forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathonnxifi_transformer.cc
1318 lines (1209 loc) · 45.1 KB
/
onnxifi_transformer.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
#include "caffe2/opt/onnxifi_transformer.h"
#include <iostream>
#include <unordered_set>
#include "onnx/proto_utils.h"
#include "caffe2/core/context.h"
#include "caffe2/core/logging.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/tensor.h"
#include "caffe2/onnx/onnx_exporter.h"
#include "caffe2/opt/backend_cutting.h"
#include "caffe2/utils/proto_utils.h"
#include "caffe2/utils/string_utils.h"
namespace caffe2 {
namespace {
const std::string kRealBatchSizeBlob = "real_batch_size";
const std::string kInitializers = "initializers";
constexpr size_t kBufferSize = 64;
// Convert ShapeInfo map to TensorShape map
std::unordered_map<std::string, TensorShape> stripShapeInfoMap(
const ShapeInfoMap& info_map) {
std::unordered_map<std::string, TensorShape> shape_map;
for (const auto& kv : info_map) {
shape_map.emplace(kv.first, kv.second.shape);
}
return shape_map;
}
uint64_t onnxifiDataType(caffe2::TensorProto::DataType t) {
#define CAFFE2_TO_ONNXIFI_TYPE(x, y) \
case (caffe2::TensorProto::x): \
return y
switch (t) {
CAFFE2_TO_ONNXIFI_TYPE(FLOAT, ONNXIFI_DATATYPE_FLOAT32);
CAFFE2_TO_ONNXIFI_TYPE(INT8, ONNXIFI_DATATYPE_INT8);
CAFFE2_TO_ONNXIFI_TYPE(UINT8, ONNXIFI_DATATYPE_UINT8);
CAFFE2_TO_ONNXIFI_TYPE(INT16, ONNXIFI_DATATYPE_INT16);
CAFFE2_TO_ONNXIFI_TYPE(UINT16, ONNXIFI_DATATYPE_UINT16);
CAFFE2_TO_ONNXIFI_TYPE(INT32, ONNXIFI_DATATYPE_INT32);
CAFFE2_TO_ONNXIFI_TYPE(INT64, ONNXIFI_DATATYPE_INT64);
CAFFE2_TO_ONNXIFI_TYPE(FLOAT16, ONNXIFI_DATATYPE_FLOAT16);
default:
LOG(WARNING) << "Unsupported Caffe2 tensor type: " << t
<< ", fallback to FLOAT";
return ONNXIFI_DATATYPE_FLOAT32;
}
#undef CAFFE2_TO_ONNXIFI_TYPE
}
std::vector<::ONNX_NAMESPACE::ValueInfoProto> convertToValueInfo(
const std::vector<std::string>& names,
const std::unordered_map<std::string, TensorShape>& shape_hints,
const std::unordered_map<std::string, ::ONNX_NAMESPACE::TypeProto>&
extra_shape_hints) {
std::vector<::ONNX_NAMESPACE::ValueInfoProto> r;
for (const auto& s : names) {
r.emplace_back();
auto& value_info = r.back();
value_info.set_name(s);
const auto it = shape_hints.find(s);
if (it == shape_hints.end()) {
const auto eit = extra_shape_hints.find(s);
if (eit == extra_shape_hints.end()) {
LOG(WARNING) << "Cannot get shape of " << s;
} else {
value_info.mutable_type()->CopyFrom(eit->second);
}
} else {
auto* tensor_type = value_info.mutable_type()->mutable_tensor_type();
tensor_type->set_elem_type(
onnx::Caffe2TypeToOnnxType(it->second.data_type()));
auto* shape = tensor_type->mutable_shape();
for (int i = 0; i < it->second.dims().size(); ++i) {
shape->add_dim()->set_dim_value(it->second.dims(i));
}
}
}
return r;
}
// Given a net, with primiary inputs and outputs defined in its
// external_inputs/outputs, and given the set of weights and extra weights
// (created during conversion to ONNX if exists), we check whether some of the
// weights are used in the net, and if so, we put it in the initialize_list and
// add it to the external_inputs too.
// \param net [in] c2 net (cutoff from a bigger net)
// \param weights_in_ws [in] all the weights in the workspace
// \param extra_weights [in] extra weights possibly generated during ONNX
// conversion \param initialization_list [out] weights that needs to be offload
// to backend \param total_inputs_vec [out] total #inputs of the net that
// doesn't have a producer
void getWeightsAndInputs(
const NetDef& net,
const std::unordered_set<std::string>& weights_in_ws,
const std::vector<std::string>& extra_weights,
std::unordered_set<std::string>* initialization_list,
std::vector<std::string>* total_inputs_vec) {
std::unordered_set<std::string> total_inputs;
// extra weights is definitely extra weights/inputs
for (const auto& extra_weight : extra_weights) {
if (total_inputs.emplace(extra_weight).second) {
total_inputs_vec->emplace_back(extra_weight);
}
initialization_list->emplace(extra_weight);
}
// Boundary inputs that should not be weights
std::unordered_set<std::string> boundary_inputs;
for (const auto& i : net.external_input()) {
boundary_inputs.emplace(i);
}
for (const auto& op : net.op()) {
for (const auto& input : op.input()) {
bool not_seen = total_inputs.emplace(input).second;
if (!not_seen) {
continue;
}
if (weights_in_ws.count(input)) {
// We add weights as inputs too
total_inputs_vec->emplace_back(input);
initialization_list->emplace(input);
VLOG(2) << "Add weights: " << input;
} else if (boundary_inputs.count(input)) {
VLOG(2) << "Adding boundary input: " << input;
total_inputs_vec->emplace_back(input);
}
}
}
}
void collectInputsAndOutputs(
const OperatorDef& op,
std::set<std::string>* inputs,
std::set<std::string>* outputs) {
for (const auto& blob : op.input()) {
inputs->emplace(blob);
}
for (const auto& blob : op.output()) {
outputs->emplace(blob);
}
}
void fetchInputsToIfOpsSubnet(NetDef* net) {
NetDef clone(*net);
clone.clear_op();
for (auto& op : net->op()) {
if (op.type() == "If" || op.type() == "AsyncIf") {
OperatorDef new_op(op);
ArgumentHelper helper(op);
std::set<std::string> subnet_inputs, subnet_outputs;
if (helper.HasSingleArgumentOfType<NetDef>("then_net")) {
auto then_net = helper.GetSingleArgument<NetDef>("then_net", NetDef());
for (const auto& nested_op : then_net.op()) {
collectInputsAndOutputs(nested_op, &subnet_inputs, &subnet_outputs);
}
}
if (helper.HasSingleArgumentOfType<NetDef>("else_net")) {
auto else_net = helper.GetSingleArgument<NetDef>("else_net", NetDef());
for (const auto& nested_op : else_net.op()) {
collectInputsAndOutputs(nested_op, &subnet_inputs, &subnet_outputs);
}
}
for (const std::string& blob : subnet_inputs) {
if (subnet_outputs.count(blob) == 0) {
new_op.add_input(blob);
}
}
clone.add_op()->CopyFrom(new_op);
} else {
clone.add_op()->CopyFrom(op);
}
}
net->Swap(&clone);
}
void fillModelInfo(::ONNX_NAMESPACE::ModelProto* model) {
model->set_ir_version(::ONNX_NAMESPACE::Version::IR_VERSION);
model->set_producer_name("caffe2");
auto* opset_id = model->add_opset_import();
opset_id->set_domain("");
opset_id->set_version(7);
}
int64_t getBlob1stDimSize(const ShapeInfo& shape_info) {
if (shape_info.shape.dims_size() == 0) {
return 0;
} else {
return shape_info.shape.dims(0);
}
}
NetDef composeResultNet(const OperatorDef& onnxifi_op) {
NetDef net_opt;
net_opt.add_op()->CopyFrom(onnxifi_op);
return net_opt;
}
void enforceFp32InputsToFp16(
const std::unordered_set<std::string>& weights,
NetDef* pred_net,
ShapeInfoMap* shape_hints) {
std::unordered_map<std::string, ShapeInfo> user_input_map;
for (const auto& i : pred_net->external_input()) {
if (weights.count(i)) {
continue;
}
auto it = shape_hints->find(i);
if (it == shape_hints->end() ||
it->second.shape.data_type() != TensorProto_DataType_FLOAT) {
continue;
}
auto& shape_info = it->second;
user_input_map[i] = shape_info;
shape_info.shape.set_data_type(TensorProto_DataType_FLOAT16);
}
if (user_input_map.empty()) {
return;
}
std::vector<OperatorDef> ops;
for (const auto& op : pred_net->op()) {
ops.emplace_back(op);
}
pred_net->clear_op();
int current_pos = ops.size();
const char kBridgeTensorSuffix[] = "_to_float_bridge";
std::vector<OperatorDef> converts;
for (const auto& elem : user_input_map) {
const auto& name = elem.first;
const auto& shape_info = elem.second;
std::string new_name = name + kBridgeTensorSuffix;
shape_hints->emplace(new_name, shape_info);
converts.emplace_back(CreateOperatorDef(
"HalfToFloat",
"",
{name},
{new_name},
{MakeArgument<int>(kNetPos, current_pos++)}));
}
for (const auto& op : converts) {
pred_net->add_op()->CopyFrom(op);
}
for (auto& op : ops) {
for (auto& input : *op.mutable_input()) {
if (user_input_map.count(input)) {
input += kBridgeTensorSuffix;
}
}
}
for (const auto& op : ops) {
pred_net->add_op()->CopyFrom(op);
}
}
void mergeFp32InputsAndConvertToFp16(
size_t batch_size,
const std::unordered_set<std::string>& weights,
NetDef* pred_net,
ShapeInfoMap* shape_hints) {
std::unordered_map<std::string, ShapeInfo> user_input_map;
for (const auto& i : pred_net->external_input()) {
if (weights.count(i)) {
continue;
}
const auto it = shape_hints->find(i);
// Heuristic: the input has to be of float type, 2-dimensional and the first
// dimension has to be of batch size
if (it == shape_hints->end() ||
it->second.shape.data_type() != TensorProto_DataType_FLOAT) {
continue;
}
auto shape_info = it->second;
if (shape_info.shape.dims_size() != 2 ||
shape_info.shape.dims(0) != batch_size) {
continue;
}
shape_info.shape.set_data_type(TensorProto_DataType_FLOAT16);
user_input_map[i] = shape_info;
}
if (user_input_map.empty()) {
return;
}
std::unordered_map<std::string, std::vector<std::string>>
user_inputs_by_partition;
std::unordered_map<std::string, std::unordered_set<std::string>>
user_input_set_by_partition;
for (const auto& op : pred_net->op()) {
for (const auto& i : op.input()) {
if (user_input_map.find(i) != user_input_map.end()) {
const auto& partition = op.device_option().node_name().empty()
? "default"
: op.device_option().node_name();
if (user_input_set_by_partition[partition].find(i) ==
user_input_set_by_partition[partition].end()) {
user_inputs_by_partition[partition].emplace_back(i);
user_input_set_by_partition[partition].insert(i);
}
}
}
}
std::vector<OperatorDef> ops;
for (const auto& op : pred_net->op()) {
ops.emplace_back(op);
}
pred_net->clear_op();
int current_pos = ops.size();
for (const auto& elem : user_inputs_by_partition) {
const auto& partition = elem.first;
const auto& user_inputs = elem.second;
const auto& user_input_set = user_input_set_by_partition[partition];
OperatorDef op1;
op1.set_type("Concat");
for (const auto& i : user_inputs) {
op1.add_input(i);
}
op1.add_output(partition + "_fp32_input_concated");
op1.add_output(partition + "_fp32_input_concated_split_info");
auto shape_info = user_input_map[user_inputs.front()];
int total = 0;
for (const auto& u : user_inputs) {
total += user_input_map[u].shape.dims(1);
}
shape_info.shape.set_dims(1, total);
AddArgument("axis", 1, &op1);
AddArgument(kNetPos, current_pos++, &op1);
pred_net->add_op()->CopyFrom(op1);
// TODO: a possible optimization is to fuse the fp16 conversion into Concat
OperatorDef op2;
op2.set_type("FloatToHalf");
op2.add_input(partition + "_fp32_input_concated");
op2.add_output(partition + "_fp16_input_concated");
AddArgument("clip", 1, &op2);
AddArgument(kNetPos, current_pos++, &op2);
shape_hints->emplace(partition + "_fp16_input_concated", shape_info);
pred_net->add_op()->CopyFrom(op2);
OperatorDef op3;
op3.set_type("Split");
op3.add_input(partition + "_fp16_input_concated");
op3.mutable_device_option()->set_node_name(partition);
std::vector<OperatorDef> converts;
for (const auto& i : user_inputs) {
std::string new_name = partition + "_" + i + "_split_fp16";
op3.add_output(new_name);
shape_hints->emplace(new_name, user_input_map[i]);
converts.emplace_back(CreateOperatorDef(
"HalfToFloat",
"",
{partition + "_" + i + "_split_fp16"},
{partition + "_" + i + "_split"},
{MakeArgument<int>(kNetPos, current_pos++)}));
converts.back().mutable_device_option()->set_node_name(partition);
auto converted_shape = user_input_map[i];
converted_shape.shape.set_data_type(TensorProto_DataType_FLOAT);
shape_hints->emplace(partition + "_" + i + "_split", converted_shape);
}
AddArgument("axis", 1, &op3);
AddArgument(kNetPos, current_pos++, &op3);
auto* arg = op3.add_arg();
arg->set_name("split");
for (const auto& u : user_inputs) {
arg->add_ints(user_input_map[u].shape.dims(1));
}
pred_net->add_op()->CopyFrom(op3);
for (const auto& op : converts) {
pred_net->add_op()->CopyFrom(op);
}
for (auto& op : ops) {
if ((!op.device_option().node_name().empty() &&
op.device_option().node_name() == partition) ||
(op.device_option().node_name().empty() && partition == "default")) {
for (auto& i : *op.mutable_input()) {
if (user_input_set.count(i)) {
i = partition + "_" + i + "_split";
}
}
}
}
}
for (const auto& op : ops) {
pred_net->add_op()->CopyFrom(op);
}
}
} // namespace
void splitSparseLengthsSumSparse(NetDef* net, const Workspace& ws) {
const static std::unordered_map<string, string> slss = {
{"SparseLengthsSum4BitRowwiseSparse", "SparseLengthsSumFused4BitRowwise"},
{"SparseLengthsWeightedSum4BitRowwiseSparse",
"SparseLengthsWeightedSumFused4BitRowwise"},
{"SparseLengthsSum8BitRowwiseSparse", "SparseLengthsSumFused8BitRowwise"},
{"SparseLengthsWeightedSum8BitRowwiseSparse",
"SparseLengthsWeightedSumFused8BitRowwise"},
{"SparseLengthsSum2BitRowwiseSparse", "SparseLengthsSumFused2BitRowwise"},
{"SparseLengthsWeightedSum2BitRowwiseSparse",
"SparseLengthsWeightedSumFused2BitRowwise"}};
NetDef new_net;
new_net.CopyFrom(*net);
new_net.mutable_op()->Clear();
for (const auto& op : net->op()) {
const auto it = slss.find(op.type());
if (it == slss.end()) {
new_net.add_op()->CopyFrom(op);
} else {
const bool is_weighted =
(op.type().find("Weighted") != std::string::npos);
const auto& compressed_mapping = op.input(is_weighted ? 4 : 3);
const auto* b = ws.GetBlob(compressed_mapping);
bool fallback = false;
if (b && b->IsType<Tensor>()) {
const auto& t = BlobGetTensor(*b, CPU);
fallback = ((t.numel() == 1) && (t.template data<int32_t>()[0] == 0));
}
if (fallback) {
// If fallback, we just replace the original slss op with a normal sls
// op
OperatorDef new_op;
new_op.CopyFrom(op);
new_op.set_type(it->second);
new_op.mutable_input()->RemoveLast();
new_net.add_op()->CopyFrom(new_op);
} else {
// Otherwise, we replace slss with slss_lookup followed by a normal sls
OperatorDef new_op;
new_op.CopyFrom(op);
new_op.set_type("SparseLengthsSumSparseLookup");
new_op.clear_input();
const auto& indices_in = is_weighted ? op.input(2) : op.input(1);
const auto& lengths_in = is_weighted ? op.input(3) : op.input(2);
const auto& compress_mapping = is_weighted ? op.input(4) : op.input(3);
const auto& weights_in = is_weighted ? op.input(1) : "";
new_op.add_input(indices_in);
new_op.add_input(lengths_in);
new_op.add_input(compress_mapping);
const auto indices_out = indices_in + "_decomp";
const auto lengths_out = lengths_in + "_decomp";
const auto weights_out = weights_in + "_decomp";
new_op.clear_output();
new_op.add_output(indices_out);
new_op.add_output(lengths_out);
if (is_weighted) {
new_op.add_input(weights_in);
new_op.add_output(weights_out);
}
new_net.add_op()->CopyFrom(new_op);
new_op.CopyFrom(op);
new_op.set_type(it->second);
new_op.mutable_input()->RemoveLast();
*new_op.mutable_input()->Mutable(is_weighted ? 2 : 1) = indices_out;
*new_op.mutable_input()->Mutable(is_weighted ? 3 : 2) = lengths_out;
if (is_weighted) {
*new_op.mutable_input()->Mutable(1) = weights_out;
}
new_net.add_op()->CopyFrom(new_op);
}
}
}
new_net.Swap(net);
}
OnnxifiTransformer::OnnxifiTransformer(const OnnxifiTransformerOptions& opts)
: BackendTransformerBase(), opts_(opts) {
lib_ = onnx::initOnnxifiLibrary();
CAFFE_ENFORCE(lib_, "Cannot initialize ONNXIFI library");
CAFFE_ENFORCE_EQ(
lib_->onnxGetBackendIDs(nullptr, &num_backends_),
ONNXIFI_STATUS_FALLBACK);
CAFFE_ENFORCE_GT(
num_backends_, 0, "At least 1 onnxifi backend should be available");
backend_ids_.resize(num_backends_);
CAFFE_ENFORCE_EQ(
lib_->onnxGetBackendIDs(backend_ids_.data(), &num_backends_),
ONNXIFI_STATUS_SUCCESS);
}
OnnxifiTransformer::~OnnxifiTransformer() {
for (unsigned i = 0; i < num_backends_; ++i) {
if (lib_->onnxReleaseBackendID(backend_ids_[i]) != ONNXIFI_STATUS_SUCCESS) {
LOG(ERROR) << "Error when calling onnxReleaseBackendID";
}
}
}
OperatorDef OnnxifiTransformer::buildOnnxifiOp(
const std::string& onnx_model_str,
const std::unordered_set<std::string>& initialization_list,
const std::vector<std::string>& external_inputs,
const std::vector<std::string>& external_outputs,
const std::unordered_map<std::string, ShapeInfo>& shape_hints) {
OperatorDef op;
op.set_type("Onnxifi");
auto* onnx_model_arg = op.add_arg();
onnx_model_arg->set_name("onnx_model");
onnx_model_arg->set_s(onnx_model_str);
// Add the names of the initializer blobs that we want to fetch from the
// workspace later
auto* initializers_arg = op.add_arg();
initializers_arg->set_name(kInitializers);
for (const auto& s : initialization_list) {
initializers_arg->add_strings(s);
}
// Add the input/output
int idx = 0;
auto* input_names = op.add_arg();
input_names->set_name("input_names");
for (const auto& input : external_inputs) {
if (!initialization_list.count(input)) {
op.add_input(input);
input_names->add_strings(input);
}
}
auto* output_names = op.add_arg();
output_names->set_name("output_names");
for (const auto& output : external_outputs) {
op.add_output(output);
output_names->add_strings(output);
}
// Find out the index of input that has a nominal batch size
const auto max_batch_size = opts_.bound_shape_spec.max_batch_size;
idx = 0;
int nominal_batch_idx{0};
for (const auto& input : external_inputs) {
if (!initialization_list.count(input)) {
const auto it = shape_hints.find(input);
CAFFE_ENFORCE(
it != shape_hints.end(), "Input shape for ", input, " not found");
const auto& info = it->second;
if (info.getDimType(0) == TensorBoundShape_DimType_BATCH &&
getBlob1stDimSize(info) == max_batch_size) {
nominal_batch_idx = idx;
break;
}
++idx;
}
}
// Add output size hints
auto* output_shape_info_arg = op.add_arg();
output_shape_info_arg->set_name("output_shape_info");
auto* output_qshape_info_arg = op.add_arg();
output_qshape_info_arg->set_name("output_qshape_info");
for (int i = 0; i < op.output_size(); ++i) {
const auto& o = op.output(i);
const auto it = shape_hints.find(o);
if (it != shape_hints.end()) {
if (!it->second.is_quantized) {
output_shape_info_arg->mutable_tensors()->Add()->CopyFrom(
wrapShapeInfoIntoTensorProto(o, it->second));
} else {
output_qshape_info_arg->mutable_qtensors()->Add()->CopyFrom(
wrapShapeInfoIntoQTensorProto(o, it->second));
}
VLOG(2) << "Adding output hint: " << o;
}
}
// Tell Onnxifi op that the model is in onnx or c2 proto format
AddArgument("use_onnx", opts_.use_onnx ? 1 : 0, &op);
// Tell Onnxifi op which backend id to use
AddArgument("backend_id", idx_, &op);
// Add model_id and net_pos to the onnxifi model
AddArgument(kModelId, model_id_, &op);
AddArgument(kNetPos, c10::to_string(onnxifi_op_id_++), &op);
// Add output resizing hints
if (opts_.adjust_batch) {
AddArgument("adjust_output_batch", 1, &op);
} else {
AddArgument("adjust_output_batch", 0, &op);
}
AddArgument("max_batch_size", opts_.bound_shape_spec.max_batch_size, &op);
AddArgument("max_seq_size", opts_.bound_shape_spec.max_seq_size, &op);
AddArgument("timeout", opts_.timeout, &op);
AddArgument("nominal_batch_idx", nominal_batch_idx, &op);
return op;
}
NetDef OnnxifiTransformer::SubnetToOnnxifiOpViaC2(
const caffe2::NetDef& net,
const std::unordered_set<std::string>& weights_in_ws,
const ShapeInfoMap& shape_hints) {
int onnxifi_op_id = onnxifi_op_id_;
if (opts_.debug) {
WriteProtoToTextFile(
net,
"debug_original_net_" + c10::to_string(onnxifi_op_id) + ".pb_txt",
false);
}
if (opts_.min_ops > net.op_size()) {
return net;
}
// We already have all the ops and external inputs and outputs!
NetDef onnxifi_net(net);
// Remove the second output of Concat/Reshape from external_output. Remove
// rest of the outputs of LayerNorm too. In addition, we remove those outputs
// from the Onnxifi op too.
// TODO: This approach is a bit hacky as we assume that the second output is
// never used. A more appropriate approach can be learned from the ONNX path,
// where we statically computes the split_info given input shape and insert a
// GivenTensorIntFill op
std::unordered_set<std::string> split_infos;
for (auto& op : *onnxifi_net.mutable_op()) {
if ((op.type() == "Concat" || op.type() == "Reshape") &&
op.output_size() == 2) {
split_infos.emplace(op.output(1));
} else if (
op.type() == "SparseLengthsSum" ||
op.type() == "SparseLengthsSumFused8BitRowwise" ||
op.type() == "SparseLengthsWeightedSum" ||
op.type() == "SparseLengthsWeightedSumFused8BitRowwise" ||
op.type() == "SparseLengthsSumFused4BitRowwise" ||
op.type() == "SparseLengthsWeightedSumFused4BitRowwise") {
int weighted = (op.type() == "SparseLengthsWeightedSum" ||
op.type() == "SparseLengthsWeightedSumFused8BitRowwise" ||
op.type() == "SparseLengthsWeightedSumFused4BitRowwise")
? 1
: 0;
const auto& indices_hint = shape_hints.at(op.input(1 + weighted));
const auto& lengths_hint = shape_hints.at(op.input(2 + weighted));
const auto& indices_shape = indices_hint.shape;
const auto& lengths_shape = lengths_hint.shape;
if ((indices_hint.getDimType(0) ==
TensorBoundShape_DimType_BATCH_OF_FEATURE_MAX ||
indices_hint.getDimType(0) ==
TensorBoundShape_DimType_BATCH_OF_FEATURE_MAX_DEFAULT) &&
indices_shape.dims_size() == 1 && lengths_shape.dims_size() == 1 &&
indices_shape.dims(0) == lengths_shape.dims(0)) {
op.add_arg()->CopyFrom(MakeArgument<int>("length1", 1));
}
} else if (op.type() == "LayerNorm" && op.output_size() > 1) {
for (int i = 1; i < op.output_size(); ++i) {
split_infos.emplace(op.output(i));
}
}
}
onnxifi_net.clear_external_output();
for (const auto& o : net.external_output()) {
if (!split_infos.count(o)) {
onnxifi_net.add_external_output(o);
}
}
// Figure out weights and add it to external_inputs too
std::unordered_set<std::string> initialization_list;
std::vector<std::string> total_inputs_vec;
getWeightsAndInputs(
net,
weights_in_ws,
std::vector<std::string>(),
&initialization_list,
&total_inputs_vec);
auto* shape_arg = onnxifi_net.add_arg();
auto* qshape_arg = onnxifi_net.add_arg();
shape_arg->set_name("input_shape_info");
qshape_arg->set_name("input_qshape_info");
std::sort(total_inputs_vec.begin(), total_inputs_vec.end());
onnxifi_net.clear_external_input();
for (const auto& i : total_inputs_vec) {
onnxifi_net.add_external_input(i);
auto info = shape_hints.at(i);
if (!info.is_quantized) {
shape_arg->mutable_tensors()->Add()->CopyFrom(
wrapShapeInfoIntoTensorProto(i, shape_hints.at(i)));
} else {
qshape_arg->mutable_qtensors()->Add()->CopyFrom(
wrapShapeInfoIntoQTensorProto(i, shape_hints.at(i)));
}
}
// Add partition info
for (const auto& p : partition_infos_) {
onnxifi_net.add_partition_info()->CopyFrom(p);
}
// Add initializers (weights) list to the net as an arg
auto* w_arg = onnxifi_net.add_arg();
w_arg->set_name(kInitializers);
for (const auto& i : initialization_list) {
w_arg->add_strings(i);
}
// Build ONNXIFI Op
std::string model_str;
onnxifi_net.SerializeToString(&model_str);
std::vector<std::string> onnxifi_net_inputs(
onnxifi_net.external_input().begin(), onnxifi_net.external_input().end());
std::vector<std::string> onnxifi_net_outputs(
onnxifi_net.external_output().begin(),
onnxifi_net.external_output().end());
auto onnxifi_op = buildOnnxifiOp(
model_str,
initialization_list,
onnxifi_net_inputs,
onnxifi_net_outputs,
shape_hints);
NetDef net_opt = composeResultNet(onnxifi_op);
// Debugging stuff
if (opts_.debug) {
WriteProtoToTextFile(
onnxifi_net,
"debug_onnxifi_net_" + c10::to_string(onnxifi_op_id) + ".pb_txt",
false);
WriteProtoToTextFile(
net_opt,
"debug_optimized_net_" + c10::to_string(onnxifi_op_id) + ".pb_txt",
false);
}
return net_opt;
}
NetDef OnnxifiTransformer::SubnetToOnnxifiOpViaOnnx(
const caffe2::NetDef& net,
const std::unordered_set<std::string>& weights_in_ws,
Workspace* ws,
onnx::OnnxExporter* exporter,
ShapeInfoMap* shape_hints) {
if (opts_.min_ops > net.op_size()) {
return net;
}
::ONNX_NAMESPACE::ModelProto onnx_model;
fillModelInfo(&onnx_model);
caffe2::NetDef onnxifi_net(net);
// Convert c2 ops to onnx ops, add const weights if there are any
DeviceOption option;
CPUContext context(option);
context.SwitchToDevice();
std::vector<std::string> extra_weights;
for (const auto& op : onnxifi_net.op()) {
const auto results = exporter->Caffe2OpToOnnxNodes(op, shape_hints_onnx_);
for (const auto& n : results.first) {
onnx_model.mutable_graph()->add_node()->CopyFrom(n);
}
for (const auto& t : results.second) {
VLOG(2) << "Adding extra init tensor: " << t.name();
TensorShape shape;
shape.mutable_dims()->CopyFrom(t.dims());
auto ret = shape_hints_onnx_.emplace(t.name(), std::move(shape));
shape_hints->emplace(
std::piecewise_construct,
std::forward_as_tuple(ret.first->first),
std::forward_as_tuple(
std::vector<TensorBoundShape::DimType>(
shape.dims_size(), TensorBoundShape_DimType_CONSTANT),
ret.first->second));
// Feed into workspace as CPU Tensors
auto* blob = ws->CreateBlob(t.name());
auto* cpu_tensor = BlobGetMutableTensor(blob, CPU);
std::vector<int64_t> dims;
for (const auto& d : t.dims()) {
dims.push_back(d);
}
cpu_tensor->Resize(dims);
if (t.data_type() == ::ONNX_NAMESPACE::TensorProto::FLOAT) {
context.CopyBytesSameDevice(
cpu_tensor->numel() * sizeof(float),
static_cast<const void*>(t.raw_data().data()),
cpu_tensor->raw_mutable_data(TypeMeta::Make<float>()));
} else if (t.data_type() == ::ONNX_NAMESPACE::TensorProto::INT64) {
context.CopyBytesSameDevice(
cpu_tensor->numel() * sizeof(int64_t),
static_cast<const void*>(t.raw_data().data()),
cpu_tensor->raw_mutable_data(TypeMeta::Make<int64_t>()));
} else {
CAFFE_THROW(
"Unsupported tensor data type for conversion: ", t.data_type());
}
context.FinishDeviceComputation();
// Add mappings
extra_weights.emplace_back(t.name());
}
}
// Convert outputs and compute output shape hints
std::vector<std::string> onnxifi_net_outputs;
for (const auto& o : net.external_output()) {
onnxifi_net_outputs.emplace_back(o);
}
auto io_vec = convertToValueInfo(
onnxifi_net_outputs,
shape_hints_onnx_,
std::unordered_map<std::string, ::ONNX_NAMESPACE::TypeProto>());
for (const auto& i : io_vec) {
onnx_model.mutable_graph()->add_output()->CopyFrom(i);
}
// Convert inputs and figure out weights
std::unordered_set<std::string> initialization_list;
std::vector<std::string> onnxifi_net_inputs;
getWeightsAndInputs(
net,
weights_in_ws,
extra_weights,
&initialization_list,
&onnxifi_net_inputs);
io_vec = convertToValueInfo(
onnxifi_net_inputs,
shape_hints_onnx_,
std::unordered_map<std::string, ::ONNX_NAMESPACE::TypeProto>());
for (const auto& i : io_vec) {
onnx_model.mutable_graph()->add_input()->CopyFrom(i);
}
// Onnx model is ready. Build ONNXIFI Op
std::string model_str;
onnx_model.SerializeToString(&model_str);
auto onnxifi_op = buildOnnxifiOp(
model_str,
initialization_list,
onnxifi_net_inputs,
onnxifi_net_outputs,
*shape_hints);
NetDef net_opt = composeResultNet(onnxifi_op);
// Debugging stuff
if (opts_.debug) {
WriteProtoToTextFile(onnx_model, "debug_onnxifi_net.onnx_txt", false);
WriteProtoToTextFile(net_opt, "debug_optimized_net.pb_txt", false);
}
return net_opt;
}
bool OnnxifiTransformer::supportOpOnnx(
const caffe2::OperatorDef& op,
onnx::OnnxExporter* exporter,
const std::unordered_set<int>& blocklisted_ops,
onnxBackendID backend_id) const {
try {
int pos =
ArgumentHelper::GetSingleArgument<OperatorDef, int>(op, kNetPos, -1);
if (blocklisted_ops.count(pos)) {
LOG(INFO) << "Skipping blacklisted op " << op.type() << " at pos " << pos;
return false;
}
const OpSchema* schema = OpSchemaRegistry::Schema(op.type());
// NB: this might not be a hard constraint as we can just export C2
// domain specific ops to ONNX
if (!schema || schema->onnx_schema().empty()) {
LOG(INFO) << "Cannot export c2 op " << op.type()
<< " to onnx as there is no corresponding ONNX schema.";
return false;
}
::ONNX_NAMESPACE::ModelProto onnx_model;
fillModelInfo(&onnx_model);
auto results = exporter->Caffe2OpToOnnxNodes(op, shape_hints_onnx_);
std::unordered_set<std::string> used_inputs;
std::unordered_set<std::string> used_outputs;
std::vector<std::string> boundary_inputs;
std::vector<std::string> boundary_outputs;
std::unordered_set<std::string> reshape_info;
// nodes are in topological order, so we just need to iterate
for (const auto& n : results.first) {
onnx_model.mutable_graph()->add_node()->CopyFrom(n);
for (const auto& i : n.input()) {
bool is_new = used_inputs.emplace(i).second;
// The input is not seen and it's not referred by any nodes before as
// output, we count it as an boundary input
if (is_new && !used_outputs.count(i)) {
boundary_inputs.emplace_back(i);
}
}
for (const auto& o : n.output()) {
used_outputs.emplace(o);
}
// For reshape node, if it has more than 1 inputs, we need to feed the
// second input which contains the shape info
if (n.op_type() == "Reshape" && n.input_size() > 1) {
reshape_info.emplace(n.input(1));
}
}
// Second iteration to account all the boundary outputs, which is a newly
// seen output and is not referred as input before
used_outputs.clear();
for (const auto& n : results.first) {
for (const auto& o : n.output()) {
bool is_new = used_outputs.emplace(o).second;
if (is_new && !used_inputs.count(o)) {
boundary_outputs.emplace_back(o);
}
}
}
std::unordered_map<std::string, ::ONNX_NAMESPACE::TypeProto>
extra_shape_hints;
for (const auto& t : results.second) {
extra_shape_hints.emplace(t.name(), onnx::ExtraTypeProto(t));
if (reshape_info.count(t.name())) {
onnx_model.mutable_graph()->add_initializer()->CopyFrom(t);
}
}
// Add input/output shape info
auto io_vec = convertToValueInfo(
boundary_inputs, shape_hints_onnx_, extra_shape_hints);
for (const auto& i : io_vec) {
onnx_model.mutable_graph()->add_input()->CopyFrom(i);
}
io_vec = convertToValueInfo(
boundary_outputs, shape_hints_onnx_, extra_shape_hints);
for (const auto& i : io_vec) {
onnx_model.mutable_graph()->add_output()->CopyFrom(i);
}
std::string onnx_model_str;
onnx_model.SerializeToString(&onnx_model_str);
auto ret = lib_->onnxGetBackendCompatibility(
backend_id, onnx_model_str.size(), onnx_model_str.c_str());
if (ret != ONNXIFI_STATUS_SUCCESS) {
LOG(INFO) << "Don't support onnx for " << op.type() << " c2 op (" << ret
<< ")";
return false;
} else {
return true;
}
} catch (const std::exception& ex) {
LOG(ERROR) << "Caught exception when converting op " << op.type()
<< ", what: " << ex.what();
return false;
}
}
bool OnnxifiTransformer::supportOpC2(
const caffe2::OperatorDef& op,
const ShapeInfoMap& shape_hints,
const std::unordered_set<std::string>& weights,
const std::unordered_set<int>& blocklisted_ops,
onnxBackendID backend_id) const {
try {
int pos =
ArgumentHelper::GetSingleArgument<OperatorDef, int>(op, kNetPos, -1);
if (blocklisted_ops.count(pos)) {
LOG(INFO) << "Skipping blocklisted op " << op.type() << " at pos " << pos;
return false;
}
// Build a c2 net with one op
NetDef net;
net.add_op()->CopyFrom(op);
std::unordered_set<std::string> seenExternalInputs;
for (const auto& i : op.input()) {
if (seenExternalInputs.count(i)) {
continue;
}
seenExternalInputs.insert(i);
net.add_external_input(i);
}
for (const auto& o : op.output()) {
net.add_external_output(o);
}
// Remove the second output of Concat/Reshape from the external_output
if ((op.type() == "Concat" || op.type() == "Reshape") &&
op.output_size() == 2) {
net.mutable_external_output()->RemoveLast();
} else if (op.type() == "LayerNorm" && op.output_size() > 1) {
int remove = op.output_size() - 1;
for (int i = 0; i < remove; ++i) {
net.mutable_external_output()->RemoveLast();
}
}
// Encode the input/output shapes to an argument
auto* shape_arg = net.add_arg();
auto* qshape_arg = net.add_arg();