-
Notifications
You must be signed in to change notification settings - Fork 22
/
cudnn_conv_test.cc
1242 lines (1124 loc) · 48.7 KB
/
cudnn_conv_test.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 2018 Google LLC
*
* 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
*
* https://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 <algorithm>
#include <cstddef>
#include <functional>
#include <memory>
#include <random>
#include <string>
#include <utility>
#include <vector>
#include "ostream_nullptr.h"
#include "glog/logging.h"
#include "google/protobuf/repeated_field.h"
#include "gtest/gtest.h"
#include "absl/strings/str_join.h"
#include "absl/types/optional.h"
#include "cudnn/include/cudnn.h"
#include "all_pairs.h"
#include "cuda_util.h"
#include "cudnn.pb.h"
#include "cudnn_test.h"
#include "cudnn_util.h"
#include "test_util.h"
// Tests that compare the outputs of all supported convolution algorithms for
// specific configuration of input, filter, etc. The configurations are loaded
// from cudnn_tests.textproto and generated using the all-pair testing approach.
//
// For cuDNN 7 and later, also tests that the cudnnGetConvolution*Algorithm_v7
// returns the same set of algorithms as GetSupportedConvolutionAlgos.
namespace nvidia_libs_test {
namespace {
class ConvolutionTest
: public ::testing::TestWithParam<proto::ConvolutionTest> {
void TearDown() override {
if (HasFatalFailure()) {
ResetDevice();
}
}
};
// Returns expected accuracy of the given algorithm. The tolerance is a mix
// between absolute and relative error, see test_util.h.
double GetTolerance(const ConvolutionAlgo& algo) {
struct Visitor {
double operator()(cudnnConvolutionFwdAlgo_t algo) {
switch (algo) {
case CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM:
return 1e-4;
case CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM:
return 1e-4;
case CUDNN_CONVOLUTION_FWD_ALGO_GEMM:
return 1e-4;
case CUDNN_CONVOLUTION_FWD_ALGO_DIRECT:
return 0.0; // Not implemented.
case CUDNN_CONVOLUTION_FWD_ALGO_FFT:
return 1e-5;
case CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING:
return 1e-5;
case CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD:
return 1e-5;
case CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED:
return 1e-2;
default:
LOG(FATAL) << "Unknown algorithm";
}
}
double operator()(cudnnConvolutionBwdDataAlgo_t algo) {
switch (algo) {
case CUDNN_CONVOLUTION_BWD_DATA_ALGO_0:
return 1e-5;
case CUDNN_CONVOLUTION_BWD_DATA_ALGO_1:
return 1e-4;
case CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT:
return 1e-5;
case CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING:
return 1e-4;
case CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD:
return 1e-5;
case CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED:
return 1e-2;
default:
LOG(FATAL) << "Unknown algorithm";
}
}
double operator()(cudnnConvolutionBwdFilterAlgo_t algo) {
switch (algo) {
case CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0:
return 1e-5;
case CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1:
return 1e-4;
case CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT:
return 1e-5;
case CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3:
return 1e-4;
case CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD:
return 0.0; // Not implemented.
case CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED:
return 1e-2;
case CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT_TILING:
return 1e-5;
default:
LOG(FATAL) << "Unknown algorithm";
}
}
};
return visit(Visitor(), algo);
}
// Returns expected accuracy of the convolution based on the data types
// configuration specified in the proto, relative to FLOAT_CONFIG (see cuDNN
// documentation).
double GetToleranceScale(const proto::ConvolutionConfig& proto) {
switch (proto.filter().data_type()) {
case proto::DATA_FLOAT:
return 1.0;
case proto::DATA_DOUBLE:
// cuDNN uses different kernels for architectures before Maxwell that have
// much lower double accuracy.
return DeviceHasAtLeastComputeCapability(5, 0) ? 1e-8 : 1e-2;
case proto::DATA_HALF:
if (proto.convolution().math_type() == proto::TENSOR_OP_MATH) {
return 1e+4; // Using tensor ops.
}
if (proto.convolution().compute_mode() == proto::DATA_FLOAT) {
return 1e+5; // Pseudo half config.
}
return 1e+5; // True half config.
default:
LOG(FATAL) << "Not yet supported";
}
}
// Returns a tensor descriptor with the same layout as the given filter
// descriptor.
proto::TensorDescriptor GetFilterTensorDescriptor(
const proto::FilterDescriptor& filter) {
proto::TensorDescriptor result;
*result.mutable_dimension() = filter.dimension();
result.set_data_type(filter.data_type());
if (result.dimension_size() == 4) {
result.set_format(filter.format());
return result;
}
// If filter rank is not 4, it must be fully packed.
CHECK_EQ(filter.format(), proto::TENSOR_NCHW);
int rank = result.dimension_size();
auto* strides = result.mutable_stride();
strides->Resize(rank, 0);
for (int i = rank - 1, stride = 1; i >= 0; --i) {
strides->Set(i, stride);
stride *= filter.dimension(i);
}
return result;
}
// Returns the algorithms specified in the proto.
std::vector<ConvolutionAlgo> GetAlgorithms(
const proto::ConvolutionConfig& proto, const CudnnHandle& handle,
const Convolution& convolution, size_t workspace_limit) {
switch (proto.algo_oneof_case()) {
case proto::ConvolutionConfig::kAllAlgos: {
return GetSupportedConvolutionAlgos(
handle, proto.all_algos(), convolution.input_desc,
convolution.filter_desc, convolution.conv_desc,
convolution.output_desc, workspace_limit);
}
case proto::ConvolutionConfig::kFwdAlgo:
return {static_cast<cudnnConvolutionFwdAlgo_t>(proto.fwd_algo())};
case proto::ConvolutionConfig::kBwdDataAlgo:
return {
static_cast<cudnnConvolutionBwdDataAlgo_t>(proto.bwd_data_algo())};
case proto::ConvolutionConfig::kBwdFilterAlgo:
return {static_cast<cudnnConvolutionBwdFilterAlgo_t>(
proto.bwd_filter_algo())};
default:
LOG(FATAL) << "Invalid algo_oneof case";
}
}
// Runs the convolution.
Status RunConvolution(double alpha, double beta, const CudnnHandle& handle,
const Convolution& convolution,
const ConvolutionAlgo& algo) {
ASSIGN_OR_RETURN_STATUS(
auto workspace_size,
GetWorkspaceSize(handle, convolution.input_desc, convolution.filter_desc,
convolution.conv_desc, convolution.output_desc, algo));
ASSIGN_OR_RETURN_STATUS(auto workspace, AllocateDeviceMemory(workspace_size));
FillWithNaNs(workspace);
RETURN_IF_ERROR_STATUS(RunConvolution(
handle, algo, alpha, beta, convolution.input_desc, convolution.input_data,
convolution.filter_desc, convolution.filter_data, convolution.conv_desc,
convolution.output_desc, convolution.output_data, workspace,
workspace_size));
// Detect and handle CUDA errors to avoid hitting the CHECK_OK_STATUS when
// deallocating the workspace.
return GetStatus(cudaDeviceSynchronize());
}
// Returns 'from' merged into 'proto', handling repeated and oneof fields.
proto::ConvolutionConfig GetMergedConvolution(
proto::ConvolutionConfig proto, const proto::ConvolutionConfig& from) {
auto process_tensor_desc = [](proto::TensorDescriptor* proto_tensor,
const proto::TensorDescriptor& from_tensor) {
if (from_tensor.dimension_size()) {
proto_tensor->clear_dimension();
}
if (from_tensor.data_type_oneof_case()) {
proto_tensor->clear_data_type();
}
if (from_tensor.format_oneof_case()) {
proto_tensor->clear_format();
}
if (from_tensor.stride_size()) {
proto_tensor->clear_stride();
}
};
// Clear all repeated and oneof fields in 'proto' that are set in 'from'.
process_tensor_desc(proto.mutable_input(), from.input());
if (from.has_output()) {
process_tensor_desc(proto.mutable_output(), from.output());
}
if (from.filter().dimension_size()) {
proto.mutable_filter()->clear_dimension();
}
if (from.filter().data_type_oneof_case()) {
proto.mutable_filter()->clear_data_type();
}
if (from.filter().format_oneof_case()) {
proto.mutable_filter()->clear_format();
}
if (from.convolution().pad_size()) {
proto.mutable_convolution()->clear_pad();
}
if (from.convolution().filter_stride_size()) {
proto.mutable_convolution()->clear_filter_stride();
}
if (from.convolution().dilation_size()) {
proto.mutable_convolution()->clear_dilation();
}
if (from.convolution().compute_mode_oneof_case()) {
proto.mutable_convolution()->clear_compute_mode();
}
if (from.algo_oneof_case()) {
proto.set_all_algos(proto::CONVOLUTION_DIRECTION_UNSPECIFIED);
}
proto.MergeFrom(from);
return proto;
}
TEST_P(ConvolutionTest, CompareResults) {
VLOG(2) << "Running all-test for\n" << GetParam().DebugString();
CudnnHandle handle = CreateCudnnHandle();
RandomGenerator rand_gen(/*seed=*/0);
// Run reference convolution first, then compare others to the reference
// result.
double data_lower = GetParam().values_lower_bound(), data_upper = 1.;
if (GetParam().values_bound_oneof_case() ==
proto::ConvolutionTest::kValuesUpperBound) {
data_upper = GetParam().values_upper_bound();
}
auto ref_proto = GetParam().reference();
ASSERT_OK_AND_ASSIGN(auto reference, CreateConvolution(ref_proto, data_lower,
data_upper, rand_gen));
ASSERT_OK_AND_ASSIGN(auto workspace_limit, GetWorkspaceLimit(ref_proto));
auto ref_algos = GetAlgorithms(ref_proto, handle, reference, workspace_limit);
ASSERT_FALSE(ref_algos.empty()) << "No algorithm found";
auto ref_algo = ref_algos.front();
// Run reference convolution with default scaling factors.
ASSERT_TRUE(IsOk(RunConvolution(1.0, 0.0, handle, reference, ref_algo)))
<< "algo: " << ref_algo;
auto ref_filter_tensor_desc =
CreateTensorDescriptor(GetFilterTensorDescriptor(ref_proto.filter()));
for (auto proto : GetParam().test()) {
proto = GetMergedConvolution(ref_proto, proto);
ASSERT_OK_AND_ASSIGN(
Convolution test,
CreateConvolution(proto, data_lower, data_upper, rand_gen));
// We now have input, filter, and output buffers for the reference and the
// test. A convolution has two read-only argument buffers and writes the
// result to the remaining buffer. The convolution direction determines
// which buffers are arguments and which buffer is the result.
//
// convolution direction: argument buffers: result buffer:
// --------------------------------------------------------------
// forward input, filter output
// backward data filter, output input
// backward filter output, input filter
//
// The reference buffers contain the arguments and ground truth result of
// the convolution. The test buffers are filled with random values at this
// point. Tensors aren't necessarily packed and we use the random value of
// buffer elements not referenced by the tensor to detect out-of-bounds
// access by cuDNN.
//
// If beta is zero, which means the result buffer values shouldn't be read,
// we populate result buffer with NaNs before sending it to the convolution.
auto direction = [&] {
switch (proto.algo_oneof_case()) {
case proto::ConvolutionConfig::kAllAlgos:
return proto.all_algos();
case proto::ConvolutionConfig::kFwdAlgo:
return proto::CONVOLUTION_FWD;
case proto::ConvolutionConfig::kBwdDataAlgo:
return proto::CONVOLUTION_BWD_DATA;
case proto::ConvolutionConfig::kBwdFilterAlgo:
return proto::CONVOLUTION_BWD_FILTER;
default:
LOG(FATAL) << "Invalid algo_oneof case";
}
}();
auto filter_tensor_desc =
CreateTensorDescriptor(GetFilterTensorDescriptor(proto.filter()));
double alpha = 1.0 - proto.one_minus_alpha();
double beta = proto.beta();
// Get a pointer to the test result buffer and its tensor descriptor.
TensorDescriptor* result_desc = nullptr;
DeviceMemory* result_data = nullptr;
struct {
double alpha = 1.0, beta = 0.0;
} input_scaling, filter_scaling, output_scaling;
switch (direction) {
case proto::CONVOLUTION_FWD:
result_desc = &test.output_desc;
result_data = &test.output_data;
output_scaling.alpha = alpha;
output_scaling.beta = beta;
break;
case proto::CONVOLUTION_BWD_DATA:
result_desc = &test.input_desc;
result_data = &test.input_data;
input_scaling.alpha = alpha;
input_scaling.beta = beta;
break;
case proto::CONVOLUTION_BWD_FILTER:
result_desc = &filter_tensor_desc;
result_data = &test.filter_data;
filter_scaling.alpha = alpha;
filter_scaling.beta = beta;
break;
default:
LOG(FATAL) << "Invalid direction: "
<< proto::ConvolutionDirection_Name(direction);
}
size_t result_size = GetTensorSizeInBytes(*result_desc);
DeviceMemory init_data(nullptr);
if (beta > 0) {
// Make a copy of the test result buffer so that we can reset it to the
// same random values before running each algorithm.
ASSERT_OK_AND_ASSIGN(init_data, AllocateDeviceMemory(result_size));
ASSERT_TRUE(IsOk(CopyDeviceMemory(init_data, *result_data, result_size)));
}
// Reset the test result buffer and run the convolution.
const auto restore_result_buffer = [&] {
if (beta > 0) {
ASSERT_TRUE(
IsOk(CopyDeviceMemory(*result_data, init_data, result_size)));
} else {
FillWithNaNs(*result_data);
}
};
// Blend the reference buffers into the corresponding test buffers. This
// may involve converting the data type (e.g. from double to float) and
// changing the data layout (e.g. from NCHW to NHWC).
ASSERT_TRUE(IsOk(ConvertAndTransformTensor(
handle, input_scaling.alpha, input_scaling.beta, reference.input_desc,
reference.input_data, test.input_desc, test.input_data)));
ASSERT_TRUE(IsOk(ConvertAndTransformTensor(
handle, filter_scaling.alpha, filter_scaling.beta,
ref_filter_tensor_desc, reference.filter_data, filter_tensor_desc,
test.filter_data)));
ASSERT_TRUE(IsOk(ConvertAndTransformTensor(
handle, output_scaling.alpha, output_scaling.beta,
reference.output_desc, reference.output_data, test.output_desc,
test.output_data)));
// The test result buffer (pointee of result_data) now contains the copy of
// the reference result buffer. Stash that away in ref_result_data and
// allocate a new buffer for the test's actual result buffer.
ASSERT_OK_AND_ASSIGN(auto ref_result_data,
AllocateDeviceMemory(result_size));
std::swap(*result_data, ref_result_data);
// Get available workspace size (after allocating all device memory).
ASSERT_OK_AND_ASSIGN(auto workspace_limit, GetWorkspaceLimit(proto));
// Run all supported algorithms for current test.
double tolerance_scale = GetToleranceScale(proto);
for (const auto& algo :
GetAlgorithms(proto, handle, test, workspace_limit)) {
restore_result_buffer();
auto get_message = [&] {
std::ostringstream oss;
oss << "format: " << proto::TensorFormat_Name(proto.input().format())
<< "\ndata_type: "
<< proto::DataType_Name(proto.input().data_type())
<< "\ncompute_mode: "
<< proto::DataType_Name(proto.convolution().compute_mode())
<< "\nmath_type: "
<< proto::MathType_Name(proto.convolution().math_type())
<< "\nalgo: " << algo;
return oss.str();
};
Status status = RunConvolution(alpha, beta, handle, test, algo);
// Some convolution configs successfully return a workspace size (or
// return from cudnnGetConvolution*Algorithm_v7 with success status) but
// then fail to run, see nvbugs/2082072. Instead of triggering test
// failures we just skip configurations that hit this bug. A targeted test
// (for 'reference' instead of 'test') is in cudnn_tests.textproto.
if (!status.ok() &&
status.message().find("CUDNN_STATUS_NOT_SUPPORTED") != string::npos) {
continue;
}
ASSERT_TRUE(IsOk(status)) << get_message();
// Check that the test result matches the reference result. This also
// compares buffer elements not referenced by the tensor.
double tolerance = tolerance_scale * GetTolerance(algo);
EXPECT_TRUE(IsOk(TensorDataEqual(ref_result_data, *result_data,
*result_desc, tolerance)))
<< get_message();
}
}
}
#if CUDNN_MAJOR >= 7
template <typename T>
std::vector<ConvolutionAlgo> ToConvolutionAlgos(std::vector<T> algo_perfs,
int num_algorithms) {
// cudnnGetConvolution*Algorithm_v7 also returns elements with non-success
// status. Filter them out here.
auto end =
std::remove_if(algo_perfs.begin(), algo_perfs.begin() + num_algorithms,
[](const T& algo_perf) {
return algo_perf.status != CUDNN_STATUS_SUCCESS;
});
std::vector<ConvolutionAlgo> result;
std::transform(algo_perfs.begin(), end, std::back_inserter(result),
[](const T& algo_perf) { return algo_perf.algo; });
return result;
}
// Compare the list of algorithms returned from GetSupportedConvolutionAlgos
// with the result from cudnnGetConvolution*Algorithm_v7.
TEST_P(ConvolutionTest, GetAlgorithm_v7) {
CudnnHandle handle = CreateCudnnHandle();
auto ref_proto = GetParam().reference();
for (auto proto : GetParam().test()) {
proto = GetMergedConvolution(ref_proto, proto);
if (proto.algo_oneof_case() != proto::ConvolutionConfig::kAllAlgos) {
continue;
}
auto direction = proto.all_algos();
auto input = CreateTensorDescriptor(proto.input());
auto filter = CreateFilterDescriptor(proto.filter());
auto convolution = CreateConvolutionDescriptor(proto.convolution());
ASSERT_OK_AND_ASSIGN(
auto output, CreateOutputDescriptor(proto, input, filter, convolution));
std::vector<ConvolutionAlgo> v7_algos;
int num_algorithms = 0;
switch (direction) {
case proto::CONVOLUTION_FWD: {
ASSERT_TRUE(IsOk(GetStatus(cudnnGetConvolutionForwardAlgorithmMaxCount(
handle.get(), &num_algorithms))));
std::vector<cudnnConvolutionFwdAlgoPerf_t> algo_perfs(num_algorithms);
ASSERT_TRUE(IsOk(GetStatus(cudnnGetConvolutionForwardAlgorithm_v7(
handle.get(), input.get(), filter.get(), convolution.get(),
output.get(), algo_perfs.size(), &num_algorithms,
algo_perfs.data()))));
v7_algos = ToConvolutionAlgos(algo_perfs, num_algorithms);
} break;
case proto::CONVOLUTION_BWD_DATA: {
ASSERT_TRUE(
IsOk(GetStatus(cudnnGetConvolutionBackwardDataAlgorithmMaxCount(
handle.get(), &num_algorithms))));
std::vector<cudnnConvolutionBwdDataAlgoPerf_t> algo_perfs(
num_algorithms);
ASSERT_TRUE(IsOk(GetStatus(cudnnGetConvolutionBackwardDataAlgorithm_v7(
handle.get(), filter.get(), output.get(), convolution.get(),
input.get(), algo_perfs.size(), &num_algorithms,
algo_perfs.data()))));
v7_algos = ToConvolutionAlgos(algo_perfs, num_algorithms);
} break;
case proto::CONVOLUTION_BWD_FILTER: {
ASSERT_TRUE(
IsOk(GetStatus(cudnnGetConvolutionBackwardFilterAlgorithmMaxCount(
handle.get(), &num_algorithms))));
std::vector<cudnnConvolutionBwdFilterAlgoPerf_t> algo_perfs(
num_algorithms);
ASSERT_TRUE(
IsOk(GetStatus(cudnnGetConvolutionBackwardFilterAlgorithm_v7(
handle.get(), input.get(), output.get(), convolution.get(),
filter.get(), algo_perfs.size(), &num_algorithms,
algo_perfs.data()))));
v7_algos = ToConvolutionAlgos(algo_perfs, num_algorithms);
} break;
default:
LOG(FATAL) << "Invalid direction: "
<< proto::ConvolutionDirection_Name(direction);
}
std::sort(v7_algos.begin(), v7_algos.end());
// If the convolution specifies TENSOR_OP_MATH, GetSupportedConvolutionAlgos
// returns algorithms that silently fall back to DEFAULT_MATH as well.
// cudnnGetConvolution*Algorithm_v7 on the other hand returns separate
// algo_perfs for DEFAULT_MATH and TENSOR_OP_MATH (but the latter may still
// fall back to DEFAULT_MATH). We can savely remove duplicates from v7_algos
// because TensorFlow tries to run every algorithm for both math types.
//
// See nvbugs/2072859 for details.
if (proto.convolution().math_type() == proto::TENSOR_OP_MATH) {
v7_algos.erase(std::unique(v7_algos.begin(), v7_algos.end()),
v7_algos.end());
}
auto supported_algos = GetSupportedConvolutionAlgos(
handle, direction, input, filter, convolution, output,
std::numeric_limits<size_t>::max());
std::sort(supported_algos.begin(), supported_algos.end());
EXPECT_EQ(supported_algos, v7_algos);
}
}
#endif // CUDNN_MAJOR >= 7
// Creates a ConvolutionConfig from the given parameters for
// proto.ConvolutionTest.reference.
//
// Note: in_channels and out_channels are multiplied by group_count.
proto::ConvolutionConfig MakeConvolutionConfig(int batch_count, int in_channels,
std::vector<int> in_sizes,
int out_channels,
std::vector<int> filter_sizes,
int group_count,
Padding padding) {
proto::TensorDescriptor input;
for (int dim : {batch_count, in_channels * group_count}) {
input.add_dimension(dim);
}
for (int dim : in_sizes) {
input.add_dimension(dim);
}
input.set_format(proto::TENSOR_NCHW);
input.set_data_type(proto::DATA_DOUBLE);
proto::FilterDescriptor filter;
for (int dim : {out_channels * group_count, in_channels}) {
filter.add_dimension(dim);
}
for (int dim : filter_sizes) {
filter.add_dimension(dim);
}
filter.set_format(proto::TENSOR_NCHW);
filter.set_data_type(proto::DATA_DOUBLE);
proto::ConvolutionDescriptor convolution;
for (auto filter_size : filter_sizes) {
// SAME padding is only this simple if stride and dilation are 1.
// TODO: Add function to compute padding for generic stride/dilation.
convolution.add_pad(padding == Padding::SAME ? filter_size / 2 : 0);
}
convolution.set_compute_mode(proto::DATA_DOUBLE);
if (group_count > 1) {
convolution.set_group_count(group_count);
}
proto::ConvolutionConfig result;
*result.mutable_input() = std::move(input);
*result.mutable_filter() = std::move(filter);
*result.mutable_convolution() = std::move(convolution);
return result;
}
struct SizeRange {
int minimum;
int maximum;
};
// Returns range ['largest 2^k+1 not greater than size', size].
SizeRange MakePowerOfTwoRange(int size) {
CHECK_GT(size, 0);
int power_of_two = 1;
while (power_of_two < size) {
power_of_two *= 2;
}
return {power_of_two / 2 + 1, size};
}
// Creates a randomized ConvolutionConfig from the given parameter ranges for
// proto.ConvolutionTest.reference. For VALID padding, the ranges of input size
// and filter size are clamped so that the filter is no larger then the inputs.
template <typename RandGen>
proto::ConvolutionConfig MakeConvolutionConfig(
RandGen&& rand_gen, SizeRange batch_count_range,
SizeRange in_channels_range, std::vector<SizeRange> in_size_ranges,
SizeRange out_channels_range, std::vector<SizeRange> filter_size_ranges,
SizeRange group_count_range, Padding padding) {
if (padding == Padding::VALID) {
for (auto& in_size_range : in_size_ranges) {
in_size_range.minimum =
std::max(in_size_range.minimum, in_size_range.minimum);
}
}
auto randomize = [&](const SizeRange& range) {
CHECK_LE(range.minimum, range.maximum);
return std::uniform_int_distribution<int>{range.minimum,
range.maximum}(rand_gen);
};
int batch_count = randomize(batch_count_range);
int in_channels = randomize(in_channels_range);
std::vector<int> in_sizes;
std::transform(in_size_ranges.begin(), in_size_ranges.end(),
std::back_inserter(in_sizes), randomize);
int out_channels = randomize(out_channels_range);
int group_count = randomize(group_count_range);
if (padding == Padding::VALID) {
for (size_t i = 0; i < filter_size_ranges.size(); ++i) {
filter_size_ranges[i].maximum =
std::min(filter_size_ranges[i].maximum, in_sizes[i]);
}
}
std::vector<int> filter_sizes;
std::transform(filter_size_ranges.begin(), filter_size_ranges.end(),
std::back_inserter(filter_sizes), randomize);
auto result =
MakeConvolutionConfig(batch_count, in_channels, in_sizes, out_channels,
filter_sizes, group_count, padding);
std::ostringstream label;
label << batch_count << 'x' << in_channels * group_count << 'x'
<< absl::StrJoin(in_sizes, "x") << '_' << out_channels * group_count
<< 'x' << in_channels << 'x' << absl::StrJoin(filter_sizes, "x") << '_'
<< (padding == Padding::SAME ? "SAME" : "VALID");
result.set_label(label.str());
return result;
}
// Set of format and data type configurations to run. For details, see
// http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#cudnnConvolutionForward
enum class Config {
NCHW_FLOAT,
NCHW_DOUBLE,
NCHW_PSEUDO_HALF,
NCHW_TRUE_HALF,
NCHW_TENSOR_OP,
NHWC_FLOAT,
NHWC_PSEUDO_HALF,
NHWC_TRUE_HALF
};
// Returns whether cuDNN supports the given config on the current device.
bool IsConfigSupported(Config config) {
switch (config) {
case Config::NCHW_PSEUDO_HALF:
return DeviceSupportsReducedPrecision();
case Config::NCHW_TRUE_HALF:
return DeviceSupportsReducedPrecision();
case Config::NCHW_TENSOR_OP:
return CUDNN_MAJOR >= 7 && DeviceSupportsTensorOpMath();
case Config::NHWC_FLOAT:
// Internal error on cudNN 6, see cudnn_tests.textproto.
return CUDNN_MAJOR >= 7;
case Config::NHWC_PSEUDO_HALF:
// Crashes on cuDNN 6, see cudnn_tests.textproto.
return CUDNN_MAJOR >= 7 && DeviceSupportsReducedPrecision();
case Config::NHWC_TRUE_HALF:
return DeviceSupportsReducedPrecision();
default:
return true;
}
}
// Creates a ConvolutionConfig from the given config for
// proto.ConvolutionTest.test.
proto::ConvolutionConfig MakeConvolutionConfig(Config config) {
proto::TensorFormat format;
proto::DataType data_type;
proto::DataType compute_mode;
proto::MathType math_type;
string label;
switch (config) {
case Config::NCHW_FLOAT:
format = proto::TENSOR_NCHW;
data_type = proto::DATA_FLOAT;
compute_mode = proto::DATA_FLOAT;
math_type = proto::DEFAULT_MATH;
label = "NCHW_FLOAT";
break;
case Config::NCHW_DOUBLE:
format = proto::TENSOR_NCHW;
data_type = proto::DATA_DOUBLE;
compute_mode = proto::DATA_DOUBLE;
math_type = proto::DEFAULT_MATH;
label = "NCHW_DOUBLE";
break;
case Config::NCHW_PSEUDO_HALF:
format = proto::TENSOR_NCHW;
data_type = proto::DATA_HALF;
compute_mode = proto::DATA_FLOAT;
math_type = proto::DEFAULT_MATH;
label = "NCHW_PSEUDO_HALF";
break;
case Config::NCHW_TRUE_HALF:
format = proto::TENSOR_NCHW;
data_type = proto::DATA_HALF;
compute_mode = proto::DATA_HALF;
math_type = proto::DEFAULT_MATH;
label = "NCHW_TRUE_HALF";
break;
case Config::NCHW_TENSOR_OP:
format = proto::TENSOR_NCHW;
data_type = proto::DATA_HALF;
compute_mode = proto::DATA_HALF;
math_type = proto::TENSOR_OP_MATH;
label = "NCHW_TENSOR_OP";
break;
case Config::NHWC_FLOAT:
format = proto::TENSOR_NHWC;
data_type = proto::DATA_FLOAT;
compute_mode = proto::DATA_FLOAT;
math_type = proto::DEFAULT_MATH;
label = "NHWC_FLOAT";
break;
case Config::NHWC_PSEUDO_HALF:
format = proto::TENSOR_NHWC;
data_type = proto::DATA_HALF;
compute_mode = proto::DATA_FLOAT;
math_type = proto::DEFAULT_MATH;
label = "NHWC_PSEUDO_HALF";
break;
case Config::NHWC_TRUE_HALF:
format = proto::TENSOR_NHWC;
data_type = proto::DATA_HALF;
compute_mode = proto::DATA_HALF;
math_type = proto::DEFAULT_MATH;
label = "NHWC_TRUE_HALF";
break;
default:
LOG(FATAL) << "Unknown config";
}
proto::ConvolutionConfig result;
result.mutable_input()->set_format(format);
result.mutable_input()->set_data_type(data_type);
result.mutable_filter()->set_format(format);
result.mutable_filter()->set_data_type(data_type);
result.mutable_convolution()->set_compute_mode(compute_mode);
result.mutable_convolution()->set_math_type(math_type);
result.set_label(label);
return result;
}
// Returns true if cuDNN should handle the given parameters for at least one
// algorithm. Changes to this function will not just drop some tests, but
// completely alter the set of generated tests!
//
// Note: in_channels and out_channels are multiplied with group_count.
bool ValidateParams(absl::optional<SizeRange> opt_batch_count,
absl::optional<SizeRange> opt_in_channels,
std::vector<absl::optional<SizeRange>> opt_in_sizes,
absl::optional<SizeRange> opt_out_channels,
std::vector<absl::optional<SizeRange>> opt_filter_sizes,
absl::optional<SizeRange> opt_group_count,
absl::optional<Padding> opt_padding,
absl::optional<Config> opt_config) {
// Default to medium sizes so we don't paint ourselves in a corner.
SizeRange tensor_range = {32, 32};
SizeRange filter_range = {5, 5};
SizeRange batch_count = opt_batch_count.value_or(tensor_range);
SizeRange out_channels = opt_out_channels.value_or(tensor_range);
SizeRange filter_in_channels = opt_in_channels.value_or(tensor_range);
SizeRange group_count = opt_group_count.value_or(tensor_range);
Padding padding = opt_padding.value_or(Padding::SAME);
Config config = opt_config.value_or(Config::NCHW_FLOAT);
int in_max_channels = filter_in_channels.maximum * group_count.maximum;
int out_max_channels = out_channels.maximum * group_count.maximum;
std::vector<int> in_max_sizes;
for (const auto& opt_in_size : opt_in_sizes) {
in_max_sizes.push_back(opt_in_size.value_or(tensor_range).maximum);
}
std::vector<int> filter_min_sizes;
std::vector<int> filter_max_sizes;
for (const auto& opt_filter_size : opt_filter_sizes) {
auto filter_size = opt_filter_size.value_or(filter_range);
filter_min_sizes.push_back(filter_size.minimum);
filter_max_sizes.push_back(filter_size.maximum);
}
size_t million = 1ull << 20; // 1 million as in MiB.
size_t max_tensor_elements = static_cast<size_t>(batch_count.maximum) *
std::max(in_max_channels, out_max_channels);
for (const auto& in_max_size : in_max_sizes) {
max_tensor_elements *= in_max_size;
}
if (max_tensor_elements >= 2048 * million) {
// cuDNN has a limit of less (!) than 2G elements per tensor.
return false;
}
if (padding == Padding::VALID) {
for (size_t i = 0; i < opt_in_sizes.size(); ++i) {
if (filter_min_sizes[i] > in_max_sizes[i]) {
// VALID padding requires the filter to be no greater than the input.
return false;
}
}
}
auto proto = MakeConvolutionConfig(
batch_count.maximum, filter_in_channels.maximum, in_max_sizes,
out_channels.maximum, filter_max_sizes, group_count.maximum, padding);
auto input = CreateTensorDescriptor(proto.input());
auto filter = CreateFilterDescriptor(proto.filter());
auto convolution = CreateConvolutionDescriptor(proto.convolution());
auto output_or =
CreateOutputDescriptor(proto::TENSOR_NCHW, input, filter, convolution);
if (!IsOk(output_or)) {
return false;
}
// Compute the size of device allocations in CompareResults and return whether
// it fits the specified memory limit.
size_t data_type_size;
switch (MakeConvolutionConfig(config).filter().data_type()) {
case proto::DATA_FLOAT:
data_type_size = sizeof(float);
break;
case proto::DATA_DOUBLE:
data_type_size = sizeof(double);
break;
case proto::DATA_HALF:
data_type_size = sizeof(__half);
break;
default:
LOG(FATAL) << "Not yet supported";
}
size_t input_num_elements = GetTensorNumElements(input);
size_t filter_num_elements = GetFilterNumElements(filter);
size_t output_num_elements = GetTensorNumElements(output_or.ValueOrDie());
size_t sum_num_elements =
input_num_elements + filter_num_elements + output_num_elements;
size_t max_num_elements =
std::max({input_num_elements, filter_num_elements, output_num_elements});
// Number of buffers for CompareResults' init_data, ref_result_data and
// (if conversion from double is needed) ConvertAndTransformTensor.
size_t num_extra_buffers = data_type_size == sizeof(double) ? 2 : 3;
size_t required_bytes = (sizeof(double) + data_type_size) * sum_num_elements +
num_extra_buffers * data_type_size * max_num_elements;
if (required_bytes > GetAvailableDeviceMemoryBytes()) {
return false;
}
return true;
}
bool ValidateParams2d(absl::optional<SizeRange> opt_batch_count,
absl::optional<SizeRange> opt_in_channels,
absl::optional<SizeRange> opt_in_height,
absl::optional<SizeRange> opt_in_width,
absl::optional<SizeRange> opt_out_channels,
absl::optional<SizeRange> opt_filter_height,
absl::optional<SizeRange> opt_filter_width,
absl::optional<Padding> opt_padding,
absl::optional<Config> opt_config) {
SizeRange group_count = {1, 1};
return ValidateParams(opt_batch_count, opt_in_channels,
{opt_in_height, opt_in_width}, opt_out_channels,
{opt_filter_height, opt_filter_width}, group_count,
opt_padding, opt_config);
}
bool ValidateParams3d(absl::optional<SizeRange> opt_batch_count,
absl::optional<SizeRange> opt_in_channels,
absl::optional<SizeRange> opt_in_depth,
absl::optional<SizeRange> opt_in_height,
absl::optional<SizeRange> opt_in_width,
absl::optional<SizeRange> opt_out_channels,
absl::optional<SizeRange> opt_filter_depth,
absl::optional<SizeRange> opt_filter_height,
absl::optional<SizeRange> opt_filter_width,
absl::optional<Padding> opt_padding,
absl::optional<Config> opt_config) {
SizeRange group_count = {1, 1};
return ValidateParams(opt_batch_count, opt_in_channels,
{opt_in_depth, opt_in_height, opt_in_width},
opt_out_channels,
{opt_filter_depth, opt_filter_height, opt_filter_width},
group_count, opt_padding, opt_config);
}
bool ValidateParamsGrouped(absl::optional<SizeRange> opt_batch_count,
absl::optional<SizeRange> in_channels,
absl::optional<SizeRange> opt_in_height,
absl::optional<SizeRange> opt_in_width,
absl::optional<SizeRange> out_channels,
absl::optional<SizeRange> opt_filter_height,
absl::optional<SizeRange> opt_filter_width,
absl::optional<SizeRange> opt_group_count,
absl::optional<Padding> opt_padding,
absl::optional<Config> opt_config) {
return ValidateParams(opt_batch_count, in_channels,
{opt_in_height, opt_in_width}, out_channels,
{opt_filter_height, opt_filter_width}, opt_group_count,
opt_padding, opt_config);
}
// Factory for proto::ConvolutionTest with one factory method per direction.
class TestFactory {
public:
template <typename RandGen>
TestFactory(RandGen&& rand_gen, SizeRange batch_count, SizeRange in_channels,
std::vector<SizeRange> in_sizes, SizeRange out_channels,
std::vector<SizeRange> filter_sizes, SizeRange group_count,
Padding padding, Config config) {
reference_ =
MakeConvolutionConfig(rand_gen, batch_count, in_channels, in_sizes,
out_channels, filter_sizes, group_count, padding);
test_ = MakeConvolutionConfig(config);
// Disable tests that are not supported on the host machine.
//
// Excluding unsupported configs from the parameter vector would make the
// supported tests depend on the machine configuration, and that seems
// worse than leaving holes in the all-pairs test space on older hardware
// and cuDNN versions.
disabled_label_ = IsConfigSupported(config) ? "" : "DISABLED_";
config_label_ = "_" + test_.label() + "_" + reference_.label();
test_.clear_label();
reference_.clear_label();
}
proto::ConvolutionTest operator()(proto::ConvolutionFwdAlgo algo) const {
// in_channels * filter_height * filter_width.
const auto& dims = reference_.filter().dimension();
int num_summands = std::accumulate(dims.begin() + 2, dims.end(), dims[1],
std::multiplies<int>());
auto result = MakeTest(proto::CONVOLUTION_FWD, num_summands);
result.mutable_reference()->set_fwd_algo(algo);
return result;
}
proto::ConvolutionTest operator()(proto::ConvolutionBwdDataAlgo algo) const {