graph_pattern_detector.cc 69.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

15
#include <algorithm>
Q
Qiao Longfei 已提交
16
#include <array>
17
#include <memory>
18
#include <string>
19 20
#include <unordered_map>
#include <unordered_set>
21 22 23
#include <vector>

#include "paddle/fluid/framework/ir/graph_helper.h"
24
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
25
#include "paddle/fluid/framework/ir/graph_traits.h"
26
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
C
chengduo 已提交
27
#include "paddle/fluid/framework/operator.h"
28
#include "paddle/fluid/platform/enforce.h"
Y
Yan Chunwei 已提交
29
#include "paddle/fluid/string/pretty_log.h"
Y
Yan Chunwei 已提交
30
#include "paddle/fluid/string/printf.h"
31

32 33 34 35
namespace paddle {
namespace framework {
namespace ir {

Y
Yan Chunwei 已提交
36 37 38 39
using string::PrettyLogEndl;
using string::PrettyLog;
using string::Style;

40 41
size_t PDPattern::id_ = 0UL;

C
chengduo 已提交
42
PDNode *PDPattern::NewNode(const std::string &name) {
Y
Yan Chunwei 已提交
43
  if (!name.empty()) {
T
tensor-tang 已提交
44
    PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL,
Y
Yan Chunwei 已提交
45 46 47 48 49
                      "PDNode's name should be unique, get duplicate [%s]",
                      name);
  }

  nodes_.emplace_back(new PDNode(this, name));
C
chengduo 已提交
50
  auto *cur = nodes_.back().get();
Y
Yan Chunwei 已提交
51 52 53 54
  node_map_[name] = cur;
  return cur;
}

C
chengduo 已提交
55
PDNode *PDPattern::NewNode(PDNode::teller_t &&teller, const std::string &name) {
56
  if (!name.empty()) {
T
tensor-tang 已提交
57
    PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL,
58 59 60 61
                      "PDNode's name should be unique, get duplicate [%s]",
                      name);
  }

62
  nodes_.emplace_back(new PDNode(std::move(teller), this, name));
C
chengduo 已提交
63
  auto *cur = nodes_.back().get();
64
  node_map_[name] = cur;
65 66 67
  return cur;
}

C
chengduo 已提交
68
PDNode *PDPattern::RetrieveNode(const std::string &id) const {
69 70 71 72 73 74 75 76
  auto it = node_map_.find(id);
  if (it == node_map_.end()) {
    return nullptr;
  }

  return it->second;
}

C
chengduo 已提交
77
void PDPattern::AddEdge(PDNode *a, PDNode *b) {
78 79 80 81 82 83
  PADDLE_ENFORCE(a);
  PADDLE_ENFORCE(b);
  PADDLE_ENFORCE(a != b, "can't connect to the same nodes.");
  edges_.emplace_back(a, b);
}

C
chengduo 已提交
84
void GraphPatternDetector::operator()(Graph *graph,
85
                                      GraphPatternDetector::handle_t handler) {
86 87 88 89
  if (!MarkPDNodesInGraph(*graph)) {
    return;
  }

90 91 92
  auto subgraphs = DetectPatterns();
  UniquePatterns(&subgraphs);
  RemoveOverlappedMatch(&subgraphs);
Y
Yan Chunwei 已提交
93
  ValidateByNodeRole(&subgraphs);
94

Y
Yan Chunwei 已提交
95
  if (subgraphs.empty()) return;
96 97
  PrettyLogEndl(Style::detail(), "---  detected %d subgraphs",
                subgraphs.size());
98
  int id = 0;
C
chengduo 已提交
99
  for (auto &g : subgraphs) {
M
minqiyang 已提交
100
    VLOG(3) << "optimizing #" << id++ << " subgraph";
101 102 103 104
    handler(g, graph);
  }
}

C
chengduo 已提交
105
bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) {
M
minqiyang 已提交
106
  VLOG(3) << "mark pdnodes in graph";
107 108
  if (graph.Nodes().empty()) return false;

C
chengduo 已提交
109 110
  for (auto &node : GraphTraits::DFS(graph)) {
    for (const auto &pdnode : pattern_.nodes()) {
111
      if (pdnode->Tell(&node)) {
112
        VLOG(4) << "Node " << node.Name() << " marked as " << pdnode->name();
113 114 115 116
        pdnodes2nodes_[pdnode.get()].insert(&node);
      }
    }
  }
Y
Yan Chunwei 已提交
117
  // Check to early stop if some PDNode can't find matched Node.
C
chengduo 已提交
118
  for (auto &pdnode : pattern_.nodes()) {
Y
Yan Chunwei 已提交
119
    if (!pdnodes2nodes_.count(pdnode.get())) {
M
minqiyang 已提交
120
      VLOG(4) << pdnode->name() << " can't find matched Node, early stop";
Y
Yan Chunwei 已提交
121
      // return false;
Y
Yan Chunwei 已提交
122 123
    }
  }
M
minqiyang 已提交
124
  VLOG(3) << pdnodes2nodes_.size() << " nodes marked";
125

126 127 128
  return !pdnodes2nodes_.empty();
}

Y
Yan Chunwei 已提交
129 130 131
// The intermediate Nodes can only link to the nodes inside the pattern, or this
// subgraph will be droped.
void GraphPatternDetector::ValidateByNodeRole(
C
chengduo 已提交
132
    std::vector<GraphPatternDetector::subgraph_t> *subgraphs) {
Y
Yan Chunwei 已提交
133 134 135 136 137
  std::vector<GraphPatternDetector::subgraph_t> result;

  subgraphs->erase(
      std::remove_if(
          subgraphs->begin(), subgraphs->end(),
C
chengduo 已提交
138
          [](const GraphPatternDetector::subgraph_t &subgraph) -> bool {
Y
Yan Chunwei 已提交
139
            // Collect the inputs and outputs.
C
chengduo 已提交
140 141
            std::unordered_set<Node *> ios;
            for (auto &item : subgraph) {
Y
Yan Chunwei 已提交
142 143 144 145
              if (!item.first->IsIntermediate()) {
                ios.insert(item.second);
              }
            }
C
chengduo 已提交
146
            for (auto &item : subgraph) {
Y
Yan Chunwei 已提交
147
              if (item.first->IsIntermediate()) {
C
chengduo 已提交
148
                for (auto *x : item.second->inputs) {
Y
Yan Chunwei 已提交
149 150 151 152
                  if (!ios.count(x)) {
                    return true;
                  }
                }
C
chengduo 已提交
153
                for (auto *x : item.second->outputs) {
Y
Yan Chunwei 已提交
154 155 156 157 158 159 160 161 162 163 164
                  if (!ios.count(x)) {
                    return true;
                  }
                }
              }
            }
            return false;
          }),
      subgraphs->end());
}

165
struct HitGroup {
C
chengduo 已提交
166
  std::unordered_map<PDNode *, Node *> roles;
167

C
chengduo 已提交
168
  bool Match(Node *node, PDNode *pat) {
169
    if (nodes_.count(node)) {
T
Tao Luo 已提交
170 171 172 173 174
      if (roles.count(pat) && roles[pat] == node) return true;
      return false;
    } else {
      if (roles.count(pat) && roles[pat] != node) return false;
      return true;
175
    }
176 177
  }

C
chengduo 已提交
178
  void Register(Node *node, PDNode *pat) {
179 180 181 182 183
    roles[pat] = node;
    nodes_.insert(node);
  }

 private:
C
chengduo 已提交
184
  std::unordered_set<Node *> nodes_;
185 186 187
};

// Tell whether Node a links to b.
C
chengduo 已提交
188 189
bool IsNodesLink(Node *a, Node *b) {
  for (auto *node : a->outputs) {
190 191 192 193 194 195 196
    if (b == node) {
      return true;
    }
  }
  return false;
}

197 198
std::vector<GraphPatternDetector::subgraph_t>
GraphPatternDetector::DetectPatterns() {
199
  // Init empty subgraphs.
200
  std::vector<GraphPatternDetector::subgraph_t> result;
201
  std::vector<HitGroup> init_groups;
202
  std::array<std::vector<HitGroup>, 2> bi_records;
C
chengduo 已提交
203
  auto *first_pnode = pattern_.edges().empty() ? pattern().nodes().front().get()
204
                                               : pattern_.edges().front().first;
205
  if (!pdnodes2nodes_.count(first_pnode)) return result;
C
chengduo 已提交
206
  for (auto *node : pdnodes2nodes_[first_pnode]) {
207 208 209 210 211 212 213 214 215 216
    HitGroup group;
    group.roles[first_pnode] = node;
    init_groups.emplace_back(group);
  }

  int step = 0;
  bi_records[0] = std::move(init_groups);

  // Extend a PDNode to subgraphs by deducing the connection relations defined
  // in edges of PDNodes.
C
chengduo 已提交
217
  for (const auto &edge : pattern_.edges()) {
M
minqiyang 已提交
218
    VLOG(4) << "check " << edge.first->name() << " -> " << edge.second->name();
Y
Yan Chunwei 已提交
219
    // TODO(Superjomn) Fix bug here, the groups might be duplicate here.
220 221
    // Each role has two PDNodes, which indicates two roles.
    // Detect two Nodes that can match these two roles and they are connected.
C
chengduo 已提交
222 223
    auto &pre_groups = bi_records[step % 2];
    auto &cur_groups = bi_records[1 - (step++ % 2)];
224
    cur_groups.clear();
225
    if (pre_groups.empty()) break;
226
    // source -> target
C
chengduo 已提交
227 228
    for (Node *source : pdnodes2nodes_[edge.first]) {
      for (Node *target : pdnodes2nodes_[edge.second]) {
M
minqiyang 已提交
229
        VLOG(8) << "check " << source->id() << " -- " << target->id();
230
        // TODO(Superjomn) add some prune strategies.
C
chengduo 已提交
231
        for (const auto &group : pre_groups) {
T
Tao Luo 已提交
232 233 234 235 236 237
          if (IsNodesLink(source, target)) {
            HitGroup new_group = group;
            bool flag = new_group.Match(source, edge.first) &&
                        new_group.Match(target, edge.second);
            if (flag) {
              new_group.Register(source, edge.first);
238 239 240 241 242 243 244 245
              new_group.Register(target, edge.second);
              cur_groups.push_back(new_group);
              // TODO(Superjomn) need to unique
            }
          }
        }
      }
    }
M
minqiyang 已提交
246
    VLOG(3) << "step " << step << " get records: " << cur_groups.size();
C
chengduo 已提交
247 248
    for (auto &group : cur_groups) {
      for (auto &item : group.roles) {
M
minqiyang 已提交
249
        VLOG(4) << "node " << item.second->id() << " as " << item.first->name();
Y
Yan Chunwei 已提交
250
      }
M
minqiyang 已提交
251
      VLOG(4) << "=========================================================";
Y
Yan Chunwei 已提交
252
    }
253 254
  }

C
chengduo 已提交
255
  for (auto &group : bi_records[step % 2]) {
256
    GraphPatternDetector::subgraph_t subgraph;
C
chengduo 已提交
257
    for (auto &role : group.roles) {
258 259 260 261 262 263 264
      subgraph.emplace(role.first, role.second);
    }
    result.emplace_back(subgraph);
  }
  return result;
}

Y
Yan Chunwei 已提交
265 266
struct GraphItemLessThan {
  bool operator()(const std::pair<PDNode *, Node *> &a,
Y
Yan Chunwei 已提交
267
                  const std::pair<PDNode *, Node *> &b) {
Y
Yan Chunwei 已提交
268 269 270 271 272
    if (a.first != b.first) {
      return a.first < b.first;
    } else {
      return a.second < b.second;
    }
Y
Yan Chunwei 已提交
273
  }
Y
Yan Chunwei 已提交
274
};
Y
Yan Chunwei 已提交
275

276 277
// TODO(Superjomn) enhance the function as it marks unique unique as duplicates
// see https://github.com/PaddlePaddle/Paddle/issues/13550
278
void GraphPatternDetector::UniquePatterns(
C
chengduo 已提交
279
    std::vector<GraphPatternDetector::subgraph_t> *subgraphs) {
280
  if (subgraphs->empty()) return;
281
  std::vector<GraphPatternDetector::subgraph_t> result;
282 283

  std::unordered_set<size_t> set;
Y
Yan Chunwei 已提交
284
  std::hash<std::string> hasher;
C
chengduo 已提交
285
  for (auto &g : *subgraphs) {
Y
Yan Chunwei 已提交
286 287
    // Sort the items in the sub-graph, and transform to a string key.
    std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
Y
Yan Chunwei 已提交
288
    std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemLessThan());
Y
Yan Chunwei 已提交
289 290 291
    std::stringstream ss;
    for (auto &item : sorted_keys) {
      ss << item.first << ":" << item.second;
292
    }
Y
Yan Chunwei 已提交
293
    auto key = hasher(ss.str());
294 295 296 297 298 299 300 301
    if (!set.count(key)) {
      result.emplace_back(g);
      set.insert(key);
    }
  }
  *subgraphs = result;
}

302
void GraphPatternDetector::RemoveOverlappedMatch(
C
chengduo 已提交
303
    std::vector<subgraph_t> *subgraphs) {
304
  std::vector<subgraph_t> result;
C
chengduo 已提交
305
  std::unordered_set<Node *> node_set;
306

C
chengduo 已提交
307
  for (const auto &subgraph : *subgraphs) {
308
    bool valid = true;
C
chengduo 已提交
309
    for (auto &item : subgraph) {
Y
Yan Chunwei 已提交
310
      if (item.first->IsIntermediate() && node_set.count(item.second)) {
311 312 313 314 315
        valid = false;
        break;
      }
    }
    if (valid) {
C
chengduo 已提交
316
      for (auto &item : subgraph) {
317 318 319 320 321 322 323 324
        node_set.insert(item.second);
      }
      result.push_back(subgraph);
    }
  }
  *subgraphs = result;
}

325 326 327 328 329
std::string PDPattern::DotString() const {
  using inference::analysis::Dot;
  Dot dot;
  int id = 0;
  // Create Nodes
C
chengduo 已提交
330 331
  std::unordered_map<PDNode *, std::string> node2dot;
  for (const auto &node : nodes()) {
332 333 334 335 336
    std::string node_id = "Node" + std::to_string(id++);
    dot.AddNode(node_id, {}, node->name());
    node2dot[node.get()] = node_id;
  }
  // Create Edges
C
chengduo 已提交
337
  for (const auto &edge : edges()) {
338 339 340 341
    if (!node2dot.count(edge.first) || !node2dot.count(edge.second)) {
      LOG(ERROR) << "no node " << edge.first << " " << edge.second;
      continue;
    }
C
chengduo 已提交
342 343
    auto &src = node2dot.at(edge.first);
    auto &trg = node2dot.at(edge.second);
344 345 346 347 348
    dot.AddEdge(src, trg, {});
  }
  return dot.Build();
}

C
chengduo 已提交
349
PDNode &PDNode::LinksTo(const std::vector<PDNode *> &others) {
350
  // extend outlinks.
C
chengduo 已提交
351
  for (PDNode *x : others) {
352 353 354 355 356
    pattern_->AddEdge(this, x);
  }
  return *this;
}

C
chengduo 已提交
357
PDNode &PDNode::LinksFrom(const std::vector<PDNode *> &others) {
358
  // extend outlinks.
C
chengduo 已提交
359
  for (PDNode *x : others) {
360 361 362 363 364
    pattern_->AddEdge(x, this);
  }
  return *this;
}

C
chengduo 已提交
365 366
PDNode *PDNode::assert_is_op() {
  asserts_.emplace_back([](Node *x) { return x && x->IsOp(); });
Y
Yan Chunwei 已提交
367 368
  return this;
}
C
chengduo 已提交
369 370 371

PDNode *PDNode::assert_is_op(const std::string &op_type) {
  asserts_.emplace_back([op_type](Node *x) {
Y
Yan Chunwei 已提交
372 373 374 375
    return x && x->IsOp() && x->Op()->Type() == op_type;
  });
  return this;
}
C
chengduo 已提交
376 377 378 379 380 381 382 383

PDNode *PDNode::assert_is_var() {
  asserts_.emplace_back([](Node *x) { return x && x->IsVar(); });
  return this;
}

PDNode *PDNode::assert_is_not_ctrl_var() {
  asserts_.emplace_back([](Node *x) { return x && !x->IsCtrlVar(); });
Y
Yan Chunwei 已提交
384 385
  return this;
}
C
chengduo 已提交
386 387

PDNode *PDNode::assert_var_not_persistable() {
Y
Yan Chunwei 已提交
388
  assert_is_var();
C
chengduo 已提交
389
  asserts_.emplace_back([](Node *x) { return !x->Var()->Persistable(); });
Y
Yan Chunwei 已提交
390 391
  return this;
}
C
chengduo 已提交
392 393

PDNode *PDNode::assert_is_persistable_var() {
Y
Yan Chunwei 已提交
394
  assert_is_var();
C
chengduo 已提交
395
  asserts_.emplace_back([=](Node *x) { return x->Var()->Persistable(); });
Y
Yan Chunwei 已提交
396 397
  return this;
}
C
chengduo 已提交
398 399 400

PDNode *PDNode::assert_is_op_nth_input(const std::string &op_type,
                                       const std::string &argument, int nth) {
Y
Yan Chunwei 已提交
401 402
  assert_is_var();
  assert_is_op_input(op_type);
C
chengduo 已提交
403 404
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->outputs) {
405 406 407
      if (op->IsOp() && op->Op()->Type() == op_type &&
          IsNthInput(x, op, argument, nth))
        return true;
Y
Yan Chunwei 已提交
408 409 410 411 412
    }
    return false;
  });
  return this;
}
C
chengduo 已提交
413 414 415

PDNode *PDNode::assert_is_op_nth_output(const std::string &op_type,
                                        const std::string &argument, int nth) {
Y
Yan Chunwei 已提交
416
  assert_is_var();
C
chengduo 已提交
417 418
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->inputs) {
419 420 421
      if (op->IsOp() && op->Op()->Type() == op_type &&
          IsNthOutput(x, op, argument, nth))
        return true;
Y
Yan Chunwei 已提交
422 423 424 425 426
    }
    return false;
  });
  return this;
}
C
chengduo 已提交
427 428

PDNode *PDNode::assert_is_only_input_of_op(const std::string &op_type) {
Y
Yan Chunwei 已提交
429
  assert_is_var();
C
chengduo 已提交
430 431
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->outputs) {
Y
Yan Chunwei 已提交
432 433 434 435 436 437 438 439 440
      if (op && op->IsOp() && op->Op() && op->Op()->Type() == op_type &&
          op->inputs.size() == 1) {
        return true;
      }
    }
    return false;
  });
  return this;
}
C
chengduo 已提交
441 442

PDNode *PDNode::assert_is_only_output_of_op(const std::string &op_type) {
Y
Yan Chunwei 已提交
443
  assert_is_var();
C
chengduo 已提交
444 445
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->inputs) {
Y
Yan Chunwei 已提交
446 447 448 449 450 451 452 453 454
      if (op && op->IsOp() && op->Op() && op->Op()->Type() == op_type &&
          op->outputs.size() == 1) {
        return true;
      }
    }
    return false;
  });
  return this;
}
C
chengduo 已提交
455 456

PDNode *PDNode::assert_is_op_output(const std::string &op_type) {
Y
Yan Chunwei 已提交
457
  assert_is_var();
C
chengduo 已提交
458 459
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->inputs) {
Y
Yan Chunwei 已提交
460 461 462 463 464 465 466 467
      if (op && op->IsOp() && op->Op() && op->Op()->Type() == op_type) {
        return true;
      }
    }
    return false;
  });
  return this;
}
C
chengduo 已提交
468 469 470

PDNode *PDNode::assert_is_op_output(const std::string &op_type,
                                    const std::string &argument) {
471 472 473 474
  assert_is_var();
  assert_is_op_nth_output(op_type, argument, 0);
  return this;
}
C
chengduo 已提交
475
PDNode *PDNode::assert_is_op_input(const std::string &op_type) {
Y
Yan Chunwei 已提交
476
  assert_is_var();
C
chengduo 已提交
477 478
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->outputs) {
Y
Yan Chunwei 已提交
479 480 481 482 483 484 485 486
      if (op && op->IsOp() && op->Op() && op->Op()->Type() == op_type) {
        return true;
      }
    }
    return false;
  });
  return this;
}
C
chengduo 已提交
487 488 489

PDNode *PDNode::assert_is_op_input(const std::string &op_type,
                                   const std::string &argument) {
490 491 492 493
  assert_is_var();
  assert_is_op_nth_input(op_type, argument, 0);
  return this;
}
C
chengduo 已提交
494 495

PDNode *PDNode::assert_op_has_n_inputs(const std::string &op_type, size_t n) {
Y
Yan Chunwei 已提交
496
  assert_is_op(op_type);
C
chengduo 已提交
497
  asserts_.emplace_back([=](Node *x) { return x->inputs.size() == n; });
Y
Yan Chunwei 已提交
498 499
  return this;
}
C
chengduo 已提交
500 501

PDNode *PDNode::assert_op_has_n_outputs(const std::string &op_type, size_t n) {
Y
Yan Chunwei 已提交
502
  assert_is_op(op_type);
C
chengduo 已提交
503
  asserts_.emplace_back([=](Node *x) { return x->outputs.size() == n; });
Y
Yan Chunwei 已提交
504 505
  return this;
}
C
chengduo 已提交
506

507 508 509 510 511 512 513 514 515 516
PDNode *PDNode::assert_has_n_inputs(size_t n) {
  asserts_.emplace_back([=](Node *x) { return x->inputs.size() == n; });
  return this;
}

PDNode *PDNode::assert_has_n_outputs(size_t n) {
  asserts_.emplace_back([=](Node *x) { return x->outputs.size() == n; });
  return this;
}

C
chengduo 已提交
517
PDNode *PDNode::assert_more(PDNode::teller_t &&teller) {
Y
Yan Chunwei 已提交
518 519 520 521
  asserts_.emplace_back(std::move(teller));
  return this;
}

C
chengduo 已提交
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
PDNode *PDNode::assert_is_ops(const std::unordered_set<std::string> &op_types) {
  asserts_.emplace_back([op_types](Node *x) {
    return x && x->IsOp() && op_types.count(x->Op()->Type());
  });
  return this;
}

PDNode *PDNode::assert_is_ops_nth_input(
    const std::unordered_set<std::string> &op_types,
    const std::string &argument, int nth) {
  assert_is_var();
  assert_is_ops_input(op_types);
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->outputs) {
      if (op->IsOp() && op_types.count(op->Op()->Type()) &&
          IsNthInput(x, op, argument, nth))
        return true;
    }
    return false;
  });
  return this;
}

PDNode *PDNode::assert_is_ops_nth_output(
    const std::unordered_set<std::string> &op_types,
    const std::string &argument, int nth) {
  assert_is_var();
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->inputs) {
      if (op->IsOp() && op_types.count(op->Op()->Type()) &&
          IsNthOutput(x, op, argument, nth))
        return true;
    }
    return false;
  });
  return this;
}
PDNode *PDNode::assert_is_ops_output(
    const std::unordered_set<std::string> &op_types) {
  assert_is_var();
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->inputs) {
      if (op && op->IsOp() && op->Op() && op_types.count(op->Op()->Type())) {
        return true;
      }
    }
    return false;
  });
  return this;
}

PDNode *PDNode::assert_is_ops_output(
    const std::unordered_set<std::string> &op_types,
    const std::string &argument) {
  assert_is_var();
  assert_is_ops_nth_output(op_types, argument, 0);
  return this;
}

PDNode *PDNode::assert_is_ops_input(
    const std::unordered_set<std::string> &op_types) {
  assert_is_var();
  asserts_.emplace_back([=](Node *x) {
    for (auto *op : x->outputs) {
      if (op && op->IsOp() && op->Op() && op_types.count(op->Op()->Type())) {
        return true;
      }
    }
    return false;
  });
  return this;
}

PDNode *PDNode::assert_is_ops_input(
    const std::unordered_set<std::string> &op_types,
    const std::string &argument) {
  assert_is_var();
  assert_is_ops_nth_input(op_types, argument, 0);
  return this;
}

bool VarLinksToOp(Node *node, const std::string &op_type) {
  for (auto *out : node->outputs) {
605 606 607 608 609 610
    if (out->IsOp() && out->Op()->Type() == op_type) {
      return true;
    }
  }
  return false;
}
C
chengduo 已提交
611 612

bool IsNthInput(Node *var, Node *op, const std::string &argument, size_t nth) {
613 614
  PADDLE_ENFORCE(var->IsVar());
  PADDLE_ENFORCE(op->IsOp());
615 616
  if (!HasInput(op, argument) || op->Op()->Input(argument).size() <= nth)
    return false;
617 618
  return var->Name() == op->Op()->Input(argument)[nth];
}
C
chengduo 已提交
619

620 621 622 623 624 625 626 627
bool HasInput(Node *op, const std::string &argument) {
  PADDLE_ENFORCE(op->IsOp());
  auto const &names = op->Op()->InputNames();
  if (std::find(names.begin(), names.end(), argument) == names.end())
    return false;
  return true;
}

C
chengduo 已提交
628
bool IsNthOutput(Node *var, Node *op, const std::string &argument, size_t nth) {
629 630 631 632 633
  PADDLE_ENFORCE(var->IsVar());
  PADDLE_ENFORCE(op->IsOp());
  if (op->Op()->Output(argument).size() <= nth) return false;
  return var->Name() == op->Op()->Output(argument)[nth];
}
C
chengduo 已提交
634 635 636 637 638

void GraphSafeRemoveNodes(Graph *graph,
                          const std::unordered_set<const Node *> &nodes) {
  for (auto *node : nodes) {
    graph->RemoveNode(const_cast<Node *>(node));
639 640
  }

C
chengduo 已提交
641
  for (auto *node : graph->Nodes()) {
642 643
    for (auto it = node->inputs.begin(); it != node->inputs.end();) {
      if (nodes.count(*it)) {
C
chengduo 已提交
644
        it = const_cast<Node *>(node)->inputs.erase(it);
645
      } else {
646
        it++;
647
      }
648 649 650
    }
    for (auto it = node->outputs.begin(); it != node->outputs.end();) {
      if (nodes.count(*it)) {
C
chengduo 已提交
651
        it = const_cast<Node *>(node)->outputs.erase(it);
652
      } else {
653
        it++;
654
      }
655 656 657
    }
  }
}
C
chengduo 已提交
658 659 660

bool VarLinksFromOp(Node *node, const std::string &op_type) {
  for (auto *out : node->inputs) {
661 662 663 664 665 666 667
    if (out->IsOp() && out->Op()->Type() == op_type) {
      return true;
    }
  }
  return false;
}

S
Sylwester Fraczek 已提交
668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
PDNode *patterns::ConvBN::operator()(paddle::framework::ir::PDNode *conv_input,
                                     bool with_eltwise_add) {
  // Create Operators
  conv_input->assert_is_op_input("conv2d", "Input");
  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");

  PDNode *eltwise_op = nullptr;
  if (with_eltwise_add) {
    eltwise_op =
        pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
  }
  auto *batch_norm_op =
      pattern->NewNode(batch_norm_repr())->assert_is_op("batch_norm");
  // Create variables
  // Conv Filter
  auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
                              ->AsInput()
                              ->assert_is_persistable_var()
                              ->assert_is_op_input("conv2d", "Filter");

  auto *conv_out_var = pattern->NewNode(conv_out_repr())
                           ->AsIntermediate()
                           ->assert_is_only_output_of_op("conv2d");

  PDNode *eltwise_y_in_var = nullptr;
  PDNode *eltwise_out_var = nullptr;
  if (with_eltwise_add) {
    // Conv output as Bias input
    conv_out_var->assert_is_op_input("elementwise_add", "X");
    // Bias
    eltwise_y_in_var = pattern->NewNode(eltwise_y_in_repr())
                           ->assert_is_op_input("elementwise_add", "Y")
                           ->AsInput();
    eltwise_out_var = pattern->NewNode(eltwise_out_repr())
                          ->AsIntermediate()
                          ->assert_is_only_output_of_op("elementwise_add");
  } else {
    // Conv output as BN input
    conv_out_var->assert_is_op_input("batch_norm", "X");
  }

  // BN Scale
  auto *bn_scale_var = pattern->NewNode(bn_scale_repr())
                           ->AsInput()
                           ->assert_is_persistable_var()
                           ->assert_is_op_input("batch_norm", "Scale");
  // BN Bias
  auto *bn_bias_var = pattern->NewNode(bn_bias_repr())
                          ->AsInput()
                          ->assert_is_persistable_var()
                          ->assert_is_op_input("batch_norm", "Bias");
  // BN Mean
  auto *bn_mean_var = pattern->NewNode(bn_mean_repr())
                          ->AsInput()
                          ->assert_is_persistable_var()
                          ->assert_is_op_input("batch_norm", "Mean");
  // BN Variance
  auto *bn_variance_var = pattern->NewNode(bn_variance_repr())
                              ->AsInput()
                              ->assert_is_persistable_var()
                              ->assert_is_op_input("batch_norm", "Variance");

  // BN output
  auto *bn_out_var = pattern->NewNode(bn_out_repr())
                         ->AsOutput()
                         ->assert_is_op_output("batch_norm");

  auto *bn_mean_out_var = pattern->NewNode(bn_mean_out_repr())
                              ->AsOutput()
                              ->assert_is_op_output("batch_norm", "MeanOut");

  auto *bn_variance_out_var =
      pattern->NewNode(bn_variance_out_repr())
          ->AsOutput()
          ->assert_is_op_output("batch_norm", "VarianceOut");

  auto *bn_saved_mean_var =
      pattern->NewNode(bn_saved_mean_repr())
          ->AsOutput()
          ->assert_is_op_output("batch_norm", "SavedMean");

  auto *bn_saved_variance_var =
      pattern->NewNode(bn_saved_variance_repr())
          ->AsOutput()
          ->assert_is_op_output("batch_norm", "SavedVariance");

  conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});

  if (with_eltwise_add) {
    eltwise_op->LinksFrom({conv_out_var, eltwise_y_in_var})
        .LinksTo({eltwise_out_var});
    batch_norm_op
        ->LinksFrom({eltwise_out_var, bn_scale_var, bn_bias_var, bn_mean_var,
                     bn_variance_var})
        .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var,
                  bn_saved_mean_var, bn_saved_variance_var});
  } else {
    batch_norm_op
        ->LinksFrom({conv_out_var, bn_scale_var, bn_bias_var, bn_mean_var,
                     bn_variance_var})
        .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var,
                  bn_saved_mean_var, bn_saved_variance_var});
  }
  return bn_out_var;
}

C
chengduo 已提交
774 775
PDNode *patterns::ConvReLU::operator()(
    paddle::framework::ir::PDNode *conv_input) {
776 777
  // Create Operators
  conv_input->assert_is_op_input("conv2d", "Input");
C
chengduo 已提交
778 779
  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
  auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu");
780 781
  // Create variables
  // Filter
C
chengduo 已提交
782
  auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
783 784 785 786
                              ->AsInput()
                              ->assert_is_persistable_var()
                              ->assert_is_op_input("conv2d", "Filter");
  // intermediate variable, will be removed in the IR after fuse.
C
chengduo 已提交
787
  auto *conv_out_var = pattern->NewNode(conv_out_repr())
788 789 790 791
                           ->AsIntermediate()
                           ->assert_is_only_output_of_op("conv2d")
                           ->assert_is_op_input("relu");
  // output
C
chengduo 已提交
792
  auto *relu_out_var = pattern->NewNode(relu_out_repr())
793 794 795
                           ->AsOutput()
                           ->assert_is_op_output("relu");

796
  conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
797 798 799 800
  relu_op->LinksFrom({conv_out_var}).LinksTo({relu_out_var});
  return relu_out_var;
}

801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827
PDNode *patterns::ConvBReLU::operator()(
    paddle::framework::ir::PDNode *conv_input) {
  // Create Operators
  conv_input->assert_is_op_input("conv2d", "Input");
  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
  auto *brelu_op = pattern->NewNode(brelu_repr())->assert_is_op("relu6");
  // Create variables
  // Filter
  auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
                              ->AsInput()
                              ->assert_is_persistable_var()
                              ->assert_is_op_input("conv2d", "Filter");
  // intermediate variable, will be removed in the IR after fuse.
  auto *conv_out_var = pattern->NewNode(conv_out_repr())
                           ->AsIntermediate()
                           ->assert_is_only_output_of_op("conv2d")
                           ->assert_is_op_input("relu6");
  // output
  auto *brelu_out_var = pattern->NewNode(brelu_out_repr())
                            ->AsOutput()
                            ->assert_is_op_output("relu6");

  conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
  brelu_op->LinksFrom({conv_out_var}).LinksTo({brelu_out_var});
  return brelu_out_var;
}

T
tensor-tang 已提交
828 829 830 831
PDNode *patterns::SeqConvEltAddRelu::operator()(
    paddle::framework::ir::PDNode *seqconv_input) {
  // Create Operators
  seqconv_input->assert_is_op_input("sequence_conv", "X");
T
tensor-tang 已提交
832 833 834 835
  auto *seqconv_op = pattern->NewNode(seqconv_repr())
                         ->assert_is_op("sequence_conv")
                         ->assert_op_attr<bool>("paddingTrainable", false)
                         ->assert_op_attr<int>("contextStride", 1);
T
tensor-tang 已提交
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872

  auto *eltadd_op =
      pattern->NewNode(eltadd_repr())->assert_is_op("elementwise_add");
  auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu");
  // Create variables
  // Filter
  auto *seqconv_weight_var =
      pattern->NewNode(seqconv_weight_repr())
          ->AsInput()
          ->assert_is_persistable_var()
          ->assert_is_op_input("sequence_conv", "Filter");
  // Bias
  auto *eltadd_bias_var = pattern->NewNode(eltadd_bias_repr())
                              ->AsInput()
                              ->assert_is_op_input("elementwise_add");
  // intermediate variable, will be removed in the IR after fuse.
  auto *seqconv_out_var = pattern->NewNode(seqconv_out_repr())
                              ->AsIntermediate()
                              ->assert_is_only_output_of_op("sequence_conv")
                              ->assert_is_op_input("elementwise_add");
  auto *eltadd_out_var = pattern->NewNode(eltadd_out_repr())
                             ->AsIntermediate()
                             ->assert_is_only_output_of_op("elementwise_add")
                             ->assert_is_only_input_of_op("relu");
  // output
  auto *relu_out_var = pattern->NewNode(relu_out_repr())
                           ->AsOutput()
                           ->assert_is_op_output("relu");

  seqconv_op->LinksFrom({seqconv_input, seqconv_weight_var})
      .LinksTo({seqconv_out_var});
  eltadd_op->LinksFrom({seqconv_out_var, eltadd_bias_var})
      .LinksTo({eltadd_out_var});
  relu_op->LinksFrom({eltadd_out_var}).LinksTo({relu_out_var});
  return relu_out_var;
}

C
chengduo 已提交
873
PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x,
Y
Yan Chunwei 已提交
874 875 876
                                 bool with_bias) {
  // Create shared nodes.
  x->assert_is_op_input("mul", "X");
C
chengduo 已提交
877
  auto *mul = pattern->NewNode(mul_repr())->assert_is_op("mul");
Y
Yan Chunwei 已提交
878

C
chengduo 已提交
879
  auto *mul_w_var = pattern->NewNode(w_repr())
Y
Yan Chunwei 已提交
880 881 882 883
                        ->AsInput()
                        ->assert_is_persistable_var()
                        ->assert_is_op_input("mul", "Y");

C
chengduo 已提交
884
  auto *mul_out_var =
Y
Yan Chunwei 已提交
885 886 887 888 889 890 891 892 893 894
      pattern->NewNode(mul_out_repr())->assert_is_op_output("mul");

  if (!with_bias) {  // not with bias
    // Add links.
    mul->LinksFrom({x, mul_w_var}).LinksTo({mul_out_var});
    return mul_out_var;

  } else {  // with bias
    mul_out_var->AsIntermediate()->assert_is_op_input("elementwise_add");
    // Create operators.
C
chengduo 已提交
895
    auto *elementwise_add = pattern->NewNode(elementwise_add_repr())
Y
Yan Chunwei 已提交
896 897
                                ->assert_is_op("elementwise_add");
    // Create variables.
C
chengduo 已提交
898
    auto *bias = pattern->NewNode(bias_repr())
Y
Yan Chunwei 已提交
899 900 901
                     ->assert_is_op_input("elementwise_add")
                     ->AsInput();

C
chengduo 已提交
902
    auto *fc_out = pattern->NewNode(Out_repr())
Y
Yan Chunwei 已提交
903 904 905 906 907 908
                       ->AsOutput()
                       ->assert_is_op_output("elementwise_add");

    mul->LinksFrom({mul_w_var, x}).LinksTo({mul_out_var});
    elementwise_add->LinksFrom({mul_out_var, bias}).LinksTo({fc_out});
    return fc_out;
909 910
  }
}
T
tensor-tang 已提交
911

912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938
PDNode *patterns::FCMKLDNN::operator()(paddle::framework::ir::PDNode *x,
                                       bool with_bias) {
  // Create shared nodes.
  x->assert_is_op_input("fc", "Input");

  auto *fc_op = pattern->NewNode(fc_repr())->assert_is_op("fc");
  // Create variables
  // Filter
  auto *fc_weight_var = pattern->NewNode(weights_repr())
                            ->AsInput()
                            ->assert_is_persistable_var()
                            ->assert_is_op_input("fc", "W");
  // Bias
  auto *fc_bias_var = pattern->NewNode(bias_repr())
                          ->AsInput()
                          ->assert_is_persistable_var()
                          ->assert_is_op_input("fc", "Bias");
  // Output
  auto *fc_out_var = pattern->NewNode(output_repr())
                         ->AsOutput()
                         ->assert_is_op_output("fc", "Out")
                         ->assert_is_only_output_of_op("fc");

  fc_op->LinksFrom({x, fc_weight_var, fc_bias_var}).LinksTo({fc_out_var});
  return fc_out_var;
}

939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956
PDNode *patterns::Embedding::operator()(PDNode *x) {
  x->assert_is_op_input("lookup_table", "Ids");
  auto *lookup_table_op =
      pattern->NewNode(lookup_table_repr())->assert_is_op("lookup_table");
#define NEW_NODE(arg__, io__)                    \
  auto *arg__ = pattern->NewNode(arg__##_repr()) \
                    ->assert_is_op_##io__("lookup_table", #arg__);

  NEW_NODE(W, input);

  NEW_NODE(Out, output);
#undef NEW_NODE

  lookup_table_op->LinksFrom({x, W});
  lookup_table_op->LinksTo({Out});
  return Out;
}

C
chengduo 已提交
957
PDNode *patterns::LSTM::operator()(PDNode *x) {
958
  x->assert_is_op_input("lstm", "Input");
C
chengduo 已提交
959
  auto *lstm_op = pattern->NewNode(lstm_repr())->assert_is_op("lstm");
Y
Yan Chunwei 已提交
960
#define NEW_NODE(arg__, io__) \
C
chengduo 已提交
961
  auto *arg__ =               \
Y
Yan Chunwei 已提交
962
      pattern->NewNode(arg__##_repr())->assert_is_op_##io__("lstm", #arg__);
963 964 965 966 967

  // Currently, the H0 and C0 are optional
  // TODO(Superjomn) upgrade the fuse framework to support optional.
  // NEW_NODE(H0, input);
  // NEW_NODE(C0, input);
Y
Yan Chunwei 已提交
968 969
  NEW_NODE(Weight, input);
  NEW_NODE(Bias, input);
970

Y
Yan Chunwei 已提交
971 972 973 974 975
  NEW_NODE(Hidden, output);
  NEW_NODE(Cell, output);
  NEW_NODE(BatchGate, output);
  NEW_NODE(BatchCellPreAct, output);
#undef NEW_NODE
976 977 978 979 980

  lstm_op->LinksFrom({x, Weight, Bias});
  lstm_op->LinksTo({Hidden, Cell, BatchGate, BatchCellPreAct});
  return Hidden;
}
T
tensor-tang 已提交
981

C
chengduo 已提交
982
PDNode *patterns::GRU::operator()(PDNode *x) {
T
tensor-tang 已提交
983
  x->assert_is_op_input("gru", "Input");
C
chengduo 已提交
984
  auto *gru_op = pattern->NewNode(gru_repr())->assert_is_op("gru");
Y
Yan Chunwei 已提交
985
#define NEW_NODE(arg__, io__) \
C
chengduo 已提交
986
  auto *arg__ =               \
Y
Yan Chunwei 已提交
987
      pattern->NewNode(arg__##_repr())->assert_is_op_##io__("gru", #arg__);
T
tensor-tang 已提交
988

Y
Yan Chunwei 已提交
989
  NEW_NODE(Weight, input);
T
tensor-tang 已提交
990 991
  // TODO(Superjomn): upgrade the fuse framework to support optional.
  // H0 and bias are optional
Y
Yan Chunwei 已提交
992
  NEW_NODE(Bias, input);  // also optional
T
tensor-tang 已提交
993 994
  // NEW_NODE(H0, input);

Y
Yan Chunwei 已提交
995
  NEW_NODE(Hidden, output);
T
tensor-tang 已提交
996
  // below are intermediate
Y
Yan Chunwei 已提交
997 998 999 1000
  NEW_NODE(BatchGate, output);
  NEW_NODE(BatchResetHiddenPrev, output);
  NEW_NODE(BatchHidden, output);
#undef NEW_NODE
T
tensor-tang 已提交
1001

T
tensor-tang 已提交
1002 1003 1004 1005
  BatchGate->AsIntermediate();
  BatchResetHiddenPrev->AsIntermediate();
  BatchHidden->AsIntermediate();

T
tensor-tang 已提交
1006 1007 1008 1009 1010
  gru_op->LinksFrom({x, Weight, Bias});
  gru_op->LinksTo({Hidden, BatchGate, BatchResetHiddenPrev, BatchHidden});
  return Hidden;
}

C
chengduo 已提交
1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104
PDNode *patterns::ActElewiseAdd::operator()(
    paddle::framework::ir::PDNode *in_var,
    std::unordered_set<std::string> act_types) {
  in_var->assert_is_ops_input(act_types, "X");

  auto *act = pattern->NewNode(act_repr())->assert_is_ops(act_types);
  auto *act_out_var = pattern->NewNode(act_out_repr())
                          ->assert_is_not_ctrl_var()
                          ->assert_is_ops_output(act_types);
  act_out_var->AsIntermediate()->assert_is_op_input("elementwise_add");

  auto *ele_x_var = pattern->NewNode(ele_x_repr())
                        ->assert_is_not_ctrl_var()
                        ->assert_is_op_input("elementwise_add")
                        ->AsInput();
  auto *elementwise_add =
      pattern->NewNode(ele_add_repr())->assert_is_op("elementwise_add");

  auto *elewise_add_out = pattern->NewNode(elewise_add_out_repr())
                              ->AsOutput()
                              ->assert_is_op_output("elementwise_add", "Out");

  act->LinksFrom({in_var}).LinksTo({act_out_var});
  elementwise_add->LinksFrom({act_out_var, ele_x_var})
      .LinksTo({elewise_add_out});

  return elewise_add_out;
}

PDNode *patterns::ElewiseAddAct::operator()(
    paddle::framework::ir::PDNode *ele_x_var,
    std::unordered_set<std::string> act_types) {
  auto *ele_y_var = pattern->NewNode(ele_y_repr())
                        ->assert_is_op_input("elementwise_add", "Y");

  auto *ele_add =
      pattern->NewNode(ele_add_repr())->assert_is_op("elementwise_add");

  auto *ele_out_var = pattern->NewNode(elewise_add_out_repr())
                          ->assert_is_op_output("elementwise_add", "Out");

  ele_out_var->AsIntermediate()->assert_is_ops_input(act_types);

  auto *act = pattern->NewNode(act_repr())->assert_is_ops(act_types);

  auto *act_out_var =
      pattern->NewNode(act_out_repr())->assert_is_ops_output(act_types, "Out");

  ele_add->LinksFrom({ele_x_var, ele_y_var}).LinksTo({ele_out_var});
  act->LinksFrom({ele_out_var}).LinksTo({act_out_var});

  return act_out_var;
}

PDNode *patterns::ElewiseAddActInplaceGrad::operator()(
    paddle::framework::ir::PDNode *d_act_out_var,
    std::unordered_set<std::string> act_types) {
  // act_grad: in["Out", "Out@GRAD"], out["X@GRAD"]
  // ele_add_grad: in["Y", "Out@GRAD"], out["X@GRAD", "Y@GRAD"]
  auto *act_grad = pattern->NewNode(act_grad_repr())->assert_is_ops(act_types);

  auto *act_out_var =
      pattern->NewNode(act_out_repr())->assert_is_ops_input(act_types, "Out");

  auto *d_intermediate_var =
      pattern->NewNode(d_itermediate_out_repr())
          ->assert_is_ops_output(act_types, GradVarName("X"));

  act_grad->LinksFrom({d_act_out_var, act_out_var})
      .LinksTo({d_intermediate_var});

  auto *ele_y_var = pattern->NewNode(ele_y_repr())
                        ->assert_is_not_ctrl_var()
                        ->assert_is_op_input("elementwise_add_grad", "Y");

  auto *ele_add_grad = pattern->NewNode(ele_add_grad_repr())
                           ->assert_is_op("elementwise_add_grad");

  auto *d_ele_x_var =
      pattern->NewNode(d_ele_x_repr())
          ->assert_is_not_ctrl_var()
          ->assert_is_op_output("elementwise_add_grad", GradVarName("X"));

  auto *d_ele_y_var =
      pattern->NewNode(d_ele_y_repr())
          ->assert_is_not_ctrl_var()
          ->assert_is_op_output("elementwise_add_grad", GradVarName("Y"));

  ele_add_grad->LinksFrom({d_intermediate_var, ele_y_var})
      .LinksTo({d_ele_x_var, d_ele_y_var});

  return ele_add_grad;
}

1105
// conv_type: conv2d, conv3d, conv2d_transpose
M
Michal Gallus 已提交
1106
PDNode *patterns::ConvBias::operator()(
1107
    paddle::framework::ir::PDNode *conv_input, std::string conv_type) {
M
Michal Gallus 已提交
1108
  // Create Operators
1109 1110
  conv_input->assert_is_op_input(conv_type, "Input");
  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op(conv_type);
M
Michal Gallus 已提交
1111 1112 1113 1114
  auto *eltiwse_op =
      pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
  // Create variables
  // Filter
Y
Yihua Xu 已提交
1115 1116 1117
  auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
                              ->AsInput()
                              ->assert_is_persistable_var()
1118
                              ->assert_is_op_input(conv_type, "Filter");
M
Michal Gallus 已提交
1119
  // intermediate variable, will be removed in the IR after fuse.
Y
Yihua Xu 已提交
1120 1121
  auto *conv_out_var = pattern->NewNode(conv_out_repr())
                           ->AsIntermediate()
1122
                           ->assert_is_only_output_of_op(conv_type)
Y
Yihua Xu 已提交
1123
                           ->assert_is_op_input("elementwise_add");
M
Michal Gallus 已提交
1124 1125 1126
  // Bias stored in elementwise_add
  auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr())
                               ->AsInput()
M
Michal Gallus 已提交
1127
                               ->assert_is_persistable_var()
M
Michal Gallus 已提交
1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138
                               ->assert_is_op_input("elementwise_add", "Y");
  // output
  auto *eltwise_out_var = pattern->NewNode(eltwise_out_repr())
                              ->AsOutput()
                              ->assert_is_op_output("elementwise_add");
  conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});
  eltiwse_op->LinksFrom({conv_out_var, eltwise_bias_var})
      .LinksTo({eltwise_out_var});
  return eltwise_out_var;
}

1139 1140 1141 1142
PDNode *patterns::Conv::operator()() {
  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");

  auto input_var = pattern->NewNode(conv_input_repr())
1143
                       ->AsInput()
1144 1145 1146
                       ->assert_is_op_input("conv2d", "Input");

  auto filter_var = pattern->NewNode(conv_filter_repr())
1147
                        ->AsInput()
1148 1149 1150
                        ->assert_is_op_input("conv2d", "Filter");

  auto output_var = pattern->NewNode(conv_output_repr())
1151
                        ->AsOutput()
1152 1153
                        ->assert_is_op_output("conv2d", "Output");

1154 1155 1156 1157 1158 1159 1160
  conv_op->LinksFrom({input_var, filter_var}).LinksTo({output_var});
  return output_var;
}

PDNode *patterns::ConvResidual::operator()(bool with_residual_data) {
  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");

1161 1162 1163 1164 1165 1166 1167 1168 1169
  if (!with_residual_data) {
    conv_op->assert_more([&](Node *x) {
      auto node_names = x->Op()->InputNames();
      if (!HasInput(x, "ResidualData") ||
          x->Op()->Input("ResidualData").size() == 0)
        return true;
      return false;
    });
  }
1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205

  auto input_var = pattern->NewNode(conv_input_repr())
                       ->AsInput()
                       ->assert_is_op_input("conv2d", "Input");

  auto filter_var = pattern->NewNode(conv_filter_repr())
                        ->AsInput()
                        ->assert_is_op_input("conv2d", "Filter");

  auto output_var = pattern->NewNode(conv_output_repr())
                        ->AsOutput()
                        ->assert_is_op_output("conv2d", "Output");

  std::vector<PDNode *> links_from{input_var, filter_var};

  if (with_residual_data) {
    auto res_conn_var = pattern->NewNode(conv_residual_data_repr())
                            ->AsInput()
                            ->assert_is_op_input("conv2d", "ResidualData");
    links_from.push_back(res_conn_var);
  }

  conv_op->LinksFrom(links_from).LinksTo({output_var});
  return output_var;
}

PDNode *patterns::Pool::operator()() {
  auto pool_op = pattern->NewNode(pool_op_repr())->assert_is_op("pool2d");

  auto input_var = pattern->NewNode(pool_input_repr())
                       ->AsInput()
                       ->assert_is_op_input("pool2d", "X");

  auto output_var = pattern->NewNode(pool_output_repr())
                        ->AsOutput()
                        ->assert_is_op_output("pool2d", "Out");
1206

1207
  pool_op->LinksFrom({input_var}).LinksTo({output_var});
1208 1209 1210
  return output_var;
}

1211
PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var, PDNode *y_var) {
1212 1213 1214
  auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
                                ->assert_is_op("elementwise_add");

1215 1216
  x_var->AsInput()->assert_is_op_input("elementwise_add", "X");
  y_var->AsInput()->assert_is_op_input("elementwise_add", "Y");
1217 1218 1219 1220
  auto out_var = pattern->NewNode(elementwise_add_out_repr())
                     ->AsOutput()
                     ->assert_is_op_output("elementwise_add", "Out");

1221
  elementwise_add_op->LinksFrom({x_var, y_var});
1222 1223 1224 1225
  elementwise_add_op->LinksTo({out_var});

  return out_var;
}
1226

1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237
PDNode *patterns::Concat::operator()() {
  auto concat_op = pattern->NewNode(concat_op_repr())->assert_is_op("concat");

  auto output_var = pattern->NewNode(concat_out_repr())
                        ->AsOutput()
                        ->assert_is_op_output("concat", "Out");

  concat_op->LinksTo({output_var});
  return output_var;
}

1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277
PDNode *patterns::ConcatReLU::operator()() {
  auto concat_op = pattern->NewNode(concat_op_repr())->assert_is_op("concat");
  auto relu_op = pattern->NewNode(relu_op_repr())->assert_is_op("relu");

  auto concat_out =
      pattern->NewNode(concat_out_repr())->assert_is_op_output("concat", "Out");

  auto relu_out = pattern->NewNode(relu_out_repr())
                      ->AsOutput()
                      ->assert_is_op_output("relu", "Out");

  concat_op->LinksTo({concat_out});
  relu_op->LinksFrom({concat_out}).LinksTo({relu_out});

  return relu_out;
}

PDNode *patterns::ConvConcatReLU::operator()() {
  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
  auto concat_op = pattern->NewNode(concat_op_repr())->assert_is_op("concat");
  auto relu_op = pattern->NewNode(relu_op_repr())->assert_is_op("relu");

  auto conv_out = pattern->NewNode(conv_out_repr())
                      ->assert_is_op_output("conv2d", "Output");

  auto concat_out = pattern->NewNode(concat_out_repr())
                        ->assert_is_op_output("concat", "Out")
                        ->assert_is_op_input("relu", "X");

  auto relu_out = pattern->NewNode(relu_out_repr())
                      ->AsOutput()
                      ->assert_is_op_output("relu", "Out");

  conv_op->LinksTo({conv_out});
  concat_op->LinksFrom({conv_out}).LinksTo({concat_out});
  relu_op->LinksFrom({concat_out}).LinksTo({relu_out});

  return relu_out;
}

1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294
PDNode *patterns::ConvRequant::operator()() {
  // Create Operators
  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
  auto requant_op =
      pattern->NewNode(requant_op_repr())->assert_is_op("requantize");
  auto conv_out = pattern->NewNode(conv_out_repr())
                      ->assert_is_op_output("conv2d", "Output");
  auto requant_out = pattern->NewNode(requant_out_repr())
                         ->AsOutput()
                         ->assert_is_op_output("requantize", "Output");

  conv_op->LinksTo({conv_out});
  requant_op->LinksFrom({conv_out}).LinksTo({requant_out});

  return requant_out;
}

1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319
PDNode *patterns::PriorBox::operator()() {
  auto prior_box_op =
      pattern->NewNode(prior_box_op_repr())->assert_is_op("prior_box");

  auto input_var = pattern->NewNode(prior_box_input_repr())
                       ->AsInput()
                       ->assert_is_op_input("prior_box", "Input");

  auto image_var = pattern->NewNode(prior_box_image_repr())
                       ->AsInput()
                       ->assert_is_op_input("prior_box", "Image");

  auto boxes_var = pattern->NewNode(prior_box_boxes_repr())
                       ->AsOutput()
                       ->assert_is_op_output("prior_box", "Boxes");

  auto variances_var = pattern->NewNode(prior_box_variances_repr())
                           ->AsOutput()
                           ->assert_is_op_output("prior_box", "Variances");

  prior_box_op->LinksFrom({input_var, image_var})
      .LinksTo({boxes_var, variances_var});
  return boxes_var;
}

H
hjchen2 已提交
1320
std::unordered_set<std::string> conv_act_set({"identity", "relu"});
1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385

PDNode *patterns::ConvElementwiseaddAct::operator()(PDNode *conv_in) {
  conv_in->AsInput();
  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
  auto conv_out = pattern->NewNode(conv_out_repr())
                      ->assert_is_op_output("conv2d")
                      ->assert_is_op_input("elementwise_add", "X")
                      ->AsIntermediate();
  auto conv_filter = pattern->NewNode(conv_filter_repr())
                         ->assert_is_op_input("conv2d", "Filter")
                         ->AsInput();
  auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
                                ->assert_is_op("elementwise_add");
  auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr())
                                  ->assert_is_op_input("elementwise_add", "Y")
                                  ->AsInput();
  auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr())
                                 ->assert_is_op_output("elementwise_add")
                                 ->AsIntermediate();

  auto act_op = pattern->NewNode(act_op_repr())
                    ->assert_is_op()
                    ->assert_more([&](Node *node) {
                      auto op_type = node->Name();
                      return conv_act_set.count(op_type);
                    });

  auto act_out = pattern->NewNode(act_out_repr())
                     ->assert_is_var()
                     // is activation op's output.
                     ->assert_more([&](Node *node) {
                       for (auto *in_op : node->inputs) {
                         if (conv_act_set.count(in_op->Name())) {
                           return true;
                         }
                       }
                       return false;
                     })
                     ->AsOutput();

  conv_op->LinksFrom({conv_in, conv_filter});
  conv_out->LinksFrom({conv_op});
  elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y})
      .LinksTo({elementwise_add_out});
  act_op->LinksFrom({elementwise_add_out}).LinksTo({act_out});

  return act_out;
}

PDNode *patterns::ConvElementwiseadd2Act::operator()(PDNode *conv_in) {
  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
  auto conv_filter = pattern->NewNode(conv_filter_repr())
                         ->assert_is_op_input("conv2d", "Filter")
                         ->AsInput();
  auto conv_out = pattern->NewNode(conv_out_repr())
                      ->assert_is_op_output("conv2d")
                      ->assert_is_op_input("elementwise_add", "X")
                      ->AsIntermediate();
  auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
                                ->assert_is_op("elementwise_add");
  auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr())
                                  ->assert_is_op_input("elementwise_add", "Y")
                                  ->AsInput();
  auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr())
                                 ->assert_is_op_output("elementwise_add")
H
hjchen2 已提交
1386
                                 ->assert_is_op_input("elementwise_add", "Y")
1387 1388 1389 1390 1391
                                 ->AsIntermediate();

  auto elementwise_add_op_1 = pattern->NewNode(elementwise_add_op_1_repr())
                                  ->assert_is_op("elementwise_add");
  auto elementwise_add_in_y_1 = pattern->NewNode(elementwise_add_in_y_1_repr())
H
hjchen2 已提交
1392
                                    ->assert_is_op_input("elementwise_add", "X")
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
                                    ->AsInput();
  auto elementwise_add_out_1 = pattern->NewNode(elementwise_add_out_1_repr())
                                   ->assert_is_op_output("elementwise_add")
                                   ->AsIntermediate();

  auto act_op = pattern->NewNode(act_op_repr())
                    ->assert_is_op()
                    ->assert_more([&](Node *node) {
                      auto op_type = node->Name();
                      return conv_act_set.count(op_type);
                    });
  auto act_out = pattern->NewNode(act_out_repr())
                     ->assert_is_var()
                     // is activation op's output.
                     ->assert_more([&](Node *node) {
                       for (auto *in_op : node->inputs) {
                         if (conv_act_set.count(in_op->Name())) {
                           return true;
                         }
                       }
                       return false;
                     })
                     ->AsOutput();

  conv_op->LinksFrom({conv_in, conv_filter}).LinksTo({conv_out});
  elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y})
      .LinksTo({elementwise_add_out});
H
hjchen2 已提交
1420 1421
  elementwise_add_op_1->LinksFrom({elementwise_add_out, elementwise_add_in_y_1})
      .LinksTo({elementwise_add_out_1});
1422 1423 1424 1425
  act_op->LinksFrom({elementwise_add_out_1}).LinksTo({act_out});
  return act_out;
}

N
nhzlx 已提交
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
PDNode *patterns::ConvElementwiseadd::operator()(PDNode *conv_in) {
  conv_in->AsInput();
  auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d");
  auto conv_out = pattern->NewNode(conv_out_repr())
                      ->assert_is_op_output("conv2d")
                      ->assert_is_op_input("elementwise_add", "X")
                      ->AsIntermediate();
  auto conv_filter = pattern->NewNode(conv_filter_repr())
                         ->assert_is_op_input("conv2d", "Filter")
                         ->AsInput();
  auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
                                ->assert_is_op("elementwise_add");
  auto elementwise_add_in_y = pattern->NewNode(elementwise_add_in_y_repr())
                                  ->assert_is_op_input("elementwise_add", "Y")
                                  ->AsInput();
  auto elementwise_add_out = pattern->NewNode(elementwise_add_out_repr())
                                 ->assert_is_op_output("elementwise_add")
                                 ->AsOutput();

  conv_op->LinksFrom({conv_in, conv_filter});
  conv_out->LinksFrom({conv_op});
  elementwise_add_op->LinksFrom({conv_out, elementwise_add_in_y})
      .LinksTo({elementwise_add_out});

  return elementwise_add_out;
}

N
nhzlx 已提交
1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498
PDNode *patterns::ConvAffineChannel::operator()(
    paddle::framework::ir::PDNode *conv_input, bool with_eltwise_add) {
  // Create Operators
  conv_input->assert_is_op_input("conv2d", "Input");
  auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");

  PDNode *eltwise_op = nullptr;
  if (with_eltwise_add) {
    eltwise_op =
        pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
  }

  auto *affine_channel_op =
      pattern->NewNode(affine_channel_repr())->assert_is_op("affine_channel");
  // Create variables
  // Conv Filter
  auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
                              ->AsInput()
                              ->assert_is_persistable_var()
                              ->assert_is_op_input("conv2d", "Filter");

  auto *conv_out_var = pattern->NewNode(conv_out_repr())
                           ->AsIntermediate()
                           ->assert_is_only_output_of_op("conv2d");

  PDNode *eltwise_y_in_var = nullptr;
  PDNode *eltwise_out_var = nullptr;
  if (with_eltwise_add) {
    // Conv output as Bias input
    conv_out_var->assert_is_op_input("elementwise_add", "X");
    // Bias
    eltwise_y_in_var = pattern->NewNode(eltwise_y_in_repr())
                           ->assert_is_op_input("elementwise_add", "Y")
                           ->AsInput();
    eltwise_out_var = pattern->NewNode(eltwise_out_repr())
                          ->AsIntermediate()
                          ->assert_is_only_output_of_op("elementwise_add");
  } else {
    // Conv output as AffineChannel input
    conv_out_var->assert_is_op_input("affine_channel", "X");
  }

  // AC Scale
  auto *ac_scale_var = pattern->NewNode(ac_scale_repr())
                           ->AsInput()
                           ->assert_is_persistable_var()
1499
                           ->assert_has_n_outputs(1)
N
nhzlx 已提交
1500 1501 1502 1503 1504
                           ->assert_is_op_input("affine_channel", "Scale");
  // AC Bias
  auto *ac_bias_var = pattern->NewNode(ac_bias_repr())
                          ->AsInput()
                          ->assert_is_persistable_var()
1505
                          ->assert_has_n_outputs(1)
N
nhzlx 已提交
1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526
                          ->assert_is_op_input("affine_channel", "Bias");

  // AC output
  auto *ac_out_var = pattern->NewNode(ac_out_repr())
                         ->AsOutput()
                         ->assert_is_op_output("affine_channel");

  conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var});

  if (with_eltwise_add) {
    eltwise_op->LinksFrom({conv_out_var, eltwise_y_in_var})
        .LinksTo({eltwise_out_var});
    affine_channel_op->LinksFrom({eltwise_out_var, ac_scale_var, ac_bias_var})
        .LinksTo({ac_out_var});
  } else {
    affine_channel_op->LinksFrom({conv_out_var, ac_scale_var, ac_bias_var})
        .LinksTo({ac_out_var});
  }
  return ac_out_var;
}

1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571
PDNode *patterns::DequantQuantAny::operator()() {
  auto *dequant_in = pattern->NewNode(dequant_in_repr())
                         ->AsInput()
                         ->assert_is_op_input("dequantize", "Input");

  auto *dequant_op =
      pattern->NewNode(dequant_op_repr())->assert_is_op("dequantize");

  auto *dequant_out = pattern->NewNode(dequant_out_repr())
                          ->AsOutput()
                          ->assert_is_op_output("dequantize", "Output");

  auto *quant_op = pattern->NewNode(quant_op_repr())
                       ->assert_is_op("quantize")
                       ->AsIntermediate();

  auto *quant_out = pattern->NewNode(quant_out_repr())
                        ->AsOutput()
                        ->assert_is_op_output("quantize");

  auto *next_op = pattern->NewNode(next_op_repr())->assert_is_op();

  dequant_op->LinksFrom({dequant_in}).LinksTo({dequant_out});
  quant_op->LinksFrom({dequant_out}).LinksTo({quant_out});
  next_op->LinksFrom({quant_out});

  return quant_out;
}

PDNode *patterns::DequantAny::operator()() {
  auto *dequant_op =
      pattern->NewNode(dequant_op_repr())->assert_is_op("dequantize");

  auto *dequant_out = pattern->NewNode(dequant_out_repr())
                          ->AsOutput()
                          ->assert_is_op_output("dequantize", "Output");

  auto *next_op = pattern->NewNode(next_op_repr())->assert_is_op();

  dequant_op->LinksTo({dequant_out});
  next_op->LinksFrom({dequant_out});

  return dequant_out;
}

1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634
// a -> transpose_op(1) -> transpose_out_a -> flatten_op(1) -> flatten_out_a
// b -> transpose_op(2) -> transpose_out_b -> flatten_op(2) -> flatten_out_b
// ...
// z -> transpose_op(n) -> transpose_out_z -> flatten_op(n) -> flatten_out_z
// flatten_out_a -> concat_op  flatten_out_b -> concat_op ... flatten_out_z ->
// concat_op
PDNode *patterns::TransposeFlattenConcat::operator()(
    std::vector<PDNode *> conv_in, int times) {
  // The times represents the repeat times of the
  // {trans, trans_out, flatten, flatten_out}
  const int kNumFields = 4;
  const int kTransOutOffset = 1;
  const int kFlattenOffset = 2;
  const int kFlattenOutOffset = 3;

  std::vector<PDNode *> nodes;

  for (int i = 0; i < times; i++) {
    nodes.push_back(
        pattern->NewNode(GetNodeName("transpose" + std::to_string(i)))
            ->assert_is_op("transpose2"));
    nodes.push_back(
        pattern->NewNode(GetNodeName("transpose_out" + std::to_string(i)))
            ->assert_is_op_output("transpose2")
            ->assert_is_op_input("flatten2", "X")
            ->AsIntermediate());
    nodes.push_back(pattern->NewNode(GetNodeName("flatten" + std::to_string(i)))
                        ->assert_is_op("flatten2"));

    nodes.push_back(
        pattern->NewNode(GetNodeName("flatten_out" + std::to_string(i)))
            ->assert_is_op_output("flatten2")
            ->assert_is_op_nth_input("concat", "X", i)
            ->AsIntermediate());
  }

  auto concat_op = pattern->NewNode(GetNodeName("concat"))
                       ->assert_is_op("concat")
                       ->assert_op_has_n_inputs("concat", times);
  auto concat_out = pattern->NewNode(GetNodeName("concat_out"))
                        ->assert_is_op_output("concat")
                        ->AsOutput();

  std::vector<PDNode *> flatten_outs;
  for (int i = 0; i < times; i++) {
    conv_in[i]->AsInput();
    // trans
    nodes[i * kNumFields]->LinksFrom({conv_in[i]});
    // trans_out
    nodes[i * kNumFields + kTransOutOffset]->LinksFrom({nodes[i * kNumFields]});
    // flatten
    nodes[i * kNumFields + kFlattenOffset]->LinksFrom(
        {nodes[i * kNumFields + kTransOutOffset]});
    // flatten_out
    nodes[i * kNumFields + kFlattenOutOffset]->LinksFrom(
        {nodes[i * kNumFields + kFlattenOffset]});
    flatten_outs.push_back(nodes[i * kNumFields + kFlattenOutOffset]);
  }

  concat_op->LinksFrom(flatten_outs).LinksTo({concat_out});
  return concat_out;
}

1635
PDNode *patterns::AnakinDetectionPattern::operator()(
N
nhzlx 已提交
1636 1637
    std::vector<PDNode *> conv_in, int times, std::string priorbox_type,
    bool is_reshape) {
1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651
  // The times represents the repeat times of the
  // {prior_box, prior_box_loc_out, flatten, prior_box_var_out, reshape}
  const int kNumFields = 7;
  const int kPriorBoxLocOffset = 1;
  const int kReshape1Offset = 2;
  const int kReshape1OutOffset = 3;
  const int kPriorBoxVarOffset = 4;
  const int kReshape2Offset = 5;
  const int kReshape2OutOffset = 6;

  const int kBoxCoderThirdInputOffset = times;
  const int kMultiClassSecondInputNmsOffset = times + 1;

  std::vector<PDNode *> nodes;
N
nhzlx 已提交
1652
  std::string op_after_priorbox = is_reshape ? "reshape2" : "flatten2";
1653 1654 1655 1656

  for (int i = 0; i < times; i++) {
    nodes.push_back(
        pattern->NewNode(GetNodeName("prior_box" + std::to_string(i)))
N
nhzlx 已提交
1657
            ->assert_is_op(priorbox_type));
1658
    nodes.push_back(pattern->NewNode(GetNodeName("box_out" + std::to_string(i)))
N
nhzlx 已提交
1659 1660
                        ->assert_is_op_output(priorbox_type, "Boxes")
                        ->assert_is_op_input(op_after_priorbox, "X")
1661 1662 1663
                        ->AsIntermediate());
    nodes.push_back(
        pattern->NewNode(GetNodeName("reshape1" + std::to_string(i)))
N
nhzlx 已提交
1664
            ->assert_is_op(op_after_priorbox));
1665 1666 1667

    nodes.push_back(
        pattern->NewNode(GetNodeName("reshape1_out" + std::to_string(i)))
N
nhzlx 已提交
1668
            ->assert_is_op_output(op_after_priorbox)
1669 1670 1671 1672 1673
            ->assert_is_op_nth_input("concat", "X", i)
            ->AsIntermediate());

    nodes.push_back(
        pattern->NewNode(GetNodeName("box_var_out" + std::to_string(i)))
N
nhzlx 已提交
1674 1675
            ->assert_is_op_output(priorbox_type, "Variances")
            ->assert_is_op_input(op_after_priorbox, "X")
1676 1677 1678
            ->AsIntermediate());
    nodes.push_back(
        pattern->NewNode(GetNodeName("reshape2" + std::to_string(i)))
N
nhzlx 已提交
1679
            ->assert_is_op(op_after_priorbox));
1680 1681 1682

    nodes.push_back(
        pattern->NewNode(GetNodeName("reshape2_out" + std::to_string(i)))
N
nhzlx 已提交
1683
            ->assert_is_op_output(op_after_priorbox)
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
            ->assert_is_op_nth_input("concat", "X", i)
            ->AsIntermediate());
  }

  auto concat_op1 = pattern->NewNode(GetNodeName("concat1"))
                        ->assert_is_op("concat")
                        ->assert_op_has_n_inputs("concat", times);
  auto concat_out1 = pattern->NewNode(GetNodeName("concat1_out"))
                         ->assert_is_op_output("concat")
                         ->AsIntermediate();

  auto concat_op2 = pattern->NewNode(GetNodeName("concat2"))
                        ->assert_is_op("concat")
                        ->assert_op_has_n_inputs("concat", times);
  auto concat_out2 = pattern->NewNode(GetNodeName("concat2_out"))
                         ->assert_is_op_output("concat")
                         ->AsIntermediate();

  auto box_coder_op = pattern->NewNode(GetNodeName("box_coder"))
                          ->assert_is_op("box_coder")
                          ->assert_op_has_n_inputs("box_coder", 3);

  auto box_coder_out = pattern->NewNode(GetNodeName("box_coder_out"))
                           ->assert_is_op_output("box_coder")
                           ->AsIntermediate();

1710 1711 1712 1713 1714 1715 1716 1717 1718 1719
  auto transpose_before_nms =
      pattern->NewNode(GetNodeName("transpose_before_nms"))
          ->assert_is_op("transpose2");

  auto transpose_before_nms_out =
      pattern->NewNode(GetNodeName("transpose_before_nms_out"))
          ->assert_is_op_output("transpose2")
          ->assert_is_op_input("multiclass_nms", "Scores")
          ->AsIntermediate();

1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769
  auto multiclass_nms_op = pattern->NewNode(GetNodeName("multiclass_nms"))
                               ->assert_is_op("multiclass_nms")
                               ->assert_op_has_n_inputs("multiclass_nms", 2);

  auto multiclass_nms_out = pattern->NewNode(GetNodeName("multiclass_nms_out"))
                                ->assert_is_op_output("multiclass_nms")
                                ->AsOutput();

  std::vector<PDNode *> reshape1_outs;
  std::vector<PDNode *> reshape2_outs;

  for (int i = 0; i < times; i++) {
    conv_in[i]->AsInput();
    // prior_box
    nodes[i * kNumFields]->LinksFrom({conv_in[i]});
    // prior_box box out
    nodes[i * kNumFields + kPriorBoxLocOffset]->LinksFrom(
        {nodes[i * kNumFields]});
    // reshape
    nodes[i * kNumFields + kReshape1Offset]->LinksFrom(
        {nodes[i * kNumFields + kPriorBoxLocOffset]});
    // reshape_out
    nodes[i * kNumFields + kReshape1OutOffset]->LinksFrom(
        {nodes[i * kNumFields + kReshape1Offset]});

    nodes[i * kNumFields + kPriorBoxVarOffset]->LinksFrom(
        {nodes[i * kNumFields]});
    // reshape
    nodes[i * kNumFields + kReshape2Offset]->LinksFrom(
        {nodes[i * kNumFields + kPriorBoxVarOffset]});
    // reshape_out
    nodes[i * kNumFields + kReshape2OutOffset]->LinksFrom(
        {nodes[i * kNumFields + kReshape2Offset]});

    reshape1_outs.push_back(nodes[i * kNumFields + kReshape1OutOffset]);
    reshape2_outs.push_back(nodes[i * kNumFields + kReshape2OutOffset]);
  }

  concat_op1->LinksFrom(reshape1_outs);
  concat_op2->LinksFrom(reshape2_outs);
  concat_out1->LinksFrom({concat_op1});
  concat_out2->LinksFrom({concat_op2});

  conv_in[kBoxCoderThirdInputOffset]->AsInput();
  conv_in[kMultiClassSecondInputNmsOffset]->AsInput();

  box_coder_op->LinksFrom(
      {concat_out1, concat_out2, conv_in[kBoxCoderThirdInputOffset]});
  box_coder_out->LinksFrom({box_coder_op});

1770 1771 1772 1773
  transpose_before_nms->LinksFrom({conv_in[kMultiClassSecondInputNmsOffset]});
  transpose_before_nms_out->LinksFrom({transpose_before_nms});

  multiclass_nms_op->LinksFrom({box_coder_out, transpose_before_nms_out})
1774 1775 1776 1777 1778
      .LinksTo({multiclass_nms_out});

  return multiclass_nms_out;
}

N
nhzlx 已提交
1779
PDNode *patterns::FillConstantElementWiseMulFuse::operator()(
1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801
    PDNode *elementwise_op_input) {
  auto fill_constant =
      pattern->NewNode(fill_constant_repr())->assert_is_op("fill_constant");

  auto fill_constant_out = pattern->NewNode(fill_constant_out_repr())
                               ->assert_is_op_output("fill_constant")
                               ->assert_is_op_input("elementwise_mul", "Y")
                               ->AsIntermediate();

  auto elementwise_mul_op =
      pattern->NewNode(elementwise_mul_repr())->assert_is_op("elementwise_mul");

  auto elementwise_mul_out = pattern->NewNode(elementwise_mul_out_repr())
                                 ->assert_is_op_output("elementwise_mul")
                                 ->AsOutput();

  fill_constant_out->LinksFrom({fill_constant});
  elementwise_mul_op->LinksFrom({elementwise_op_input, fill_constant_out});
  elementwise_mul_out->LinksFrom({elementwise_mul_op});
  return elementwise_mul_out;
}

N
nhzlx 已提交
1802 1803 1804
void patterns::QuantDequantOpFuse::operator()(PDNode *quant_op_input,
                                              const std::string &op_type,
                                              const std::string &weight_name,
1805
                                              int times,
1806 1807 1808
                                              const std::string &quant_type,
                                              const std::string &dequant_type) {
  int kNumFields = 5;
N
nhzlx 已提交
1809 1810 1811 1812 1813
  const int kQuantizedWeightOffset = 0;
  const int kQuantizedOpOffset = 1;
  const int kQuantizedOpOutOffset = 2;
  const int kDequantOpOffset = 3;
  const int kDequantOpOutOffset = 4;
1814 1815
  const int kDequantOpWeightScaleOffset = 5;

N
nhzlx 已提交
1816
  // the quant op always be one.
1817 1818 1819 1820 1821
  auto quant_op_in_scale = pattern->NewNode(GetNodeName("quant_op_in_scale"))
                               ->assert_is_op_input(quant_type, "InScale")
                               ->AsInput();
  auto quant_op =
      pattern->NewNode(GetNodeName("quant_op"))->assert_is_op(quant_type);
N
nhzlx 已提交
1822

1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835
  PDNode *quant_op_out_scale = nullptr;
  if (dequant_type == "fake_channel_wise_dequantize_max_abs") {
    kNumFields += 1;
    quant_op_out_scale = pattern->NewNode(GetNodeName("quant_op_out_scale"))
                             ->assert_is_op_output(quant_type, "OutScale")
                             ->assert_is_op_nth_input(dequant_type, "Scales", 1)
                             ->AsIntermediate();
  } else {
    quant_op_out_scale = pattern->NewNode(GetNodeName("quant_op_out_scale"))
                             ->assert_is_op_output(quant_type, "OutScale")
                             ->assert_is_op_input(dequant_type, "Scale")
                             ->AsIntermediate();
  }
N
nhzlx 已提交
1836

1837 1838 1839 1840
  auto quant_op_out = pattern->NewNode(GetNodeName("quant_op_out"))
                          ->assert_is_op_output(quant_type, "Out")
                          ->assert_is_op_input(op_type)
                          ->AsIntermediate();
N
nhzlx 已提交
1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855

  // there are 'times' quantized and dequant op
  std::vector<PDNode *> nodes;
  for (int i = 0; i < times; i++) {
    nodes.push_back(
        pattern->NewNode(GetNodeName("quantized_op_weight") + std::to_string(i))
            ->assert_is_op_input(op_type, weight_name)
            ->AsInput());
    nodes.push_back(
        pattern->NewNode(GetNodeName("quantized_op") + std::to_string(i))
            ->assert_is_op(op_type));

    nodes.push_back(
        pattern->NewNode(GetNodeName("quantized_op_out") + std::to_string(i))
            ->assert_is_op_output(op_type)
1856
            ->assert_is_op_input(dequant_type, "X")
N
nhzlx 已提交
1857 1858 1859 1860
            ->AsIntermediate());

    nodes.push_back(
        pattern->NewNode(GetNodeName("dequant_op") + std::to_string(i))
1861 1862
            ->assert_is_op(dequant_type));

N
nhzlx 已提交
1863 1864
    nodes.push_back(
        pattern->NewNode(GetNodeName("dequant_op_out") + std::to_string(i))
1865
            ->assert_is_op_output(dequant_type, "Out")
N
nhzlx 已提交
1866
            ->AsOutput());
1867 1868 1869 1870 1871 1872 1873 1874

    if (dequant_type == "fake_channel_wise_dequantize_max_abs") {
      nodes.push_back(pattern
                          ->NewNode(GetNodeName("dequant_channel_scale") +
                                    std::to_string(i))
                          ->assert_is_op_nth_input(dequant_type, "Scales", 0)
                          ->AsInput());
    }
N
nhzlx 已提交
1875 1876 1877 1878 1879 1880 1881 1882 1883
  }

  quant_op->LinksFrom({quant_op_input, quant_op_in_scale});
  quant_op_out->LinksFrom({quant_op});
  for (int i = 0; i < times; i++) {
    nodes[i * kNumFields + kQuantizedOpOffset]->LinksFrom(
        {quant_op_out, nodes[i * kNumFields + kQuantizedWeightOffset]});
    nodes[i * kNumFields + kQuantizedOpOutOffset]->LinksFrom(
        {nodes[i * kNumFields + kQuantizedOpOffset]});
1884 1885 1886 1887 1888 1889 1890 1891
    if (dequant_type == "fake_channel_wise_dequantize_max_abs") {
      nodes[i * kNumFields + kDequantOpOffset]->LinksFrom(
          {nodes[i * kNumFields + kQuantizedOpOutOffset], quant_op_out_scale,
           nodes[i * kNumFields + kDequantOpWeightScaleOffset]});
    } else {
      nodes[i * kNumFields + kDequantOpOffset]->LinksFrom(
          {nodes[i * kNumFields + kQuantizedOpOutOffset], quant_op_out_scale});
    }
N
nhzlx 已提交
1892 1893 1894 1895 1896
    nodes[i * kNumFields + kDequantOpOutOffset]->LinksFrom(
        {nodes[i * kNumFields + kDequantOpOffset]});
  }
}

1897 1898 1899
void patterns::ShuffleChannelPattern::operator()(PDNode *reshape1_in) {
  auto reshape1_op =
      pattern->NewNode(reshape1_op_repr())->assert_is_op("reshape2");
1900 1901 1902
  reshape1_op->assert_more([&](Node *x) {
    return boost::get<std::vector<int>>(x->Op()->GetAttr("shape")).size() == 5;
  });
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

  auto reshape1_out = pattern->NewNode(reshape1_out_repr())
                          ->assert_is_op_output("reshape2", "Out")
                          ->assert_is_op_input("transpose2")
                          ->AsIntermediate();

  auto transpose_op =
      pattern->NewNode(transpose_op_repr())->assert_is_op("transpose2");

  auto transpose_out = pattern->NewNode(transpose_out_repr())
                           ->assert_is_op_output("transpose2", "Out")
                           ->assert_is_op_input("reshape2")
                           ->AsIntermediate();

  auto reshape2_op =
      pattern->NewNode(reshape2_op_repr())->assert_is_op("reshape2");
  auto reshape2_out = pattern->NewNode(reshape2_out_repr())
                          ->assert_is_op_output("reshape2", "Out")
                          ->AsOutput();

  reshape1_op->LinksFrom({reshape1_in});
  reshape1_out->LinksFrom({reshape1_op});
  transpose_op->LinksFrom({reshape1_out});
  transpose_out->LinksFrom({transpose_op});
  reshape2_op->LinksFrom({transpose_out});
  reshape2_out->LinksFrom({reshape2_op});
}

1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965
void patterns::DeleteQuantDequantOpPattern::operator()() {
  auto any_op_out =
      pattern->NewNode(any_op_out_repr())
          ->assert_is_op_input(
              "fake_quantize_dequantize_moving_average_abs_max", "X")
          ->AsInput();

  auto quant_dequant_op_inscale =
      pattern->NewNode(quant_dequant_op_inscale_repr())
          ->assert_is_op_input(
              "fake_quantize_dequantize_moving_average_abs_max", "InScale")
          ->AsInput();
  auto quant_dequant_op =
      pattern->NewNode(quant_dequant_op_repr())
          ->assert_is_op("fake_quantize_dequantize_moving_average_abs_max");

  auto quant_dequant_out =
      pattern->NewNode(quant_dequant_op_out_repr())
          ->assert_is_op_output(
              "fake_quantize_dequantize_moving_average_abs_max", "Out")
          ->AsIntermediate();

  auto quant_dequant_op_outscale =
      pattern->NewNode(quant_dequant_op_outscale_repr())
          ->assert_is_op_output(
              "fake_quantize_dequantize_moving_average_abs_max", "OutScale")
          ->AsOutput();
  auto any_op2 = pattern->NewNode(any_op2_repr())->assert_is_op()->AsOutput();

  quant_dequant_op->LinksFrom({any_op_out, quant_dequant_op_inscale});
  quant_dequant_op_outscale->LinksFrom({quant_dequant_op});
  quant_dequant_out->LinksFrom({quant_dequant_op});
  any_op2->LinksFrom({quant_dequant_out});
}

1966 1967 1968
}  // namespace ir
}  // namespace framework
}  // namespace paddle