-
Notifications
You must be signed in to change notification settings - Fork 74k
/
lookup_table_op.cc
1247 lines (1105 loc) · 49 KB
/
lookup_table_op.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 2015 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/core/kernels/lookup_table_op.h"
#define EIGEN_USE_THREADS
#include <string>
#include <type_traits>
#include <utility>
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/framework/variant.h"
#include "tensorflow/core/kernels/initializable_lookup_table.h"
#include "tensorflow/core/lib/gtl/inlined_vector.h"
#include "tensorflow/core/lib/hash/hash.h"
#include "tensorflow/core/platform/random.h"
namespace tensorflow {
namespace lookup {
std::string UniqueNodeName(const std::string& base) {
static std::atomic<int64_t> counter(0);
return strings::StrCat(base, "/", counter.fetch_add(1), "/", random::New64());
}
// Lookup table that wraps an unordered_map, where the key and value data type
// is specified. Each individual value must be a scalar. If vector values are
// required, use MutableHashTableOfTensors.
//
// This table is mutable and thread safe - Insert can be called at any time.
//
// Sample use case:
//
// MutableHashTableOfScalars<int64, int64> table; // int64 -> int64.
// // Populate the table, elements could be added in one or multiple calls.
// table.Insert(key_tensor, value_tensor); // Populate the table.
//
// table.Find(in_t, &out_t, default_t)
//
template <class K, class V>
class MutableHashTableOfScalars final : public LookupInterface {
public:
MutableHashTableOfScalars(OpKernelContext* ctx, OpKernel* kernel) {}
size_t size() const override {
tf_shared_lock l(mu_);
return table_.size();
}
Status Find(OpKernelContext* ctx, const Tensor& key, Tensor* value,
const Tensor& default_value) override {
const auto key_values = key.flat<K>();
auto value_values = value->flat<V>();
const auto default_flat = default_value.flat<V>();
int64_t total = value_values.size();
int64_t default_total = default_flat.size();
bool is_full_size_default = (total == default_total);
tf_shared_lock l(mu_);
for (int64_t i = 0; i < key_values.size(); ++i) {
// is_full_size_default is true:
// Each key has an independent default value, key_values(i)
// corresponding uses default_flat(i) as its default value.
//
// is_full_size_default is false:
// All keys will share the default_flat(0) as default value.
value_values(i) = gtl::FindWithDefault(
table_, SubtleMustCopyIfIntegral(key_values(i)),
is_full_size_default ? default_flat(i) : default_flat(0));
}
return OkStatus();
}
Status DoInsert(bool clear, const Tensor& keys, const Tensor& values) {
const auto key_values = keys.flat<K>();
const auto value_values = values.flat<V>();
mutex_lock l(mu_);
if (clear) {
table_.clear();
}
for (int64_t i = 0; i < key_values.size(); ++i) {
gtl::InsertOrUpdate(&table_, SubtleMustCopyIfIntegral(key_values(i)),
SubtleMustCopyIfIntegral(value_values(i)));
}
return OkStatus();
}
Status Insert(OpKernelContext* ctx, const Tensor& keys,
const Tensor& values) override {
return DoInsert(false, keys, values);
}
Status Remove(OpKernelContext* ctx, const Tensor& keys) override {
const auto key_values = keys.flat<K>();
mutex_lock l(mu_);
for (int64_t i = 0; i < key_values.size(); ++i) {
table_.erase(SubtleMustCopyIfIntegral(key_values(i)));
}
return OkStatus();
}
Status ImportValues(OpKernelContext* ctx, const Tensor& keys,
const Tensor& values) override {
return DoInsert(true, keys, values);
}
Status ExportValues(OpKernelContext* ctx) override {
tf_shared_lock l(mu_);
int64_t size = table_.size();
Tensor* keys;
Tensor* values;
TF_RETURN_IF_ERROR(
ctx->allocate_output("keys", TensorShape({size}), &keys));
TF_RETURN_IF_ERROR(
ctx->allocate_output("values", TensorShape({size}), &values));
ExportKeysAndValues(keys, values);
return OkStatus();
}
DataType key_dtype() const override { return DataTypeToEnum<K>::v(); }
DataType value_dtype() const override { return DataTypeToEnum<V>::v(); }
TensorShape key_shape() const final { return TensorShape(); }
TensorShape value_shape() const override { return TensorShape(); }
int64_t MemoryUsed() const override {
int64_t ret = 0;
tf_shared_lock l(mu_);
for (unsigned i = 0; i < table_.bucket_count(); ++i) {
size_t bucket_size = table_.bucket_size(i);
if (bucket_size == 0) {
ret++;
} else {
ret += bucket_size;
}
}
return sizeof(MutableHashTableOfScalars) + ret;
}
Status AsGraphDef(GraphDefBuilder* builder, Node** out) const override {
tf_shared_lock l(mu_);
int64_t size = table_.size();
Tensor keys(key_dtype(), TensorShape({size}));
Tensor values(value_dtype(), TensorShape({size}));
ExportKeysAndValues(&keys, &values);
// We set use_node_name_sharing with a unique node name so that the resource
// can outlive the MutableHashTableV2 kernel. This means that the lifetime
// of the resource will be tied to the lifetime of the resource manager it
// is created in.
// TODO(b/181695913): Provide a mechanism for deleting this resource
// earlier when appropriate.
Node* table = ops::SourceOp(
"MutableHashTableV2",
builder->opts()
.WithName(UniqueNodeName("MutableHashTableFromGraphDef"))
.WithAttr("use_node_name_sharing", true)
.WithAttr("key_dtype", key_dtype())
.WithAttr("value_dtype", value_dtype()));
Node* keys_node = ops::SourceOp(
"Const",
builder->opts().WithAttr("dtype", key_dtype()).WithAttr("value", keys));
Node* values_node =
ops::SourceOp("Const", builder->opts()
.WithAttr("dtype", value_dtype())
.WithAttr("value", values));
Node* import_table =
ops::TernaryOp("LookupTableImportV2", table, keys_node, values_node,
builder->opts()
.WithAttr("Tin", key_dtype())
.WithAttr("Tout", value_dtype()));
*out = ops::UnaryOp("Identity", table,
builder->opts().WithControlInput(import_table));
return OkStatus();
}
private:
// Writes all keys and values into `keys` and `values`. `keys` and `values`
// must point to tensors of size `table_.size()`.
void ExportKeysAndValues(Tensor* keys, Tensor* values) const
TF_SHARED_LOCKS_REQUIRED(mu_) {
auto keys_data = keys->flat<K>();
auto values_data = values->flat<V>();
int64_t i = 0;
for (auto it = table_.begin(); it != table_.end(); ++it, ++i) {
keys_data(i) = it->first;
values_data(i) = it->second;
}
}
mutable mutex mu_;
std::unordered_map<K, V> table_ TF_GUARDED_BY(mu_);
};
// Lookup table that wraps an unordered_map. Behaves identical to
// MutableHashTableOfScalars except that each value must be a vector.
template <class K, class V>
class MutableHashTableOfTensors final : public LookupInterface {
public:
MutableHashTableOfTensors(OpKernelContext* ctx, OpKernel* kernel) {
OP_REQUIRES_OK(ctx,
GetNodeAttr(kernel->def(), "value_shape", &value_shape_));
OP_REQUIRES(
ctx, TensorShapeUtils::IsVector(value_shape_),
errors::InvalidArgument("Default value must be a vector, got shape ",
value_shape_.DebugString()));
}
size_t size() const override {
tf_shared_lock l(mu_);
return table_.size();
}
Status Find(OpKernelContext* ctx, const Tensor& key, Tensor* value,
const Tensor& default_value) override {
const auto default_flat = default_value.flat_inner_dims<V, 2>();
const auto key_values = key.flat<K>();
auto value_values = value->flat_inner_dims<V, 2>();
int64_t value_dim = value_shape_.dim_size(0);
int64_t total = value_values.size();
int64_t default_total = default_flat.size();
bool is_full_size_default = (total == default_total);
tf_shared_lock l(mu_);
for (int64_t i = 0; i < key_values.size(); ++i) {
ValueArray* value_vec =
gtl::FindOrNull(table_, SubtleMustCopyIfIntegral(key_values(i)));
if (value_vec != nullptr) {
for (int64_t j = 0; j < value_dim; j++) {
value_values(i, j) = value_vec->at(j);
}
} else {
// is_full_size_default is true:
// Each key has an independent default value, key_values(i)
// corresponding uses default_flat(i) as its default value.
//
// is_full_size_default is false:
// All keys will share the default_flat(0) as default value.
for (int64_t j = 0; j < value_dim; j++) {
value_values(i, j) =
is_full_size_default ? default_flat(i, j) : default_flat(0, j);
}
}
}
return OkStatus();
}
Status DoInsert(bool clear, const Tensor& keys, const Tensor& values) {
const auto key_values = keys.flat<K>();
const auto value_values = values.flat_inner_dims<V, 2>();
int64_t value_dim = value_shape_.dim_size(0);
mutex_lock l(mu_);
if (clear) {
table_.clear();
}
for (int64_t i = 0; i < key_values.size(); ++i) {
ValueArray value_vec;
for (int64_t j = 0; j < value_dim; j++) {
V value = value_values(i, j);
value_vec.push_back(value);
}
gtl::InsertOrUpdate(&table_, SubtleMustCopyIfIntegral(key_values(i)),
value_vec);
}
return OkStatus();
}
Status Insert(OpKernelContext* ctx, const Tensor& keys,
const Tensor& values) override {
return DoInsert(false, keys, values);
}
Status Remove(OpKernelContext* ctx, const Tensor& keys) override {
const auto key_values = keys.flat<K>();
mutex_lock l(mu_);
for (int64_t i = 0; i < key_values.size(); ++i) {
table_.erase(SubtleMustCopyIfIntegral(key_values(i)));
}
return OkStatus();
}
Status ImportValues(OpKernelContext* ctx, const Tensor& keys,
const Tensor& values) override {
return DoInsert(true, keys, values);
}
Status ExportValues(OpKernelContext* ctx) override {
tf_shared_lock l(mu_);
int64_t size = table_.size();
int64_t value_dim = value_shape_.dim_size(0);
Tensor* keys;
Tensor* values;
TF_RETURN_IF_ERROR(
ctx->allocate_output("keys", TensorShape({size}), &keys));
TF_RETURN_IF_ERROR(ctx->allocate_output(
"values", TensorShape({size, value_dim}), &values));
ExportKeysAndValues(keys, values);
return OkStatus();
}
DataType key_dtype() const override { return DataTypeToEnum<K>::v(); }
DataType value_dtype() const override { return DataTypeToEnum<V>::v(); }
TensorShape key_shape() const final { return TensorShape(); }
TensorShape value_shape() const override { return value_shape_; }
int64_t MemoryUsed() const override {
int64_t ret = 0;
tf_shared_lock l(mu_);
for (unsigned i = 0; i < table_.bucket_count(); ++i) {
size_t bucket_size = table_.bucket_size(i);
if (bucket_size == 0) {
ret++;
} else {
ret += bucket_size;
}
}
return sizeof(MutableHashTableOfTensors) + ret;
}
Status AsGraphDef(GraphDefBuilder* builder, Node** out) const override {
tf_shared_lock l(mu_);
int64_t size = table_.size();
Tensor keys(key_dtype(), TensorShape({size}));
Tensor values(value_dtype(), TensorShape({size, value_shape_.dim_size(0)}));
ExportKeysAndValues(&keys, &values);
// We set use_node_name_sharing with a unique node name so that the resource
// can outlive the MutableHashTableOfTensorsV2 kernel. This means that the
// lifetime of the resource will be tied to the lifetime of the resource
// manager it is created in.
// TODO(b/181695913): Provide a mechanism for deleting this resource
// earlier when appropriate.
Node* table =
ops::SourceOp("MutableHashTableOfTensorsV2",
builder->opts()
.WithName(UniqueNodeName("MutableHashTableOfTensors"))
.WithAttr("use_node_name_sharing", true)
.WithAttr("key_dtype", key_dtype())
.WithAttr("value_dtype", value_dtype())
.WithAttr("value_shape", value_shape_));
Node* keys_node = ops::SourceOp(
"Const",
builder->opts().WithAttr("dtype", key_dtype()).WithAttr("value", keys));
Node* values_node =
ops::SourceOp("Const", builder->opts()
.WithAttr("dtype", value_dtype())
.WithAttr("value", values));
Node* import_table =
ops::TernaryOp("LookupTableImportV2", table, keys_node, values_node,
builder->opts()
.WithAttr("Tin", key_dtype())
.WithAttr("Tout", value_dtype()));
*out = ops::UnaryOp("Identity", table,
builder->opts().WithControlInput(import_table));
return OkStatus();
}
private:
// Writes all keys and values into `keys` and `values`. `keys` and `values`
// must point to tensors of size `table_.size()`.
void ExportKeysAndValues(Tensor* keys, Tensor* values) const
TF_SHARED_LOCKS_REQUIRED(mu_) {
int64_t value_dim = value_shape_.dim_size(0);
auto keys_data = keys->flat<K>();
auto values_data = values->matrix<V>();
int64_t i = 0;
for (auto it = table_.begin(); it != table_.end(); ++it, ++i) {
K key = it->first;
ValueArray value = it->second;
keys_data(i) = key;
for (int64_t j = 0; j < value_dim; j++) {
values_data(i, j) = value[j];
}
}
}
TensorShape value_shape_;
mutable mutex mu_;
typedef gtl::InlinedVector<V, 4> ValueArray;
std::unordered_map<K, ValueArray> table_ TF_GUARDED_BY(mu_);
};
namespace {
template <typename T>
inline uint64 HashScalar(const T& key) {
return static_cast<uint64>(key);
}
inline uint64 HashScalar(const tstring& key) { return Hash64(key); }
// If the given shape is a scalar return {1} instead. Otherwise leave it alone.
TensorShape MaybeVectorizeShape(const TensorShape& shape) {
if (shape.dims() == 0) {
return TensorShape({1});
}
return shape;
}
} // namespace
// Modeled after densehashtable in https://github.com/sparsehash/sparsehash
template <class K, class V>
class MutableDenseHashTable final : public LookupInterface {
public:
MutableDenseHashTable(OpKernelContext* ctx, OpKernel* kernel) {
OP_REQUIRES_OK(
ctx, GetNodeAttr(kernel->def(), "max_load_factor", &max_load_factor_));
OP_REQUIRES(ctx, max_load_factor_ > 0 && max_load_factor_ < 1,
errors::InvalidArgument(
"max_load_factor must be between 0 and 1, got: ",
max_load_factor_));
OP_REQUIRES_OK(ctx,
GetNodeAttr(kernel->def(), "value_shape", &value_shape_));
OP_REQUIRES(ctx,
TensorShapeUtils::IsScalar(value_shape_) ||
TensorShapeUtils::IsVector(value_shape_),
errors::InvalidArgument(
"Empty value must be a scalar or a vector, got shape ",
value_shape_.DebugString()));
const Tensor* empty_key_input;
OP_REQUIRES_OK(ctx, ctx->input("empty_key", &empty_key_input));
key_shape_ = empty_key_input->shape();
OP_REQUIRES(ctx,
TensorShapeUtils::IsScalar(key_shape_) ||
TensorShapeUtils::IsVector(key_shape_),
errors::InvalidArgument(
"Empty key must be a scalar or a vector, got shape ",
key_shape_.DebugString()));
empty_key_ = *empty_key_input;
empty_key_hash_ = HashKey(
empty_key_input->template shaped<K, 2>({1, key_shape_.num_elements()}),
0);
const Tensor* deleted_key_input;
OP_REQUIRES_OK(ctx, ctx->input("deleted_key", &deleted_key_input));
OP_REQUIRES(ctx, key_shape_.IsSameSize(deleted_key_input->shape()),
errors::InvalidArgument(
"Empty and deleted keys must have same shape, got shapes: ",
key_shape_.DebugString(), " and ",
deleted_key_input->shape().DebugString()));
deleted_key_ = *deleted_key_input;
deleted_key_hash_ = HashKey(deleted_key_input->template shaped<K, 2>(
{1, key_shape_.num_elements()}),
0);
if (empty_key_hash_ == deleted_key_hash_) {
const int64_t key_size = key_shape_.num_elements();
const auto empty_key_matrix =
empty_key_.template shaped<K, 2>({1, key_size});
const auto deleted_key_matrix =
deleted_key_.template shaped<K, 2>({1, key_size});
OP_REQUIRES(
ctx, !IsEqualKey(empty_key_matrix, 0, deleted_key_matrix, 0),
errors::InvalidArgument("Empty and deleted keys cannot be equal"));
}
int64_t initial_num_buckets;
OP_REQUIRES_OK(ctx, GetNodeAttr(kernel->def(), "initial_num_buckets",
&initial_num_buckets));
OP_REQUIRES_OK(ctx, AllocateBuckets(ctx, initial_num_buckets));
}
size_t size() const override TF_LOCKS_EXCLUDED(mu_) {
tf_shared_lock l(mu_);
return num_entries_;
}
Status Find(OpKernelContext* ctx, const Tensor& key, Tensor* value,
const Tensor& default_value) override TF_LOCKS_EXCLUDED(mu_) {
const int64_t num_elements = (key.dims() == 0) ? 1 : key.dim_size(0);
const int64_t key_size = key_shape_.num_elements();
const int64_t value_size = value_shape_.num_elements();
if (key.NumElements() != num_elements * key_size) {
TensorShape expected_shape({num_elements});
expected_shape.AppendShape(key_shape_);
return errors::InvalidArgument("Expected key shape ",
expected_shape.DebugString(), " got ",
key.shape().DebugString());
}
const auto key_matrix = key.shaped<K, 2>({num_elements, key_size});
auto value_matrix = value->shaped<V, 2>({num_elements, value_size});
const auto default_flat = default_value.flat<V>();
tf_shared_lock l(mu_);
const auto key_buckets_matrix = key_buckets_.template matrix<K>();
const auto value_buckets_matrix = value_buckets_.template matrix<V>();
const auto empty_key_matrix =
empty_key_.template shaped<K, 2>({1, key_size});
const auto deleted_key_matrix =
deleted_key_.template shaped<K, 2>({1, key_size});
const int64_t bit_mask = num_buckets_ - 1;
// TODO(andreasst): parallelize using work_sharder
for (int64_t i = 0; i < num_elements; ++i) {
const uint64 key_hash = HashKey(key_matrix, i);
if (empty_key_hash_ == key_hash &&
IsEqualKey(empty_key_matrix, 0, key_matrix, i)) {
return errors::InvalidArgument(
"Using the empty_key as a table key is not allowed");
}
if (deleted_key_hash_ == key_hash &&
IsEqualKey(deleted_key_matrix, 0, key_matrix, i)) {
return errors::InvalidArgument(
"Using the deleted_key as a table key is not allowed");
}
int64_t bucket_index = key_hash & bit_mask;
int64_t num_probes = 0;
while (true) {
if (IsEqualKey(key_buckets_matrix, bucket_index, key_matrix, i)) {
for (int64_t j = 0; j < value_size; ++j) {
// TODO(andreasst): check if we can get rid of SubtleMustCopy
// here and elsewhere in this file.
value_matrix(i, j) =
SubtleMustCopyIfIntegral(value_buckets_matrix(bucket_index, j));
}
break;
}
if (IsEqualKey(key_buckets_matrix, bucket_index, empty_key_matrix, 0)) {
for (int64_t j = 0; j < value_size; ++j) {
value_matrix(i, j) = SubtleMustCopyIfIntegral(default_flat(j));
}
break;
}
++num_probes;
bucket_index =
(bucket_index + num_probes) & bit_mask; // quadratic probing
if (num_probes >= num_buckets_) {
return errors::Internal(
"Internal error in MutableDenseHashTable lookup");
}
}
}
return OkStatus();
}
Status Insert(OpKernelContext* ctx, const Tensor& key,
const Tensor& value) override TF_LOCKS_EXCLUDED(mu_) {
const int64_t batch_size = (key.dims() == 0) ? 1 : key.dim_size(0);
if (key.NumElements() != batch_size * key_shape_.num_elements()) {
TensorShape expected_shape({batch_size});
expected_shape.AppendShape(key_shape_);
return errors::InvalidArgument("Expected key shape ",
expected_shape.DebugString(), " got ",
key.shape().DebugString());
}
mutex_lock l(mu_);
// For simplicity we assume that all keys in the input result in inserts
// rather than updates. That means we may grow the table even though we
// don't need to. As long as the number of keys inserted in one call is
// small compared to the size of the map, the impact of this is minimal.
const int64_t pending_num_entries = num_entries_ + batch_size;
if (pending_num_entries > num_buckets_ * max_load_factor_) {
int64_t new_num_buckets = num_buckets_;
do {
new_num_buckets <<= 1;
} while (pending_num_entries > new_num_buckets * max_load_factor_);
TF_RETURN_IF_ERROR(Rebucket(ctx, new_num_buckets));
}
return DoInsert(ctx, key, value, false);
}
Status Remove(OpKernelContext* ctx, const Tensor& key) override
TF_LOCKS_EXCLUDED(mu_) {
if (key.NumElements() != key.dim_size(0) * key_shape_.num_elements()) {
TensorShape expected_shape({key.dim_size(0)});
expected_shape.AppendShape(key_shape_);
return errors::InvalidArgument("Expected key shape ",
expected_shape.DebugString(), " got ",
key.shape().DebugString());
}
mutex_lock l(mu_);
return DoRemove(ctx, key);
}
Status ImportValues(OpKernelContext* ctx, const Tensor& keys,
const Tensor& values) override TF_LOCKS_EXCLUDED(mu_) {
mutex_lock l(mu_);
num_buckets_ = keys.dim_size(0);
key_buckets_ = keys;
value_buckets_ = values;
// Count the number of keys that are not the empty_key or deleted_key.
// This requires iterating through the whole table but that is OK as we
// only execute it during checkpoint restore.
num_entries_ = 0;
const auto empty_key_tensor =
empty_key_.template shaped<K, 2>({1, key_shape_.num_elements()});
const auto deleted_key_tensor =
deleted_key_.template shaped<K, 2>({1, key_shape_.num_elements()});
const auto key_buckets_tensor = key_buckets_.template matrix<K>();
for (int64_t i = 0; i < num_buckets_; ++i) {
if (!IsEqualKey(key_buckets_tensor, i, empty_key_tensor, 0) &&
!IsEqualKey(key_buckets_tensor, i, deleted_key_tensor, 0)) {
++num_entries_;
}
}
return OkStatus();
}
Status ExportValues(OpKernelContext* ctx) override TF_LOCKS_EXCLUDED(mu_) {
tf_shared_lock l(mu_);
TF_RETURN_IF_ERROR(ctx->set_output("keys", key_buckets_));
TF_RETURN_IF_ERROR(ctx->set_output("values", value_buckets_));
return OkStatus();
}
Status CheckKeyAndValueTensorsForImport(const Tensor& keys,
const Tensor& values) override {
TF_RETURN_IF_ERROR(CheckKeyAndValueTypes(keys, values));
TF_RETURN_IF_ERROR(CheckKeyShape(keys.shape()));
// The storage format in key_buckets_ and value_buckets_ is always vectors,
// even if the inputs are scalars. This is what eventually gets exported
// and is expected by the import method as well.
TensorShape key_shape = MaybeVectorizeShape(key_shape_);
TensorShape value_shape = MaybeVectorizeShape(value_shape_);
// Compute the final expected shape of the value by starting with the shape
// of all keys, removing the dimensions particular to each key and then
// appending the shape of a single value.
TensorShape expected_value_shape = keys.shape();
expected_value_shape.RemoveLastDims(key_shape.dims());
expected_value_shape.AppendShape(value_shape);
if (values.shape() != expected_value_shape) {
return errors::InvalidArgument(
"Expected shape ", expected_value_shape.DebugString(),
" for value, got ", values.shape().DebugString());
}
return OkStatus();
}
DataType key_dtype() const override { return DataTypeToEnum<K>::v(); }
DataType value_dtype() const override { return DataTypeToEnum<V>::v(); }
TensorShape key_shape() const override { return key_shape_; }
TensorShape value_shape() const override { return value_shape_; }
int64_t MemoryUsed() const override TF_LOCKS_EXCLUDED(mu_) {
tf_shared_lock l(mu_);
return sizeof(MutableDenseHashTable) + key_buckets_.AllocatedBytes() +
value_buckets_.AllocatedBytes() + empty_key_.AllocatedBytes();
}
private:
Status DoInsert(OpKernelContext* ctx, const Tensor& key, const Tensor& value,
bool ignore_empty_and_deleted_key)
TF_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
const int64_t num_elements = (key.dims() == 0) ? 1 : key.dim_size(0);
const int64_t value_size = value_shape_.num_elements();
const int64_t key_size = key_shape_.num_elements();
const auto key_matrix = key.shaped<K, 2>({num_elements, key_size});
auto value_matrix = value.shaped<V, 2>({num_elements, value_size});
auto key_buckets_matrix = key_buckets_.template matrix<K>();
auto value_buckets_matrix = value_buckets_.template matrix<V>();
const auto empty_key_tensor =
empty_key_.template shaped<K, 2>({1, key_size});
const auto deleted_key_tensor =
deleted_key_.template shaped<K, 2>({1, key_size});
const int64_t bit_mask = num_buckets_ - 1;
for (int64_t i = 0; i < num_elements; ++i) {
const uint64 key_hash = HashKey(key_matrix, i);
if (empty_key_hash_ == key_hash &&
IsEqualKey(empty_key_tensor, 0, key_matrix, i)) {
if (ignore_empty_and_deleted_key) {
continue;
}
return errors::InvalidArgument(
"Using the empty_key as a table key is not allowed");
}
if (deleted_key_hash_ == key_hash &&
IsEqualKey(deleted_key_tensor, 0, key_matrix, i)) {
if (ignore_empty_and_deleted_key) {
continue;
}
return errors::InvalidArgument(
"Using the deleted_key as a table key is not allowed");
}
int64_t bucket_index = key_hash & bit_mask;
int64_t num_probes = 0;
while (true) {
if (IsEqualKey(key_buckets_matrix, bucket_index, key_matrix, i)) {
for (int64_t j = 0; j < value_size; ++j) {
value_buckets_matrix(bucket_index, j) =
SubtleMustCopyIfIntegral(value_matrix(i, j));
}
break;
}
if (IsEqualKey(key_buckets_matrix, bucket_index, empty_key_tensor, 0) ||
IsEqualKey(key_buckets_matrix, bucket_index, deleted_key_tensor,
0)) {
++num_entries_;
for (int64_t j = 0; j < key_size; ++j) {
key_buckets_matrix(bucket_index, j) =
SubtleMustCopyIfIntegral(key_matrix(i, j));
}
for (int64_t j = 0; j < value_size; ++j) {
value_buckets_matrix(bucket_index, j) =
SubtleMustCopyIfIntegral(value_matrix(i, j));
}
break;
}
++num_probes;
bucket_index =
(bucket_index + num_probes) & bit_mask; // quadratic probing
if (num_probes >= num_buckets_) {
return errors::Internal(
"Internal error in MutableDenseHashTable insert");
}
}
}
return OkStatus();
}
Status DoRemove(OpKernelContext* ctx, const Tensor& key)
TF_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
const int64_t num_elements = key.dim_size(0);
const int64_t key_size = key_shape_.num_elements();
const auto key_matrix = key.shaped<K, 2>({num_elements, key_size});
auto key_buckets_matrix = key_buckets_.template matrix<K>();
const auto empty_key_tensor =
empty_key_.template shaped<K, 2>({1, key_size});
const auto deleted_key_tensor =
deleted_key_.template shaped<K, 2>({1, key_size});
const auto deleted_key_flat = deleted_key_.template flat<K>();
const int64_t bit_mask = num_buckets_ - 1;
for (int64_t i = 0; i < num_elements; ++i) {
const uint64 key_hash = HashKey(key_matrix, i);
if (empty_key_hash_ == key_hash &&
IsEqualKey(empty_key_tensor, 0, key_matrix, i)) {
return errors::InvalidArgument(
"Using the empty_key as a table key is not allowed");
}
if (deleted_key_hash_ == key_hash &&
IsEqualKey(deleted_key_tensor, 0, key_matrix, i)) {
return errors::InvalidArgument(
"Using the deleted_key as a table key is not allowed");
}
int64_t bucket_index = key_hash & bit_mask;
int64_t num_probes = 0;
while (true) {
if (IsEqualKey(key_buckets_matrix, bucket_index, key_matrix, i)) {
--num_entries_;
for (int64_t j = 0; j < key_size; ++j) {
key_buckets_matrix(bucket_index, j) =
SubtleMustCopyIfIntegral(deleted_key_flat(j));
}
break;
}
if (IsEqualKey(key_buckets_matrix, bucket_index, empty_key_tensor, 0)) {
break;
}
++num_probes;
bucket_index =
(bucket_index + num_probes) & bit_mask; // quadratic probing
if (num_probes >= num_buckets_) {
return errors::Internal(
"Internal error in MutableDenseHashTable remove");
}
}
}
return OkStatus();
}
Status AllocateBuckets(OpKernelContext* ctx, int64_t new_num_buckets)
TF_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
if (new_num_buckets < 4 ||
((new_num_buckets & (new_num_buckets - 1)) != 0)) {
return errors::InvalidArgument(
"Number of buckets must be at least 4 and a power of 2, got: ",
new_num_buckets);
}
num_buckets_ = new_num_buckets;
num_entries_ = 0;
const int64_t key_size = key_shape_.num_elements();
TF_RETURN_IF_ERROR(ctx->allocate_temp(
key_dtype(), TensorShape({num_buckets_, key_size}), &key_buckets_));
auto key_buckets_matrix = key_buckets_.matrix<K>();
const auto empty_key_flat = empty_key_.template flat<K>();
for (int64_t i = 0; i < num_buckets_; ++i) {
for (int64_t j = 0; j < key_size; ++j) {
key_buckets_matrix(i, j) = empty_key_flat(j);
}
}
const int64_t value_size = value_shape_.num_elements();
TF_RETURN_IF_ERROR(ctx->allocate_temp(
value_dtype(), TensorShape({num_buckets_, value_size}),
&value_buckets_));
auto value_buckets_matrix = value_buckets_.matrix<V>();
for (int64_t i = 0; i < num_buckets_; ++i) {
for (int64_t j = 0; j < value_size; ++j) {
// Initialize values to the default value for the type to avoid
// exposing uninitialized memory in ExportValues().
value_buckets_matrix(i, j) = V();
}
}
return OkStatus();
}
Status Rebucket(OpKernelContext* ctx, int64_t num_new_buckets)
TF_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
Tensor old_key_buckets = key_buckets_;
Tensor old_value_buckets = value_buckets_;
TF_RETURN_IF_ERROR(AllocateBuckets(ctx, num_new_buckets));
return DoInsert(ctx, old_key_buckets, old_value_buckets, true);
}
uint64 HashKey(typename TTypes<K>::ConstMatrix key, int64_t index) const {
if (key_shape_.num_elements() == 1) {
return HashScalar(key(index, 0));
}
uint64 result = 0;
for (int64_t i = 0; i < key_shape_.num_elements(); ++i) {
result = Hash64Combine(result, HashScalar(key(index, i)));
}
return result;
}
// Use a template to allow this function to be used both with Matrix and
// ConstMatrix types.
template <typename MT2>
bool IsEqualKey(typename TTypes<K>::Matrix tensor1, int64_t index1,
MT2 tensor2, int64_t index2) const {
for (int64_t i = 0; i < key_shape_.num_elements(); ++i) {
if (tensor1(index1, i) != tensor2(index2, i)) {
return false;
}
}
return true;
}
TensorShape key_shape_;
TensorShape value_shape_;
float max_load_factor_;
mutable mutex mu_;
int64_t num_entries_ TF_GUARDED_BY(mu_);
int64_t num_buckets_ TF_GUARDED_BY(mu_);
Tensor key_buckets_ TF_GUARDED_BY(mu_);
Tensor value_buckets_ TF_GUARDED_BY(mu_);
Tensor empty_key_;
uint64 empty_key_hash_;
Tensor deleted_key_;
uint64 deleted_key_hash_;
};
} // namespace lookup
// Base class for kernels that take a LookupTable handle as the 0th input.
class LookupTableOpKernel : public OpKernel {
public:
explicit LookupTableOpKernel(OpKernelConstruction* ctx)
: OpKernel(ctx),
expected_input_0_(ctx->input_type(0) == DT_RESOURCE ? DT_RESOURCE
: DT_STRING_REF) {}
protected:
Status GetTable(OpKernelContext* ctx, lookup::LookupInterface** table) {
if (expected_input_0_ == DT_RESOURCE) {
return GetResourceLookupTable("table_handle", ctx, table);
} else {
return GetReferenceLookupTable("table_handle", ctx, table);
}
}
// Input 0 could be a STRING_REF or a RESOURCE
const DataType expected_input_0_;
};
// Table lookup op. Perform the lookup operation on the given table.
class LookupTableFindOp : public LookupTableOpKernel {
public:
using LookupTableOpKernel::LookupTableOpKernel;
void Compute(OpKernelContext* ctx) override {
lookup::LookupInterface* table;
OP_REQUIRES_OK(ctx, GetTable(ctx, &table));
core::ScopedUnref unref_me(table);
DataTypeVector expected_inputs = {expected_input_0_, table->key_dtype(),
table->value_dtype()};
DataTypeVector expected_outputs = {table->value_dtype()};
OP_REQUIRES_OK(ctx, ctx->MatchSignature(expected_inputs, expected_outputs));
const Tensor& key = ctx->input(1);
const Tensor& default_value = ctx->input(2);
OP_REQUIRES_OK(ctx, table->CheckFindArguments(key, default_value));
TensorShape output_shape = key.shape();
output_shape.RemoveLastDims(table->key_shape().dims());
output_shape.AppendShape(table->value_shape());
Tensor* out;
OP_REQUIRES_OK(ctx, ctx->allocate_output("values", output_shape, &out));
OP_REQUIRES_OK(ctx, table->Find(ctx, key, out, default_value));
}
};
REGISTER_KERNEL_BUILDER(Name("LookupTableFind").Device(DEVICE_CPU),
LookupTableFindOp);
REGISTER_KERNEL_BUILDER(Name("LookupTableFindV2").Device(DEVICE_CPU),
LookupTableFindOp);
// Table insert op.
class LookupTableInsertOp : public LookupTableOpKernel {
public:
using LookupTableOpKernel::LookupTableOpKernel;
void Compute(OpKernelContext* ctx) override {
lookup::LookupInterface* table;
OP_REQUIRES_OK(ctx, GetTable(ctx, &table));
core::ScopedUnref unref_me(table);
DataTypeVector expected_inputs = {expected_input_0_, table->key_dtype(),
table->value_dtype()};
OP_REQUIRES_OK(ctx, ctx->MatchSignature(expected_inputs, {}));
const Tensor& keys = ctx->input(1);
const Tensor& values = ctx->input(2);
OP_REQUIRES_OK(ctx, table->CheckKeyAndValueTensorsForInsert(keys, values));
int64_t memory_used_before = 0;
if (ctx->track_allocations()) {
memory_used_before = table->MemoryUsed();
}
OP_REQUIRES_OK(ctx, table->Insert(ctx, keys, values));
if (ctx->track_allocations()) {
ctx->record_persistent_memory_allocation(table->MemoryUsed() -
memory_used_before);
}
}
};
REGISTER_KERNEL_BUILDER(Name("LookupTableInsert").Device(DEVICE_CPU),
LookupTableInsertOp);
REGISTER_KERNEL_BUILDER(Name("LookupTableInsertV2").Device(DEVICE_CPU),
LookupTableInsertOp);
// Table remove op.
class LookupTableRemoveOp : public LookupTableOpKernel {
public:
using LookupTableOpKernel::LookupTableOpKernel;
void Compute(OpKernelContext* ctx) override {
lookup::LookupInterface* table;
OP_REQUIRES_OK(ctx, GetTable(ctx, &table));
core::ScopedUnref unref_me(table);
DataTypeVector expected_inputs = {expected_input_0_, table->key_dtype()};
OP_REQUIRES_OK(ctx, ctx->MatchSignature(expected_inputs, {}));
const Tensor& key = ctx->input(1);
OP_REQUIRES_OK(ctx, table->CheckKeyTensorForRemove(key));
int64_t memory_used_before = 0;
if (ctx->track_allocations()) {
memory_used_before = table->MemoryUsed();
}
OP_REQUIRES_OK(ctx, table->Remove(ctx, key));
if (ctx->track_allocations()) {
ctx->record_persistent_memory_allocation(table->MemoryUsed() -
memory_used_before);
}
}
};