graph_pattern_detector.h 27.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#pragma once

#ifdef PADDLE_WITH_TESTING
#include <gtest/gtest_prod.h>
#endif

21
#include <memory>
22
#include <numeric>
23
#include <string>
24 25
#include <unordered_map>
#include <unordered_set>
26 27
#include <utility>
#include <vector>
28 29
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/node.h"
30
#include "paddle/fluid/inference/analysis/dot.h"
31 32 33 34

namespace paddle {
namespace framework {
namespace ir {
35
class PDPattern;
36

37
// Some basic terminologies:
38 39 40 41 42 43 44 45 46
//   - PDPattern: a pattern defined as a data flow graph.
//   - PDNode: the node in the pattern, each PDNode represents an `ir::Node`
//     that meets some conditions defined in `PDNode.teller`.
//   - A pattern is defined with PDNodes with edges.

// Pattern detector node. This node helps to build a pattern.
struct PDNode {
  // tell whether an ir::Node* is a candidation for a PDNode.
  using teller_t = std::function<bool(Node*)>;
47
  enum class Type { kOp, kVar };
Y
Yan Chunwei 已提交
48 49 50 51 52 53
  enum class Role {
    kUnknown,      // No role,
    kInput,        // an input and will be retained,
    kOutput,       // an output and will be retained,
    kIntermediate  // will be removed after handler.
  };
54

55 56 57
  // this link to others
  PDNode& LinksTo(const std::vector<PDNode*>& others);
  PDNode& LinksFrom(const std::vector<PDNode*>& others);
58 59

  bool Tell(Node* node) const {
Y
Yan Chunwei 已提交
60 61 62 63 64 65
    if (teller_) return teller_(node);

    for (auto& asrt : asserts_) {
      if (!asrt(node)) return false;
    }
    return true;
66 67
  }

68 69 70
  bool IsOp() const { return type_ == Type::kOp; }
  bool IsVar() const { return type_ == Type::kVar; }

71 72 73
  const std::string& name() const { return name_; }

  PDNode& operator=(const PDNode&) = delete;
Y
Yan Chunwei 已提交
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
  PDNode(const PDNode&) = delete;

  // Mark this node is an Input of a subgraph and will be retained.
  PDNode* AsInput() {
    role_ = Role::kInput;
    return this;
  }
  // Mark this node is an Output of a subgraph and will be retained.
  PDNode* AsOutput() {
    role_ = Role::kOutput;
    return this;
  }
  // Mark this node will be removed, so all the links should be inside a matched
  // sub-graph.
  PDNode* AsIntermediate() {
    role_ = Role::kIntermediate;
    return this;
  }

  bool IsIntermediate() const { return role_ == Role::kIntermediate; }
  bool IsInput() const { return role_ == Role::kInput; }
  bool IsOutput() const { return role_ == Role::kOutput; }

  // Assertions, helper functions to simplify the pattern definition.
  PDNode* assert_is_op();
  PDNode* assert_is_op(const std::string& op_type);
  PDNode* assert_is_var();
C
chengduo 已提交
101
  PDNode* assert_is_not_ctrl_var();
Y
Yan Chunwei 已提交
102 103 104
  PDNode* assert_var_not_persistable();
  PDNode* assert_is_persistable_var();
  PDNode* assert_is_op_output(const std::string& op_type);
105 106
  PDNode* assert_is_op_output(const std::string& op_type,
                              const std::string& argument);
Y
Yan Chunwei 已提交
107
  PDNode* assert_is_op_input(const std::string& op_type);
108 109
  PDNode* assert_is_op_input(const std::string& op_type,
                             const std::string& argument);
Y
Yan Chunwei 已提交
110 111 112 113 114 115 116 117 118
  PDNode* assert_is_op_nth_input(const std::string& op_type,
                                 const std::string& argument, int nth);
  PDNode* assert_is_op_nth_output(const std::string& op_type,
                                  const std::string& argument, int nth);
  PDNode* assert_is_only_input_of_op(const std::string& op_type);
  PDNode* assert_is_only_output_of_op(const std::string& op_type);
  PDNode* assert_op_has_n_inputs(const std::string& op_type, size_t n);
  PDNode* assert_op_has_n_outputs(const std::string& op_type, size_t n);
  PDNode* assert_more(teller_t&& teller);
119

C
chengduo 已提交
120 121 122 123 124 125 126 127 128 129 130 131 132 133
  PDNode* assert_is_ops_output(const std::unordered_set<std::string>& op_types);
  PDNode* assert_is_ops(const std::unordered_set<std::string>& op_types);
  PDNode* assert_is_ops_output(const std::unordered_set<std::string>& op_types,
                               const std::string& argument);
  PDNode* assert_is_ops_nth_input(
      const std::unordered_set<std::string>& op_types,
      const std::string& argument, int nth);
  PDNode* assert_is_ops_input(const std::unordered_set<std::string>& op_types);
  PDNode* assert_is_ops_input(const std::unordered_set<std::string>& op_types,
                              const std::string& argument);
  PDNode* assert_is_ops_nth_output(
      const std::unordered_set<std::string>& op_types,
      const std::string& argument, int nth);

T
tensor-tang 已提交
134 135 136 137 138 139 140 141 142
  template <typename T>
  PDNode* assert_op_attr(const std::string& attr_name, const T& attr) {
    asserts_.emplace_back([=](Node* x) {
      return x && x->IsOp() && x->Op()->HasAttr(attr_name) &&
             boost::get<T>(x->Op()->GetAttr(attr_name)) == attr;
    });
    return this;
  }

143
 private:
Y
Yan Chunwei 已提交
144 145 146
  PDNode(PDPattern* pattern, const std::string& name = "",
         Type type = Type::kVar)
      : pattern_(pattern), name_(name), type_(type) {}
147 148 149 150 151 152 153 154 155 156 157 158 159
  PDNode(teller_t&& teller, PDPattern* pattern, const std::string& name = "",
         Type type = Type::kVar)
      : teller_(std::move(teller)),
        pattern_(pattern),
        name_(name),
        type_(type) {
    PADDLE_ENFORCE(teller_ != nullptr, "invalid teller functer is set.");
  }

  PDNode(PDNode&& other) = default;

  friend class PDPattern;

Y
Yan Chunwei 已提交
160
  // Will removed latter.
161
  teller_t teller_;
Y
Yan Chunwei 已提交
162
  std::vector<teller_t> asserts_;
163
  PDPattern* pattern_;
164
  std::string name_;
165
  Type type_;
Y
Yan Chunwei 已提交
166
  Role role_{Role::kUnknown};
167 168 169 170 171 172 173 174 175 176 177 178
};

/*
 * A pattern in a graph, which defined with PDNode and edges. Most graph
 * patterns can be divided into PDNodes and link relations between them.
 *
 * For example, the FC fusion need to filter the MUL and ELEMENTWISE_ADD
 * operators from the computation graph, the MUL's output should have only one
 * consumer which is the ELEMENTWISE_ADD.
 * This pattern can be defined as with the following pseudo codes
 *
 *     // Create two operator PDNodes.
Y
Yan Chunwei 已提交
179 180
 *     MUL = PDPattern.NewNode().assert_is_op("mul");
 *     ELE = PDPattern.NewNode().assert_is_op("elementwise_add");
181
 *     // Create the variable PDNodes.
Y
Yan Chunwei 已提交
182 183 184 185 186 187
 *     MUL_out = PDPattern.NewNode().assert_is_op_output("mul") \
 *                                  .assert_is_op_input("elementwise_add") \
 *                                  .AsIntermediate();
 *     // Add relations.
 *     MUL->LinksTo({MUL_out});
 *     MUL_out->LinksTo({ELE});
188
 *
Y
Yan Chunwei 已提交
189 190
 * One can add more specific asserts for PDNodes or edges, both the Operator
 * and Variable Nodes can be ruled in PDNode.assert_more(...).
191 192 193 194 195 196 197 198 199 200 201
 *
 * PDPattern can record the general patterns, such as the pattern represents
 *   - Op in CPU -> Op in GPU -> Op in CPU, to findout the IO abnormal place.
 *   - Ops whose inputs and outputs share the same variables
 */
class PDPattern {
 public:
  using edge_t = std::pair<PDNode*, PDNode*>;

  void AddEdge(PDNode* a, PDNode* b);

202
  PDNode* NewNode(PDNode::teller_t&& teller, const std::string& name = NewID());
Y
Yan Chunwei 已提交
203
  PDNode* NewNode(const std::string& name = NewID());
204 205 206
  PDNode* NewNode(const std::string& prefix, const std::string& name) {
    return NewNode(prefix + "/" + name);
  }
Y
Yan Chunwei 已提交
207
  PDNode* RetrieveNode(const std::string& id) const;
208 209 210 211

  const std::vector<std::unique_ptr<PDNode>>& nodes() const { return nodes_; }
  const std::vector<edge_t>& edges() const { return edges_; }

212 213
  std::string DotString() const;

214 215 216 217 218 219
 private:
#ifdef PADDLE_WITH_TESTING
  FRIEND_TEST(PDPattern, AddEdge);
  FRIEND_TEST(PDPattern, NewNode);
#endif

220 221
  static std::string NewID() { return "pdnode-" + std::to_string(id_++); }

222 223
  std::vector<std::unique_ptr<PDNode>> nodes_;
  std::vector<edge_t> edges_;
224 225
  std::unordered_map<std::string, PDNode*> node_map_;
  static size_t id_;
226 227 228
};

/*
229
 * GraphPatternDetector helps to detect the specific patterns in the graph.
230 231 232 233 234 235 236 237 238 239 240
 * Input a pattern, output a list of the matched subgraphs/nodes.
 * This helper can be used to support fuse(conv+batchnorm => batchnorm e.g.).
 *
 * The algorithm has three phases:
 *   1. Mark the nodes that match the defined PDNodes in a PDPattern,
 *   2. Extend a PDNode to subgraphs by deducing the connection relation defined
 *      in PAPattern(the edges),
 *   3. Get the filtered subgraphs and treat them with a pre-defined handler.
 *
 * Usage:
 *    // Create a detector
241
 *    GraphPatternDetector detector;
242 243 244 245 246 247 248 249
 *    // Define the detector's pattern, by adding PDNode and define the edges.
 *    auto* node0 = detector.mutable_pattern().AddNode(...)
 *    auto* node1 = detector.mutable_pattern().AddNode(...)
 *    node0->teller = some lambda.
 *    node1->teller = some lambda.
 *    detector.mutable_pattern().AddEdge(node0, node1);
 *    // Create an handler, to define the behavior of treating the filtered
 *    // subgraphs that comply with the patterns.
250
 *    GraphPatternDetector::handle_t handler = some labmda
251 252 253
 *    // Execute the detector.
 *    detector(&graph, handler);
 */
254
class GraphPatternDetector {
255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
 public:
  using subgraph_t = std::unordered_map<PDNode*, Node*>;

  // Operate on the detected pattern.
  using handle_t =
      std::function<void(const subgraph_t& /*hitted pattern*/, Graph*)>;

  void operator()(Graph* graph, handle_t handler);

  const PDPattern& pattern() const { return pattern_; }
  PDPattern* mutable_pattern() { return &pattern_; }

 private:
  // Mark the nodes that fits the pattern.
  bool MarkPDNodesInGraph(const ir::Graph& graph);

  // Detect all the pattern and output the hit records.
  std::vector<subgraph_t> DetectPatterns();

  // Remove duplicate patterns.
  void UniquePatterns(std::vector<subgraph_t>* subgraphs);

  // Remove overlapped match subgraphs, when overlapped, keep the previous one.
Y
Yan Chunwei 已提交
278 279
  // The intermediate PDNodes will be removed, so can't shared by multiple
  // patterns.
280 281
  void RemoveOverlappedMatch(std::vector<subgraph_t>* subgraphs);

Y
Yan Chunwei 已提交
282 283 284
  // Validate whether the intermediate nodes are linked by external nodes.
  void ValidateByNodeRole(std::vector<subgraph_t>* subgraphs);

285 286 287 288 289 290 291 292 293 294 295 296
#ifdef PADDLE_WITH_TESTING
  FRIEND_TEST(GraphPatternDetecter, MarkPDNodesInGraph);
  FRIEND_TEST(GraphPatternDetecter, DetectPatterns);
#endif

 private:
  using hit_rcd_t =
      std::pair<Node* /*node in graph*/, PDNode* /*node in pattern*/>;
  PDPattern pattern_;
  std::unordered_map<const PDNode*, std::unordered_set<Node*>> pdnodes2nodes_;
};

297 298
// some helper methods.

299 300 301 302 303
// Tell if a var links to an Op
bool VarLinksToOp(Node* node, const std::string& op_type);

// Tell if an op links to a var
bool VarLinksFromOp(Node* node, const std::string& op_type);
304 305

// Check whether a var node is a op node's nth input.
306
bool IsNthInput(Node* var, Node* op, const std::string& argument, size_t nth);
307

308 309 310 311 312 313 314 315
// Tell whether a var node is a op node's nth output.
bool IsNthOutput(Node* var, Node* op, const std::string& argument, size_t nth);

// Graph safely remove some nodes, will automatically clean up the edges.
void GraphSafeRemoveNodes(Graph* graph,
                          const std::unordered_set<const Node*>& nodes);

// Some pre-defined patterns those can be reused in multiple passes.
316 317
// The related Fluid Layer or Op should be one pattern here for better re-usage
// across different fusion.
318 319
namespace patterns {

Y
Yan Chunwei 已提交
320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
struct KeyCounter {
  static KeyCounter& Instance() {
    static KeyCounter x;
    return x;
  }

  int IncCounter(const std::string& key) { return dic_[key]++; }

 private:
  std::unordered_map<std::string, size_t> dic_;
};

// Generate a unique PDNode's name with name_scope and id.
// The format is {name_scope}/{repr}/{id}/{name}
static std::string PDNodeName(const std::string& name_scope,
                              const std::string& repr, size_t id,
                              const std::string& name) {
  return string::Sprintf("%s/%s/%d/%s", name_scope, repr, id, name);
}
// Generate a unique PDNode's name.
// The format is {name_scope}/{repr}/{id}
static std::string PDNodeName(const std::string& name_scope,
                              const std::string& repr) {
  return string::Sprintf("%s/%s/%d", name_scope, repr,
                         KeyCounter::Instance().IncCounter(repr));
}
// Generate a unique key. It can be used for a universally unique temporary
// name.
// The format is {repr}/{id}
static std::string UniqueKey(const std::string& repr) {
  return string::Sprintf("%s/%d", repr,
                         KeyCounter::Instance().IncCounter(repr));
}

// Declare a PDNode in a pattern, will create two methods:
// std::string xxx_repr(); return this PDNode's string id.
// PDNode* xxx_n(); return the corresponding PDNode.
#define PATTERN_DECL_NODE(name__)                        \
  std::string name__##_repr() const {                    \
    return PDNodeName(name_scope_, repr_, id_, #name__); \
  }                                                      \
  PDNode* name__##_n() const { return pattern->RetrieveNode(name__##_repr()); }

// Get an ir::Node* from the matched subgraph.
// var: variable.
// arg: the argument declared by PATTERN_DECL_NODE in a pattern definition.
// pat: the pattern object.
#define GET_IR_NODE_FROM_SUBGRAPH(var, arg, pat)                    \
  PADDLE_ENFORCE(subgraph.count(pat.arg##_n()),                     \
                 "Node not found for PDNode %s", pat.arg##_repr()); \
  Node* var = subgraph.at(pat.arg##_n());                           \
  PADDLE_ENFORCE(var, "node %s not exists in the sub-graph", #arg)

// The base class of all the patterns.
struct PatternBase {
  PatternBase(PDPattern* pattern, const std::string& name_scope,
              const std::string& repr)
      : pattern(pattern),
        name_scope_(name_scope),
        repr_(repr),
        id_(KeyCounter::Instance().IncCounter(repr)) {}

  PDPattern* pattern;

 protected:
  std::string name_scope_;
  std::string repr_;
  size_t id_;
};

S
Sylwester Fraczek 已提交
390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
// Conv with batch norm
// op: conv + (elementwise_add +) batch_norm
// named nodes:
// conv_weight, conv_out, conv,
// bn_x, bn_scale, bn_bias, bn_mean,  bn_variance,
// bn_batch_norm, bn_y, bn_mean_out, bn_variance_out,
// bn_saved_mean, bn_saved_variance
struct ConvBN : public PatternBase {
  ConvBN(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_bn") {}

  PDNode* operator()(PDNode* conv_input, bool with_eltwise_add);

  // declare operator node's name
  PATTERN_DECL_NODE(conv);
  PATTERN_DECL_NODE(batch_norm);
  PATTERN_DECL_NODE(eltwise);  // ELEMENTWISE_ADD
  // CONV inputs
  PATTERN_DECL_NODE(conv_weight);  // Filter
  // CONV outputs
  PATTERN_DECL_NODE(conv_out);  // tmp
  // ELTWISE inputs
  PATTERN_DECL_NODE(eltwise_y_in);
  // ELTWISE outputs
  PATTERN_DECL_NODE(eltwise_out);  // tmp
  // BN inputs
  PATTERN_DECL_NODE(bn_scale);
  PATTERN_DECL_NODE(bn_bias);
  PATTERN_DECL_NODE(bn_mean);
  PATTERN_DECL_NODE(bn_variance);
  // BN outputs
  PATTERN_DECL_NODE(bn_out);  // Out
  PATTERN_DECL_NODE(bn_mean_out);
  PATTERN_DECL_NODE(bn_variance_out);
  PATTERN_DECL_NODE(bn_saved_mean);
  PATTERN_DECL_NODE(bn_saved_variance);
};

428 429 430 431
// CONV with ReLU
// op: conv + relu
// named nodes:
// conv_input, conv_weight,
432
// conv_out, conv,
433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448
// relu_out, relu
struct ConvReLU : public PatternBase {
  ConvReLU(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_relu") {}

  PDNode* operator()(PDNode* conv_input);

  // declare operator node's name
  PATTERN_DECL_NODE(conv);
  PATTERN_DECL_NODE(relu);
  // declare variable node's name
  PATTERN_DECL_NODE(conv_weight);
  PATTERN_DECL_NODE(conv_out);
  PATTERN_DECL_NODE(relu_out);
};

T
tensor-tang 已提交
449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473
// SEQCONV with Elementwise_Add ReLU
// op: seqconv + elementwise_add + relu
// named nodes:
// seqconv_input, seqconv_weight,
// seqconv_out, seqconv,
// elementwise_add_bias, elementwise_add_out, elementwise_add
// relu_out, relu
struct SeqConvEltAddRelu : public PatternBase {
  SeqConvEltAddRelu(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "seqconv_eltadd_relu") {}

  PDNode* operator()(PDNode* seqconv_input);

  // declare operator node's name
  PATTERN_DECL_NODE(seqconv);
  PATTERN_DECL_NODE(eltadd);
  PATTERN_DECL_NODE(relu);
  // declare variable node's name
  PATTERN_DECL_NODE(seqconv_weight);
  PATTERN_DECL_NODE(seqconv_out);
  PATTERN_DECL_NODE(eltadd_bias);
  PATTERN_DECL_NODE(eltadd_out);
  PATTERN_DECL_NODE(relu_out);
};

474 475 476 477 478
// FC with bias
// op: mul + elementwise_add
// named nodes:
// mul, elementwise_add
// w, mul_out, bias, fc_out
Y
Yan Chunwei 已提交
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495
struct FC : public PatternBase {
  FC(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "fc") {}

  PDNode* operator()(PDNode* x, bool with_bias);

  // declare operator node's name
  PATTERN_DECL_NODE(fc);
  PATTERN_DECL_NODE(mul);
  PATTERN_DECL_NODE(elementwise_add);
  // declare variable node's name
  PATTERN_DECL_NODE(w);
  PATTERN_DECL_NODE(mul_out);  // (x,w) -> mul_out
  PATTERN_DECL_NODE(bias);
  PATTERN_DECL_NODE(Out);
};

496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512
// Embedding
struct Embedding : public PatternBase {
  Embedding(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "embedding") {}

  PDNode* operator()(PDNode* x);

  // declare operator node's name
  PATTERN_DECL_NODE(lookup_table);
  // Inputs
  //
  PATTERN_DECL_NODE(Ids);
  PATTERN_DECL_NODE(W);  // embeddings
  // Outputs
  PATTERN_DECL_NODE(Out);
};

Y
Yan Chunwei 已提交
513 514 515
struct LSTM : public PatternBase {
  LSTM(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "lstm") {}
516

Y
Yan Chunwei 已提交
517
  PDNode* operator()(PDNode* x);
518

Y
Yan Chunwei 已提交
519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537
  // Operators
  PATTERN_DECL_NODE(lstm);

  // Inputs
  PATTERN_DECL_NODE(Input);
  PATTERN_DECL_NODE(H0);
  PATTERN_DECL_NODE(C0);
  PATTERN_DECL_NODE(Weight);
  PATTERN_DECL_NODE(Bias);

  // Outputs
  PATTERN_DECL_NODE(Hidden);
  PATTERN_DECL_NODE(Cell);
  PATTERN_DECL_NODE(BatchGate);
  PATTERN_DECL_NODE(BatchCellPreAct);
};

struct GRU : public PatternBase {
  GRU(PDPattern* pattern, const std::string& name_scope)
S
superjomn 已提交
538
      : PatternBase(pattern, name_scope, "gru") {}
Y
Yan Chunwei 已提交
539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554

  PDNode* operator()(PDNode* x);

  // Operators
  PATTERN_DECL_NODE(gru);

  // Inputs
  PATTERN_DECL_NODE(Bias);
  PATTERN_DECL_NODE(Weight);

  // Outputs
  PATTERN_DECL_NODE(BatchGate);
  PATTERN_DECL_NODE(BatchResetHiddenPrev);
  PATTERN_DECL_NODE(BatchHidden);
  PATTERN_DECL_NODE(Hidden);
};
T
tensor-tang 已提交
555

C
chengduo 已提交
556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
// The following patterns are used to fuse elewise_add and act
// formula: act(ele_add(x, y))
// op: elementwise_add + act
// named nodes: elementwise_add, act
//              ele_x, ele_y, elewise_add_out, act_out
struct ElewiseAddAct : public PatternBase {
  ElewiseAddAct(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "elewise_add_act") {}

  PDNode* operator()(PDNode* x, std::unordered_set<std::string> acts);

  // declare operator node's name
  PATTERN_DECL_NODE(ele_add);
  PATTERN_DECL_NODE(act);
  // declare variable node's name
  PATTERN_DECL_NODE(elewise_add_out);
  PATTERN_DECL_NODE(ele_y);
  PATTERN_DECL_NODE(act_out);
};

// formula: ele_add(x, act(y))
// op: elementwise_add + act
// named nodes: elementwise_add, act
//              act_in, act_out, ele_x, elewise_add_out
struct ActElewiseAdd : public PatternBase {
  ActElewiseAdd(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "act_elewise_add") {}

  PDNode* operator()(PDNode* x, std::unordered_set<std::string> acts);

  // declare operator node's name
  PATTERN_DECL_NODE(act);
  PATTERN_DECL_NODE(ele_add);
  // declare variable node's name
  PATTERN_DECL_NODE(act_out);
  PATTERN_DECL_NODE(ele_x);
  PATTERN_DECL_NODE(elewise_add_out);
};

// the backward of act(ele_add(x, y))
// the act is inplace.
// op: elementwise_add_grad + act_grad
// named nodes: elementwise_add_grad, act_grad
//              act_out, act_out_g, ele_y, d_itermediate_out, d_ele_x, d_ele_y
struct ElewiseAddActInplaceGrad : public PatternBase {
  ElewiseAddActInplaceGrad(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "elewise_add_act_grad1") {}

  // act_grad: in["Out", "Out@GRAD"], out["X@GRAD"]
  // ele_add_grad: in["Y", "Out@GRAD"], out["X@GRAD", "Y@GRAD"]
  PDNode* operator()(PDNode* x, std::unordered_set<std::string> acts);

  // declare operator node's name
  PATTERN_DECL_NODE(act_grad);
  PATTERN_DECL_NODE(ele_add_grad);
  // declare variable node's name
  PATTERN_DECL_NODE(act_out);
  PATTERN_DECL_NODE(d_itermediate_out);
  PATTERN_DECL_NODE(d_ele_x);
  PATTERN_DECL_NODE(d_ele_y);
  PATTERN_DECL_NODE(ele_y);
};
M
Michal Gallus 已提交
618 619 620 621 622 623 624 625 626 627 628

// Conv with Elementwise_add as bias
// op: conv + elementwise_add
// named nodes:
// conv_input, conv_weight,
// conv_out, conv,
// eltwise_bias, eltwise_out,
// elementwise_add
struct ConvBias : public PatternBase {
  ConvBias(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_bias") {}
629
  PDNode* operator()(PDNode* conv_input, bool is_conv3d = false);
M
Michal Gallus 已提交
630 631 632 633 634 635 636 637 638
  // declare operator node's name
  PATTERN_DECL_NODE(conv);
  PATTERN_DECL_NODE(eltwise);
  // declare variable node's name
  PATTERN_DECL_NODE(conv_weight);
  PATTERN_DECL_NODE(conv_out);
  PATTERN_DECL_NODE(eltwise_bias);
  PATTERN_DECL_NODE(eltwise_out);
};
639

640 641 642 643 644 645 646 647 648
// Convolution op
// Forward pass for convolution.
// conv_input, conv_bias and conv_filter are inputs.
// conv_output is a result of the operator.
// residual_data is data used by skip connection.
// If residual connection fusion is on, the formula is:
// conv_output = conv_op(conv_filter, conv_input, conv_bias)
//             + conv_residual_data
// If the fusion is off, conv_residual_data is not added.
649 650 651 652 653 654 655 656 657 658 659 660 661
struct Conv : public PatternBase {
  Conv(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "convolution") {}

  PDNode* operator()();

  PATTERN_DECL_NODE(conv_op);
  PATTERN_DECL_NODE(conv_input);
  PATTERN_DECL_NODE(conv_filter);
  PATTERN_DECL_NODE(conv_residual_data);
  PATTERN_DECL_NODE(conv_output);
};

662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690
// Convolution op with residual data
struct ConvResidual : public PatternBase {
  ConvResidual(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_residual") {}

  PDNode* operator()(bool with_residual_data);

  PATTERN_DECL_NODE(conv_op);
  PATTERN_DECL_NODE(conv_input);
  PATTERN_DECL_NODE(conv_filter);
  PATTERN_DECL_NODE(conv_residual_data);
  PATTERN_DECL_NODE(conv_output);
};

// Pool op
// Forward pass for pooling.
// pool_input is the input.
// pool_output is a result of the operator.
struct Pool : public PatternBase {
  Pool(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "pooling") {}

  PDNode* operator()();

  PATTERN_DECL_NODE(pool_op);
  PATTERN_DECL_NODE(pool_input);
  PATTERN_DECL_NODE(pool_output);
};

691 692 693 694
// ElementwiseAdd used in residual connections.
// y_var is used and convolution output.
// The operator is removed, when residual
// connection fusion is on.
695 696 697 698
struct ElementwiseAdd : public PatternBase {
  ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "elementwise_add") {}

699
  PDNode* operator()(PDNode* x_var, PDNode* y_var);
700 701 702 703 704 705

  PATTERN_DECL_NODE(elementwise_add_op);
  PATTERN_DECL_NODE(elementwise_add_x);
  PATTERN_DECL_NODE(elementwise_add_y);
  PATTERN_DECL_NODE(elementwise_add_out);
};
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

// Conv + ElementwiseAdd + an activation
// This pattern can futher fuse the conv related ops after the conv+bn fusion.
struct ConvElementwiseaddAct : public PatternBase {
  ConvElementwiseaddAct(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_elementwiseadd_act") {}

  PDNode* operator()(PDNode* conv_in);

  PATTERN_DECL_NODE(conv_op);
  PATTERN_DECL_NODE(conv_out);
  PATTERN_DECL_NODE(conv_filter);

  PATTERN_DECL_NODE(elementwise_add_op);
  PATTERN_DECL_NODE(elementwise_add_in_y);  // input
  PATTERN_DECL_NODE(elementwise_add_out);

  PATTERN_DECL_NODE(act_op);
  PATTERN_DECL_NODE(act_out);
};

// Conv + ElementwiseAdd + ElementwiseAdd + Activation
struct ConvElementwiseadd2Act : public PatternBase {
  ConvElementwiseadd2Act(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope,
                    "conv_elementwiseadd2_elementwiseadd_act") {}

  PDNode* operator()(PDNode* conv_in);

  PATTERN_DECL_NODE(conv_op);
  PATTERN_DECL_NODE(conv_filter);
  PATTERN_DECL_NODE(conv_out);

  PATTERN_DECL_NODE(elementwise_add_op);
  PATTERN_DECL_NODE(elementwise_add_in_y);  // input
  PATTERN_DECL_NODE(elementwise_add_out);

  PATTERN_DECL_NODE(elementwise_add_op_1);
  PATTERN_DECL_NODE(elementwise_add_in_y_1);  // input
  PATTERN_DECL_NODE(elementwise_add_out_1);

  PATTERN_DECL_NODE(act_op);
  PATTERN_DECL_NODE(act_out);
};

N
nhzlx 已提交
751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768
// Conv + ElementwiseAdd
// This pattern should be used after ConvElementwiseadd2Act or
// ConvElementwiseadd pass
struct ConvElementwiseadd : public PatternBase {
  ConvElementwiseadd(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_elementwiseadd") {}

  PDNode* operator()(PDNode* conv_in);

  PATTERN_DECL_NODE(conv_op);
  PATTERN_DECL_NODE(conv_out);
  PATTERN_DECL_NODE(conv_filter);

  PATTERN_DECL_NODE(elementwise_add_op);
  PATTERN_DECL_NODE(elementwise_add_in_y);
  PATTERN_DECL_NODE(elementwise_add_out);
};

N
nhzlx 已提交
769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800
// Conv with affine_channel
// op: conv + (elementwise_add +) affine_channel
// named nodes:
// conv_weight, conv_out, conv,
// ac_x, ac_scale, ac_bias
// affine_channel, ac_out
struct ConvAffineChannel : public PatternBase {
  ConvAffineChannel(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_affine_channel") {}

  PDNode* operator()(PDNode* conv_input, bool with_eltwise_add);

  // declare operator node's name
  PATTERN_DECL_NODE(conv);
  PATTERN_DECL_NODE(affine_channel);
  PATTERN_DECL_NODE(eltwise);  // ELEMENTWISE_ADD
  // CONV inputs
  PATTERN_DECL_NODE(conv_weight);  // Filter
  // CONV outputs
  PATTERN_DECL_NODE(conv_out);  // tmp
  // ELTWISE inputs
  PATTERN_DECL_NODE(eltwise_y_in);
  // ELTWISE outputs
  PATTERN_DECL_NODE(eltwise_out);  // tmp

  // AC(Affine_Channel) inputs
  PATTERN_DECL_NODE(ac_scale);
  PATTERN_DECL_NODE(ac_bias);
  // AC outputs
  PATTERN_DECL_NODE(ac_out);  // Out
};

801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828
// Dequantize + Quantize + anyOP
// This pattern is used for squashing the dequantize-quantize pairs.
struct DequantQuantAny : public PatternBase {
  DequantQuantAny(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "dequant_quant_any") {}
  PDNode* operator()();

  PATTERN_DECL_NODE(dequant_in);
  PATTERN_DECL_NODE(dequant_op);
  PATTERN_DECL_NODE(dequant_out);
  PATTERN_DECL_NODE(quant_op);
  PATTERN_DECL_NODE(quant_out);
  PATTERN_DECL_NODE(next_op);
};

// Dequantize + anyOP
// This quantize is used for getting number of ops the Dequantize's
// output is an input to.
struct DequantAny : public PatternBase {
  DequantAny(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "dequant_any") {}
  PDNode* operator()();

  PATTERN_DECL_NODE(dequant_op);
  PATTERN_DECL_NODE(dequant_out);
  PATTERN_DECL_NODE(next_op);
};

829 830 831 832 833 834 835 836 837 838 839 840 841 842 843
struct TransposeFlattenConcat : public PatternBase {
  TransposeFlattenConcat(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "transpose_flatten_concat") {}

  PDNode* operator()(std::vector<PDNode*> conv_inputs, int times);

  std::string GetNodeName(const std::string& op_type) {
    return PDNodeName(name_scope_, repr_, id_, op_type);
  }

  PDNode* GetPDNode(const std::string& op_type) {
    return pattern->RetrieveNode(GetNodeName(op_type));
  }
};

844
}  // namespace patterns
845

Y
Yan Chunwei 已提交
846
// Link two ir::Nodes from each other.
847 848 849 850
#define IR_NODE_LINK_TO(a, b) \
  a->outputs.push_back(b);    \
  b->inputs.push_back(a);

C
chengduo 已提交
851 852 853 854 855 856
// Set the out_var as the output of the op
#define IR_OP_VAR_LINK(op, out_var) \
  op->outputs.push_back(out_var);   \
  out_var->inputs.clear();          \
  out_var->inputs.push_back(op);

857 858 859
}  // namespace ir
}  // namespace framework
}  // namespace paddle