/
convolutionocl.cpp
3809 lines (3365 loc) · 158 KB
/
convolutionocl.cpp
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
/*******************************************************************************
*
* MIT License
*
* Copyright (c) 2020 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*******************************************************************************/
#include <miopen/algorithm.hpp>
#include <miopen/conv_algo_name.hpp>
#include <miopen/check_numerics.hpp>
#include <miopen/config.h>
#include <miopen/convolution.hpp>
#include <miopen/conv_algo_name.hpp>
#include <miopen/db.hpp>
#include <miopen/db_record.hpp>
#include <miopen/env.hpp>
#include <miopen/find_db.hpp>
#include <miopen/finddb_kernel_cache_key.hpp>
#include <miopen/find_controls.hpp>
#include <miopen/float_equal.hpp>
#include <miopen/invoker.hpp>
#include <miopen/kernel.hpp>
#include <miopen/solver.hpp>
#include <miopen/tensor_ops.hpp>
#include <miopen/tensor.hpp>
#include <miopen/util.hpp>
#include <miopen/visit_float.hpp>
#include <miopen/datatype.hpp>
#include <miopen/any_solver.hpp>
#include <miopen/conv/tensors.hpp>
#include <miopen/conv/compiled_in_parameters.hpp>
#include <miopen/conv/data_invoke_params.hpp>
#include <miopen/conv/wrw_invoke_params.hpp>
#if MIOPEN_USE_GEMM
#include <miopen/gemm_v2.hpp>
#endif
#include <cassert>
#include <type_traits>
#include <boost/range/adaptors.hpp>
namespace miopen {
MIOPEN_DECLARE_ENV_VAR(MIOPEN_DEBUG_CONV_GEMM)
MIOPEN_DECLARE_ENV_VAR(MIOPEN_DEBUG_CONV_DIRECT)
MIOPEN_DECLARE_ENV_VAR(MIOPEN_DEBUG_CONV_WINOGRAD)
MIOPEN_DECLARE_ENV_VAR(MIOPEN_DEBUG_CONV_IMPLICIT_GEMM)
MIOPEN_DECLARE_ENV_VAR(MIOPEN_CONV_PRECISE_ROCBLAS_TIMING)
MIOPEN_DECLARE_ENV_VAR(MIOPEN_DEBUG_CONV_FFT)
MIOPEN_DECLARE_ENV_VAR(MIOPEN_DEVICE_ARCH)
MIOPEN_DECLARE_ENV_VAR(MIOPEN_DEBUG_CONV_IMMED_FALLBACK)
#if MIOPEN_USE_GEMM
#ifdef CPPCHECK
// Keep the value unknown in cppcheck since this can differ between opencl and hip
static bool IsUseRocBlas;
#else
static const bool IsUseRocBlas = (MIOPEN_USE_ROCBLAS == 1);
#endif
static inline bool IsAnyBufferBF16(const TensorDescriptor& xDesc,
const TensorDescriptor& yDesc,
const TensorDescriptor& wDesc)
{
return xDesc.GetType() == miopenBFloat16 || yDesc.GetType() == miopenBFloat16 ||
wDesc.GetType() == miopenBFloat16;
}
#endif
size_t GetKernelGlobalWorkDim(const KernelInvoke& kernel, int dim)
{
#if(MIOPEN_BACKEND_HIP)
return kernel.gdims[dim];
#else
return kernel.global_work_dim[dim];
#endif
}
size_t GetKernelLocalWorkDim(const KernelInvoke& kernel, int dim)
{
#if(MIOPEN_BACKEND_HIP)
return kernel.ldims[dim];
#else
// sometimes local_work_dim = {0,0,0} look in issue #1724
return kernel.local_work_dim[dim];
#endif
}
static inline void AddKernels(const Handle& handle,
const std::string& algorithm_name,
const std::string& network_config,
const miopen::solver::ConvSolution& s,
std::vector<KernelInvoke>* const kernels)
{
if(!algorithm_name.empty() && !network_config.empty())
{
handle.ClearKernels(algorithm_name, network_config);
}
else
{
assert(algorithm_name.empty() && network_config.empty());
}
int i = 0;
for(auto& k : s.construction_params)
{
MIOPEN_LOG_I2(k.kernel_name);
auto kernel = handle.AddKernel(algorithm_name,
network_config,
k.kernel_file,
k.kernel_name,
k.l_wk,
k.g_wk,
k.comp_options,
i);
if(kernels != nullptr)
{
kernels->push_back(kernel);
}
++i;
}
}
static inline void ValidateGroupCount(const TensorDescriptor& xDesc,
const TensorDescriptor& wDesc,
const ConvolutionDescriptor& conv)
{
if(conv.group_count == 1)
{
if(xDesc.GetLengths()[1] != wDesc.GetLengths()[1])
MIOPEN_THROW(miopenStatusBadParm, "Invalid filter channel number");
}
if(conv.group_count > 1)
{
if(xDesc.GetLengths()[1] % conv.group_count != 0 ||
wDesc.GetLengths()[0] % conv.group_count != 0 ||
conv.group_count > xDesc.GetLengths()[1] || conv.group_count > wDesc.GetLengths()[0] ||
conv.group_count < 1)
MIOPEN_THROW(miopenStatusBadParm, "Invalid group number");
if(xDesc.GetLengths()[1] / conv.group_count != wDesc.GetLengths()[1])
MIOPEN_THROW(miopenStatusBadParm, "Invalid filter channel number");
}
}
std::vector<miopen::solver::ConvSolution>
ConvolutionDescriptor::FindWinogradSolutions(const ConvolutionContext& ctx,
const AnyInvokeParams& invoke_ctx) const
{
if(miopen::IsDisabled(MIOPEN_DEBUG_CONV_WINOGRAD{}))
return {};
try
{
return FindAllWinogradSolutions(ctx, invoke_ctx);
}
catch(miopen::Exception& ex)
{
MIOPEN_LOG_WE(ex.what());
return {};
}
}
std::vector<miopen::solver::ConvSolution>
ConvolutionDescriptor::FindDataDirectSolutions(Handle& handle,
const TensorDescriptor& xDesc,
const TensorDescriptor& wDesc,
const TensorDescriptor& yDesc,
bool exhaustiveSearch,
bool isForward,
const ConvolutionUserBuffers& bufs,
const AnyInvokeParams& invoke_ctx) const
{
if(miopen::IsDisabled(MIOPEN_DEBUG_CONV_DIRECT{}))
return {};
const auto dir = isForward ? conv::Direction::Forward : conv::Direction::BackwardData;
auto ctx = ConvolutionContext{xDesc, wDesc, yDesc, *this, dir};
ctx.skip_solutions_that_take_long_time_to_build_and_have_narrow_coverage =
miopen::FindMode(ctx).IsFastHybrid();
ctx.use_dynamic_solutions_only = miopen::FindMode(ctx).IsDynamicHybrid();
ctx.do_search = exhaustiveSearch;
ctx.save_srch_req = true;
ctx.general_compile_options = "";
ctx.SetStream(&handle);
ctx.SetBufs(bufs);
ctx.DetectRocm();
ctx.SetupFloats();
try
{
return FindAllDirectSolutions(ctx, invoke_ctx);
}
catch(miopen::Exception& ex)
{
MIOPEN_LOG_WE(ex.what());
return {};
}
}
std::vector<miopen::solver::ConvSolution>
ConvolutionDescriptor::FindDataImplicitGemmSolutions(Handle& handle,
const TensorDescriptor& xDesc,
const TensorDescriptor& wDesc,
const TensorDescriptor& yDesc,
bool exhaustiveSearch,
bool isForward,
const ConvolutionUserBuffers& bufs,
const AnyInvokeParams& invoke_ctx) const
{
if(miopen::IsDisabled(MIOPEN_DEBUG_CONV_IMPLICIT_GEMM{}))
return {};
const auto dir = isForward ? conv::Direction::Forward : conv::Direction::BackwardData;
auto ctx = ConvolutionContext{xDesc, wDesc, yDesc, *this, dir};
ctx.skip_solutions_that_take_long_time_to_build_and_have_narrow_coverage =
miopen::FindMode(ctx).IsFastHybrid();
ctx.use_dynamic_solutions_only = miopen::FindMode(ctx).IsDynamicHybrid();
ctx.do_search = exhaustiveSearch;
ctx.save_srch_req = true;
ctx.general_compile_options = "";
ctx.SetStream(&handle);
ctx.SetBufs(bufs);
ctx.DetectRocm();
ctx.SetupFloats();
try
{
return FindAllImplicitGemmSolutions(ctx, invoke_ctx);
}
catch(miopen::Exception& ex)
{
MIOPEN_LOG_WE(ex.what());
return {};
}
}
template <class InvokeParams>
static void EvaluateInvokers(Handle& handle,
const std::vector<solver::ConvSolution>& solutions,
const AlgorithmName& algorithm_name,
const NetworkConfig& network_config,
const InvokeParams& invoke_ctx,
DbRecord& record)
{
const char* const arch = miopen::GetStringEnv(MIOPEN_DEVICE_ARCH{});
if(arch != nullptr && strlen(arch) > 0)
{
return;
}
miopen::solver::ConvSolution selected{miopenStatusUnknownError};
float best = std::numeric_limits<float>::max();
Invoker best_invoker;
for(const auto& sol : solutions)
{
if(sol.workspce_sz > 0)
{
if(invoke_ctx.workSpace == nullptr)
{
MIOPEN_LOG_I("Warning: skipping solver <" << sol.solver_id
<< "> due to no workspace provided ("
<< sol.workspce_sz
<< " required)");
continue;
}
if(invoke_ctx.workSpaceSize < sol.workspce_sz)
{
MIOPEN_LOG_I("Warning: skipping solver <" << sol.solver_id
<< "> due to insufficient workspace ("
<< invoke_ctx.workSpaceSize
<< " < "
<< sol.workspce_sz
<< ")");
continue;
}
}
if(!sol.invoker_factory)
MIOPEN_THROW("Invoker is not provided by solver " + sol.solver_id);
const auto invoker = handle.PrepareInvoker(*sol.invoker_factory, sol.construction_params);
invoker(handle, invoke_ctx);
const auto elapsed = handle.GetKernelTime();
MIOPEN_LOG_I(sol << ": " << elapsed << (elapsed < best ? " < " : " >= ") << best);
if(elapsed < best)
{
best = elapsed;
selected = sol;
best_invoker = invoker;
}
}
if(selected.Succeeded())
{
handle.RegisterInvoker(best_invoker, network_config, selected.solver_id, algorithm_name);
MIOPEN_LOG_I(
"Selected: " << selected << ": " << best << ", workspce_sz = " << selected.workspce_sz);
record.SetValues(algorithm_name,
FindDbData{selected.solver_id,
best,
selected.workspce_sz,
FindDbKCacheKey::MakeUnused(algorithm_name)});
}
}
static void DirConvFindCore(Handle& handle,
const TensorDescriptor& xDesc,
ConstData_t x,
const TensorDescriptor& wDesc,
ConstData_t w,
const TensorDescriptor& yDesc,
Data_t y,
Data_t workSpace,
size_t workSpaceSize,
const ConvolutionDescriptor& conv,
bool exhaustiveSearch,
DbRecord& record,
const ConvolutionContext& ctx,
bool use_winograd_only)
{
AutoEnableProfiling enableProfiling{handle};
ValidateGroupCount(xDesc, wDesc, conv);
#if MIOPEN_USE_GEMM
if(!use_winograd_only && !miopen::IsDisabled(MIOPEN_DEBUG_CONV_GEMM{}) &&
!(IsAnyBufferBF16(xDesc, yDesc, wDesc) && !IsUseRocBlas))
{ // GEMM algo
std::size_t in_n, in_c;
std::tie(in_n, in_c) = tie_pick<0, 1>()(xDesc.GetLengths());
std::size_t wei_k = wDesc.GetLengths()[0];
std::size_t spatial_dim = conv.GetSpatialDimension();
auto in_spatial = boost::adaptors::slice(xDesc.GetLengths(), 2, 2 + spatial_dim);
auto wei_spatial = boost::adaptors::slice(wDesc.GetLengths(), 2, 2 + spatial_dim);
auto out_spatial = boost::adaptors::slice(yDesc.GetLengths(), 2, 2 + spatial_dim);
float time_gemm = 0;
const bool time_precision = (!IsDisabled(MIOPEN_CONV_PRECISE_ROCBLAS_TIMING{}));
// Use transpose path 1x1, stride=2
if(conv.GetSpatialDimension() == 2 &&
miopen::all_of(wei_spatial, [](auto v) { return v == 1; }) &&
miopen::all_of(conv.GetConvPads(), [](auto v) { return v == 0; }) &&
miopen::all_of(conv.GetConvStrides(), [](auto v) { return v == 2; }))
{
size_t workspace_req = conv.ForwardGetWorkSpaceSizeGEMMTranspose(xDesc, yDesc);
if(workSpace != nullptr && workSpaceSize >= workspace_req)
{
if(conv.group_count > 1)
{
MIOPEN_LOG_FUNCTION("groupconv, 1x1 u2xv2");
}
else
{
MIOPEN_LOG_FUNCTION("convolution, 1x1 u2xv2");
}
// y = CNHW2NCHW(w * NCHW2CNHW(x))
transpose_NCHW2CNHW(handle,
in_n,
in_c,
in_spatial[0],
in_spatial[1],
out_spatial[0],
out_spatial[1],
x,
workSpace,
0,
0,
conv.GetConvStrides()[0],
conv.GetConvStrides()[1],
xDesc.GetType());
time_gemm = handle.GetKernelTime();
std::size_t out_spatial_size = std::accumulate(out_spatial.begin(),
out_spatial.end(),
std::size_t(1),
std::multiplies<std::size_t>());
std::size_t x_t_size = in_n * in_c * out_spatial_size;
std::size_t wksp_offset = 0;
if(wDesc.GetType() == miopenInt8)
{
wksp_offset = x_t_size;
transpose_packed_MN2NM(handle,
in_c,
static_cast<int>(in_n * out_spatial_size),
0,
wksp_offset,
workSpace,
workSpace,
xDesc.GetType());
time_gemm += handle.GetKernelTime();
x_t_size *= 2;
}
if((wDesc.GetType() == miopenInt8 || wDesc.GetType() == miopenInt8x4) &&
(yDesc.GetType() == miopenInt32 || yDesc.GetType() == miopenFloat))
x_t_size /= 4;
FindDbKCacheKey kcache_key;
GemmDescriptor gemm_desc =
conv.group_count > 1 ? CreateGemmDescriptorGroupConvCNHWFwd(
wDesc, xDesc, yDesc, conv.group_count)
: CreateGemmDescriptorConvCNHWFwd(wDesc, xDesc, yDesc);
miopenStatus_t gemm_status =
CallGemmTimeMeasure(handle,
gemm_desc,
w,
0,
workSpace,
wksp_offset,
workSpace,
x_t_size,
&kcache_key,
time_precision,
conv.group_count > 1 ? callGemmStridedBatched : callGemm);
time_gemm += handle.GetKernelTime();
transpose_CNHW2NCHW(handle,
in_n,
wei_k,
out_spatial[0],
out_spatial[1],
out_spatial[0],
out_spatial[1],
workSpace,
y,
x_t_size,
0,
1,
1,
yDesc.GetType());
time_gemm += handle.GetKernelTime();
if((wDesc.GetType() == miopenInt8 || wDesc.GetType() == miopenInt8x4) &&
yDesc.GetType() != miopenInt32)
{
TensorDescriptor ygemmDesc(miopenInt32, yDesc.GetLengths(), yDesc.GetStrides());
CastTensor(handle, &conv.lowp_quant, ygemmDesc, y, yDesc, y, 0, 0);
time_gemm += handle.GetKernelTime();
}
if(gemm_status == miopenStatusSuccess)
record.SetValues(
"miopenConvolutionFwdAlgoGEMM",
FindDbData{
"gemm", time_gemm, workspace_req, kcache_key}); // Todo: gemm solver id?
}
}
// 1x1_stride=1 with GEMM and zero workspace
else if(miopen::all_of(wei_spatial, [](auto v) { return v == 1; }) &&
miopen::all_of(conv.GetConvPads(), [](auto v) { return v == 0; }) &&
miopen::all_of(conv.GetConvStrides(), [](auto v) { return v == 1; }))
{
if(conv.group_count > 1)
{
MIOPEN_LOG_FUNCTION("groupconv, 1x1");
}
else
{
MIOPEN_LOG_FUNCTION("convolution, 1x1");
}
// y = w * x
FindDbKCacheKey kcache_key;
miopenStatus_t gemm_status = miopenStatusNotInitialized;
size_t workspace_req = 0;
if(wDesc.GetType() == miopenInt8)
{
workspace_req = conv.ForwardGetWorkSpaceSizeGEMM(wDesc, yDesc);
GemmDescriptor gemm_desc = CreateGemmDescriptorConvFwd(wDesc, xDesc, yDesc);
std::size_t out_offset = 0;
std::size_t in_offset = 0;
std::size_t in_spatial_size = std::accumulate(in_spatial.begin(),
in_spatial.end(),
std::size_t(1),
std::multiplies<std::size_t>());
transpose_packed_MN2NM(
handle, in_c, in_spatial_size, in_offset, 0, x, workSpace, xDesc.GetType());
time_gemm += (in_n * handle.GetKernelTime());
gemm_status = CallGemmTimeMeasure(handle,
gemm_desc,
w,
0,
workSpace,
0,
y,
out_offset,
&kcache_key,
time_precision,
callGemm);
time_gemm += (in_n * handle.GetKernelTime());
}
else
{
GemmDescriptor gemm_desc =
conv.group_count > 1
? CreateGemmDescriptorGroupConvFwd(wDesc, xDesc, yDesc, conv.group_count)
: CreateGemmStridedBatchedDescriptorConv1x1Fwd(wDesc, xDesc, yDesc);
gemm_status = CallGemmTimeMeasure(handle,
gemm_desc,
w,
0,
x,
0,
y,
0,
&kcache_key,
time_precision,
callGemmStridedBatched);
time_gemm = handle.GetKernelTime();
if(conv.group_count > 1)
time_gemm *= in_n;
}
if((wDesc.GetType() == miopenInt8 || wDesc.GetType() == miopenInt8x4) &&
yDesc.GetType() != miopenInt32)
{
TensorDescriptor ygemmDesc(miopenInt32, yDesc.GetLengths(), yDesc.GetStrides());
CastTensor(handle, &conv.lowp_quant, ygemmDesc, y, yDesc, y, 0, 0);
time_gemm += handle.GetKernelTime();
}
if(gemm_status == miopenStatusSuccess)
record.SetValues(
"miopenConvolutionFwdAlgoGEMM",
FindDbData{
"gemm", time_gemm, workspace_req, kcache_key}); // Todo: gemm solver id?
}
// if not 1x1
else if(workSpace != nullptr &&
workSpaceSize >= (conv.ForwardGetWorkSpaceSizeGEMM(wDesc, yDesc)))
{
if(conv.group_count > 1)
{
MIOPEN_LOG_FUNCTION("groupconv, non 1x1");
}
else
{
MIOPEN_LOG_FUNCTION("convolution, non 1x1");
}
// y = w * Im2Col(x)
float time_im2col = 0;
int in_offset = 0;
time_im2col = Im2ColGPU(handle,
conv.GetSpatialDimension(),
x,
in_offset,
in_c,
in_spatial,
wei_spatial,
out_spatial,
conv.GetConvPads(),
conv.GetConvStrides(),
conv.GetConvDilations(),
workSpace,
xDesc.GetType());
std::size_t wksp_offset = 0;
if(wDesc.GetType() == miopenInt8)
{
std::size_t wei_spatial_size = std::accumulate(wei_spatial.begin(),
wei_spatial.end(),
std::size_t(1),
std::multiplies<std::size_t>());
std::size_t out_spatial_size = std::accumulate(out_spatial.begin(),
out_spatial.end(),
std::size_t(1),
std::multiplies<std::size_t>());
wksp_offset = in_c * wei_spatial_size * out_spatial_size;
transpose_packed_MN2NM(handle,
static_cast<int>(in_c * wei_spatial_size),
out_spatial_size,
0,
wksp_offset,
workSpace,
workSpace,
xDesc.GetType());
time_gemm += (in_n * handle.GetKernelTime());
}
FindDbKCacheKey kcache_key;
GemmDescriptor gemm_desc =
conv.group_count > 1
? CreateGemmDescriptorGroupConvFwd(wDesc, xDesc, yDesc, conv.group_count)
: CreateGemmDescriptorConvFwd(wDesc, xDesc, yDesc);
miopenStatus_t gemm_status = CallGemmTimeMeasure(
handle,
gemm_desc,
w,
0,
workSpace,
wksp_offset,
y,
0,
&kcache_key,
time_precision,
conv.group_count > 1 ? callGemmStridedBatched : callGemm,
(conv.group_count > 1 || wDesc.GetType() == miopenInt8 ||
wDesc.GetType() == miopenInt8x4 || wDesc.GetType() == miopenBFloat16)
? GemmBackend_t::rocblas
: GemmBackend_t::miopengemm);
time_gemm += (in_n * (time_im2col + handle.GetKernelTime()));
if((wDesc.GetType() == miopenInt8 || wDesc.GetType() == miopenInt8x4) &&
yDesc.GetType() != miopenInt32)
{
TensorDescriptor ygemmDesc(miopenInt32, yDesc.GetLengths(), yDesc.GetStrides());
CastTensor(handle, &conv.lowp_quant, ygemmDesc, y, yDesc, y, 0, 0);
time_gemm += handle.GetKernelTime();
}
if(gemm_status == miopenStatusSuccess)
record.SetValues("miopenConvolutionFwdAlgoGEMM",
FindDbData{"gemm",
time_gemm,
(conv.ForwardGetWorkSpaceSizeGEMM(wDesc, yDesc)),
kcache_key}); // Todo: gemm solver id?
}
}
#endif
const auto network_config = ctx.BuildConfKey();
const auto invoke_ctx =
conv::DataInvokeParams{{xDesc, x, wDesc, w, yDesc, y}, workSpace, workSpaceSize};
// Winograd algo
{
const auto all = conv.FindWinogradSolutions(ctx, invoke_ctx);
PrecompileSolutions(handle, all);
const auto algorithm_name = AlgorithmName{"miopenConvolutionFwdAlgoWinograd"};
EvaluateInvokers(handle, all, algorithm_name, network_config, invoke_ctx, record);
}
// Direct algo
if(!use_winograd_only)
{
ConvolutionUserBuffers bufs(workSpace, workSpaceSize);
bufs.SetFwd(x, w, y);
const auto all = conv.FindDataDirectSolutions(
handle, xDesc, wDesc, yDesc, exhaustiveSearch, true, bufs, invoke_ctx);
PrecompileSolutions(handle, all);
const auto algorithm_name = AlgorithmName{"miopenConvolutionFwdAlgoDirect"};
EvaluateInvokers(handle, all, algorithm_name, network_config, invoke_ctx, record);
}
// Implicit GEMM algo
if(!use_winograd_only)
{
ConvolutionUserBuffers bufs(workSpace, workSpaceSize);
bufs.SetFwd(x, w, y);
const auto all = conv.FindDataImplicitGemmSolutions(
handle, xDesc, wDesc, yDesc, exhaustiveSearch, true, bufs, invoke_ctx);
PrecompileSolutions(handle, all);
const auto algorithm_name = AlgorithmName{"miopenConvolutionFwdAlgoImplicitGEMM"};
EvaluateInvokers(handle, all, algorithm_name, network_config, invoke_ctx, record);
}
// FFT algo
if(!use_winograd_only && conv.GetSpatialDimension() == 2 &&
miopen::all_of(conv.GetConvDilations(), [](auto v) { return v == 1; }) &&
conv.group_count == 1 && wDesc.GetType() != miopenInt8 && wDesc.GetType() != miopenInt8x4)
{
std::vector<KernelInvoke> kernels_fft;
size_t workspace_fft = conv.ForwardGetWorkSpaceSizeFFT(wDesc, xDesc, yDesc);
if(conv.FindFwdFFTKernel(
handle, xDesc, wDesc, yDesc, workspace_fft, kernels_fft, network_config) == 0)
{
(void)kernels_fft; // not used now, but needed as fft coverage widens
if(workSpace != nullptr && workSpaceSize >= workspace_fft)
{
float time_fft = conv.ExecuteFwdFFTKernel(handle,
xDesc,
x,
wDesc,
w,
yDesc,
y,
workSpace,
workSpaceSize,
network_config,
true);
record.SetValues("miopenConvolutionFwdAlgoFFT",
FindDbData{"fft",
time_fft,
workspace_fft,
{"miopenConvolutionFwdAlgoFFT",
network_config}}); // Todo: fft solver id?
}
}
}
}
void ConvolutionDescriptor::FindConvFwdAlgorithm(Handle& handle,
const TensorDescriptor& xDesc,
ConstData_t x,
const TensorDescriptor& wDesc,
ConstData_t w,
const TensorDescriptor& yDesc,
Data_t y,
const int requestAlgoCount,
int* const returnedAlgoCount,
miopenConvAlgoPerf_t* perfResults,
Data_t workSpace,
size_t workSpaceSize,
bool exhaustiveSearch) const
{
MIOPEN_LOG_I("requestAlgoCount = " << requestAlgoCount << ", workspace = " << workSpaceSize);
if(x == nullptr || w == nullptr || y == nullptr)
MIOPEN_THROW(miopenStatusBadParm, "Buffers cannot be NULL");
if(returnedAlgoCount == nullptr)
MIOPEN_THROW(miopenStatusBadParm, "returnedAlgoCount cannot be nullptr");
if(perfResults == nullptr)
MIOPEN_THROW(miopenStatusBadParm, "perfResults cannot be nullptr");
if(requestAlgoCount < 1)
MIOPEN_THROW(miopenStatusBadParm, "requestAlgoCount cannot be < 1");
*returnedAlgoCount = 0;
const ProblemDescription problem(xDesc, wDesc, yDesc, *this, conv::Direction::Forward);
auto ctx = ConvolutionContext{problem};
ctx.SetStream(&handle);
ctx.DetectRocm();
ConvolutionUserBuffers bufs(workSpace, workSpaceSize);
bufs.SetFwd(x, w, y);
ctx.SetBufs(bufs);
const bool use_winograd_only = IsWinograd3x3SupportedAndFast(ctx);
std::vector<PerfField> perf_db;
const miopen::FindMode fm(ctx);
/// \section ffind_special_cases
/// Fast Find mode: Let's allow known fast-to-build special cases
/// (this is only Winograd 3x3 so far) to override switching to Immediate mode.
/// This minimizes performance drop in Fast Find mode at for free.
/// Otherwise we can hit Immediate mode fallback (which is just GEMM
/// right now) in many cases. -- atamazov 21 Nov 2019.
///
/// \todo Revise this (and similar cases) when Immediate mode
/// will be better elaborated.
bool use_immediate_solution = false;
miopenConvSolution_t sol;
if((fm.IsFast() || fm.IsHybrid()) && !use_winograd_only)
{
size_t count;
GetForwardSolutions(handle, wDesc, xDesc, yDesc, 1, &count, &sol);
use_immediate_solution = (count > 0) && !(fm.IsHybrid() && sol.time < 0);
// In Hybrid Find mode, we use Normal Find instead of Immediate fallback kernels.
}
if(use_immediate_solution)
{
CompileForwardSolution(handle, wDesc, xDesc, yDesc, sol.solution_id);
/// It is possible to measure actual execution time and return it to the caller.
/// \todo Consider if we need (and want to spend time) for this.
const auto id = solver::Id(sol.solution_id);
perf_db.push_back(
{id.GetAlgo(conv::Direction::Forward), id.ToString(), sol.time, sol.workspace_size});
}
else
{
ctx.skip_solutions_that_take_long_time_to_build_and_have_narrow_coverage =
miopen::FindMode(ctx).IsFastHybrid();
ctx.use_dynamic_solutions_only = miopen::FindMode(ctx).IsDynamicHybrid();
perf_db = UserFindDbRecord::TryLoad(handle, problem, [&](DbRecord& record) {
DirConvFindCore(handle,
xDesc,
x,
wDesc,
w,
yDesc,
y,
workSpace,
workSpaceSize,
*this,
exhaustiveSearch,
record,
ctx,
use_winograd_only);
});
}
if(perf_db.empty())
MIOPEN_THROW("Forward Convolution cannot be executed due to incorrect params");
std::sort(begin(perf_db), end(perf_db));
for(const auto& entry : perf_db)
MIOPEN_LOG_I(entry.name << "\t" << entry.time << "\t" << entry.workspace);
*returnedAlgoCount = std::min(requestAlgoCount, static_cast<int>(perf_db.size()));
for(int i = 0; i < *returnedAlgoCount; i++)
{
perfResults[i].fwd_algo = StringToConvolutionFwdAlgo(perf_db[i].name);
perfResults[i].time = perf_db[i].time;
perfResults[i].memory = perf_db[i].workspace;
}
MIOPEN_LOG_I("FW Chosen Algorithm: " << perf_db[0].solver_id << " , " << perf_db[0].workspace
<< ", "
<< perf_db[0].time);
}
void ValidateConvTensors(const ConvTensors& tensors)
{
const auto invalid_buffers =
tensors.x == nullptr || tensors.w == nullptr || tensors.y == nullptr;
const auto tensor_sizes_not_matched = tensors.xDesc.GetSize() != tensors.yDesc.GetSize() ||
tensors.xDesc.GetSize() != tensors.wDesc.GetSize();
const auto tensor_types_not_matched =
(tensors.xDesc.GetType() != tensors.yDesc.GetType() &&
tensors.xDesc.GetType() != miopenInt8 && tensors.xDesc.GetType() != miopenInt8x4) ||
tensors.xDesc.GetType() != tensors.wDesc.GetType();
// if(xDesc.GetLengths()[1] != wDesc.GetLengths()[1]) {
// MIOPEN_THROW(miopenStatusBadParm);
//}
const auto x_tensor_invalid = tensors.xDesc.GetSize() < 3;
const auto bad_parameters =
invalid_buffers || tensor_sizes_not_matched || tensor_types_not_matched || x_tensor_invalid;
if(bad_parameters)
MIOPEN_THROW(miopenStatusBadParm);
}
void ValidateAlphaBeta(const void* alpha, const void* beta)
{
if(!float_equal(*(static_cast<const float*>(alpha)), 1.0) ||
!float_equal(*(static_cast<const float*>(beta)), 0))
{
MIOPEN_THROW(miopenStatusNotImplemented, "Only alpha=1 and beta=0 is supported");
}
}
static void ConvForwardCheckNumerics(const Handle& handle,
const ConvFwdTensors& tensors,
std::function<void()>&& worker)
{
if(!miopen::CheckNumericsEnabled())
{
worker();
return;
}
miopen::checkNumericsInput(handle, tensors.xDesc, tensors.x);
miopen::checkNumericsInput(handle, tensors.wDesc, tensors.w);
worker();
miopen::checkNumericsOutput(handle, tensors.yDesc, tensors.y);
}
void ConvolutionDescriptor::ConvolutionForward(Handle& handle,
const void* alpha,
const TensorDescriptor& xDesc,
ConstData_t x,
const TensorDescriptor& wDesc,
ConstData_t w,
miopenConvFwdAlgorithm_t algo,
const void* beta,
const TensorDescriptor& yDesc,
Data_t y,
Data_t workSpace,
size_t workSpaceSize) const
{
MIOPEN_LOG_I("algo = " << algo << ", workspace = " << workSpaceSize);
const auto tensors = ConvFwdTensors{xDesc, x, wDesc, w, yDesc, y};
ValidateConvTensors(tensors);
ValidateAlphaBeta(alpha, beta);
if(algo != miopenConvolutionFwdAlgoGEMM &&
(xDesc.GetType() == miopenInt8 || xDesc.GetType() == miopenInt8x4))
{
MIOPEN_THROW(miopenStatusBadParm);
}
ConvForwardCheckNumerics(handle, tensors, [&]() {
ValidateGroupCount(xDesc, wDesc, *this);
const auto algorithm_name = AlgorithmName{ConvolutionAlgoToDirectionalString(
static_cast<miopenConvAlgorithm_t>(algo), conv::Direction::Forward)};
auto ctx =
ConvolutionContext{xDesc, wDesc, yDesc, *this, conv::Direction::Forward}; // forward
ctx.SetStream(&handle);
const auto network_config = ctx.BuildConfKey();
const auto& invoker = handle.GetInvoker(network_config, {}, algorithm_name);
if(invoker)
{
const auto& invoke_ctx = conv::DataInvokeParams{tensors, workSpace, workSpaceSize};
(*invoker)(handle, invoke_ctx);
return;
}
switch(algo)
{
case miopenConvolutionFwdAlgoDirect:
case miopenConvolutionFwdAlgoWinograd:
case miopenConvolutionFwdAlgoImplicitGEMM:
MIOPEN_THROW("No invoker was registered for convolution forward. Was find executed?");
case miopenConvolutionFwdAlgoGEMM:
ConvFwdGemm(handle, tensors, workSpace, workSpaceSize);
break;
case miopenConvolutionFwdAlgoFFT:
ConvFwdFFT(handle, tensors, workSpace, workSpaceSize, network_config);
break;
}
});
}
void ConvolutionDescriptor::ConvFwdGemm(Handle& handle,
const ConvFwdTensors& tensors,
Data_t workSpace,
std::size_t workSpaceSize) const
{
#if MIOPEN_USE_GEMM
if(miopen::IsDisabled(MIOPEN_DEBUG_CONV_GEMM{}))
{
MIOPEN_THROW("GEMM convolution is disabled");
}
if(IsAnyBufferBF16(tensors.xDesc, tensors.yDesc, tensors.wDesc) && !IsUseRocBlas)
{
MIOPEN_THROW("GEMM convolution is unsupported");
}
std::size_t in_n, in_c;
std::tie(in_n, in_c) = tie_pick<0, 1>()(tensors.xDesc.GetLengths());
std::size_t wei_k = tensors.wDesc.GetLengths()[0];
std::size_t spatial_dim = GetSpatialDimension();
auto in_spatial = boost::adaptors::slice(tensors.xDesc.GetLengths(), 2, 2 + spatial_dim);
auto wei_spatial = boost::adaptors::slice(tensors.wDesc.GetLengths(), 2, 2 + spatial_dim);
auto out_spatial = boost::adaptors::slice(tensors.yDesc.GetLengths(), 2, 2 + spatial_dim);
// Use transpose path for 1x1, stride=2
if(GetSpatialDimension() == 2 && miopen::all_of(wei_spatial, [](auto v) { return v == 1; }) &&
miopen::all_of(GetConvPads(), [](auto v) { return v == 0; }) &&
miopen::all_of(GetConvStrides(), [](auto v) { return v == 2; }))
{
if(group_count > 1)
{
MIOPEN_LOG_FUNCTION("groupconv, 1x1 u2xv2");
}