/
graph_pattern_detector.h
1905 lines (1613 loc) · 60.3 KB
/
graph_pattern_detector.h
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
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
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
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
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
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
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
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
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
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
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
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
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
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
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
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
// 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
#include <map>
#include <memory>
#include <numeric>
#include <set>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/node.h"
#include "paddle/fluid/inference/analysis/dot.h"
namespace paddle {
namespace framework {
namespace ir {
class Graph;
class Node;
} // namespace ir
} // namespace framework
} // namespace paddle
namespace paddle {
namespace framework {
namespace ir {
class PDPattern;
// Some basic terminologies:
// - 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*)>;
enum class Type { kOp, kVar };
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.
};
// this link to others
PDNode& LinksTo(const std::vector<PDNode*>& others);
PDNode& LinksFrom(const std::vector<PDNode*>& others);
bool Tell(Node* node) const {
if (teller_) return teller_(node);
for (auto& asrt : asserts_) {
if (!asrt(node)) return false;
}
return true;
}
bool IsOp() const { return type_ == Type::kOp; }
bool IsVar() const { return type_ == Type::kVar; }
const std::string& name() const { return name_; }
const PDPattern* pdpattern() const { return pattern_; }
PDNode& operator=(const PDNode&) = delete;
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_not_op_type(const std::string& op_type);
PDNode* assert_is_var();
PDNode* assert_var_dtype(proto::VarType::Type dtype);
PDNode* assert_is_not_ctrl_var();
PDNode* assert_var_not_persistable();
PDNode* assert_is_persistable_var();
PDNode* assert_is_op_output(const std::string& op_type);
PDNode* assert_is_op_output(const std::string& op_type,
const std::string& argument);
PDNode* assert_is_op_input(const std::string& op_type);
PDNode* assert_is_op_input(const std::string& op_type,
const std::string& argument);
PDNode* assert_is_op_nth_input(const std::string& op_type,
const std::string& argument,
int nth);
PDNode* assert_is_not_op_input(const std::string& argument);
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);
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);
PDNode* assert_is_only_input_of_ops(
const std::unordered_set<std::string>& op_types);
PDNode* assert_is_only_output_of_ops(
const std::unordered_set<std::string>& op_types);
PDNode* assert_has_n_inputs(size_t n);
PDNode* assert_has_n_outputs(size_t n);
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) &&
PADDLE_GET_CONST(T, x->Op()->GetAttr(attr_name)) == attr;
});
return this;
}
private:
PDNode(PDPattern* pattern,
const std::string& name = "",
Type type = Type::kVar)
: pattern_(pattern), name_(name), type_(type) {}
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_NOT_NULL(
teller_,
platform::errors::NotFound("invalid teller is set, teller is null"));
}
PDNode(PDNode&& other) = default;
friend class PDPattern;
// Will removed latter.
teller_t teller_;
std::vector<teller_t> asserts_;
PDPattern* pattern_;
std::string name_;
Type type_;
Role role_{Role::kUnknown};
};
/*
* 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.
* MUL = PDPattern.NewNode().assert_is_op("mul");
* ELE = PDPattern.NewNode().assert_is_op("elementwise_add");
* // Create the variable PDNodes.
* 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});
*
* One can add more specific asserts for PDNodes or edges, both the Operator
* and Variable Nodes can be ruled in PDNode.assert_more(...).
*
* 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);
PDNode* NewNode(PDNode::teller_t&& teller, const std::string& name = NewID());
PDNode* NewNode(const std::string& name = NewID());
PDNode* NewNode(const std::string& prefix, const std::string& name) {
return NewNode(prefix + "/" + name);
}
PDNode* RetrieveNode(const std::string& id) const;
const std::vector<std::unique_ptr<PDNode>>& nodes() const { return nodes_; }
const std::vector<edge_t>& edges() const { return edges_; }
std::string DotString() const;
private:
#ifdef PADDLE_WITH_TESTING
FRIEND_TEST(PDPattern, AddEdge);
FRIEND_TEST(PDPattern, NewNode);
#endif
static std::string NewID() { return "pdnode-" + std::to_string(id_++); }
std::vector<std::unique_ptr<PDNode>> nodes_;
std::vector<edge_t> edges_;
std::map<std::string, PDNode*> node_map_;
static size_t id_;
};
/*
* GraphPatternDetector helps to detect the specific patterns in the graph.
* 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
* GraphPatternDetector detector;
* // 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.
* GraphPatternDetector::handle_t handler = some labmda
* // Execute the detector.
* detector(&graph, handler);
*/
class GraphPatternDetector {
public:
struct NodeIdCompare {
bool operator()(Node* node1, Node* node2) const {
return node1->id() < node2->id();
}
};
struct PDNodeCompare {
bool operator()(const PDNode* node1, const PDNode* node2) const {
auto& nodes1 = node1->pdpattern()->nodes();
auto& nodes2 = node2->pdpattern()->nodes();
if (nodes1.size() != nodes2.size()) {
return nodes1.size() < nodes2.size();
} else {
std::string pdnode_hash_key1 = "";
std::string pdnode_hash_key2 = "";
for (auto& node : nodes1) {
pdnode_hash_key1 += node.get()->name();
pdnode_hash_key1 += "#";
}
pdnode_hash_key1 += node1->name();
for (auto& node : nodes2) {
pdnode_hash_key2 += node.get()->name();
pdnode_hash_key2 += "#";
}
pdnode_hash_key2 += node2->name();
auto pdnode_key1 =
std::to_string(std::hash<std::string>()(pdnode_hash_key1));
auto pdnode_key2 =
std::to_string(std::hash<std::string>()(pdnode_hash_key2));
return pdnode_key1 < pdnode_key2;
}
return false;
}
};
using subgraph_t = std::map<PDNode*, Node*, PDNodeCompare>;
// 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);
// Sort subgraphs, sort subgraphs by the specified node so that
// the removed forward and backward subgraphs are corresponding
// when two subgraphs are overlapped. Note: this function is
// currently only used for bn_add_act, refer to PR28196 for details.
void SortSubgraphs(std::vector<subgraph_t>* subgraphs);
// Remove overlapped match subgraphs, when overlapped, keep the previous one.
// The intermediate PDNodes will be removed, so can't shared by multiple
// patterns.
void RemoveOverlappedMatch(std::vector<subgraph_t>* subgraphs);
// Validate whether the intermediate nodes are linked by external nodes.
void ValidateByNodeRole(std::vector<subgraph_t>* subgraphs);
#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::map<const PDNode*, std::set<Node*, NodeIdCompare>, PDNodeCompare>
pdnodes2nodes_;
};
// some helper methods.
// 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);
// Check whether a var node is a op node's nth input.
bool IsNthInput(Node* var, Node* op, const std::string& argument, size_t nth);
// Check whether the op node has input of given name.
bool HasInput(Node* op, const std::string& argument);
// Check whether the op node has output of given name.
bool HasOutput(Node* op, const std::string& argument);
// 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,
std::unordered_set<std::shared_ptr<Node>>* saved_nodes = nullptr);
// Some pre-defined patterns those can be reused in multiple passes.
// The related Fluid Layer or Op should be one pattern here for better re-usage
// across different fusion.
namespace patterns {
struct KeyCounter {
static KeyCounter& Instance() {
static KeyCounter x;
return x;
}
#ifdef PADDLE_WITH_TENSORRT
static int IncCounter(const std::string& key) { return dic_[key]++; }
static void CleanCounter() { dic_.clear(); }
private:
static thread_local std::unordered_map<std::string, size_t> dic_;
#else
int IncCounter(const std::string& key) { return dic_[key]++; }
private:
std::unordered_map<std::string, size_t> dic_;
#endif
};
// 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_NE(subgraph.count(pat.arg##_n()), \
0UL, \
platform::errors::NotFound("Node not found for PDNode %s", \
pat.arg##_repr())); \
Node* var = subgraph.at(pat.arg##_n()); \
PADDLE_ENFORCE_NOT_NULL(var, \
platform::errors::NotFound( \
"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_;
};
// 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,
const std::string& conv_type,
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);
};
struct OperatorActivation : public PatternBase {
OperatorActivation(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "operator_activation") {}
PDNode* operator()(const std::string& operator_type,
const std::string& activation_type);
PATTERN_DECL_NODE(preceding_op);
PATTERN_DECL_NODE(preceding_op_out);
PATTERN_DECL_NODE(activation);
PATTERN_DECL_NODE(activation_out);
};
// 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);
};
// FC with bias
// op: mul + elementwise_add
// named nodes:
// mul, elementwise_add
// w, mul_out, bias, fc_out
struct FC : public PatternBase {
FC(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "fc") {}
PDNode* operator()(PDNode* x, bool with_bias, bool with_relu);
// declare operator node's name
PATTERN_DECL_NODE(fc);
PATTERN_DECL_NODE(mul);
PATTERN_DECL_NODE(elementwise_add);
PATTERN_DECL_NODE(relu);
// 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(elementwise_add_out);
PATTERN_DECL_NODE(relu_out);
};
// 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(input);
PATTERN_DECL_NODE(weights);
PATTERN_DECL_NODE(bias);
PATTERN_DECL_NODE(output);
};
// 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);
};
struct LSTM : public PatternBase {
LSTM(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "lstm") {}
PDNode* operator()(PDNode* x);
// 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)
: PatternBase(pattern, name_scope, "gru") {}
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);
};
// The following pattern is used to fuse batch_norm and act
// formula: act(bn(x))
// op: batch_norm + act
struct BatchNormAct : public PatternBase {
BatchNormAct(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "bn_act") {}
PDNode* operator()(PDNode* x, std::unordered_set<std::string> acts);
// declare operator node's name
PATTERN_DECL_NODE(batch_norm);
PATTERN_DECL_NODE(act);
// declare variable node's name
// BN inputs
PATTERN_DECL_NODE(bn_scale);
PATTERN_DECL_NODE(bn_bias);
PATTERN_DECL_NODE(bn_variance);
PATTERN_DECL_NODE(bn_mean);
// BN outputs
PATTERN_DECL_NODE(bn_mean_out);
PATTERN_DECL_NODE(bn_variance_out);
PATTERN_DECL_NODE(bn_saved_variance);
PATTERN_DECL_NODE(bn_saved_mean);
PATTERN_DECL_NODE(bn_reserve_space);
PATTERN_DECL_NODE(bn_out);
// ACT output
PATTERN_DECL_NODE(act_out);
};
// the backward of act(bn(x))
// op: batch_norm_grad + act_grad
struct BatchNormActGrad : public PatternBase {
BatchNormActGrad(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "bn_act_grad") {}
// act_grad: in["Out", "Out@GRAD"], out["X@GRAD"]
// bn_grad: in["X", "Y@GRAD", "Scale", "Bias", "SavedMean", "SavedVariance",
// "ReserveSpace"],
// out["X@GRAD", "Scale@GRAD", "Bias@GRAD"]
PDNode* operator()(PDNode* x, std::unordered_set<std::string> act_grad_types);
// declare operator node's name
PATTERN_DECL_NODE(act_grad);
PATTERN_DECL_NODE(batch_norm_grad);
// declare variable node's name
PATTERN_DECL_NODE(act_out);
PATTERN_DECL_NODE(d_itermediate_out);
PATTERN_DECL_NODE(bn_x);
PATTERN_DECL_NODE(bn_scale);
PATTERN_DECL_NODE(bn_bias);
PATTERN_DECL_NODE(bn_saved_mean);
PATTERN_DECL_NODE(bn_saved_variance);
PATTERN_DECL_NODE(bn_reserve_space);
PATTERN_DECL_NODE(d_bn_x);
PATTERN_DECL_NODE(d_bn_scale);
PATTERN_DECL_NODE(d_bn_bias);
};
//
// \brief Pattern looking for batch_norm and a directly following activation
// operator.
//
// \note Currently only ReLU is supported as an activation function.
// Formula: act(bn(x))
// Op: batch_norm + act
struct BatchNormActOneDNN : public PatternBase {
BatchNormActOneDNN(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "bn_act_onednn") {}
PDNode* operator()(const std::string& act_type);
// declare operator node's name
PATTERN_DECL_NODE(bn_in);
PATTERN_DECL_NODE(batch_norm);
PATTERN_DECL_NODE(act);
PATTERN_DECL_NODE(bn_out);
PATTERN_DECL_NODE(act_out);
};
// The following pattern is used to fuse batch_norm, elewise_add, and act
// formula: act(bn(x) + z)
// op: batch_norm + elewise_add + act
struct BatchNormAddAct : public PatternBase {
BatchNormAddAct(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "bn_add_act") {}
PDNode* operator()(PDNode* x, std::unordered_set<std::string> acts);
// declare operator node's name
PATTERN_DECL_NODE(batch_norm);
PATTERN_DECL_NODE(elewise_add);
PATTERN_DECL_NODE(act);
// declare variable node's name
// BN inputs
PATTERN_DECL_NODE(bn_scale);
PATTERN_DECL_NODE(bn_bias);
// BN outputs
PATTERN_DECL_NODE(bn_mean_out);
PATTERN_DECL_NODE(bn_variance_out);
PATTERN_DECL_NODE(bn_saved_variance);
PATTERN_DECL_NODE(bn_saved_mean);
PATTERN_DECL_NODE(bn_reserve_space);
PATTERN_DECL_NODE(bn_out);
// Elewise_Add input
PATTERN_DECL_NODE(elewise_add_in);
// Elewise_Add output
PATTERN_DECL_NODE(elewise_add_out);
// ACT output
PATTERN_DECL_NODE(act_out);
};
// the backward of act(bn(x) + z)
// op: batch_norm_grad + elewise_add_grad + act_grad
struct BatchNormAddActGrad : public PatternBase {
BatchNormAddActGrad(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "bn_add_act_grad") {}
// act_grad: in["Out", "Out@GRAD"], out["X@GRAD"]
// elewise_add_grad: in["Out@GRAD"], out["X@GRAD", "Y@GRAD"]
// bn_grad: in["X", "Z", "Y@GRAD", "Scale", "Bias", "SavedMean",
// "SavedVariance",
// "ReserveSpace"],
// out["X@GRAD", "Z@GRAD", "Scale@GRAD", "Bias@GRAD"]
PDNode* operator()(PDNode* x, std::unordered_set<std::string> act_grad_types);
// declare operator node's name
PATTERN_DECL_NODE(act_grad);
PATTERN_DECL_NODE(elewise_add_grad);
PATTERN_DECL_NODE(batch_norm_grad);
// declare variable node's name
PATTERN_DECL_NODE(act_out);
PATTERN_DECL_NODE(d_act_x);
PATTERN_DECL_NODE(d_elewise_add_in);
PATTERN_DECL_NODE(d_bn_out);
PATTERN_DECL_NODE(bn_x);
PATTERN_DECL_NODE(bn_scale);
PATTERN_DECL_NODE(bn_bias);
PATTERN_DECL_NODE(bn_saved_mean);
PATTERN_DECL_NODE(bn_saved_variance);
PATTERN_DECL_NODE(bn_reserve_space);
PATTERN_DECL_NODE(d_bn_x);
PATTERN_DECL_NODE(d_bn_scale);
PATTERN_DECL_NODE(d_bn_bias);
};
// 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);
};
// The following patterns are used to fuse linear and act (ReLu or GeLU)
// formula: act(F.linear(x))
// op: matmul_v2 + elementwise_add + act
// named nodes: matmul, elementwise_add, act
// matmul_w, matmul_out
// ele_bias, elewise_add_out, act_out
struct LinearAct : public PatternBase {
LinearAct(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "linear_act") {}
PDNode* operator()(PDNode* x,
const std::unordered_set<std::string>& act_types,
bool with_grad_link,
bool is_act_grad_x_from_act);
// declare operator node's name
PATTERN_DECL_NODE(matmul);
PATTERN_DECL_NODE(ele_add);
PATTERN_DECL_NODE(act);
PATTERN_DECL_NODE(act_grad);
// declare variable node's name
PATTERN_DECL_NODE(matmul_w);
PATTERN_DECL_NODE(matmul_out);
PATTERN_DECL_NODE(elewise_add_out);
PATTERN_DECL_NODE(ele_bias);
PATTERN_DECL_NODE(act_out);
};
// The following patterns are used to fuse linear_grad and act_grad (ReLu or
// GeLU)
// formula: the backward of F.linear( act(x) )
// op: elementwise_add_grad + matmul_v2_grad + act_grad
// named nodes: ele_add_grad, matmul_grad, act_grad
// ele_grad_bias, ele_grad_dx, ele_grad_dbias
// matmul_grad_x, matmul_grad_dx, matmul_grad_dx
// matmul_grad_dw, act_grad_dx
struct ElewiseAddMatmulAct : public PatternBase {
ElewiseAddMatmulAct(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "elewiseadd_matmul_act") {}
PDNode* operator()(PDNode* x,
const std::unordered_set<std::string>& act_grad_types,
bool without_x_gradient,
bool is_act_grad_x_from_act);
// declare operator node's name
PATTERN_DECL_NODE(ele_add_grad);
PATTERN_DECL_NODE(matmul_grad);
PATTERN_DECL_NODE(act_grad);
// declare variable node's name
PATTERN_DECL_NODE(ele_out);
PATTERN_DECL_NODE(ele_grad_bias);
PATTERN_DECL_NODE(ele_grad_dx);
PATTERN_DECL_NODE(ele_grad_dbias);
PATTERN_DECL_NODE(matmul_grad_x);
PATTERN_DECL_NODE(matmul_grad_w);
PATTERN_DECL_NODE(matmul_grad_dx);
PATTERN_DECL_NODE(matmul_grad_dw);
PATTERN_DECL_NODE(act_grad_dx);
};
// 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") {}
PDNode* operator()(PDNode* conv_input, std::string conv_type = "conv2d");
// 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);
};
// 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.
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_output);
};
// 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 {