-
Notifications
You must be signed in to change notification settings - Fork 74k
/
copy_insertion.cc
2196 lines (2037 loc) · 90.4 KB
/
copy_insertion.cc
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 2017 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/compiler/xla/service/copy_insertion.h"
#include <algorithm>
#include <cstddef>
#include <optional>
#include <sstream>
#include "absl/container/flat_hash_map.h"
#include "absl/container/flat_hash_set.h"
#include "absl/container/inlined_vector.h"
#include "absl/functional/function_ref.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_join.h"
#include "absl/types/any.h"
#include "tensorflow/compiler/xla/frontend_attributes.h"
#include "tensorflow/compiler/xla/hlo/ir/hlo_computation.h"
#include "tensorflow/compiler/xla/hlo/ir/hlo_instruction.h"
#include "tensorflow/compiler/xla/hlo/ir/hlo_module.h"
#include "tensorflow/compiler/xla/hlo/ir/hlo_opcode.h"
#include "tensorflow/compiler/xla/service/compile_time_cap.h"
#include "tensorflow/compiler/xla/service/dump.h"
#include "tensorflow/compiler/xla/service/hlo_alias_analysis.h"
#include "tensorflow/compiler/xla/service/hlo_buffer.h"
#include "tensorflow/compiler/xla/service/hlo_dce.h"
#include "tensorflow/compiler/xla/service/hlo_graph_dumper.h"
#include "tensorflow/compiler/xla/service/hlo_ordering.h"
#include "tensorflow/compiler/xla/service/logical_buffer.h"
#include "tensorflow/compiler/xla/service/tuple_simplifier.h"
#include "tensorflow/compiler/xla/status_macros.h"
#include "tensorflow/compiler/xla/statusor.h"
#include "tensorflow/compiler/xla/types.h"
#include "tensorflow/compiler/xla/util.h"
#include "tensorflow/tsl/platform/logging.h"
namespace xla {
namespace {
using absl::StrAppend;
bool IsReadonlyEntryParameterValue(const HloValue& value) {
const HloComputation* computation = value.defining_instruction()->parent();
return value.defining_instruction()->opcode() == HloOpcode::kParameter &&
computation == computation->parent()->entry_computation() &&
!computation->parent()->input_output_alias_config().ParameterHasAlias(
value.defining_instruction()->parameter_number(), value.index());
}
bool IsConstantValue(const HloValue& value) {
return value.defining_instruction()->opcode() == HloOpcode::kConstant;
}
bool ValueIsReadOnly(const HloValue& value) {
return IsConstantValue(value) || IsReadonlyEntryParameterValue(value);
}
// Data structure describing the action which should be taken on parts of a
// computation buffers, with respect to the adding of special case copies.
struct SpecialCaseCopyPolicy {
// Insert a copy if the same buffer is found at multiple indices within the
// output tuple.
bool copy_root_replicated_buffers = false;
// If true, insert a copy if a buffer coming from a constant or a parameter
// is found within the output tuple.
bool copy_parameters_and_constants = false;
};
SpecialCaseCopyPolicy GetSpecialCaseCopyPolicy(const CallGraphNode& node,
HloModule* module,
HloComputation* computation) {
SpecialCaseCopyPolicy policy;
if (computation == module->entry_computation()) {
policy.copy_parameters_and_constants = true;
policy.copy_root_replicated_buffers = true;
}
return policy;
}
bool ShouldCopyRootValue(const HloValue& value,
const SpecialCaseCopyPolicy& policy) {
if (policy.copy_parameters_and_constants) {
return ValueIsReadOnly(value);
}
return false;
}
// Deep copy the given instructions 'from' and 'to' at the ShapeIndexes given in
// 'indices_to_copy'. Add control edges from the respective kCopy instructions
// in deep copy of 'from' to the respective kCopy instruction in the deep copy
// of 'to'.
//
// Requirements: 'from' and 'to' must have compatible shapes.
//
// For example, suppose 'from' and 'to' are two-element tuples where index 0 is
// the only index to copy. Prior to deep-copying we have:
//
//
// 'from'
// |
// ...
// |
// 'to'
//
// DeepCopyAndAddControlEdges produces:
//
// 'from'
// / \
// GTE GTE
// | |
// Copy |
// / \ /
// | Tuple
// | |
// ctrl ...
// edge |
// | |
// | 'to'
// | / \
// | GTE GTE
// \ | |
// Copy |
// \ /
// Tuple
//
StatusOr<std::pair<HloInstruction*, HloInstruction*>>
DeepCopyAndAddControlEdges(HloInstruction* from, HloInstruction* to,
const ShapeTree<bool>& indices_to_copy) {
DCHECK(ShapeUtil::Compatible(from->shape(), to->shape()));
// to/from_copy_tree hold the kCopy instruction produces by the deep
// copies. Elements which are not copied (indices_to_copy.element(index) ==
// false) have nullptr at that index.
ShapeTree<HloInstruction*> from_copy_tree(from->shape(),
/*init_value=*/nullptr);
TF_ASSIGN_OR_RETURN(HloInstruction * from_deep_copy,
from->parent()->DeepCopyInstruction(
from, &indices_to_copy, &from_copy_tree));
ShapeTree<HloInstruction*> to_copy_tree(to->shape(), /*init_value=*/nullptr);
TF_ASSIGN_OR_RETURN(
HloInstruction * to_deep_copy,
to->parent()->DeepCopyInstruction(to, &indices_to_copy, &to_copy_tree));
// Add control edges between the respective kCopy instructions.
for (const auto& pair : from_copy_tree) {
const ShapeIndex& index = pair.first;
HloInstruction* from_copy = pair.second;
HloInstruction* to_copy = to_copy_tree.element(index);
if (from_copy == nullptr) {
TF_RET_CHECK(to_copy == nullptr);
continue;
}
TF_RET_CHECK(to_copy != nullptr);
TF_RETURN_IF_ERROR(from_copy->AddControlDependencyTo(to_copy));
}
return std::make_pair(from_deep_copy, to_deep_copy);
}
// Compute the indices of the loop state which need copies in order to avoid
// live range interference. Generally, an element in the loop state does not
// need to be copied if the element is passed through transparently through the
// body.
//
// Returns whether any indices need to be copied.
bool IndicesToCopyForWhile(const HloDataflowAnalysis& dataflow,
const HloInstruction* xla_while,
ShapeTree<bool>* indices_to_copy) {
DCHECK(ShapeUtil::Compatible(indices_to_copy->shape(), xla_while->shape()));
bool any_copies = false;
const HloInstruction* init = xla_while->operand(0);
for (auto& pair : *indices_to_copy) {
const ShapeIndex& index = pair.first;
bool& should_copy = pair.second;
// If there is any ambiguity, then loop state must be copied.
if (dataflow.GetValueSet(init, index).values().size() > 1 ||
dataflow.GetValueSet(xla_while, index).values().size() > 1) {
should_copy = true;
} else {
// If the output of the while instruction is not the same as the init
// value of the while, then this element is not passed through the body
// transparently and must be copied.
should_copy = dataflow.GetUniqueValueAt(xla_while, index) !=
dataflow.GetUniqueValueAt(init, index);
}
any_copies |= should_copy;
}
return any_copies;
}
// Compute the indices of the conditional outputs which need copies. Umambiguous
// buffers(buffer with only one value) don't need copies.
bool IndicesToCopyForConditional(const HloDataflowAnalysis& dataflow,
const HloInstruction* xla_conditional,
ShapeTree<bool>* indices_to_copy) {
DCHECK(ShapeUtil::Compatible(indices_to_copy->shape(),
xla_conditional->shape()));
bool any_copies = false;
for (auto& pair : *indices_to_copy) {
const ShapeIndex& index = pair.first;
bool& should_copy = pair.second;
CHECK_EQ(dataflow.GetValueSet(xla_conditional, index).values().size(), 1);
auto value = dataflow.GetValueSet(xla_conditional, index).values()[0];
// The conditional must be copied if the value is a phi.
should_copy =
value->is_phi() && value->defining_instruction() == xla_conditional;
any_copies |= should_copy;
}
return any_copies;
}
// Add kCopy instructions around the given kWhile instruction to eliminate any
// possible live range interference of HLO values assuming a dependency-based
// ordering. Copies are added conservatively. There likely are copies which are
// not strictly necessary, but they are removed later in the pass via
// RemoveUnnecessaryCopies.
//
// Elements (each ShapeIndex) in the loop state are considered independently. A
// copy is added to each element of the loop state which is modified in the
// while body. For each such element, a total of three kCopy instructions are
// added at following locations:
//
// (1) The init value is copied before the kWhile instruction. Before:
//
// (Init)
// |
// kWhile
// |
// ...
//
// After:
//
// (Init)
// |
// kCopy
// |
// kWhile
// |
// ...
//
// This copy is necessary in case the init value is simultaneously live
// with the kWhile.
//
// (2) Copies are added to the parameter and root of the while body
// computation. Before:
//
// kParameter
// |
// ...
// |
// (body root)
//
// After:
//
// kParameter
// |
// kCopy ----------+
// | |
// ... ctrl
// | edge
// (body root) |
// | |
// kCopy <---------+
//
// The root kCopy becomes the new root of the computation. Both copies are
// necessary to any potential interference between the parameter value and
// the root value. The control edge prevents potential interference
// between the copies themselves.
//
// If the loop state is a tuple then the above kCopy instructions are a deep
// copy constructed of kCopy, kGetTupleElement, and kTuple instruction as
// constructed by HloInstruction::DeepCopyInstruction.
Status AddCopiesForWhile(const HloAliasAnalysis& alias_analysis,
HloInstruction* xla_while) {
VLOG(2) << "Adding copies for kWhile instruction " << xla_while->name();
TF_RET_CHECK(xla_while->opcode() == HloOpcode::kWhile);
ShapeTree<bool> indices_to_copy(xla_while->shape());
if (!IndicesToCopyForWhile(alias_analysis.dataflow_analysis(), xla_while,
&indices_to_copy)) {
VLOG(2) << "No copies necessary for kWhile instruction "
<< xla_while->name();
return OkStatus();
}
VLOG(2) << "Adding copies for " << xla_while->name() << " at indices:";
for (auto& pair : indices_to_copy) {
if (pair.second) {
VLOG(2) << " " << pair.first;
}
}
// Deep copy init.
HloInstruction* while_init = xla_while->mutable_operand(0);
TF_ASSIGN_OR_RETURN(
HloInstruction * while_init_copy,
xla_while->parent()->DeepCopyInstruction(while_init, &indices_to_copy));
TF_RETURN_IF_ERROR(while_init->ReplaceUseWith(xla_while, while_init_copy));
// Deep copy the parameter and the root. Extend a control edge from the copy
// of the parameter value to the corresponding copy value of the root.
HloComputation* body = xla_while->while_body();
HloInstruction* param = body->parameter_instruction(0);
HloInstruction* root = body->root_instruction();
// If param is the root then all indices should have been passed through the
// while body and we should have returned early above.
TF_RET_CHECK(param != root);
// Copy users before making a deep copy of the parameter as the deep copy
// will create new users of the parameter (eg, the GTE instructions of the
// deep copy).
std::vector<HloInstruction*> param_users = param->users();
TF_ASSIGN_OR_RETURN(auto pair,
DeepCopyAndAddControlEdges(param, root, indices_to_copy));
HloInstruction* param_copy = pair.first;
HloInstruction* root_copy = pair.second;
for (HloInstruction* user : param_users) {
TF_RETURN_IF_ERROR(param->ReplaceUseWith(user, param_copy));
}
body->set_root_instruction(root_copy);
return OkStatus();
}
// Add copies for the operands of in-place operations. RemoveUnnecessaryCopies
// will remove the unnecessary copies.
Status AddCopiesForInPlaceOperation(const HloAliasAnalysis& alias_analysis,
HloInstruction* in_place_op,
int64_t operand_number) {
VLOG(2) << "Adding copies for in-place operation " << in_place_op->name();
HloInstruction* operand = in_place_op->mutable_operand(operand_number);
TF_ASSIGN_OR_RETURN(HloInstruction * deep_copy,
in_place_op->parent()->DeepCopyInstruction(operand));
TF_RETURN_IF_ERROR(
operand->ReplaceUseWith(in_place_op, operand_number, deep_copy));
return OkStatus();
}
// Conservatively adds copies before root instruction of entry computation and
// each aliased parameter to resolve interference of aliased input and output
// buffer. We later rely on RemoveUnnecessaryCopies to drop the unnecessary
// ones.
Status AddCopiesForAliasedInputOutputs(
HloModule* module,
const absl::flat_hash_set<absl::string_view>& execution_threads) {
HloComputation* entry = module->entry_computation();
if (!HloInstruction::IsThreadIncluded(entry->execution_thread(),
execution_threads)) {
return OkStatus();
}
HloInstruction* root = entry->root_instruction();
ShapeTree<bool> output_indices_to_copy(root->shape());
std::vector<std::optional<ShapeTree<HloInstruction*>>> copied_parameters(
entry->num_parameters());
bool has_alias = false;
for (auto* param : entry->parameter_instructions()) {
bool param_has_alias = false;
ShapeTree<bool> param_indices_to_copy(param->shape());
module->input_output_alias_config().ForEachAlias(
[&](const ShapeIndex& output_index,
const HloInputOutputAliasConfig::Alias& alias) {
if (alias.parameter_number == param->parameter_number()) {
param_has_alias = true;
*(param_indices_to_copy.mutable_element(alias.parameter_index)) =
true;
*(output_indices_to_copy.mutable_element(output_index)) = true;
}
});
if (!param_has_alias) {
continue;
}
TF_RET_CHECK(param->parameter_number() < entry->num_parameters());
TF_RET_CHECK(!copied_parameters[param->parameter_number()]);
has_alias = true;
// Store a snapshot of users before DeepCopyInstruction, as
// DeepCopyInstruction introduces new users of the instruction.
std::vector<HloInstruction*> users = param->users();
ShapeTree<HloInstruction*> param_copy_tree(param->shape(),
/*init_value=*/nullptr);
TF_ASSIGN_OR_RETURN(HloInstruction * copied,
entry->DeepCopyInstruction(
param, ¶m_indices_to_copy, ¶m_copy_tree));
if (param == root) {
entry->set_root_instruction(copied);
root = copied;
}
for (HloInstruction* user : users) {
TF_RETURN_IF_ERROR(param->ReplaceUseWith(user, copied));
}
copied_parameters[param->parameter_number()] = param_copy_tree;
}
if (!has_alias) {
return OkStatus();
}
// Add copies before root instruction.
ShapeTree<HloInstruction*> output_copy_tree(root->shape(),
/*init_value=*/nullptr);
TF_ASSIGN_OR_RETURN(HloInstruction * root_copied,
root->parent()->DeepCopyInstruction(
root, &output_indices_to_copy, &output_copy_tree));
// Add control dependencies between the input/output copies.
TF_RETURN_IF_ERROR(module->input_output_alias_config().ForEachAliasWithStatus(
[&](const ShapeIndex& output_index,
const HloInputOutputAliasConfig::Alias& alias) -> Status {
if (!copied_parameters[alias.parameter_number]) {
return OkStatus();
}
HloInstruction* from =
copied_parameters[alias.parameter_number]->element(
alias.parameter_index);
HloInstruction* to = output_copy_tree.element(output_index);
TF_RET_CHECK(from != nullptr);
TF_RET_CHECK(to != nullptr);
TF_RETURN_IF_ERROR(from->AddControlDependencyTo(to));
return OkStatus();
}));
entry->set_root_instruction(root_copied);
return OkStatus();
}
// Removes any control dependencies to or from the given instruction.
Status StripControlDependenciesFrom(HloInstruction* instruction) {
while (!instruction->control_successors().empty()) {
TF_RETURN_IF_ERROR(instruction->RemoveControlDependencyTo(
instruction->control_successors().front()));
}
while (!instruction->control_predecessors().empty()) {
TF_RETURN_IF_ERROR(
instruction->control_predecessors().front()->RemoveControlDependencyTo(
instruction));
}
return OkStatus();
}
class LiveRangeRegions {
public:
struct InstructionInfo {
InstructionInfo() : value_definition(nullptr), is_definition(false) {}
// The instruction that defines the value being used. It basically saves
// the defining instruction of each HloValue.
HloInstruction* value_definition;
// Whether the instruction defines a new value (or merely uses one). This
// basically remembers whether the instruction actually creates an HloValue
// or merely uses one, from a collection of given HloValues. Note that if
// is_definition = true, it merely says the instruction creates a new
// HloValue with or without defining a new one. For example, kAdd create a
// new HloValue (can be value_definition), but tuples or get-tuple-element,
// create a new HloValue aliasing without defining a new value (cannot be
// value_definition).
bool is_definition;
};
// Map instructions that use a value to the defining instruction of the value.
// Because all values must belong to the same live range, an instruction can
// have at most a single value-defining instruction; otherwise the multiple
// incoming active values would share a single buffer, which is not allowed.
// The value-defining and value-use instructions do not have to belong to the
// same computation, but the value use needs to be nested within the defining
// computation.
typedef HloInstructionMap<InstructionInfo> InstructionMap;
typedef std::pair<HloInstruction*, InstructionInfo> InstructionEntry;
// Map each computation to its immediately contained instructions.
typedef absl::flat_hash_map<const HloComputation*, InstructionMap>
ComputationMap;
InstructionMap& operator[](const HloComputation* computation) {
if (computation_map_.find(computation) == computation_map_.end()) {
computation_vector_.push_back(computation);
}
return computation_map_[computation];
}
const InstructionMap& operator[](const HloComputation* computation) const {
ComputationMap::const_iterator p = computation_map_.find(computation);
CHECK(p != computation_map_.end());
return p->second;
}
absl::InlinedVector<const HloComputation*, 5>::const_iterator begin() const {
return computation_vector_.begin();
}
absl::InlinedVector<const HloComputation*, 5>::const_iterator end() const {
return computation_vector_.end();
}
int64_t size() const {
CHECK_EQ(computation_vector_.size(), computation_map_.size());
return computation_vector_.size();
}
bool empty() const { return size() == 0; }
const HloComputation* Computation(int64_t index) const {
return computation_vector_[index];
}
bool contains(HloInstruction* instr) const {
CHECK_NE(instr, nullptr);
auto* computation = instr->parent();
auto p = computation_map_.find(computation);
if (p == computation_map_.end()) {
return false;
}
auto instr_map = (*p).second;
return instr_map.find(instr) != instr_map.end();
}
private:
ComputationMap computation_map_;
absl::InlinedVector<const HloComputation*, 5> computation_vector_;
};
namespace {
// Represent relations between the locations of two regions of instructions,
// each region can include 0-n instructions.
class Relation {
public:
enum RuntimeOrder {
// Indicate that there is no overlap whatsoever between the two regions.
kNoOverlap = 0,
// Indicate that the first region includes the same set of instructions as
// the second region.
kSameInstr = 1,
// Indicate that the first region is entirely before the second region
// starts.
kBeforeStart = 2,
// Indicate that the first region is before the second region ends.
kBeforeStartOrSameInstr = kBeforeStart | kSameInstr,
// Indicate that the first region is entirely after the second region ends.
kAfterEnd = 4,
// Indicate that the first region is after the second region
// starts, with some instructions before the second region ends.
kAfterEndOrSameInstr = kAfterEnd | kSameInstr,
// Indicate that the first region overlaps with the second one, but share no
// common instructions.
kBeforeStartOrAfterEnd = kBeforeStart | kAfterEnd,
// Indicate that the first region overlaps with the second one, and have
// some common instructions.
kBeforeOrAfterOrOverlap = kBeforeStart | kAfterEnd | kSameInstr,
};
Relation() : intercept_def_use_(false) {}
explicit Relation(RuntimeOrder order, bool intercept_def_use = false)
: intercept_def_use_(intercept_def_use) {
orders_.push_back(order);
}
Relation(const Relation& that)
: intercept_def_use_(that.intercept_def_use_), orders_(that.orders_) {}
bool operator==(const Relation& that) const {
return intercept_def_use_ == that.intercept_def_use_ &&
absl::c_equal(orders_, that.orders_);
}
// Return whether the runtime ordering may imply interception, assuming it
// models the relation between a modifying and a use instruction.
bool UseImpliesInterception() const {
CHECK_EQ(orders_.size(), 1);
return UseImpliesInterception(orders_[0]);
}
// Return whether the runtime ordering may imply interception, assuming it
// models the relation between a modifying and a definition instruction.
bool DefinitionImpliesInterception() const {
CHECK_EQ(orders_.size(), 1);
return DefinitionImpliesInterception(orders_[0]);
}
// Return whether the current relation models a modifying instruction that
// intercepts the dataflow of another live range region.
bool InterceptDefUse() const { return intercept_def_use_; }
// Update interception state to the given value.
void UpdateInterception(bool value) {
CHECK_EQ(orders_.size(), 1);
intercept_def_use_ = value;
}
Relation::RuntimeOrder GetRuntimeOrder() const {
if (orders_.empty()) {
return Relation::kNoOverlap;
}
CHECK_EQ(orders_.size(), 1);
return orders_[0];
}
// Return whether the current relation implies two overlapping regions.
bool RuntimeOrderOverlap() const {
return absl::c_any_of(orders_, ImpliesOverlap);
}
bool RuntimeOrderIsUnordered() const {
return orders_.size() == 1 && orders_[0] == kBeforeStartOrAfterEnd;
}
bool RuntimeOrderIsNoOverlap() const {
return orders_.empty() || (orders_.size() == 1 && orders_[0] == kNoOverlap);
}
bool RuntimeOrderIsRunBefore() const {
return orders_.size() == 1 && orders_[0] == kBeforeStart;
}
bool RuntimeOrderIsRunAfter() const {
return orders_.size() == 1 && orders_[0] == kAfterEnd;
}
std::string ToString() const {
return absl::StrCat("Interception = ", intercept_def_use_, ";",
absl::StrJoin(orders_, ","));
}
static bool DefinitionImpliesInterception(RuntimeOrder definition) {
return (definition == kAfterEnd || definition == kBeforeStartOrAfterEnd);
}
static bool UseImpliesInterception(RuntimeOrder use) {
return (use == kBeforeStart || use == kBeforeStartOrAfterEnd);
}
// Summarize additional relations into a single runtime ordering, assuming
// both relations are modeling constraints of the same source instruction.
void UnionRelationFromSameSource(const Relation& rel) {
CHECK_LE(orders_.size(), 1);
CHECK_EQ(rel.orders_.size(), 1);
if (orders_.empty()) {
orders_.push_back(rel.orders_[0]);
} else {
orders_[0] = Union(orders_[0], rel.orders_[0]);
}
intercept_def_use_ = intercept_def_use_ || rel.intercept_def_use_;
}
// Summarize additional relations into disjoint runtime orderings, assuming
// the relations are modeling constraints of different source instructions.
void UnionRelationFromDifferentSource(const Relation& rel) {
if (rel.orders_.empty()) {
return;
}
CHECK_EQ(rel.orders_.size(), 1);
intercept_def_use_ = intercept_def_use_ || rel.intercept_def_use_;
for (auto& local_order : orders_) {
if (OverwriteIfSubsume(rel.orders_[0], &local_order)) {
return;
}
}
orders_.push_back(rel.orders_[0]);
}
static Relation::RuntimeOrder ReverseRuntimeOrder(RuntimeOrder order) {
switch (order) {
case kNoOverlap:
case kSameInstr:
case kBeforeStartOrAfterEnd:
case kBeforeOrAfterOrOverlap:
return order;
case kBeforeStart:
return kAfterEnd;
case kBeforeStartOrSameInstr:
return kAfterEndOrSameInstr;
case kAfterEnd:
return kBeforeStart;
case kAfterEndOrSameInstr:
return kBeforeStartOrSameInstr;
}
}
private:
// Indicate that the second region may intercept the def-use dataflow of the
// first region, if their buffers are combined.
bool intercept_def_use_;
// Remember the different runtime orderings of different instructions.
absl::InlinedVector<RuntimeOrder, 4> orders_;
static RuntimeOrder Union(RuntimeOrder o1, RuntimeOrder o2) {
return static_cast<Relation::RuntimeOrder>(o1 | o2);
}
static bool ImpliesOverlap(RuntimeOrder o) {
return o >= RuntimeOrder::kBeforeStartOrAfterEnd;
}
// Returns whether ordering constraint o1 includes o2 as a subset, when they
// represent runtime orderings (interleavings) of two different regions.
static bool Subsume(RuntimeOrder o1, RuntimeOrder o2) {
return Union(o1, o2) == o1;
}
// Overwrites o1 with o2 if o2 subsumes o1 (as defined above by the Subsume
// function). Return whether o2 is subsumed by the new value in o1.
static bool OverwriteIfSubsume(RuntimeOrder o2, RuntimeOrder* o1) {
if (*o1 == o2) {
return true;
}
CHECK_NE(o1, nullptr);
// Overwrite o1 with o2 if it is subsumed by o2.
if (Subsume(o2, *o1)) {
*o1 = o2;
return true;
} else if (Subsume(*o1, o2)) {
// If o2 is already subsumed by o1, do nothing.
return true;
}
// If neither o1 nor o2 is subsumed by the other, return false, so that o2
// will be inserted as a separate entry representing all possible orderings.
return false;
}
};
class ComputeRelativeLocation {
public:
typedef LiveRangeRegions::InstructionEntry InstructionEntry;
explicit ComputeRelativeLocation(HloOrdering* ordering)
: ordering_(ordering) {
VLOG(3) << "New analysis\n";
}
// Compute locationing constraints between two instructions. Here entry2 is
// the source instruction, in that the returned value describes the relation
// of entry2 in terms of whether it is before or after entry1, and whether it
// can intercept the def-use data flow of entry1.
Relation Compute(const InstructionEntry& entry1,
const InstructionEntry& entry2, bool instr2_can_modify) {
auto def = entry1.second.value_definition;
auto use = entry1.first;
Relation::RuntimeOrder order =
ComputeRuntimeOrdering(entry2.first, entry1.first);
if (order == Relation::kSameInstr &&
entry1.second.is_definition != entry2.second.is_definition) {
if (entry1.second.is_definition) {
order = Relation::kBeforeStart;
} else {
order = Relation::kAfterEnd;
}
}
bool intercept = AlwaysForceInterception(entry2.first);
if (def == nullptr || !instr2_can_modify) {
return Relation(order, intercept);
}
// If the definition and use are parameter and return (root) of the parent
// computation, then any modification is considered intercepting.
if (def->opcode() == HloOpcode::kParameter &&
use == use->parent()->root_instruction()) {
VLOG(3) << "Setting interception due to parameter/root relation\n";
return Relation(order, true);
}
if (Relation::UseImpliesInterception(order)) {
auto order2 = ComputeRuntimeOrdering(entry2.first, def);
if (Relation::DefinitionImpliesInterception(order2)) {
VLOG(3) << "Setting interception for " << def->ToString()
<< " with use:" << entry1.first->ToString() << "\n";
intercept = true;
}
}
return Relation(order, intercept);
}
// Return the relative locations (defined above) of range2 in relation to
// instructions in range1. Return kNoOverlap if range2 is outside of range1.
Relation Compute(const LiveRangeRegions& range1,
const LiveRangeRegions& range2) {
Relation dir_src_dest;
for (int64_t index = 0; index < range1.size(); index++) {
auto* computation1 = range1.Computation(index);
for (const auto* computation2 : range2) {
for (auto instr_entry2 : range2[computation2]) {
if (!ordering_->call_graph().Dominates(computation1, computation2)) {
continue;
}
VLOG(3) << "Locationing " << instr_entry2.first->ToString();
// Saves relations between instr2 and other instructions in range1.
bool instr2_can_modify =
InstructionCanIntercept(instr_entry2, range1);
Relation instr2_relation;
std::vector<InstructionEntry> unordered_ops;
bool unordered_intercept = false;
for (auto instr_entry1 : range1[computation1]) {
auto rel = Compute(instr_entry1, instr_entry2, instr2_can_modify);
VLOG(3) << "new relation with:" << instr_entry1.first->ToString()
<< " = " << rel.ToString() << "\n";
if (!rel.RuntimeOrderIsUnordered()) {
instr2_relation.UnionRelationFromSameSource(rel);
} else {
unordered_ops.push_back(instr_entry1);
unordered_intercept |= rel.InterceptDefUse();
}
VLOG(3) << "instr2 relation:" << instr2_relation.ToString() << "\n";
}
// Here instru2_relation is guaranteed to have at most a single entry,
// because it was initialized to be empty, and has been updated only
// via instr2_relation.UnionRelationFromSameSource(rel), which
// maintains that the updated result has only a single entry.
if (!ForceRuntimeOrder(unordered_ops, instr_entry2,
instr2_relation.GetRuntimeOrder())) {
VLOG(3) << "Unable to force ordering of unordered ops\n";
instr2_relation.UnionRelationFromSameSource(Relation(
Relation::kBeforeStartOrAfterEnd, unordered_intercept));
}
dir_src_dest.UnionRelationFromDifferentSource(instr2_relation);
VLOG(3) << "Resulting relation : " << dir_src_dest.ToString() << "\n";
}
}
}
return dir_src_dest;
}
// Return whether control dependences, if exist, are added successfully.
bool AddControlDependenceForUnorderedOps() {
if (ctrl_deps_.empty()) {
return true;
}
PredecessorHloOrdering* ordering =
dynamic_cast<PredecessorHloOrdering*>(ordering_);
if (ordering == nullptr) {
// Support force ordering of unordered-ops only when using predecssor
// ordering.
return false;
}
for (const auto& comp_it : ctrl_deps_) {
HloComputation* parent = comp_it.first;
HloReachabilityMap& reachability_map = ordering->reachability_map(parent);
for (const auto& instr_it : comp_it.second) {
HloInstruction* entry1 = instr_it.first;
for (HloInstruction* entry2 : instr_it.second) {
VLOG(3) << "Add control dependence between " << entry2->ToString();
VLOG(3) << "\n vs " << entry1->ToString() << "\n";
TF_CHECK_OK(entry2->AddControlDependencyTo(entry1));
}
reachability_map.UpdateReachabilityThroughInstruction(entry1);
for (HloInstruction* entry2 : instr_it.second) {
DCHECK(ordering_->GetExecutionConstraint(entry1, entry2) ==
HloOrdering::ExecutionConstraint::kRunAfter);
}
}
}
return true;
}
private:
enum ComputeStatus {
kFullyComputed,
kPartiallyComputed,
kNotComputed,
};
typedef std::pair<ComputeStatus, Relation::RuntimeOrder> SavedRelation;
// Returns whether it is safe to force the desired_relation ordering between
// all operations in unordered_ops and entry2. If safe, save the new enforced
// ordering relations.
bool ForceRuntimeOrder(absl::Span<const InstructionEntry> unordered_ops,
const InstructionEntry entry2,
Relation::RuntimeOrder desired_relation) {
if (unordered_ops.empty()) {
return true;
}
if (desired_relation != Relation::kBeforeStart &&
desired_relation != Relation::kAfterEnd) {
return false;
}
auto ModifiesNonCopy = [](HloInstruction* instr, const HloInstruction* op) {
auto in_place = HloDataflowAnalysis::GetInPlaceInputOutputPairs(instr);
if (in_place.empty()) {
return false;
}
return absl::c_any_of(
in_place, [&](const std::pair<HloOperandIndex, ShapeIndex>&
operand_and_output_index) {
auto* op2 =
instr->operand(operand_and_output_index.first.operand_number);
return (op == nullptr) ? (op2->opcode() == HloOpcode::kCopy)
: (op2 == op);
});
};
for (const InstructionEntry& entry1 : unordered_ops) {
// Only consider instructions in the same computation.
if (entry1.first->parent() != entry2.first->parent()) {
return false;
}
HloInstruction* pred = (desired_relation == Relation::kBeforeStart)
? entry2.first
: entry1.first;
HloInstruction* succ = (desired_relation == Relation::kBeforeStart)
? entry1.first
: entry2.first;
if (pred == pred->parent()->root_instruction()) {
return false;
}
if (succ->opcode() == HloOpcode::kCopy &&
ModifiesNonCopy(pred, succ->operand(0))) {
VLOG(3) << "Failed to force unordered op ordering due to copy ordering "
<< " between " << pred->ToString() << "\n";
VLOG(3) << " vs. " << succ->ToString() << "\n";
return false;
}
}
for (const InstructionEntry& entry1 : unordered_ops) {
Save(entry2.first, entry1.first, desired_relation, true);
}
return true;
}
static bool AlwaysForceInterception(HloInstruction* instr) {
// The following communication operations can have some unexpected side
// effects, when synchronizing across processes. Therefore, we
// conservatively try provide dedicated buffers to these operations instead
// of allowing them to share buffers with other operations, as the reuse may
// cause unexpected interferences.
if (HloDataflowAnalysis::IsAsynchronousOperationStart(instr->opcode()) ||
HloDataflowAnalysis::IsAsynchronousOperationDone(instr->opcode())) {
return true;
}
switch (instr->opcode()) {
// TODO(b/190903339): It appears that collectivePermute needs to be
// followed by a copy when escaping through a computation root.
case HloOpcode::kCollectivePermute:
return true;
default:
return false;
}
}
// Returns whether the given instr may intercept the def-use flow of another
// ongoing live range if its buffer is combined with the other live range.
// The function should return true if instr creates a new HloValue that could
// overwrite an existing HloValue in the combined buffer.
// More specifically, here we are looking for operations that create new
// values, e.g., add, subtract, in contrast to HLOs that merely create
// aliasings among existing values, e.g., tuple, get-tuple-element. Any of the
// new values created by operations such as add or subtract, when included as
// definition operations in a live range, are aliases of the buffer to be
// allocated to the live range and so are treated as they may be modifying the
// targeting buffer.
bool InstructionCanIntercept(const InstructionEntry& entry,
const LiveRangeRegions& region) {
auto instr = entry.first;
if (!entry.second.is_definition) {
// If the instruction only uses the value, it can intercept only if it
// modifies the buffer in place.
return !HloDataflowAnalysis::GetInPlaceInputOutputPairs(instr).empty();
}
switch (instr->opcode()) {
// If the copy instruction is used to connect two live range regions,
// it does not overwrite the combined buffer with new values.
case HloOpcode::kCopy:
// Checking the copy simply copies from the other live range with no
// layout conflicts.
if (region.contains(instr->mutable_operand(0)) &&
ShapeUtil::Equal(instr->shape(), instr->operand(0)->shape())) {
return false; // Cannot intercept.
}
return true;
// The following operations merely create aliases among the HloValues.
case HloOpcode::kParameter:
case HloOpcode::kTuple:
case HloOpcode::kGetTupleElement:
// Here we consider all the compound operations (e.g., conditionals and
// while loops) as if they do not modify any HloValue, with the argument
// being that any value modifying operation contained inside will be
// considered separately to make sure the kIntercept relation being
// recorded as appropriate. Since the compound operations may or may not
// modify, not treating them as value modifying would make the algorithm
// less conservative.
case HloOpcode::kWhile:
case HloOpcode::kCall:
case HloOpcode::kConditional:
return false;
default:
return true;
}
return true;
}
SavedRelation AlreadyComputed(HloInstruction* op1, HloInstruction* op2) {
auto p2 = saved_relations_.find(op2);
if (p2 != saved_relations_.end()) {
auto p1 = (*p2).second.find(op1);
if (p1 != (*p2).second.end()) {
return SavedRelation(kFullyComputed, (*p1).second);
}
}
p2 = saved_relations_.find(op1);
if (p2 != saved_relations_.end()) {
auto p1 = (*p2).second.find(op2);
if (p1 != (*p2).second.end()) {
return SavedRelation(kPartiallyComputed,