“def272cf42e9b2ebf529b39f183874a1dede9c2a”上不存在“git@gitcode.net:Crayonxin2000/Paddle.git”
graph_pattern_detector.h 33.4 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);

134 135 136
  PDNode* assert_has_n_inputs(size_t n);
  PDNode* assert_has_n_outputs(size_t n);

T
tensor-tang 已提交
137 138 139 140 141 142 143 144 145
  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;
  }

146
 private:
Y
Yan Chunwei 已提交
147 148 149
  PDNode(PDPattern* pattern, const std::string& name = "",
         Type type = Type::kVar)
      : pattern_(pattern), name_(name), type_(type) {}
150 151 152 153 154 155 156 157 158 159 160 161 162
  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
163
  // Will removed latter.
164
  teller_t teller_;
Y
Yan Chunwei 已提交
165
  std::vector<teller_t> asserts_;
166
  PDPattern* pattern_;
167
  std::string name_;
168
  Type type_;
Y
Yan Chunwei 已提交
169
  Role role_{Role::kUnknown};
170 171 172 173 174 175 176 177 178 179 180 181
};

/*
 * 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
182 183
 *     MUL = PDPattern.NewNode().assert_is_op("mul");
 *     ELE = PDPattern.NewNode().assert_is_op("elementwise_add");
184
 *     // Create the variable PDNodes.
Y
Yan Chunwei 已提交
185 186 187 188 189 190
 *     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});
191
 *
Y
Yan Chunwei 已提交
192 193
 * One can add more specific asserts for PDNodes or edges, both the Operator
 * and Variable Nodes can be ruled in PDNode.assert_more(...).
194 195 196 197 198 199 200 201 202 203 204
 *
 * 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);

205
  PDNode* NewNode(PDNode::teller_t&& teller, const std::string& name = NewID());
Y
Yan Chunwei 已提交
206
  PDNode* NewNode(const std::string& name = NewID());
207 208 209
  PDNode* NewNode(const std::string& prefix, const std::string& name) {
    return NewNode(prefix + "/" + name);
  }
Y
Yan Chunwei 已提交
210
  PDNode* RetrieveNode(const std::string& id) const;
211 212 213 214

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

215 216
  std::string DotString() const;

217 218 219 220 221 222
 private:
#ifdef PADDLE_WITH_TESTING
  FRIEND_TEST(PDPattern, AddEdge);
  FRIEND_TEST(PDPattern, NewNode);
#endif

223 224
  static std::string NewID() { return "pdnode-" + std::to_string(id_++); }

225 226
  std::vector<std::unique_ptr<PDNode>> nodes_;
  std::vector<edge_t> edges_;
227 228
  std::unordered_map<std::string, PDNode*> node_map_;
  static size_t id_;
229 230 231
};

/*
232
 * GraphPatternDetector helps to detect the specific patterns in the graph.
233 234 235 236 237 238 239 240 241 242 243
 * 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
244
 *    GraphPatternDetector detector;
245 246 247 248 249 250 251 252
 *    // 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.
253
 *    GraphPatternDetector::handle_t handler = some labmda
254 255 256
 *    // Execute the detector.
 *    detector(&graph, handler);
 */
257
class GraphPatternDetector {
258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
 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
281 282
  // The intermediate PDNodes will be removed, so can't shared by multiple
  // patterns.
283 284
  void RemoveOverlappedMatch(std::vector<subgraph_t>* subgraphs);

Y
Yan Chunwei 已提交
285 286 287
  // Validate whether the intermediate nodes are linked by external nodes.
  void ValidateByNodeRole(std::vector<subgraph_t>* subgraphs);

288 289 290 291 292 293 294 295 296 297 298 299
#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_;
};

300 301
// some helper methods.

302 303 304 305 306
// 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);
307 308

// Check whether a var node is a op node's nth input.
309
bool IsNthInput(Node* var, Node* op, const std::string& argument, size_t nth);
310

311 312 313
// Check whether the op node has input of given name.
bool HasInput(Node* op, const std::string& argument);

314 315 316 317 318 319 320 321
// 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.
322 323
// The related Fluid Layer or Op should be one pattern here for better re-usage
// across different fusion.
324 325
namespace patterns {

Y
Yan Chunwei 已提交
326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395
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
396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433
// 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);
};

434 435 436 437
// CONV with ReLU
// op: conv + relu
// named nodes:
// conv_input, conv_weight,
438
// conv_out, conv,
439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454
// 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);
};

455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475
// CONV with ReLU6
// op: conv + relu6
// named nodes:
// conv_input, conv_weight,
// conv_out, conv,
// relu6_out, relu6
struct ConvBReLU : public PatternBase {
  ConvBReLU(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_bounded_relu") {}

  PDNode* operator()(PDNode* conv_input);

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

T
tensor-tang 已提交
476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500
// 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);
};

501 502 503 504 505
// FC with bias
// op: mul + elementwise_add
// named nodes:
// mul, elementwise_add
// w, mul_out, bias, fc_out
Y
Yan Chunwei 已提交
506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522
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);
};

523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
// MKL-DNN's FC with bias
// op: fc
// named node:
// fc
// w, bias, output
struct FCMKLDNN : public PatternBase {
  FCMKLDNN(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "fc_mkldnn") {}

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

  // declare operator node's name
  PATTERN_DECL_NODE(fc);
  // declare variable node's name
  PATTERN_DECL_NODE(weights);
  PATTERN_DECL_NODE(bias);
  PATTERN_DECL_NODE(output);
};

542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558
// 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
559 560 561
struct LSTM : public PatternBase {
  LSTM(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "lstm") {}
562

Y
Yan Chunwei 已提交
563
  PDNode* operator()(PDNode* x);
564

Y
Yan Chunwei 已提交
565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583
  // 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
584
      : PatternBase(pattern, name_scope, "gru") {}
Y
Yan Chunwei 已提交
585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600

  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
601

C
chengduo 已提交
602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663
// 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
664 665 666 667 668 669 670 671 672 673 674

// 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") {}
675
  PDNode* operator()(PDNode* conv_input, std::string conv_type = "conv2d");
M
Michal Gallus 已提交
676 677 678 679 680 681 682 683 684
  // 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);
};
685

686 687 688 689 690 691 692 693 694
// 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.
695 696 697 698 699 700 701 702 703 704 705 706 707
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);
};

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
// 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);
};

737 738 739 740
// ElementwiseAdd used in residual connections.
// y_var is used and convolution output.
// The operator is removed, when residual
// connection fusion is on.
741 742 743 744
struct ElementwiseAdd : public PatternBase {
  ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "elementwise_add") {}

745
  PDNode* operator()(PDNode* x_var, PDNode* y_var);
746 747 748 749 750 751

  PATTERN_DECL_NODE(elementwise_add_op);
  PATTERN_DECL_NODE(elementwise_add_x);
  PATTERN_DECL_NODE(elementwise_add_y);
  PATTERN_DECL_NODE(elementwise_add_out);
};
752

753 754 755 756 757 758 759 760 761 762 763 764 765
// Concat op
// Forward pass for concat.
// concat_out is a result of the operator.
struct Concat : public PatternBase {
  Concat(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "concat") {}

  PDNode* operator()();

  PATTERN_DECL_NODE(concat_op);
  PATTERN_DECL_NODE(concat_out);
};

766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798
// Concat + ReLU
// named nodes:
// concat_op, concat_out, relu_op, relu_out
struct ConcatReLU : public PatternBase {
  ConcatReLU(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "concat_relu") {}

  PDNode* operator()();

  PATTERN_DECL_NODE(concat_op);
  PATTERN_DECL_NODE(concat_out);
  PATTERN_DECL_NODE(relu_op);
  PATTERN_DECL_NODE(relu_out);
};

// Conv + Concat + ReLU
// named nodes:
// conv_op, conv_out
// concat_op, concat_out, relu_op, relu_out
struct ConvConcatReLU : public PatternBase {
  ConvConcatReLU(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "conv_concat_relu") {}

  PDNode* operator()();

  PATTERN_DECL_NODE(conv_op);
  PATTERN_DECL_NODE(conv_out);
  PATTERN_DECL_NODE(concat_op);
  PATTERN_DECL_NODE(concat_out);
  PATTERN_DECL_NODE(relu_op);
  PATTERN_DECL_NODE(relu_out);
};

799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815
// PriorBox operator
// operator: prior_box_op
// inputs: prior_box_input, prior_box_image
// outputs: prior_box_boxes, prior_box_variances
struct PriorBox : public PatternBase {
  PriorBox(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "PriorBox") {}

  PDNode* operator()();

  PATTERN_DECL_NODE(prior_box_op);
  PATTERN_DECL_NODE(prior_box_input);
  PATTERN_DECL_NODE(prior_box_image);
  PATTERN_DECL_NODE(prior_box_boxes);
  PATTERN_DECL_NODE(prior_box_variances);
};

816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859
// 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
860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877
// 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
878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909
// 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
};

910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937
// 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);
};

938 939 940 941 942 943 944 945 946 947 948 949 950 951 952
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));
  }
};

953 954 955 956
struct AnakinDetectionPattern : public PatternBase {
  AnakinDetectionPattern(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "anakin_detect_pattern") {}

N
nhzlx 已提交
957 958
  PDNode* operator()(std::vector<PDNode*> conv_inputs, int times,
                     std::string priorbox_type, bool is_reshape);
959 960 961 962 963 964 965 966 967 968

  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));
  }
};

N
nhzlx 已提交
969 970 971
struct FillConstantElementWiseMulFuse : public PatternBase {
  FillConstantElementWiseMulFuse(PDPattern* pattern,
                                 const std::string& name_scope)
972 973 974 975 976 977 978 979 980 981 982 983
      : PatternBase(pattern, name_scope,
                    "anakin_fillconstant_elementwisemul_fuse") {}

  PDNode* operator()(PDNode* elementwise_op_input);

  // declare operator node's name
  PATTERN_DECL_NODE(fill_constant);
  PATTERN_DECL_NODE(fill_constant_out);
  PATTERN_DECL_NODE(elementwise_mul);
  PATTERN_DECL_NODE(elementwise_mul_out);
};

N
nhzlx 已提交
984 985 986 987 988
struct QuantDequantOpFuse : public PatternBase {
  QuantDequantOpFuse(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "quant_dequant_fuse") {}

  void operator()(PDNode* quant_op_input, const std::string& op_name,
989
                  const std::string& weight_name, int times,
990 991
                  const std::string& quant_type,
                  const std::string& dequant_type);
N
nhzlx 已提交
992 993 994 995 996 997 998 999 1000 1001

  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));
  }
};

1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016
struct ShuffleChannelPattern : public PatternBase {
  ShuffleChannelPattern(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "shufflechannel_pattern") {}

  void operator()(PDNode* reshape1_in);

  PATTERN_DECL_NODE(reshape1_op);
  PATTERN_DECL_NODE(reshape1_out);

  PATTERN_DECL_NODE(transpose_op);
  PATTERN_DECL_NODE(transpose_out);
  PATTERN_DECL_NODE(reshape2_op);
  PATTERN_DECL_NODE(reshape2_out);
};

1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030
struct DeleteQuantDequantOpPattern : public PatternBase {
  DeleteQuantDequantOpPattern(PDPattern* pattern, const std::string& name_scope)
      : PatternBase(pattern, name_scope, "delete_quantdequant_op_pattern") {}

  void operator()();

  PATTERN_DECL_NODE(any_op_out);
  PATTERN_DECL_NODE(quant_dequant_op_inscale);
  PATTERN_DECL_NODE(quant_dequant_op);
  PATTERN_DECL_NODE(quant_dequant_op_outscale);
  PATTERN_DECL_NODE(quant_dequant_op_out);
  PATTERN_DECL_NODE(any_op2);
};

1031
}  // namespace patterns
1032

Y
Yan Chunwei 已提交
1033
// Link two ir::Nodes from each other.
1034 1035 1036 1037
#define IR_NODE_LINK_TO(a, b) \
  a->outputs.push_back(b);    \
  b->inputs.push_back(a);

C
chengduo 已提交
1038 1039 1040 1041 1042 1043
// 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);

1044 1045 1046
}  // namespace ir
}  // namespace framework
}  // namespace paddle