ngraph_engine.cc 22.6 KB
Newer Older
B
baojun 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
/* 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. */

#include <glog/logging.h>

#include <algorithm>
#include <map>
19
#include <memory>
B
baojun 已提交
20
#include <string>
21 22
#include <unordered_set>
#include <utility>
B
baojun 已提交
23 24 25 26 27 28 29 30 31 32
#include <vector>

#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/var_type.h"
33
#include "paddle/fluid/operators/ngraph/ngraph_bridge.h"
B
baojun 已提交
34 35 36 37 38 39 40 41 42 43
#include "paddle/fluid/operators/ngraph/ngraph_engine.h"

namespace paddle {
namespace operators {

static ngraph::Shape Ddim2Shape(const framework::DDim& dims) {
  ngraph::Shape sp;
  for (int i = 0; i < dims.size(); ++i) {
    int k = dims[i];
    k = k == 0 ? 1 : k;
44
    sp.emplace_back(k);
B
baojun 已提交
45 46 47 48
  }
  return sp;
}

49 50 51 52 53 54 55 56 57
static framework::DDim Shape2Ddim(const ngraph::Shape& shape) {
  std::vector<int64_t> dims;
  for (size_t i = 0; i < shape.size(); ++i) {
    int64_t k = shape[i];
    dims.emplace_back(k);
  }
  return framework::make_ddim(dims);
}

B
baojun 已提交
58 59 60 61 62 63
static std::map<framework::proto::VarType::Type, ngraph::element::Type>
    pd2ng_type_map = {
        {framework::proto::VarType::FP32, ngraph::element::f32},
        {framework::proto::VarType::FP64, ngraph::element::f64},
        {framework::proto::VarType::INT32, ngraph::element::i32},
        {framework::proto::VarType::INT64, ngraph::element::i64},
64
        {framework::proto::VarType::UINT8, ngraph::element::u8},
65 66 67 68 69 70 71 72
        {framework::proto::VarType::BOOL, ngraph::element::boolean}};

static std::map<ngraph::element::Type, framework::proto::VarType::Type>
    ng2pd_type_map = {
        {ngraph::element::f32, framework::proto::VarType::FP32},
        {ngraph::element::f64, framework::proto::VarType::FP64},
        {ngraph::element::i32, framework::proto::VarType::INT32},
        {ngraph::element::i64, framework::proto::VarType::INT64},
73
        {ngraph::element::u8, framework::proto::VarType::UINT8},
74 75 76 77 78 79
        {ngraph::element::boolean, framework::proto::VarType::BOOL}};

std::vector<std::string> NgraphEngine::feed_vars = {};
std::vector<std::string> NgraphEngine::fetch_vars = {};
framework::Variable* NgraphEngine::pre_var_ptr = nullptr;
const framework::BlockDesc* NgraphEngine::p_bdesc = nullptr;
80
bool NgraphEngine::is_training = false;
81 82 83 84 85

std::unordered_map<std::string, EngineCache> NgraphEngine::engine_cache = {};
std::unordered_map<std::string,
                   std::vector<std::shared_ptr<ngraph::runtime::Tensor>>>
    NgraphEngine::t_in_cache_ = {};
B
baojun 已提交
86 87 88 89 90

std::shared_ptr<ngraph::runtime::Backend> NgraphEngine::backend_ =
    ngraph::runtime::Backend::create("CPU");

static std::vector<std::vector<int>> NgraphOpIntervals(
91 92 93
    std::vector<std::unique_ptr<framework::OperatorBase>>* ops) {
  NgraphEngine::feed_vars.clear();
  NgraphEngine::fetch_vars.clear();
B
baojun 已提交
94
  std::vector<std::vector<int>> intervals;
95 96

  int size = ops->size();
B
baojun 已提交
97
  int left = 0;
B
baojun 已提交
98
  while (left < size && ops->at(left)->Type() != framework::kFeedOpType &&
99
         ops->at(left)->Type() != "read" &&
B
baojun 已提交
100
         ops->at(left)->Type() != framework::kFetchOpType) {
B
baojun 已提交
101 102
    ++left;
  }
103

104 105
  while (left < size && (ops->at(left)->Type() == framework::kFeedOpType ||
                         ops->at(left)->Type() == "read")) {
106 107 108 109 110
    for (auto& var_name_item : ops->at(left)->Outputs()) {
      for (auto& var_name : var_name_item.second) {
        NgraphEngine::feed_vars.emplace_back(var_name);
      }
    }
B
baojun 已提交
111 112 113 114
    ++left;
  }

  int right = left;
115
  while (right < size && ops->at(right)->Type() != framework::kFetchOpType) {
B
baojun 已提交
116 117 118
    ++right;
  }

119 120 121 122 123 124 125 126 127 128
  int index = right;
  while (index < size && ops->at(index)->Type() == framework::kFetchOpType) {
    for (auto& var_name_item : ops->at(index)->Inputs()) {
      for (auto& var_name : var_name_item.second) {
        NgraphEngine::fetch_vars.emplace_back(var_name);
      }
    }
    ++index;
  }

B
baojun 已提交
129 130 131 132
  if (left == size || ops->at(left)->Type() == framework::kFetchOpType) {
    left = 0;
  }

B
baojun 已提交
133 134 135
  // (left, right - 1) represents indices between feed and fetch
  int pivot = left;
  while (pivot < right) {
136
    auto op_type = ops->at(pivot)->Type();
137
    if (!NgraphBridge::isSupported(ops->at(pivot))) {
B
baojun 已提交
138 139 140
      ++pivot;
    } else {
      int start = pivot, end = start;
141
      while (pivot < right && (NgraphBridge::isSupported(ops->at(pivot)))) {
B
baojun 已提交
142 143 144 145
        ++pivot;
        ++end;
      }
      std::vector<int> interval = {start, end};
146
      intervals.emplace_back(interval);
B
baojun 已提交
147 148 149 150 151
    }
  }  // end while
  return intervals;
}

152 153 154 155 156 157 158 159
static void SubstituteNgraphOp(
    std::vector<std::unique_ptr<framework::OperatorBase>>* ops,
    std::string engine_key, std::string block_str, std::vector<int> interval) {
  framework::OpDesc ng_op_desc(nullptr);
  ng_op_desc.SetType("ngraph_engine");
  ng_op_desc.SetAttr("interval", interval);
  ng_op_desc.SetAttr("engine_key", engine_key);
  ng_op_desc.SetAttr("graph", block_str);
160 161
  ng_op_desc.SetInput("Xs", std::vector<std::string>(0));
  ng_op_desc.SetOutput("Ys", std::vector<std::string>(0));
162 163 164 165

  ops->erase(ops->begin() + interval[0], ops->begin() + interval[1]);
  ops->insert(ops->begin() + interval[0],
              framework::OpRegistry::CreateOp(ng_op_desc));
B
baojun 已提交
166 167
}

168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219
std::string SerializedBlock(const std::vector<framework::OpDesc*>& op_descs) {
  framework::proto::BlockDesc block_proto;
  framework::BlockDesc block_desc(nullptr, &block_proto);
  block_desc.Proto()->set_parent_idx(-1);
  block_desc.Proto()->set_idx(0);

  for (auto* op_desc : op_descs) {
    auto* op = block_desc.AppendOp();
    *op->Proto() = *op_desc->Proto();
  }
  return block_desc.Proto()->SerializeAsString();
}

std::string GenerateEngineKey(const framework::BlockDesc& bdesc) {
  framework::proto::BlockDesc block_proto;
  framework::BlockDesc block_desc(nullptr, &block_proto);
  block_desc.Proto()->set_parent_idx(-1);
  block_desc.Proto()->set_idx(0);

  for (auto& op_desc : bdesc.AllOps()) {
    auto* op = block_desc.AppendOp();
    *op->Proto() = *op_desc->Proto();
  }
  auto engine_key = std::to_string(
      std::hash<std::string>()(block_desc.Proto()->SerializeAsString()));
  return engine_key;
}

std::string GenerateEngineKey(const std::vector<std::string>& engine_inputs,
                              const std::vector<std::string>& engine_outputs,
                              int size) {
  std::string engine_hash_key = "";
  for (auto name : engine_inputs) {
    engine_hash_key += name;
  }
  for (auto name : engine_outputs) {
    engine_hash_key += name;
  }
  engine_hash_key += std::to_string(size);
  auto engine_key = std::to_string(std::hash<std::string>()(engine_hash_key));
  return engine_key;
}

void NgraphEngine::FuseNgraphOps(
    const framework::BlockDesc& block_desc,
    std::vector<std::unique_ptr<framework::OperatorBase>>* ops) {
  NgraphEngine::p_bdesc = &block_desc;
  auto intervals = NgraphOpIntervals(ops);
  std::string engine_key =
      GenerateEngineKey(feed_vars, fetch_vars, ops->size());
  for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
    SubstituteNgraphOp(ops, engine_key, "", *it);
B
baojun 已提交
220 221 222 223 224
  }
}

NgraphEngine::NgraphEngine(const framework::Scope& scope,
                           const platform::Place& place,
225
                           const framework::ExecutionContext& ctx)
B
baojun 已提交
226 227 228 229 230 231 232
    : scope_(scope), place_(place) {
  var_in_node_map_ = std::make_shared<
      std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>();

  var_node_map_ = std::make_shared<
      std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>();

233
  GetNgFunction(ctx);
B
baojun 已提交
234 235
}

236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256
void NgraphEngine::Prepare(const framework::ExecutionContext& ctx) {
  auto interval = ctx.Attr<std::vector<int>>("interval");
  std::string serialized_graph = ctx.Attr<std::string>("graph");

  auto input_vars = ctx.Inputs("Xs");
  if (!input_vars.empty()) {
    feed_vars = input_vars;
    var_in_ = input_vars;
  }
  auto output_vars = ctx.Outputs("Ys");
  if (!output_vars.empty()) {
    var_out_ = output_vars;
  }

  framework::proto::BlockDesc block_proto;
  if (!serialized_graph.empty()) block_proto.ParseFromString(serialized_graph);
  framework::BlockDesc block_desc(nullptr, &block_proto);
  if (!serialized_graph.empty()) {
    NgraphEngine::p_bdesc = &block_desc;
  }

B
baojun 已提交
257
  bool has_fetch = false, is_full = false;
258
  for (auto& var : p_bdesc->AllVars()) {
B
baojun 已提交
259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
    if (!(var->GetType() == framework::proto::VarType::SELECTED_ROWS ||
          var->GetType() == framework::proto::VarType::LOD_TENSOR ||
          var->GetType() == framework::proto::VarType::LOD_TENSOR_ARRAY)) {
      continue;
    }

    auto var_name = var->Name();
    if (var->Name() == framework::kEmptyVarName) {
      continue;
    }

    if (var_name != framework::kFeedOpType &&
        var_name != framework::kFetchOpType) {
      auto pd_type = var->GetDataType();
      if (pd2ng_type_map.find(pd_type) == pd2ng_type_map.end()) {
        PADDLE_THROW("Data type of var %s not found in pd2ng_type_map",
                     var_name);
      }
      var_type_map_[var_name] = pd2ng_type_map[pd_type];
    }

    if (var->Persistable()) {
      persistables_.insert(var->Name());
    }
  }

285 286 287
  std::vector<paddle::framework::OpDesc*> ops_desc;
  for (auto op_desc : p_bdesc->AllOps()) {
    ops_desc.emplace_back(op_desc);
B
baojun 已提交
288 289 290
    if (op_desc->Type() == framework::kFetchOpType) {
      has_fetch = true;
    }
B
baojun 已提交
291 292
  }

293
  for (auto op_desc : ops_desc) {
B
baojun 已提交
294
    if (op_desc->Type().find("_grad") != std::string::npos) {
295
      is_training = true;
296
      this->is_test_ = false;
B
baojun 已提交
297 298 299 300
      break;
    }
  }

301 302 303
  if (interval[0] > 0 &&
      ops_desc.at(interval[0] - 1)->Type() == framework::kFeedOpType &&
      interval[1] < static_cast<int>(ops_desc.size()) &&
B
baojun 已提交
304 305
      ops_desc.at(interval[1])->Type() == framework::kFetchOpType) {
    is_full = true;
B
baojun 已提交
306 307
  }

B
baojun 已提交
308
  if (is_full) {
309 310 311 312
    this->op_state_ = this->is_test_ ? OpState::FULL_TEST : OpState::FULL_TRAIN;
  } else {
    this->op_state_ =
        this->is_test_ ? OpState::PARTIAL_TEST : OpState::PARTIAL_TRAIN;
B
baojun 已提交
313 314
  }

315 316 317 318 319 320
  int idx = interval[0];
  while (idx < interval[1]) {
    this->fused_ops_.emplace_back(
        framework::OpRegistry::CreateOp(*(ops_desc[idx])));
    ++idx;
  }
B
baojun 已提交
321 322
  while (idx < static_cast<int>(ops_desc.size()) &&
         ops_desc.at(idx)->Type() != framework::kFetchOpType) {
323 324
    auto op_desc = ops_desc.at(idx);
    for (auto& var_name_item : op_desc->Inputs()) {
B
baojun 已提交
325
      for (auto& var_name : var_name_item.second) {
326
        this->post_op_inputs_.insert(var_name);
B
baojun 已提交
327 328
      }
    }
329
    ++idx;
B
baojun 已提交
330
  }
331

B
baojun 已提交
332 333 334 335
  if (!has_fetch) {
    op_state_ = OpState::UNKNOWN;
  }

336 337 338 339 340 341 342 343 344
  if (var_in_.empty() && var_out_.empty()) {
    BuildNgIO(ops_desc, interval);
  }
  for (size_t i = 0; i < var_in_.size(); ++i) {
    auto var_name = var_in_[i];
    if (persistables_.find(var_name) == persistables_.end()) {
      var_in_updates_.emplace_back(i);
    }
  }
B
baojun 已提交
345 346
}

347 348
void NgraphEngine::BuildNgIO(const std::vector<framework::OpDesc*>& ops_desc,
                             const std::vector<int>& interval) {
B
baojun 已提交
349 350
  std::unordered_set<std::string> inputs;
  std::unordered_set<std::string> outputs;
351 352
  for (int i = interval[0]; i < interval[1]; ++i) {
    auto op = ops_desc[i];
B
baojun 已提交
353 354 355 356 357 358
    for (auto& var_name_item : op->Inputs()) {
      for (auto& var_name : var_name_item.second) {
        inputs.insert(var_name);
        const bool is_output = outputs.find(var_name) != outputs.end();
        if (!is_output &&
            std::find(var_in_.begin(), var_in_.end(), var_name) ==
359 360
                var_in_.end() &&
            scope_.FindVar(var_name)) {
B
baojun 已提交
361
          // fill var_in here to keep lhs and rhs order
362
          this->var_in_.emplace_back(var_name);
B
baojun 已提交
363 364 365 366 367 368 369 370 371 372 373 374 375 376 377
        }
      }
    }

    for (auto& var_name_item : op->Outputs()) {
      PADDLE_ENFORCE_LE(var_name_item.second.size(), 1,
                        "op %s has more than 1 output - Not handling yet",
                        op->Type());
      for (auto& var_name : var_name_item.second) {
        outputs.insert(var_name);
      }
    }
  }

  // var_out.clear();
378 379
  for (int i = interval[0]; i < interval[1]; ++i) {
    auto op = ops_desc[i];
B
baojun 已提交
380 381 382 383 384
    for (auto& var_name_item : op->Outputs()) {
      PADDLE_ENFORCE_LE(var_name_item.second.size(), 1,
                        "op %s has more than 1 output - Not handling yet",
                        op->Type());
      for (auto& var_name : var_name_item.second) {
385
        switch (this->op_state_) {
B
baojun 已提交
386 387
          case OpState::PARTIAL_TEST:
            if (post_op_inputs_.find(var_name) != post_op_inputs_.end() ||
388 389 390
                find(fetch_vars.begin(), fetch_vars.end(), var_name) !=
                    fetch_vars.end()) {
              this->var_out_.emplace_back(var_name);
B
baojun 已提交
391 392 393
            }
            break;
          case OpState::FULL_TEST:
394 395 396
            if (find(fetch_vars.begin(), fetch_vars.end(), var_name) !=
                fetch_vars.end()) {
              this->var_out_.emplace_back(var_name);
B
baojun 已提交
397 398 399
            }
            break;
          case OpState::PARTIAL_TRAIN:
400 401
            if (find(fetch_vars.begin(), fetch_vars.end(), var_name) !=
                    fetch_vars.end() ||
B
baojun 已提交
402 403
                post_op_inputs_.find(var_name) != post_op_inputs_.end() ||
                persistables_.find(var_name) != persistables_.end()) {
404
              this->var_out_.emplace_back(var_name);
B
baojun 已提交
405 406 407
            }
            break;
          case OpState::FULL_TRAIN:
408 409
            if (find(fetch_vars.begin(), fetch_vars.end(), var_name) !=
                    fetch_vars.end() ||
B
baojun 已提交
410
                persistables_.find(var_name) != persistables_.end()) {
411
              this->var_out_.emplace_back(var_name);
B
baojun 已提交
412 413 414
            }
            break;
          default:
415
            this->var_out_.emplace_back(var_name);
B
baojun 已提交
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
void NgraphEngine::GetNgInputShape() {
  for (auto& var_name : var_in_) {
    auto* var = scope_.FindVar(var_name);
    if (var && var->IsType<framework::LoDTensor>()) {
      auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
      auto sp = Ddim2Shape(tensor_pd->dims());
      auto ng_type = var_type_map_[var_name];
      auto prm = std::make_shared<ngraph::op::Parameter>(ng_type, sp, true);
      (*var_node_map_)[var_name] = prm;
      (*var_in_node_map_)[var_name] = prm;
    }
  }
}

void NgraphEngine::BuildNgNodes() {
  for (auto& op : fused_ops_) {
    for (auto& var_name_item : op->Outputs()) {
      for (auto& var_name : var_name_item.second) {
        if (var_node_map_->find(var_name) == var_node_map_->end()) {
          auto* var = scope_.FindVar(var_name);
          if (var && var->IsType<framework::LoDTensor>()) {
            auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
            auto& ddim = tensor_pd->dims();
            auto ng_shape = Ddim2Shape(ddim);
            auto ng_type = var_type_map_[var_name];
            auto prm = std::make_shared<ngraph::op::Parameter>(ng_type,
                                                               ng_shape, true);
            (*var_node_map_)[var_name] = prm;
          }
        }
      }
    }
  }
  NgraphBridge ngb(var_node_map_);
  for (auto& op : fused_ops_) {
    ngb.BuildNgNode(op);
  }
}

void NgraphEngine::RunInferShape() {
  for (auto& op : fused_ops_) {
    framework::RuntimeContext ctx(op->Inputs(), op->Outputs(), scope_);
    op->RuntimeInferShape(scope_, place_, ctx);
  }
}

468 469
void NgraphEngine::BuildNgFunction(const framework::ExecutionContext& ctx) {
  Prepare(ctx);
470 471
  RunInferShape();
  GetNgInputShape();
B
baojun 已提交
472 473 474 475 476 477
  BuildNgNodes();
  ngraph_function_ = nullptr;
  ngraph::NodeVector func_outputs;
  ngraph::ParameterVector func_inputs;

  for (auto& vo : var_out_) {
478
    func_outputs.emplace_back(var_node_map_->at(vo));
B
baojun 已提交
479 480 481 482 483 484
  }

  for (auto& vi : var_in_) {
    std::shared_ptr<ngraph::op::Parameter> prm =
        std::dynamic_pointer_cast<ngraph::op::Parameter>(
            var_in_node_map_->at(vi));
485
    func_inputs.emplace_back(prm);
B
baojun 已提交
486 487 488 489 490 491
  }

  ngraph_function_ =
      std::make_shared<ngraph::Function>(func_outputs, func_inputs);
}

492 493 494
void NgraphEngine::GetNgFunction(const framework::ExecutionContext& ctx) {
  auto interval = ctx.Attr<std::vector<int>>("interval");
  std::string engine_key = ctx.Attr<std::string>("engine_key");
495 496 497
  bool use_cache = true;
  if (use_cache) {
    this->func_cache_key_ = "";
498
    for (int i = 0; i < static_cast<int>(feed_vars.size()); ++i) {
499 500 501 502 503 504 505
      auto* var = scope_.FindVar(feed_vars[i]);
      if (var && var->IsType<framework::LoDTensor>()) {
        auto* tensor_pd = GetLoDTensorOrSelectedRowsValueFromVar(*var);
        auto dims = tensor_pd->dims();
        for (int j = 0; j < dims.size(); ++j) {
          func_cache_key_ += std::to_string(dims[j]);
        }
B
baojun 已提交
506 507
      }
    }
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527
    func_cache_key_ += std::to_string(interval[0]) + "_" +
                       std::to_string(interval[1]) + engine_key;
    func_cache_key_ = std::to_string(std::hash<std::string>()(func_cache_key_));

    if (engine_cache.find(func_cache_key_) != engine_cache.end()) {
      if (engine_cache[func_cache_key_].persistables.size() == 0) {
        engine_cache.clear();
        t_in_cache_.clear();
      } else {
        auto var_name = engine_cache[func_cache_key_].persistables.begin();
        framework::Variable* var = scope_.FindVar(*var_name);
        if (var != pre_var_ptr) {
          engine_cache.clear();
          t_in_cache_.clear();
        }
        pre_var_ptr = var;
      }
    }

    if (engine_cache.find(func_cache_key_) == engine_cache.end()) {
528
      BuildNgFunction(ctx);
529 530 531 532 533 534
      engine_cache[func_cache_key_].ngraph_function = this->ngraph_function_;
      engine_cache[func_cache_key_].persistables = this->persistables_;
      engine_cache[func_cache_key_].var_in_updates = this->var_in_updates_;
      engine_cache[func_cache_key_].var_in = this->var_in_;
      engine_cache[func_cache_key_].var_out = this->var_out_;
      engine_cache[func_cache_key_].is_test = this->is_test_;
B
baojun 已提交
535 536
    }
  } else {
537
    BuildNgFunction(ctx);
B
baojun 已提交
538 539 540 541 542
  }
}

void NgraphEngine::Run(const framework::Scope& scope,
                       const platform::Place& place) const {
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
  std::shared_ptr<ngraph::Function> ng_func;
  const std::set<std::string>* p_persistables;
  const std::vector<size_t>* p_var_in_updates;
  const std::vector<std::string>* p_var_in;
  const std::vector<std::string>* p_var_out;
  bool is_test;

  bool use_cache = true;
  if (use_cache) {
    PADDLE_ENFORCE(engine_cache.find(func_cache_key_) != engine_cache.end(),
                   "Cannot find cached data to run ngraph function");
    ng_func = engine_cache[func_cache_key_].ngraph_function;
    p_persistables = &(engine_cache[func_cache_key_].persistables);
    p_var_in_updates = &(engine_cache[func_cache_key_].var_in_updates);
    p_var_in = &(engine_cache[func_cache_key_].var_in);
    p_var_out = &(engine_cache[func_cache_key_].var_out);
    is_test = engine_cache[func_cache_key_].is_test;
  } else {
    ng_func = ngraph_function_;
    p_persistables = &this->persistables_;
    p_var_in_updates = &this->var_in_updates_;
    p_var_in = &this->var_in_;
    p_var_out = &this->var_out_;
    is_test = this->is_test_;
  }
B
baojun 已提交
568

569 570 571 572 573
  std::vector<std::shared_ptr<ngraph::runtime::Tensor>>* p_t_in;
  std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_in = {};

  auto m_parameters = ng_func->get_parameters();
  auto m_results = ng_func->get_results();
574 575 576 577 578
  // Due to optimization backend may produce results in other layouts,
  // make sure we get default layout for results.
  for (auto& r : m_results) {
    r->set_needs_default_layout(true);
  }
579 580 581 582 583 584 585 586 587 588 589 590 591 592 593
  if (is_test && use_cache &&
      t_in_cache_.find(func_cache_key_) != t_in_cache_.end()) {
    p_t_in = &(t_in_cache_[func_cache_key_]);
    for (size_t i = 0; i < p_var_in_updates->size(); ++i) {
      int index = p_var_in_updates->at(i);
      auto vi = p_var_in->at(index);
      auto sp = m_parameters[index]->get_shape();
      auto ng_type = m_parameters[index]->get_element_type();
      std::shared_ptr<ngraph::runtime::Tensor> ti;
      auto* var = scope.FindVar(vi);
      if (var && var->IsType<framework::LoDTensor>()) {
        auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var);
        void* pd_arr = tensor_pd->mutable_data(place, ng2pd_type_map[ng_type]);
        ti = backend_->create_tensor(ng_type, sp, pd_arr);
        (*p_t_in)[index] = ti;
B
baojun 已提交
594
      } else {
595
        PADDLE_THROW("Cannot find var or tensor with var name %s", vi);
B
baojun 已提交
596
      }
597 598 599 600
    }
  } else {
    if (is_test && use_cache) {
      p_t_in = &(t_in_cache_[func_cache_key_]);
B
baojun 已提交
601
    } else {
602
      p_t_in = &t_in;
B
baojun 已提交
603
    }
604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621

    for (size_t i = 0; i < p_var_in->size(); ++i) {
      auto vi = p_var_in->at(i);
      auto sp = m_parameters[i]->get_shape();
      auto ng_type = m_parameters[i]->get_element_type();
      std::shared_ptr<ngraph::runtime::Tensor> ti;
      auto* var = scope.FindVar(vi);
      if (var && var->IsType<framework::LoDTensor>()) {
        auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var);
        void* pd_arr = tensor_pd->mutable_data(place, ng2pd_type_map[ng_type]);
        PADDLE_ENFORCE(sp == Ddim2Shape(tensor_pd->dims()),
                       "Ensure ngraph tensor layout align with paddle tensor");
        ti = backend_->create_tensor(ng_type, sp, pd_arr);
      } else {
        PADDLE_THROW("Cannot find var or tensor with var name %s", vi);
      }
      bool is_persistable =
          (p_persistables->find(vi) != p_persistables->end()) ? true : false;
622
      if (!is_training && is_test && is_persistable) {
623 624 625
        ti->set_stale(false);
      }
      (*p_t_in).emplace_back(ti);
B
baojun 已提交
626 627 628
    }
  }

629 630 631
  std::vector<std::shared_ptr<ngraph::runtime::Tensor>> t_out = {};
  for (size_t i = 0; i < p_var_out->size(); ++i) {
    auto vo = p_var_out->at(i);
B
baojun 已提交
632 633
    auto* var = scope.FindVar(vo);
    if (var && var->IsType<framework::LoDTensor>()) {
634 635
      auto sp = m_results[i]->get_shape();
      var->GetMutable<framework::LoDTensor>()->Resize(Shape2Ddim(sp));
B
baojun 已提交
636
      auto* tensor_pd = GetMutableLoDTensorOrSelectedRowsValueFromVar(var);
637 638 639 640 641
      auto ng_type = m_results[i]->get_element_type();
      void* pd_arr = tensor_pd->mutable_data(place, ng2pd_type_map[ng_type]);
      std::shared_ptr<ngraph::runtime::Tensor> to =
          backend_->create_tensor(ng_type, sp, pd_arr);
      t_out.emplace_back(to);
B
baojun 已提交
642 643 644 645 646
    } else {
      PADDLE_THROW("Cannot find var or tensor with var name %s", vo);
    }
  }

647 648
  auto handle = backend_->compile(ng_func);
  handle->call_with_validate(t_out, *p_t_in);
B
baojun 已提交
649 650 651
}  // NgraphEngine::Run
}  // namespace operators
}  // namespace paddle