 83
namespace paddle {

N
nhzlx 已提交
84
using inference::Singleton;
N
nhzlx 已提交
85
#if PADDLE_WITH_TENSORRT
N
nhzlx 已提交
86
using inference::tensorrt::TRTInt8Calibrator;
N
nhzlx 已提交
87 88
using inference::tensorrt::TRTCalibratorEngine;
using inference::tensorrt::TRTCalibratorEngineManager;
N
nhzlx 已提交
89
#endif
90

91 92
int AnalysisPredictor::clone_num_ = 1;

93 94 95 96
namespace {
bool IsPersistable(const framework::VarDesc *var) {
  if (var->Persistable() &&
      var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
97 98
      var->GetType() != framework::proto::VarType::FETCH_LIST &&
      var->GetType() != framework::proto::VarType::RAW) {
99 100 101 102 103 104
    return true;
  }
  return false;
}
}  // namespace

105 106
bool PaddleTensorToLoDTensor(const PaddleTensor &pt, framework::LoDTensor *t,
                             const platform::Place &place) {
107
  framework::DDim ddim = phi::make_ddim(pt.shape);
108 109 110 111 112 113 114
  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);
115 116
  } else if (pt.dtype == PaddleDType::FLOAT16) {
    input_ptr = t->mutable_data<float16>(ddim, place);
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134
  } 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.
    std::memcpy(static_cast<void *>(input_ptr), pt.data.data(),
                pt.data.length());
J
jianghaicheng 已提交
135 136 137 138 139 140 141 142
  } else if (platform::is_ipu_place(place)) {
#ifdef PADDLE_WITH_IPU
    std::memcpy(static_cast<void *>(input_ptr), pt.data.data(),
                pt.data.length());
#else
    PADDLE_THROW(paddle::platform::errors::Fatal(
        "Not compile with WITH_IPU, should not reach here."));
#endif
143 144 145 146
  } else if (platform::is_gpu_place(place)) {
    PADDLE_ENFORCE_EQ(platform::is_xpu_place(place), false,
                      platform::errors::InvalidArgument(
                          "Only one choice can be made between CPU and XPU."));
147
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
148 149 150
    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
    auto *dev_ctx =
        static_cast<const platform::CUDADeviceContext *>(pool.Get(place));
151
    auto dst_gpu_place = place;
152 153 154 155 156 157 158
    memory::Copy(dst_gpu_place, static_cast<void *>(input_ptr),
                 platform::CPUPlace(), pt.data.data(), pt.data.length(),
                 dev_ctx->stream());
#else
    PADDLE_THROW(paddle::platform::errors::Fatal(
        "Not compile with CUDA, should not reach here."));
#endif
159 160
  } else if (platform::is_xpu_place(place)) {
#ifdef PADDLE_WITH_XPU
161
    auto dst_xpu_place = place;
162 163 164 165 166 167 168 169 170
    memory::Copy(dst_xpu_place, static_cast<void *>(input_ptr),
                 platform::CPUPlace(), pt.data.data(), pt.data.length());
#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."));
171 172 173 174 175 176 177 178 179 180
  }
  // 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
181
bool AnalysisPredictor::Init(
182 183
    const std::shared_ptr<framework::Scope> &parent_scope,
    const std::shared_ptr<framework::ProgramDesc> &program) {
M
minqiyang 已提交
184
  VLOG(3) << "Predictor::init()";
185 186
  if (config_.with_profile_) {
    LOG(WARNING) << "Profiler is activated, which might affect the performance";
187 188
    auto tracking_device = config_.use_gpu() ? platform::ProfilerState::kAll
                                             : platform::ProfilerState::kCPU;
T
tensor-tang 已提交
189
    platform::EnableProfiler(tracking_device);
190
  } else {
191 192
    VLOG(2) << "Profiler is deactivated, and no profiling report will be "
               "generated.";
T
tensor-tang 已提交
193 194
  }

195
  // no matter with or without MKLDNN
L
luotao1 已提交
196
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
197

198 199 200 201 202 203 204 205 206 207
  if (!PrepareScope(parent_scope)) {
    return false;
  }
  if (!CreateExecutor()) {
    return false;
  }
  if (!PrepareProgram(program)) {
    return false;
  }

208 209 210
  // Get the feed_target_names and fetch_target_names
  PrepareFeedFetch();

211 212 213
  // Prepare executor, create local variables.
  if (!PrepareExecutor()) {
    return true;
Y
Yan Chunwei 已提交
214
  }
215 216 217 218 219 220

  return true;
}

bool AnalysisPredictor::PrepareScope(
    const std::shared_ptr<framework::Scope> &parent_scope) {
Y
Yan Chunwei 已提交
221
  if (parent_scope) {
222 223
    PADDLE_ENFORCE_NOT_NULL(
        parent_scope,
224 225
        platform::errors::PreconditionNotMet(
            "Both program and parent_scope should be set in Clone mode."));
Y
Yan Chunwei 已提交
226
    scope_ = parent_scope;
227
    status_is_cloned_ = true;
Y
Yan Chunwei 已提交
228
  } else {
229
    paddle::framework::InitDevices();
230
    paddle::framework::InitDefaultKernelSignatureMap();
W
Wilber 已提交
231 232
    // TODO(wilber): we need to release memory occupied by weights.
    scope_.reset(new paddle::framework::Scope());
233
    status_is_cloned_ = false;
Y
Yan Chunwei 已提交
234
  }
235 236 237 238 239
  sub_scope_ = &scope_->NewScope();
  return true;
}
bool AnalysisPredictor::PrepareProgram(
    const std::shared_ptr<framework::ProgramDesc> &program) {
240 241
  if (!program) {
    if (!LoadProgramDesc()) return false;
242 243 244 245 246 247 248 249 250
    // 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_);

251 252 253 254
    // if enable_ir_optim_ is false,
    // the analysis pass(op fuse, graph analysis, trt subgraph, mkldnn etc) will
    // not be executed.
    OptimizeInferenceProgram();
Y
Yan Chunwei 已提交
255
  } else {
256 257
    // If the program is passed from external, no need to optimize it, this
    // logic is used in the clone scenario.
258 259
    inference_program_ = program;
  }
M
Michal Gallus 已提交
260

261 262 263 264 265
  executor_->CreateVariables(*inference_program_, 0, false, sub_scope_);

  return true;
}
bool AnalysisPredictor::CreateExecutor() {
266
  if (config_.use_gpu()) {
267 268 269
    PADDLE_ENFORCE_EQ(config_.use_xpu(), false,
                      platform::errors::InvalidArgument(
                          "Only one choice can be made between CPU and XPU."));
270
    place_ = paddle::platform::CUDAPlace(config_.gpu_device_id());
271
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
272 273 274 275 276 277 278 279
    if (config_.thread_local_stream_enabled()) {
      auto *ctx = static_cast<platform::CUDADeviceContext *>(
          platform::DeviceContextPool::Instance().Get(place_));
      VLOG(3) << "The prediction process will be completed using a separate "
                 "normal-priority stream on each thread.";
      ctx->ResetThreadContext(platform::stream::Priority::kNormal);
    }
#endif
280
  } else if (config_.use_xpu()) {
281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
    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
304 305 306 307 308 309 310 311
  } 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
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
  } 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
328 329 330 331 332 333 334
  } 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."));
335 336 337 338 339 340 341 342 343
#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
344
#endif
345 346 347 348 349 350
  } else {
    place_ = paddle::platform::CPUPlace();
  }
  executor_.reset(new paddle::framework::NaiveExecutor(place_));
  return true;
}
W
wenbin 已提交
351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383

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(
    std::shared_ptr<framework::ProgramDesc> inference_program, int block,
    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");
      DisablePrepareDataOpt(inference_program, blockID,
                            disable_opt || pre_disable_opt);
    }
    // disable prepare data if unfriendly op is found
W
wenbin 已提交
384 385 386
    if (!disable_opt) {
      disable_opt = IsPrepareDataOptTargetOp(op);
    }
W
wenbin 已提交
387 388 389
  }
}

390
bool AnalysisPredictor::PrepareExecutor() {
391
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
392 393 394 395 396
  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
397 398
  DisablePrepareDataOpt(inference_program_, 0, false);

399
  executor_->Prepare(sub_scope_, *inference_program_, 0,
400
                     config_.use_feed_fetch_ops_);
401

402 403 404
  PADDLE_ENFORCE_NOT_NULL(sub_scope_,
                          platform::errors::PreconditionNotMet(
                              "The sub_scope should not be nullptr."));
Y
Yan Chunwei 已提交
405

406 407 408
  return true;
}

409
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627
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(),
                   *(inference_program_.get()), scope_.get(), place_, 1,
                   {task_node_.get()}, id_to_rank, feed_fetch_vars);
  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]);
    }
    InsertCommOp(var_name_base + std::to_string(order), ranks_in_group,
                 rank_in_group, peer_endpoints, comm_init_block, ring_id);
    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(
    std::string tmp_var_name, int nranks, int rank,
    const std::vector<std::string> &peer_endpoints, framework::BlockDesc *block,
    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(
      static_cast<bool>(fin.is_open()), true,
      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

628 629
void AnalysisPredictor::MkldnnPreSet(const std::vector<PaddleTensor> &inputs) {
#ifdef PADDLE_WITH_MKLDNN
W
Wilber 已提交
630 631 632 633 634 635 636 637 638 639 640 641
  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="
642
          << platform::MKLDNNDeviceContext::tls().get_cur_mkldnn_session_id();
643 644 645
  // In cache clearing mode.
  if (config_.mkldnn_cache_capacity_ > 0) {
    VLOG(2) << "In mkldnn cache clear mode.";
646 647 648
    platform::MKLDNNDeviceContext::tls().set_cur_mkldnn_session_id(
        platform::MKLDNNDeviceContextThreadLocals::
            kMKLDNNSessionID_CacheClearing);
649 650
    // Set current_input_shape for caching dynamic shape.
    std::stringstream ss;
W
Wilber 已提交
651 652 653
    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] << "-";
654 655 656
      }
    }
    VLOG(2) << "Set input shape=" << ss.str();
657
    platform::MKLDNNDeviceContext::tls().set_cur_input_shape_str(ss.str());
658
  }
659 660 661
  platform::MKLDNNDeviceContext::tls().set_cur_input_shape_cache_capacity(
      config_.mkldnn_cache_capacity_);

662 663 664 665 666 667
#endif
}

void AnalysisPredictor::MkldnnPostReset() {
#ifdef PADDLE_WITH_MKLDNN
  // In cache clearing mode.
668 669 670 671
  if (config_.mkldnn_cache_capacity_ > 0 &&
      static_cast<platform::MKLDNNDeviceContext *>(
          (&platform::DeviceContextPool::Instance())->Get(platform::CPUPlace()))
              ->GetCachedObjectsNumber() > 0) {
672 673 674 675 676 677 678 679
    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_));
    }
680 681 682
    // We cannot reset to the default cache settings
    // as there maybe CopyToCPU method used and oneDNN
    // primitives are used there so cache would grow
683 684 685 686
  }
#endif
}

687 688 689
bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
                            std::vector<PaddleTensor> *output_data,
                            int batch_size) {
690
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
691 692 693
#ifdef PADDLE_WITH_MKLDNN
  if (config_.use_mkldnn_) MkldnnPreSet(inputs);
#endif
M
minqiyang 已提交
694
  VLOG(3) << "Predictor::predict";
695 696 697 698
  inference::Timer timer;
  timer.tic();
  // set feed variable
  framework::Scope *scope = sub_scope_ ? sub_scope_ : scope_.get();
699 700
  PADDLE_ENFORCE_NOT_NULL(scope, platform::errors::PreconditionNotMet(
                                     "The scope should not be nullptr."));
701 702
  if (!SetFeed(inputs, scope)) {
    LOG(ERROR) << "fail to set feed";
Y
Yan Chunwei 已提交
703
    return false;
704
  }
M
Michal Gallus 已提交
705

706 707 708
  // Run the inference program
  // if share variables, we need not create variables
  executor_->Run();
709

710 711 712 713
  // get fetch variable
  if (!GetFetch(output_data, scope)) {
    LOG(ERROR) << "fail to get fetches";
    return false;
T
tensor-tang 已提交
714
  }
Y
Yan Chunwei 已提交
715

M
minqiyang 已提交
716
  VLOG(3) << "predict cost: " << timer.toc() << "ms";
Y
Yan Chunwei 已提交
717

Y
Yan Chunwei 已提交
718 719 720 721 722
  // 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.
723 724 725
  if (sub_scope_) {
    tensor_array_batch_cleaner_.CollectNoTensorVars(sub_scope_);
  }
Y
Yan Chunwei 已提交
726
  tensor_array_batch_cleaner_.ResetNoTensorVars();
727 728 729 730

  // 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);
731 732
#ifdef PADDLE_WITH_MKLDNN
  if (config_.use_mkldnn_) MkldnnPostReset();
T
Tao Luo 已提交
733
#endif
734
#if defined(PADDLE_WITH_MKLML)
T
Tao Luo 已提交
735 736 737 738
  // 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();
739
#endif
740 741
  return true;
}
742

743 744
bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                                framework::Scope *scope) {
M
minqiyang 已提交
745
  VLOG(3) << "Predictor::set_feed";
746 747 748 749 750 751 752 753 754 755
  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) {
756 757
    framework::LoDTensor *input = &feed_tensors_[i];
    if (!PaddleTensorToLoDTensor(inputs[i], input, place_)) {
758 759 760
      return false;
    }
    int idx = -1;
761
    if (config_.specify_input_name_) {
T
tensor-tang 已提交
762 763
      auto name = inputs[i].name;
      if (feed_names_.find(name) == feed_names_.end()) {
T
tensor-tang 已提交
764 765
        LOG(ERROR) << "feed names from program do not have name: [" << name
                   << "] from specified input";
T
tensor-tang 已提交
766 767
      }
      idx = feed_names_[name];
768
    } else {
769
      idx = BOOST_GET_CONST(int, feeds_[i]->GetAttr("col"));
770
    }
771
    framework::SetFeedVariable(scope, *input, "feed", idx);
772 773 774 775 776 777 778 779
  }
  return true;
}

template <typename T>
void AnalysisPredictor::GetFetchOne(const framework::LoDTensor &fetch,
                                    PaddleTensor *output) {
  // set shape.
780
  auto shape = phi::vectorize(fetch.dims());
781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797
  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
798
  VLOG(3) << "Predictor::get_fetch";
Y
Yan Chunwei 已提交
799 800
  outputs->resize(fetches_.size());
  for (size_t i = 0; i < fetches_.size(); ++i) {
801
    int idx = BOOST_GET_CONST(int, fetches_[i]->GetAttr("col"));
802 803 804 805 806
    PADDLE_ENFORCE_EQ(
        static_cast<size_t>(idx), i,
        platform::errors::InvalidArgument(
            "Fetch op's col attr(%d) should be equal to the index(%d)", idx,
            i));
807
    framework::FetchType &fetch_var =
808
        framework::GetFetchVariable(*scope, "fetch", idx);
809
    auto &fetch = BOOST_GET(framework::LoDTensor, fetch_var);
810
    auto type = framework::TransToProtoVarType(fetch.dtype());
811
    auto output = &(outputs->at(i));
Y
Yan Chunwei 已提交
812
    output->name = fetches_[idx]->Input("X")[0];
Y
Yu Yang 已提交
813
    if (type == framework::proto::VarType::FP32) {
814 815
      GetFetchOne<float>(fetch, output);
      output->dtype = PaddleDType::FLOAT32;
Y
Yu Yang 已提交
816
    } else if (type == framework::proto::VarType::INT64) {
817 818
      GetFetchOne<int64_t>(fetch, output);
      output->dtype = PaddleDType::INT64;
819 820 821
    } else if (type == framework::proto::VarType::INT32) {
      GetFetchOne<int32_t>(fetch, output);
      output->dtype = PaddleDType::INT32;
822 823 824
    } else if (type == framework::proto::VarType::FP16) {
      GetFetchOne<float16>(fetch, output);
      output->dtype = PaddleDType::FLOAT16;
825
    } else {
826 827
      LOG(ERROR) << "unknown type, only support float32, float16, int64 and "
                    "int32 now.";
828 829
    }
  }
Y
Yan Chunwei 已提交
830 831
  return true;
}
832

833
void AnalysisPredictor::PrepareArgument() {
834
  argument_.SetUseGPU(config_.use_gpu());
835
  argument_.SetUseFcPadding(config_.use_fc_padding());
836
  argument_.SetGPUDeviceId(config_.gpu_device_id());
837
  argument_.SetEnableAnalysisOptim(config_.enable_ir_optim_);
Y
Yan Chunwei 已提交
838
  argument_.SetEnableMemoryOptim(config_.enable_memory_optim());
T
Tao Luo 已提交
839
  argument_.SetModelFromMemory(config_.model_from_memory_);
Y
Yan Chunwei 已提交
840
  // Analyze inference_program
841
  argument_.SetPredictorID(predictor_id_);
842
  argument_.SetOptimCacheDir(config_.opt_cache_dir_);
843 844
  if (!config_.model_dir().empty()) {
    argument_.SetModelDir(config_.model_dir());
T
Tao Luo 已提交
845
  } else {
846 847 848
    PADDLE_ENFORCE_EQ(config_.prog_file().empty(), false,
                      platform::errors::PreconditionNotMet(
                          "Either model_dir or prog_file should be set."));
N
nhzlx 已提交
849
    std::string dir = inference::analysis::GetDirRoot(config_.prog_file());
N
nhzlx 已提交
850

851 852
    argument_.SetModelProgramPath(config_.prog_file());
    argument_.SetModelParamsPath(config_.params_file());
Y
Yan Chunwei 已提交
853
  }
854

855 856 857 858 859 860 861 862
  argument_.SetTensorRtPrecisionMode(config_.tensorrt_precision_mode_);
  argument_.SetTensorRtUseOSS(config_.trt_use_oss_);
  argument_.SetTensorRtWithInterleaved(config_.trt_with_interleaved_);
  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());
863
  if (config_.use_gpu() && config_.tensorrt_engine_enabled()) {
Y
Yan Chunwei 已提交
864
    LOG(INFO) << "TensorRT subgraph engine is enabled";
865 866 867
    argument_.SetUseTensorRT(true);
    argument_.SetTensorRtWorkspaceSize(config_.tensorrt_workspace_size_);
    argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
868
    argument_.SetTensorRtMinSubgraphSize(config_.tensorrt_min_subgraph_size_);
869
    argument_.SetTensorRtDisabledOPs(config_.trt_disabled_ops_);
870 871
    argument_.SetTensorRtUseDLA(config_.trt_use_dla_);
    argument_.SetTensorRtDLACore(config_.trt_dla_core_);
N
nhzlx 已提交
872
    argument_.SetTensorRtUseStaticEngine(config_.trt_use_static_engine_);
873
    argument_.SetTensorRtUseCalibMode(config_.trt_use_calib_mode_);
874
    argument_.SetCloseTrtPluginFp16(config_.disable_trt_plugin_fp16_);
875 876 877
    argument_.SetTensorRtShapeRangeInfoPath(config_.shape_range_info_path());
    argument_.SetTensorRtAllowBuildAtRuntime(
        config_.trt_allow_build_at_runtime());
878
    argument_.SetTensorRtUseInspector(config_.trt_use_inspector_);
W
Wojciech Uss 已提交
879
  }
880

D
denglin-github 已提交
881 882 883 884 885 886
  if (config_.dlnne_enabled()) {
    LOG(INFO) << "Dlnne subgraph is enabled";
    argument_.SetUseDlnne(true);
    argument_.SetDlnneMinSubgraphSize(config_.dlnne_min_subgraph_size_);
  }

887 888 889 890 891
  if (config_.gpu_fp16_enabled()) {
    argument_.SetUseGPUFp16(true);
    argument_.SetGpuFp16DisabledOpTypes(config_.gpu_fp16_disabled_op_types_);
  }

石晓伟 已提交
892
  if (config_.lite_engine_enabled()) {
W
Wilber 已提交
893 894
    argument_.SetCpuMathLibraryNumThreads(
        config_.cpu_math_library_num_threads());
石晓伟 已提交
895 896 897
    argument_.SetLitePrecisionMode(config_.lite_precision_mode_);
    argument_.SetLitePassesFilter(config_.lite_passes_filter_);
    argument_.SetLiteOpsFilter(config_.lite_ops_filter_);
898 899 900
    argument_.SetLiteZeroCopy(config_.lite_zero_copy_);
    argument_.SetUseXpu(config_.use_xpu_);
    argument_.SetXpuL3WorkspaceSize(config_.xpu_l3_workspace_size_);
W
Wilber 已提交
901 902 903 904 905
    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_);
906
    argument_.SetXpuDeviceId(config_.xpu_device_id_);
907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926
    // 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
927 928 929
    LOG(INFO) << "Lite subgraph engine is enabled";
  }

930
#ifdef PADDLE_WITH_IPU
J
jianghaicheng 已提交
931 932
  argument_.SetUseIpu(config_.use_ipu_);
  argument_.SetIpuDeviceNum(config_.ipu_device_num());
933
  argument_.SetIpuMicroBatchSize(config_.ipu_micro_batch_size_);
J
jianghaicheng 已提交
934 935
  argument_.SetIpuEnablePipelining(config_.ipu_enable_pipelining_);
  argument_.SetIpuBatchesPerStep(config_.ipu_batches_per_step_);
936 937 938 939 940 941
  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
942

943 944 945
  argument_.SetUseNpu(config_.use_npu_);
  argument_.SetNPUDeviceId(config_.npu_device_id());

946
  if (config_.use_mkldnn_) {
Y
Yan Chunwei 已提交
947
    LOG(INFO) << "MKLDNN is enabled";
948 949 950
    argument_.SetMKLDNNEnabledOpTypes(config_.mkldnn_enabled_op_types_);
  }

951 952 953 954 955 956 957 958
#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());
  }
959 960 961 962
  if (config_.use_mkldnn_bfloat16_) {
    LOG(INFO) << "Bfloat16 is enabled";
    argument_.SetBfloat16EnabledOpTypes(config_.bfloat16_enabled_op_types_);
  }
B
baoachun 已提交
963 964 965 966 967 968 969

  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({});
  }
970 971
#endif

972
  auto passes = config_.pass_builder()->AllPasses();
Y
Yan Chunwei 已提交
973 974 975 976
  if (!config_.ir_optim()) {
    passes.clear();
    LOG(INFO) << "ir_optim is turned off, no IR pass will be executed";
  }
977
  argument_.SetDisableLogs(config_.glog_info_disabled());
978
  argument_.SetIrAnalysisPasses(passes);
Y
Yan Chunwei 已提交
979
  argument_.SetAnalysisPasses(config_.pass_builder()->AnalysisPasses());
980
  argument_.SetScopeNotOwned(scope_.get());
981 982 983 984 985
}

// NOTE All the members in AnalysisConfig should be copied to Argument.
void AnalysisPredictor::OptimizeInferenceProgram() {
  PrepareArgument();
986 987
  Analyzer().Run(&argument_);

988 989 990
  PADDLE_ENFORCE_EQ(
      argument_.scope_valid(), true,
      platform::errors::InvalidArgument("The argument scope should be valid."));
991 992
  VLOG(5) << "to prepare executor";
  ARGUMENT_CHECK_FIELD((&argument_), ir_analyzed_program);
Y
Yan Chunwei 已提交
993
  inference_program_.reset(
994 995 996 997 998
      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
999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016
        auto &block = prog->Block(0);
        for (auto &op_desc : block.AllOps()) {
          if (op_desc->Type() == "tensorrt_engine") {
            std::string engine_key =
                BOOST_GET_CONST(std::string, op_desc->GetAttr("engine_key"));
            int engine_predictor_id =
                BOOST_GET_CONST(int, op_desc->GetAttr("predictor_id"));
            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);
            }
          }
        }
1017 1018 1019
#endif
        delete prog;
      });
1020 1021 1022 1023
  // The config and argument take a lot of storage,
  // when the predictor settings are complete, we release these stores.
  argument_.PartiallyRelease();
  config_.PartiallyRelease();
1024
  LOG(INFO) << "======= optimize end =======";
Y
Yan Chunwei 已提交
1025
}
1026 1027

template <>
1028 1029
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
    AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) {
W
Wilber 已提交
1030 1031
  // TODO(NHZlX): Should add the link to the doc of
  // paddle_infer::CreatePredictor<paddle_infer::Config>
P
Pei Yang 已提交
1032 1033 1034 1035
  if (config.glog_info_disabled()) {
    FLAGS_logtostderr = 1;
    FLAGS_minloglevel = 2;  // GLOG_ERROR
  }
M
minqiyang 已提交
1036
  VLOG(3) << "create AnalysisConfig";
1037 1038 1039 1040
  PADDLE_ENFORCE_EQ(
      config.is_valid(), true,
      platform::errors::InvalidArgument(
          "Note: Each config can only be used for one predictor."));
1041

1042 1043 1044 1045
  // 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,
1046
                 []() { inference::RegisterAllCustomOperator(); });
1047

1048
  if (config.use_gpu()) {
1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072
    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(
          config.memory_pool_init_size_mb(), 0.f,
          platform::errors::InvalidArgument(
              "The size of memory pool should be greater than 0."));
      PADDLE_ENFORCE_GE(
          config.gpu_device_id(), 0,
          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(...)";
      }
1073

1074 1075 1076 1077 1078 1079 1080
      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);
      }

1081 1082 1083 1084 1085 1086 1087 1088 1089
      // 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;
      }

J
JingZhuangzhuang 已提交
1090 1091 1092 1093 1094 1095
      // TODO(Jingzhuangzhuang): Fix trt error when allocator_strategy is
      // auto_growth
      if (config.tensorrt_engine_enabled()) {
        gflags.push_back("--allocator_strategy=naive_best_fit");
      }

1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110
      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) {
1111 1112 1113 1114 1115 1116
      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."));
1117 1118 1119 1120
    }
  }

  std::unique_ptr<PaddlePredictor> predictor(new AnalysisPredictor(config));
1121 1122
  // Each config can only be used for one predictor.
  config.SetInValid();
1123 1124 1125 1126 1127 1128 1129
  auto predictor_p = dynamic_cast<AnalysisPredictor *>(predictor.get());

  if (!predictor_p->Init(nullptr)) {
    return nullptr;
  }

  if (config.mkldnn_quantizer_enabled() && !predictor_p->MkldnnQuantize()) {
1130 1131
    return nullptr;
  }
1132

G
Gabor Buella 已提交
1133
  return predictor;
1134 1135
}

1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
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
}

1148
void AnalysisPredictor::PrepareFeedFetch() {
1149 1150 1151
  PADDLE_ENFORCE_NOT_NULL(sub_scope_,
                          platform::errors::InvalidArgument(
                              "The sub_scope should not be nullptr."));
1152
  CreateFeedFetchVar(sub_scope_);
1153 1154
  for (auto *op : inference_program_->Block(0).AllOps()) {
    if (op->Type() == "feed") {
1155
      int idx = BOOST_GET_CONST(int, op->GetAttr("col"));
1156 1157 1158 1159 1160
      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
1161
      idx2feeds_[idx] = op->Output("Out")[0];
1162
    } else if (op->Type() == "fetch") {
1163
      int idx = BOOST_GET_CONST(int, op->GetAttr("col"));
Y
Yan Chunwei 已提交
1164 1165
      if (fetches_.size() <= static_cast<size_t>(idx)) {
        fetches_.resize(idx + 1);
1166
      }
Y
Yan Chunwei 已提交
1167
      fetches_[idx] = op;
N
nhzlx 已提交
1168
      idx2fetches_[idx] = op->Input("X")[0];
1169 1170 1171 1172
    }
  }
}

1173
void AnalysisPredictor::CreateFeedFetchVar(framework::Scope *scope) {
1174 1175
  PADDLE_ENFORCE_NOT_NULL(scope, platform::errors::InvalidArgument(
                                     "The scope should not be nullptr."));
1176
  auto *var = scope->Var("feed");
1177
  var->GetMutable<framework::FeedList>();
1178
  var = scope->Var("fetch");
1179
  var->GetMutable<framework::FetchList>();
1180 1181
}

N
nhzlx 已提交
1182 1183 1184 1185 1186 1187 1188 1189
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;
}

1190 1191 1192 1193 1194 1195
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);
1196 1197
    PADDLE_ENFORCE_NOT_NULL(var, platform::errors::PreconditionNotMet(
                                     "Input %s does not exist.", name));
1198 1199 1200 1201 1202
    input_shapes[name] = var->GetShape();
  }
  return input_shapes;
}

N
nhzlx 已提交
1203 1204 1205 1206 1207 1208 1209 1210
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;
}

1211 1212
std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
    const std::string &name) {
1213
  framework::Scope *scope;
1214
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
1215 1216 1217 1218 1219 1220 1221 1222
  if (config_.dist_config().use_dist_model()) {
    scope = scope_.get();
  } else {
    scope = executor_->scope();
  }
#else
  scope = executor_->scope();
#endif
1223
  PADDLE_ENFORCE_NOT_NULL(
1224
      scope->FindVar(name),
1225
      platform::errors::PreconditionNotMet(
1226
          "The variable named %s is not found in the scope of the executor.",
1227
          name));
1228
  std::unique_ptr<ZeroCopyTensor> res(
1229
      new ZeroCopyTensor(static_cast<void *>(scope)));
1230 1231
  res->input_or_output_ = true;
  res->SetName(name);
N
nhzlx 已提交
1232 1233
  if (platform::is_cpu_place(place_)) {
    res->SetPlace(PaddlePlace::kCPU);
J
jianghaicheng 已提交
1234 1235 1236 1237
  } 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);
1238
  } else if (platform::is_xpu_place(place_)) {
1239 1240 1241 1242 1243 1244 1245 1246
    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 {
1247
      auto xpu_place = place_;
1248 1249
      res->SetPlace(PaddlePlace::kXPU, xpu_place.GetDeviceId());
    }
W
Wilber 已提交
1250
  } else if (platform::is_npu_place(place_)) {
1251
    auto npu_place = place_;
W
Wilber 已提交
1252
    res->SetPlace(PaddlePlace::kNPU, npu_place.GetDeviceId());
1253 1254 1255 1256 1257 1258
  } 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
1259
  } else {
1260
    auto gpu_place = place_;
N
nhzlx 已提交
1261 1262
    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
  }
1263 1264 1265 1266 1267
  return res;
}

std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
    const std::string &name) {
1268
  framework::Scope *scope;
1269
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
1270 1271 1272 1273 1274 1275 1276 1277
  if (config_.dist_config().use_dist_model()) {
    scope = scope_.get();
  } else {
    scope = executor_->scope();
  }
#else
  scope = executor_->scope();
#endif
1278
  PADDLE_ENFORCE_NOT_NULL(
1279
      scope->FindVar(name),
1280
      platform::errors::PreconditionNotMet(
1281
          "The variable named %s is not found in the scope of the executor.",
1282
          name));
1283
  std::unique_ptr<ZeroCopyTensor> res(
1284
      new ZeroCopyTensor(static_cast<void *>(scope)));
1285 1286
  res->input_or_output_ = false;
  res->SetName(name);
N
nhzlx 已提交
1287 1288
  if (platform::is_cpu_place(place_)) {
    res->SetPlace(PaddlePlace::kCPU);
J
jianghaicheng 已提交
1289 1290 1291 1292
  } 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);
1293
  } else if (platform::is_xpu_place(place_)) {
1294 1295 1296 1297 1298 1299 1300 1301
    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 {
1302
      auto xpu_place = place_;
1303 1304
      res->SetPlace(PaddlePlace::kXPU, xpu_place.GetDeviceId());
    }
W
Wilber 已提交
1305
  } else if (platform::is_npu_place(place_)) {
1306
    auto npu_place = place_;
W
Wilber 已提交
1307
    res->SetPlace(PaddlePlace::kNPU, npu_place.GetDeviceId());
1308 1309 1310 1311 1312 1313
  } 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
1314
  } else {
1315
    auto gpu_place = place_;
N
nhzlx 已提交
1316 1317
    res->SetPlace(PaddlePlace::kGPU, gpu_place.GetDeviceId());
  }
1318 1319 1320 1321
  return res;
}

bool AnalysisPredictor::ZeroCopyRun() {
1322
#if defined(PADDLE_WITH_DISTRIBUTE) && defined(PADDLE_WITH_PSCORE)
1323 1324 1325 1326 1327 1328 1329 1330 1331 1332
  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
1333
  paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
W
Wilber 已提交
1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344
#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
1345
  executor_->Run();
1346 1347 1348 1349 1350

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

Y
Yan Chunwei 已提交
1351
  // Fix TensorArray reuse not cleaned bug.
Y
Yan Chunwei 已提交
1352
  tensor_array_batch_cleaner_.CollectTensorArrays(sub_scope_);
Y
Yan Chunwei 已提交
1353
  tensor_array_batch_cleaner_.ResetTensorArray();
1354 1355 1356 1357

  // 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);
W
Wilber 已提交
1358 1359 1360
#ifdef PADDLE_WITH_MKLDNN
  if (config_.use_mkldnn_) MkldnnPostReset();
#endif
1361
#if defined(PADDLE_WITH_MKLML)
T
Tao Luo 已提交
1362 1363 1364 1365 1366
  // 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
1367 1368 1369
  return true;
}

W
Wilber 已提交
1370 1371 1372 1373 1374
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
bool AnalysisPredictor::ExpRunWithExternalStream(const gpuStream_t stream) {
  if (stream != nullptr) {
    paddle::platform::DeviceContextPool &pool =
        paddle::platform::DeviceContextPool::Instance();
1375
    auto gpu_place = place_;
W
Wilber 已提交
1376 1377 1378 1379 1380 1381 1382 1383
    auto *dev_ctx = reinterpret_cast<paddle::platform::CUDADeviceContext *>(
        pool.Get(gpu_place));
    dev_ctx->SetThreadLocalStream(stream);
  }
  return ZeroCopyRun();
}
#endif

1384 1385 1386 1387 1388 1389
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();
1390
    auto gpu_place = place_;
1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455
    auto *dev_ctx = static_cast<const paddle::platform::CUDADeviceContext *>(
        pool.Get(gpu_place));
#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(
          counter.begin(), counter.end(),
          [](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;
  }

  inference::SerializeShapeRangeInfo(config_.shape_range_info_path(),
                                     min_shapes, max_shapes, opt_shapes);
}

1456 1457
bool AnalysisPredictor::LoadProgramDesc() {
  // Initialize the inference program
1458
  std::string filename;
1459 1460
  if (!config_.model_dir().empty()) {
    filename = config_.model_dir() + "/__model__";
1461
  } else if (!config_.prog_file().empty()) {
1462 1463 1464
    // 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`.
1465
    filename = config_.prog_file();
1466
  } else {
1467
    if (config_.model_dir().empty() && config_.prog_file().empty()) {
1468 1469 1470 1471
      LOG(ERROR)
          << "Either model_dir or (prog_file, param_file) should be set.";
      return false;
    }
1472
    LOG(ERROR) << string::Sprintf(
1473 1474
        "not valid model path '%s' or program path '%s'.", config_.model_dir(),
        config_.params_file());
1475 1476
    return false;
  }
1477 1478 1479

  // Create ProgramDesc
  framework::proto::ProgramDesc proto;
T
Tao Luo 已提交
1480
  if (!config_.model_from_memory()) {
T
Tao Luo 已提交
1481 1482 1483
    std::string pb_content;
    // Read binary
    std::ifstream fin(filename, std::ios::in | std::ios::binary);
1484 1485 1486 1487 1488
    PADDLE_ENFORCE_EQ(
        static_cast<bool>(fin.is_open()), true,
        platform::errors::NotFound(
            "Cannot open file %s, please confirm whether the file is normal.",
            filename));
T
Tao Luo 已提交
1489 1490 1491 1492 1493 1494 1495 1496
    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 {
1497
    proto.ParseFromString(config_.prog_file());
T
Tao Luo 已提交
1498
  }
1499 1500 1501 1502 1503 1504
  inference_program_.reset(new framework::ProgramDesc(proto));
  return true;
}

bool AnalysisPredictor::LoadParameters() {
  PADDLE_ENFORCE_NOT_NULL(inference_program_.get(),
1505 1506
                          platform::errors::PreconditionNotMet(
                              "The inference program should be loaded first."));
T
Tao Luo 已提交
1507

1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527
  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);

1528
      if (!config_.params_file().empty()) {
1529 1530 1531 1532 1533 1534
        params.push_back(new_var->Name());
      } else {
        // append_op
        framework::OpDesc *op = load_block->AppendOp();
        op->SetType("load");
        op->SetOutput("Out", {new_var->Name()});
1535
        op->SetAttr("file_path", {config_.model_dir() + "/" + new_var->Name()});
1536 1537 1538 1539 1540
        op->CheckAttrs();
      }
    }
  }

1541
  if (!config_.params_file().empty()) {
1542 1543 1544 1545 1546 1547
    // 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);
1548
    op->SetAttr("file_path", {config_.params_file()});
1549 1550 1551 1552
    op->CheckAttrs();
  }

  // Use NaiveExecutor to Load parameters.
S
superjomn 已提交
1553
  framework::NaiveExecutor e(place_);
1554 1555 1556 1557
  e.Prepare(scope_.get(), *load_program, 0, false);
  e.Run();
  VLOG(3) << "get " << scope_->LocalVarNames().size() << " vars after load";

1558 1559
  return true;
}
1560

1561 1562 1563 1564 1565
uint64_t AnalysisPredictor::TryShrinkMemory() {
  ClearIntermediateTensor();
  return paddle::memory::Release(place_);
}

1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584
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
1585
#if PADDLE_WITH_TENSORRT
N
nhzlx 已提交
1586
bool AnalysisPredictor::SaveTrtCalibToDisk() {
1587 1588 1589
  PADDLE_ENFORCE_EQ(config_.tensorrt_engine_enabled(), true,
                    platform::errors::PreconditionNotMet(
                        "This func can be invoked only in trt mode"));
N
nhzlx 已提交
1590 1591 1592
  auto &block = inference_program_->Block(0);
  for (auto &op_desc : block.AllOps()) {
    if (op_desc->Type() == "tensorrt_engine") {
1593 1594
      std::string engine_name = BOOST_GET_CONST(
          std::string, op_desc->GetAttr("calibration_engine_key"));
N
nhzlx 已提交
1595
      if (!Singleton<TRTCalibratorEngineManager>::Global().Has(engine_name)) {
N
nhzlx 已提交
1596 1597 1598 1599
        LOG(ERROR) << "You should run the predictor(with trt) on the real data "
                      "to generate calibration info";
        return false;
      }
N
nhzlx 已提交
1600 1601
      TRTCalibratorEngine *calib_engine =
          Singleton<TRTCalibratorEngineManager>::Global().Get(engine_name);
N
nhzlx 已提交
1602
      LOG(INFO) << "Wait for calib threads done.";
N
nhzlx 已提交
1603
      calib_engine->calib_->waitAndSetDone();
N
nhzlx 已提交
1604 1605
      LOG(INFO) << "Generating TRT Calibration table data, this may cost a lot "
                   "of time...";
N
nhzlx 已提交
1606 1607 1608
      calib_engine->thr_->join();
      std::string calibration_table_data =
          calib_engine->calib_->getCalibrationTableAsString();
N
nhzlx 已提交
1609

N
nhzlx 已提交
1610
      if (calibration_table_data.empty()) {
N
nhzlx 已提交
1611 1612 1613
        LOG(ERROR) << "the calibration table is empty.";
        return false;
      }
N
nhzlx 已提交
1614

N
nhzlx 已提交
1615 1616 1617 1618 1619
      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
1620
      std::string calibration_table_data_path =
N
nhzlx 已提交
1621 1622 1623 1624
          inference::analysis::GetTrtCalibPath(
              inference::analysis::GetOrCreateModelOptCacheDir(
                  model_opt_cache_dir),
              engine_name);
N
nhzlx 已提交
1625 1626 1627 1628 1629

      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
1630 1631 1632 1633
      ofile.close();
    }
  }
  // Free all calibrator resources.
N
nhzlx 已提交
1634
  Singleton<TRTCalibratorEngineManager>::Global().DeleteALL();
N
nhzlx 已提交
1635 1636
  return true;
}
N
nhzlx 已提交
1637
#endif
N
nhzlx 已提交
1638

1639
AnalysisPredictor::~AnalysisPredictor() {
N
nhzlx 已提交
1640
#if PADDLE_WITH_TENSORRT
N
nhzlx 已提交
1641
  if (config_.tensorrt_engine_enabled() &&
N
nhzlx 已提交
1642 1643
      config_.tensorrt_precision_mode_ == AnalysisConfig::Precision::kInt8 &&
      Singleton<TRTCalibratorEngineManager>::Global().Has()) {
N
nhzlx 已提交
1644 1645
    SaveTrtCalibToDisk();
  }
N
nhzlx 已提交
1646
#endif
1647
  if (config_.with_profile_) {
1648 1649 1650 1651 1652 1653
    platform::DisableProfiler(platform::EventSortingKey::kTotal,
                              "./profile.log");
  }
  if (sub_scope_) {
    scope_->DeleteScope(sub_scope_);
  }
Y
Yan Chunwei 已提交
1654

1655 1656 1657 1658 1659 1660
#if PADDLE_WITH_MKLDNN
  if (mkldnn_quantizer_) {
    delete mkldnn_quantizer_;
    mkldnn_quantizer_ = nullptr;
  }
#endif
1661

1662 1663 1664 1665
  if (config_.shape_range_info_collected()) {
    StatisticShapeRangeInfo();
  }

W
Wilber 已提交
1666 1667 1668
  if (place_.GetType() != phi::AllocationType::UNDEFINED) {
    memory::Release(place_);
  }
1669 1670
}

1671
std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
Y
Yan Chunwei 已提交
1672
  std::lock_guard<std::mutex> lk(clone_mutex_);
1673 1674
  auto *x = new AnalysisPredictor(config_);
  x->Init(scope_, inference_program_);
1675
  x->executor_->ResetTrtOps(++AnalysisPredictor::clone_num_);
1676 1677 1678
  return std::unique_ptr<PaddlePredictor>(x);
}

1679
std::string AnalysisPredictor::GetSerializedProgram() const {
Y
Yan Chunwei 已提交
1680 1681 1682
  return inference_program_->Proto()->SerializeAsString();
}

1683 1684 1685 1686 1687 1688 1689 1690 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 1716 1717 1718 1719 1720 1721
// 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
1722
template <>
1723 1724
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<AnalysisConfig>(
    const AnalysisConfig &config) {
W
Wilber 已提交
1725
  LOG(WARNING) << "Deprecated. Please use CreatePredictor instead.";
1726 1727
  return CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
      config);
Y
Yan Chunwei 已提交
1728 1729
}

1730
}  // namespace paddle
1731 1732 1733

#if PADDLE_WITH_TENSORRT
USE_TRT_CONVERTER(elementwise_add_weight);
S
shentanyue 已提交
1734 1735 1736 1737
USE_TRT_CONVERTER(elementwise_sub_weight);
USE_TRT_CONVERTER(elementwise_mul_weight);
USE_TRT_CONVERTER(elementwise_div_weight);
USE_TRT_CONVERTER(elementwise_pow_weight);
1738 1739 1740 1741 1742 1743 1744
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);
1745 1746
USE_TRT_CONVERTER(transpose);
USE_TRT_CONVERTER(flatten);
1747
USE_TRT_CONVERTER(flatten_contiguous_range);
1748
USE_TRT_CONVERTER(matmul);
1749 1750
USE_TRT_CONVERTER(conv2d);
USE_TRT_CONVERTER(relu);
Z
zhupengyang 已提交
1751 1752
USE_TRT_CONVERTER(exp);
USE_TRT_CONVERTER(log);
1753 1754 1755 1756 1757 1758 1759 1760 1761
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);
1762 1763
USE_TRT_CONVERTER(hard_sigmoid);
USE_TRT_CONVERTER(hard_swish);
1764
USE_TRT_CONVERTER(split);
1765 1766
USE_TRT_CONVERTER(prelu);
USE_TRT_CONVERTER(conv2d_transpose);
H
hjchen2 已提交
1767
USE_TRT_CONVERTER(leaky_relu);
1768 1769
USE_TRT_CONVERTER(shuffle_channel);
USE_TRT_CONVERTER(swish);
1770
USE_TRT_CONVERTER(group_norm);
1771
USE_TRT_CONVERTER(instance_norm);
P
Pei Yang 已提交
1772 1773 1774
USE_TRT_CONVERTER(layer_norm);
USE_TRT_CONVERTER(gelu);
USE_TRT_CONVERTER(multihead_matmul);
1775 1776
USE_TRT_CONVERTER(fused_embedding_eltwise_layernorm);
USE_TRT_CONVERTER(skip_layernorm);
1777
USE_TRT_CONVERTER(slice);
1778
USE_TRT_CONVERTER(scale);
1779
USE_TRT_CONVERTER(stack);
P
Pei Yang 已提交
1780
USE_TRT_CONVERTER(clip);
1781
USE_TRT_CONVERTER(gather);
1782
USE_TRT_CONVERTER(anchor_generator);
Z
zlsh80826 已提交
1783
USE_TRT_CONVERTER(yolo_box);
1784
USE_TRT_CONVERTER(yolo_box_head);
1785
USE_TRT_CONVERTER(roi_align);
1786
USE_TRT_CONVERTER(affine_channel);
Z
zlsh80826 已提交
1787
USE_TRT_CONVERTER(multiclass_nms);
1788
USE_TRT_CONVERTER(multiclass_nms3);
1789
USE_TRT_CONVERTER(nearest_interp);
1790
USE_TRT_CONVERTER(nearest_interp_v2);
W
Wangzheee 已提交
1791
USE_TRT_CONVERTER(reshape);
1792 1793
USE_TRT_CONVERTER(reduce_sum);
USE_TRT_CONVERTER(gather_nd);
W
wenbin 已提交
1794
USE_TRT_CONVERTER(reduce_mean);
W
wenbin 已提交
1795
USE_TRT_CONVERTER(tile);
W
wenbin 已提交
1796 1797
USE_TRT_CONVERTER(conv3d);
USE_TRT_CONVERTER(conv3d_transpose);
W
wangxinxin08 已提交
1798
USE_TRT_CONVERTER(mish);
W
wangxinxin08 已提交
1799
USE_TRT_CONVERTER(deformable_conv);
F
feng_shuai 已提交
1800
USE_TRT_CONVERTER(pool3d)
1801 1802
USE_TRT_CONVERTER(fused_preln_embedding_eltwise_layernorm)
USE_TRT_CONVERTER(preln_skip_layernorm)
F
feng_shuai 已提交
1803
USE_TRT_CONVERTER(roll)
F
feng_shuai 已提交
1804
USE_TRT_CONVERTER(strided_slice)
1805
#endif
W
Wilber 已提交
1806 1807 1808 1809 1810 1811

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
1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832
  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 {
      predictor_ = paddle::CreatePaddlePredictor<
          Config, paddle::PaddleEngineKind::kONNXRuntime>(config);
      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
  }
W
Wilber 已提交
1833 1834 1835 1836 1837 1838 1839 1840 1841
  predictor_ = paddle::CreatePaddlePredictor<
      Config, paddle::PaddleEngineKind::kAnalysis>(config);
}

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

std::unique_ptr<Tensor> Predictor::GetInputHandle(const std::string &name) {
1842
  return predictor_->GetInputTensor(name);
W
Wilber 已提交
1843 1844 1845 1846 1847 1848 1849
}

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

std::unique_ptr<Tensor> Predictor::GetOutputHandle(const std::string &name) {
1850
  return predictor_->GetOutputTensor(name);
W
Wilber 已提交
1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864
}

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

std::unique_ptr<Predictor> Predictor::Clone() {
  auto analysis_pred = predictor_->Clone();
  std::unique_ptr<Predictor> pred(new Predictor(std::move(analysis_pred)));
  return pred;
}

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

1865 1866
uint64_t Predictor::TryShrinkMemory() { return predictor_->TryShrinkMemory(); }

W
Wilber 已提交
1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884
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(); }

1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900
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
1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944
std::string UpdateDllFlag(const char *name, const char *value) {
  return paddle::UpdateDllFlag(name, value);
}

}  // 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(
      size, 1UL,
      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(
      idx, preds_.size() + 1,
      paddle::platform::errors::InvalidArgument(
          "There are (%d) predictors in the pool, but the idx is (%d)", idx,
          preds_.size() + 1));
  if (idx == 0) {
    return main_pred_.get();
  }
  return preds_[idx - 1].get();
}
}  // namespace services
W
Wilber 已提交
1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964

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
1965

1966 1967 1968 1969 1970 1971
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
1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988

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();
  auto *dev_ctx = reinterpret_cast<paddle::platform::CUDADeviceContext *>(
      pool.Get(pred->place_));
  cudaStreamSynchronize(dev_ctx->stream());
#endif
}
void InternalUtils::SyncStream(cudaStream_t stream) {
#ifdef PADDLE_WITH_CUDA
  cudaStreamSynchronize(stream);
#endif
}

W
Wilber 已提交
1989
}  // namespace experimental
W
Wilber 已提交
1990
}  // namespace paddle_infer