forked from NVIDIA/cutlass
-
Notifications
You must be signed in to change notification settings - Fork 20
/
gemm_grouped.cu
1578 lines (1256 loc) · 49.7 KB
/
gemm_grouped.cu
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 (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* 3. Neither the name of the copyright holder nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
/*! \file
\brief GEMM Grouped Example.
This workload computes a batch of GEMM operations with distinct problem sizes. Pointers to matrices
in Global Memory are passed to the kernel in array (also held in Global Memory). Similarly,
leading dimensions and problem sizes are stored in arrays in GMEM.
This differs from "Batched Array" GEMM because the size of each GEMM problem in the Grouped GEMM
concept may be distinct.
This benchmark program initializes a workspace with random problem sizes for a given number of
groups. Command line options enable overriding M, N, and/or K dimensions with uniform values to
model problems more similar to the traditional batched GEMM.
Additionally, problem sizes are collected and binned to compute the same problem as a series of
conventional batched GEMMs (setup for this problem is not timed). This demonstrates the performance
enhancement achieved by implementing a specialized grouped GEMM kernel.
Examples:
# Runs a grouped GEMM with 100 random problem sizes
$ ./examples/24_gemm_grouped/24_gemm_grouped --groups=100
# Runs a grouped GEMM with 100 random problem sizes (with GEMM-K dimension equal to 1024)
$ ./examples/24_gemm_grouped/24_gemm_grouped --groups=100 --k=1024 --verbose=true
# Runs a grouped GEMM that is equivalent to a batched GEMM
$ ./examples/24_gemm_grouped/24_gemm_grouped --groups=100 --m=2048 --n=1024 --k=1024 --verbose=true
# Execute Grouped GEMM and profile with NSight
$ nv-nsight-cu-cli ./examples/24_gemm_grouped/24_gemm_grouped --m=256 --n=256 --k=256 --verbose=true \
--iterations=1 --reference-check=false
*/
/////////////////////////////////////////////////////////////////////////////////////////////////
#include <chrono>
#include <iostream>
#include <fstream>
#include <sstream>
#include <vector>
#include <map>
#include <unordered_map>
#include "cutlass/cutlass.h"
#include "cutlass/gemm/gemm.h"
#include "cutlass/gemm/kernel/gemm_grouped.h"
#include "cutlass/gemm/kernel/default_gemm_grouped.h"
#include "cutlass/gemm/device/gemm_grouped.h"
#include "cutlass/gemm/device/gemm_universal.h"
#include "cutlass/util/command_line.h"
#include "cutlass/util/distribution.h"
#include "cutlass/util/device_memory.h"
#include "cutlass/util/tensor_view_io.h"
#include "cutlass/util/host_tensor.h"
#include "cutlass/util/reference/host/gemm_complex.h"
#include "cutlass/util/reference/device/gemm_complex.h"
#include "cutlass/util/reference/host/tensor_compare.h"
#include "cutlass/util/reference/host/tensor_copy.h"
#include "cutlass/util/reference/device/tensor_fill.h"
#include "cutlass/util/reference/host/tensor_norm.h"
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Result structure
struct Result {
double runtime_ms;
double initialization_time_ms;
double gflops;
cutlass::Status status;
cudaError_t error;
bool passed;
//
// Methods
//
Result(
double runtime_ms = 0,
double initialization_time_ms = 0,
double gflops = 0,
cutlass::Status status = cutlass::Status::kSuccess,
cudaError_t error = cudaSuccess
):
runtime_ms(runtime_ms), initialization_time_ms(initialization_time_ms), gflops(gflops),
status(status), error(error), passed(true) { }
};
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Hash function for cutlass::gemm::GemmCoord
struct HashGemmCoord {
size_t operator()(cutlass::gemm::GemmCoord const &problem) const {
std::hash<int> hasher;
return (hasher(problem.m() * 3)) ^ (hasher(1 + problem.n() * 5)) ^ (hasher(2 + problem.k() * 7));
}
};
/////////////////////////////////////////////////////////////////////////////////////////////////
// Command line options parsing
struct Options {
bool help;
bool error;
bool reference_check;
bool profile_initialization;
bool sort_problems;
std::vector<cutlass::gemm::GemmCoord> problem_sizes;
// problem size bins
std::unordered_map<
cutlass::gemm::GemmCoord,
std::vector<int32_t>,
HashGemmCoord> problem_bins;
int alignment;
int problem_count;
int iterations;
int cuda_streams;
bool verbose;
float alpha;
float beta;
std::string benchmark_path;
std::string output_tag;
std::ofstream output_file;
using GroupScheduleMode = cutlass::gemm::kernel::GroupScheduleMode;
std::vector<GroupScheduleMode> scheduler_modes;
std::unordered_map<std::string, GroupScheduleMode>
str_to_scheduler_mode = {
{"kDeviceOnly", GroupScheduleMode::kDeviceOnly},
{"kHostPrecompute", GroupScheduleMode::kHostPrecompute}
};
struct GroupScheduleModeHash {
size_t operator()(GroupScheduleMode m) const {
return static_cast<size_t>(m);
}
};
std::unordered_map<GroupScheduleMode, std::string, GroupScheduleModeHash>
scheduler_mode_to_str = {
{GroupScheduleMode::kDeviceOnly, "kDeviceOnly"},
{GroupScheduleMode::kHostPrecompute, "kHostPrecompute"}
};
std::vector<GroupScheduleMode> all_scheduler_modes = {GroupScheduleMode::kDeviceOnly, GroupScheduleMode::kHostPrecompute};
//
// Methods
//
Options():
help(false),
error(false),
alignment(8),
reference_check(true),
profile_initialization(false),
sort_problems(false),
problem_count(15),
iterations(20),
cuda_streams(0),
verbose(false),
alpha(1),
beta(),
scheduler_modes({GroupScheduleMode::kDeviceOnly})
{ }
// Parses the command line
void parse(int argc, char const **args) {
cutlass::CommandLine cmd(argc, args);
if (cmd.check_cmd_line_flag("help")) {
help = true;
return;
}
cmd.get_cmd_line_argument("alignment", alignment, 8);
cmd.get_cmd_line_argument("groups", problem_count, 15);
cmd.get_cmd_line_argument("alpha", alpha, 1.0f);
cmd.get_cmd_line_argument("beta", beta, 0.0f);
cmd.get_cmd_line_argument("iterations", iterations, 20);
cmd.get_cmd_line_argument("streams", cuda_streams, 0);
cmd.get_cmd_line_argument("verbose", verbose, false);
cmd.get_cmd_line_argument("reference-check", reference_check, true);
cmd.get_cmd_line_argument("profile-initialization", profile_initialization, false);
cmd.get_cmd_line_argument("sort-problems", sort_problems, false);
cmd.get_cmd_line_argument("benchmark", benchmark_path);
std::vector<std::string> scheduler_mode_strs;
cmd.get_cmd_line_arguments("scheduler-modes", scheduler_mode_strs);
if (!scheduler_mode_strs.empty()) {
scheduler_modes.clear();
if (scheduler_mode_strs.size() == 1 && scheduler_mode_strs[0] == "all") {
scheduler_modes = all_scheduler_modes;
} else {
for (std::string precomp_str : scheduler_mode_strs) {
auto it = str_to_scheduler_mode.find(precomp_str);
if (it != str_to_scheduler_mode.end()) {
scheduler_modes.push_back(it->second);
} else if (precomp_str == "all") {
std::cerr << "Flag --scheduler-modes=all must not contain other scheduler modes in list." << std::endl;
error = true;
return;
} else {
std::cerr << "Unrecognized scheduler mode '" << precomp_str << "'" << std::endl;
error = true;
return;
}
}
}
}
std::string output_path;
cmd.get_cmd_line_argument("tag", output_tag);
cmd.get_cmd_line_argument("output_file", output_path);
if (!output_path.empty()) {
std::ios_base::openmode open_mode = std::ios_base::out;
std::ifstream input_file(output_path.c_str());
if (input_file.good()) {
open_mode = std::ios_base::app;
input_file.close();
}
output_file.open(output_path.c_str(), open_mode);
if (output_file.good() && open_mode != std::ios_base::app) {
output_file << "Tag,Provider,Kind,Groups,Runtime,GFLOPs\n";
}
}
// Decide how to initialize the problems
if (!benchmark_path.empty()) {
if (!benchmark_problems()) {
error = true;
problem_sizes.clear();
return;
}
}
else {
randomize_problems(cmd);
}
// Post-process the problem sizes
bin_problems();
}
void randomize_problems(cutlass::CommandLine &cmd) {
//
// For now, randomly choose the problem sizes.
//
int cmd_line_m = -1;
int cmd_line_n = -1;
int cmd_line_k = -1;
cmd.get_cmd_line_argument("m", cmd_line_m);
cmd.get_cmd_line_argument("n", cmd_line_n);
cmd.get_cmd_line_argument("k", cmd_line_k);
problem_sizes.reserve(problem_count);
for (int i = 0; i < problem_count; ++i) {
int m = cmd_line_m;
int n = cmd_line_n;
int k = cmd_line_k;
if (m < 1) {
m = alignment * ((rand() % 256) + 1);
}
if (n < 1) {
n = alignment * ((rand() % 256) + 1);
}
if (k < 1) {
k = alignment * ((rand() % 256) + 1);
}
cutlass::gemm::GemmCoord problem(m, n, k);
problem_sizes.push_back(problem);
}
}
/// Load a benchmark
bool benchmark_problems() {
std::ifstream file(benchmark_path);
if (!file.good()) {
return false;
}
while (file.good()) {
int idx = -1;
std::string extent_str;
file >> idx >> extent_str;
if (idx < 0 || extent_str.empty()) {
break;
}
cutlass::gemm::GemmCoord extent;
std::vector<std::string> tokens;
cutlass::CommandLine::tokenize(tokens, extent_str, 'x');
for (int i = 0; i < int(tokens.size()); ++i) {
int x = std::atoi(tokens.at(i).c_str());
// round up
if (x % alignment) {
x += (alignment - (x % alignment));
}
extent.at(i) = x;
}
if (extent.product()) {
problem_sizes.push_back(extent);
}
}
return true;
}
/// Post processes the problems
void bin_problems() {
problem_bins.clear();
problem_count = int(problem_sizes.size());
//
// Insert the problem sizes into a sorted container class. This is *NOT* necessary
// to run the CUTLASS kernel, but it enables the execution of cublas's batched GEMM.
//
for (int i = 0; i < int(problem_sizes.size()); ++i) {
auto it = problem_bins.find(problem_sizes.at(i));
if (it == problem_bins.end()) {
problem_bins.insert({problem_sizes.at(i), std::vector<int32_t>({i}) });
}
else {
it->second.push_back(i);
}
}
}
/// Prints the usage statement.
std::ostream & print_usage(std::ostream &out) const {
out << "24_gemm_grouped\n\n"
<< " This example profiles the performance of a 'grouped' GEMM kernel. This is similar to batched GEMM\n"
<< " in that multiple, independent GEMMs are computed by one grid launch. It differs in that each\n"
<< " 'group' may compute a unique problem size. Problem sizes and pointers to matrices are both stored\n"
<< " in device Global Memory and loaded by the kernel.\n\n"
<< "Options:\n\n"
<< " --help If specified, displays this usage statement.\n\n"
<< " --benchmark=<str> Executes a benchmark problem size.\n"
<< " --output_file=<str> Path to a CSV file to output results. If it exists already, results are appended.\n"
<< " --tag=<str> String tag to prepend to the CSV file.\n"
<< " --groups=<int> Number of individual GEMM problems (default: --groups=15)\n"
<< " --m=<int> Sets the M dimension for all groups. Otherwise, it is selected randomly\n"
<< " --n=<int> Sets the N dimension for all groups. Otherwise, it is selected randomly\n"
<< " --k=<int> Sets the K dimension for all groups. Otherwise, it is selected randomly\n"
<< " --alpha=<f32> Epilogue scalar alpha (real part)\n"
<< " --beta=<f32> Epilogue scalar beta (real part)\n"
<< " --scheduler-modes=<str> List of scheduler modes to be profile for grouped GEMM scheduler (default: --scheduler_modes=kDeviceOnly)\n"
<< " --iterations=<int> Number of profiling iterations to perform.\n"
<< " --reference-check=<bool> If true, performs reference check.\n"
<< " --verbose=<bool> If true, prints problem sizes and batching structure.\n"
<< " --profile-initialization=<bool> If true, profiles the device-level kernel's initialization.\n"
<< " --sort-problems=<bool> If true, sorts problem sizes in descending order of GEMM-K dimension.\n";
out << "\n\nExamples:\n\n"
<< "# Runs a grouped GEMM with 100 random problem sizes\n"
<< "$ ./examples/24_gemm_grouped/24_gemm_grouped --groups=100\n\n"
<< "# Runs a grouped GEMM with 100 random problem sizes (with GEMM-K dimension equal to 1024)\n"
<< "$ ./examples/24_gemm_grouped/24_gemm_grouped --groups=100 --k=1024 --verbose=true\n\n"
<< "# Runs a grouped GEMM that is equivalent to a batched GEMM\n"
<< "$ ./examples/24_gemm_grouped/24_gemm_grouped --groups=100 --m=2048 --n=1024 --k=1024 --verbose=true\n\n"
<< "# Runs a grouped GEMM with each different scheduler mode\n"
<< "$ ./examples/24_gemm_grouped/24_gemm_grouped --scheduler-modes=all\n\n"
<< "# Runs a grouped GEMM with each different scheduler mode and profiles host-side initialization time\n"
<< "$ ./examples/24_gemm_grouped/24_gemm_grouped --scheduler-modes=all --profile-initialization=true\n\n"
<< "# Runs a grouped GEMM problem given an externally supplied benchmark file. This is a text file in which\n"
<< "# Each line contains a unique group index and an MxNxK triple indicating problemsize.\n"
<< "#\n"
<< "# For example, assume the following are the contents of 'problems.txt'\n"
<< "#\n"
<< "# 0 1024x256x520\n"
<< "# 1 520x264x1024\n"
<< "# 2 96x48x1024\n"
<< "#\n"
<< "$ ./examples/24_gemm_grouped/24_gemm_grouped --benchmark=problems.txt\n\n"
<< "# Execute Grouped GEMM and profile with NSight\n"
<< "$ nv-nsight-cu-cli ./examples/24_gemm_grouped/24_gemm_grouped --m=256 --n=256 --k=256 --verbose=true --iterations=1 --reference-check=false\n\n";
return out;
}
/// Compute performance in GFLOP/s
double gflops(double runtime_s) const {
// Number of real-valued multiply-adds
int64_t fmas = int64_t();
for (auto const & problem : problem_sizes) {
fmas += problem.product();
}
// Two flops per multiply-add
return 2.0 * double(fmas) / double(1.0e9) / runtime_s;
}
};
///////////////////////////////////////////////////////////////////////////////////////////////////
template <typename Gemm>
class BaseTestbed {
public:
//
// Type definitions
//
using ElementA = typename Gemm::ElementA;
using ElementB = typename Gemm::ElementB;
using ElementC = typename Gemm::ElementC;
using ElementAccumulator = typename Gemm::ElementAccumulator;
using EpilogueOutputOp = typename Gemm::GemmKernel::Epilogue::OutputOp;
using ElementCompute = typename EpilogueOutputOp::ElementCompute;
using LayoutA = typename Gemm::LayoutA;
using LayoutB = typename Gemm::LayoutB;
using LayoutC = typename Gemm::LayoutC;
using MatrixCoord = typename LayoutC::TensorCoord;
//
// Data members
//
Options & options;
/// Initialization
cutlass::Distribution::Kind init_A;
cutlass::Distribution::Kind init_B;
cutlass::Distribution::Kind init_C;
uint32_t seed;
cutlass::DeviceAllocation<cutlass::gemm::GemmCoord> problem_sizes_device;
std::vector<int64_t> offset_A;
std::vector<int64_t> offset_B;
std::vector<int64_t> offset_C;
std::vector<int64_t> offset_D;
std::vector<int64_t> lda_host;
std::vector<int64_t> ldb_host;
std::vector<int64_t> ldc_host;
std::vector<int64_t> ldd_host;
cutlass::DeviceAllocation<int64_t> lda;
cutlass::DeviceAllocation<int64_t> ldb;
cutlass::DeviceAllocation<int64_t> ldc;
cutlass::DeviceAllocation<int64_t> ldd;
cutlass::DeviceAllocation<ElementA> block_A;
cutlass::DeviceAllocation<ElementB> block_B;
cutlass::DeviceAllocation<ElementC> block_C;
cutlass::DeviceAllocation<ElementC> block_D;
cutlass::DeviceAllocation<ElementA *> ptr_A;
cutlass::DeviceAllocation<ElementB *> ptr_B;
cutlass::DeviceAllocation<ElementC *> ptr_C;
cutlass::DeviceAllocation<ElementC *> ptr_D;
BaseTestbed(
Options &options_,
cutlass::Distribution::Kind init_A_ = cutlass::Distribution::Uniform,
cutlass::Distribution::Kind init_B_ = cutlass::Distribution::Uniform,
cutlass::Distribution::Kind init_C_ = cutlass::Distribution::Uniform,
uint32_t seed_ = 3080
):
options(options_), init_A(init_A_), init_B(init_B_), init_C(init_C_), seed(seed_) { }
int problem_count() const {
return options.problem_count;
}
/// Helper to initialize a tensor view
template <typename Element>
void initialize_tensor(
Element *ptr,
size_t capacity,
cutlass::Distribution::Kind dist_kind,
uint32_t seed) {
if (dist_kind == cutlass::Distribution::Uniform) {
Element scope_max, scope_min;
int bits_input = cutlass::sizeof_bits<Element>::value;
int bits_output = cutlass::sizeof_bits<typename Gemm::ElementC>::value;
if (bits_input == 1) {
scope_max = 2;
scope_min = 0;
} else if (bits_input <= 8) {
scope_max = 2;
scope_min = -2;
} else if (bits_output == 16) {
if (cutlass::sizeof_bits<ElementAccumulator>::value <= 16) {
scope_max = 5;
scope_min = -5;
}
else {
scope_max = 8;
scope_min = -8;
}
} else {
scope_max = 8;
scope_min = -8;
}
cutlass::reference::device::BlockFillRandomUniform(
ptr, capacity, seed, scope_max, scope_min, 0);
}
else if (dist_kind == cutlass::Distribution::Gaussian) {
cutlass::reference::device::BlockFillRandomGaussian(
ptr, capacity, seed, Element(), Element(0.5f));
}
else if (dist_kind == cutlass::Distribution::Sequential) {
// Fill with increasing elements
cutlass::reference::device::BlockFillSequential(
ptr, capacity, Element(1), Element());
}
else {
// Fill with all 1s
cutlass::reference::device::BlockFillSequential(
ptr, capacity, Element(), Element(1));
}
}
/// Allocates device-side data
void allocate() {
int64_t total_elements_A = 0;
int64_t total_elements_B = 0;
int64_t total_elements_C = 0;
int64_t total_elements_D = 0;
lda_host.resize(problem_count());
ldb_host.resize(problem_count());
ldc_host.resize(problem_count());
ldd_host.resize(problem_count());
for (int32_t i = 0; i < problem_count(); ++i) {
auto problem = options.problem_sizes.at(i);
lda_host.at(i) = LayoutA::packed({problem.m(), problem.k()}).stride(0);
ldb_host.at(i) = LayoutB::packed({problem.k(), problem.n()}).stride(0);
ldc_host.at(i) = LayoutC::packed({problem.m(), problem.n()}).stride(0);
ldd_host.at(i) = LayoutC::packed({problem.m(), problem.n()}).stride(0);
offset_A.push_back(total_elements_A);
offset_B.push_back(total_elements_B);
offset_C.push_back(total_elements_C);
offset_D.push_back(total_elements_D);
int64_t elements_A = problem.m() * problem.k();
int64_t elements_B = problem.k() * problem.n();
int64_t elements_C = problem.m() * problem.n();
int64_t elements_D = problem.m() * problem.n();
total_elements_A += elements_A;
total_elements_B += elements_B;
total_elements_C += elements_C;
total_elements_D += elements_D;
}
lda.reset(problem_count());
ldb.reset(problem_count());
ldc.reset(problem_count());
ldd.reset(problem_count());
block_A.reset(total_elements_A);
block_B.reset(total_elements_B);
block_C.reset(total_elements_C);
block_D.reset(total_elements_D);
}
/// Initializes device-side data
void initialize() {
problem_sizes_device.reset(problem_count());
problem_sizes_device.copy_from_host(options.problem_sizes.data());
lda.copy_from_host(lda_host.data());
ldb.copy_from_host(ldb_host.data());
ldc.copy_from_host(ldc_host.data());
ldd.copy_from_host(ldd_host.data());
//
// Assign pointers
//
std::vector<ElementA *> ptr_A_host(problem_count());
std::vector<ElementB *> ptr_B_host(problem_count());
std::vector<ElementC *> ptr_C_host(problem_count());
std::vector<ElementC *> ptr_D_host(problem_count());
for (int32_t i = 0; i < problem_count(); ++i) {
ptr_A_host.at(i) = block_A.get() + offset_A.at(i);
ptr_B_host.at(i) = block_B.get() + offset_B.at(i);
ptr_C_host.at(i) = block_C.get() + offset_C.at(i);
ptr_D_host.at(i) = block_D.get() + offset_D.at(i);
}
ptr_A.reset(problem_count());
ptr_A.copy_from_host(ptr_A_host.data());
ptr_B.reset(problem_count());
ptr_B.copy_from_host(ptr_B_host.data());
ptr_C.reset(problem_count());
ptr_C.copy_from_host(ptr_C_host.data());
ptr_D.reset(problem_count());
ptr_D.copy_from_host(ptr_D_host.data());
//
// Initialize the problems of the workspace
//
initialize_tensor(block_A.get(), block_A.size(), init_A, seed * 2021);
initialize_tensor(block_B.get(), block_B.size(), init_B, seed * 2022);
initialize_tensor(block_C.get(), block_C.size(), init_C, seed * 2023);
cutlass::reference::device::BlockFillSequential(
block_D.get(), block_D.size(), ElementC(), ElementC());
}
/// Verifies the result is a GEMM
bool verify() {
bool passed = true;
for (int32_t i = 0; i < problem_count(); ++i) {
cutlass::gemm::GemmCoord problem = options.problem_sizes.at(i);
LayoutA layout_A(lda_host.at(i));
LayoutB layout_B(ldb_host.at(i));
LayoutC layout_C(ldc_host.at(i));
LayoutC layout_D(ldd_host.at(i));
MatrixCoord extent_A{problem.m(), problem.k()};
MatrixCoord extent_B{problem.k(), problem.n()};
MatrixCoord extent_C{problem.m(), problem.n()};
cutlass::TensorView<ElementA, LayoutA> view_A(block_A.get() + offset_A.at(i), layout_A, extent_A);
cutlass::TensorView<ElementB, LayoutB> view_B(block_B.get() + offset_B.at(i), layout_B, extent_B);
cutlass::TensorView<ElementC, LayoutC> view_C(block_C.get() + offset_C.at(i), layout_C, extent_C);
cutlass::DeviceAllocation<ElementC> block_Ref(layout_D.capacity(extent_C));
cutlass::TensorView<ElementC, LayoutC> view_Ref_device(block_Ref.get(), layout_D, extent_C);
// Reference GEMM
cutlass::reference::device::GemmComplex<
ElementA, LayoutA,
ElementB, LayoutB,
ElementC, LayoutC,
ElementCompute, ElementAccumulator
>(
problem,
options.alpha,
view_A,
Gemm::kTransformA,
view_B,
Gemm::kTransformB,
options.beta,
view_C,
view_Ref_device,
ElementAccumulator(0)
);
// Copy to host memory
std::vector<ElementC> matrix_D(layout_D.capacity(extent_C));
std::vector<ElementC> matrix_Ref(layout_D.capacity(extent_C));
cutlass::device_memory::copy_to_host(matrix_D.data(), block_D.get() + offset_D.at(i), matrix_D.size());
cutlass::device_memory::copy_to_host(matrix_Ref.data(), block_Ref.get(), matrix_D.size());
cutlass::TensorView<ElementC, LayoutC> view_D( matrix_D.data(), layout_D, extent_C);
cutlass::TensorView<ElementC, LayoutC> view_Ref(matrix_Ref.data(), layout_D, extent_C);
// Reference check
passed = cutlass::reference::host::TensorEquals(view_D, view_Ref);
if (!passed) {
std::cerr << "\n***\nError - problem " << i << " failed the QA check\n***\n" << std::endl;
return passed;
}
}
return passed;
}
};
template <typename Gemm>
class TestbedBatched : BaseTestbed<Gemm> {
public:
TestbedBatched(
Options &options_,
cutlass::Distribution::Kind init_A_ = cutlass::Distribution::Uniform,
cutlass::Distribution::Kind init_B_ = cutlass::Distribution::Uniform,
cutlass::Distribution::Kind init_C_ = cutlass::Distribution::Uniform,
uint32_t seed_ = 3080
): BaseTestbed<Gemm>(options_, init_A_, init_B_, init_C_, seed_) {}
void print_problem_sizes() {
std::cout << std::endl;
size_t bin_idx = 0;
size_t problem_count_check = 0;
std::cout << "Conventionally executed as " << this->options.problem_bins.size() << " batched GEMMs:\n";
for (auto const & bin : this->options.problem_bins) {
std::cout << " [" << bin_idx << "]: "
<< bin.first.m() << "-by-" << bin.first.n() << "-by-" << bin.first.k()
<< ", batch count: " << bin.second.size() << "\n";
++bin_idx;
problem_count_check += bin.second.size();
}
if (problem_count_check != size_t(this->problem_count())) {
std::cout << "\n***\nERROR in BINNING LOGIC!\n***\n" << std::endl;
}
std::cout << std::endl;
}
/// Executes a batched kernel and measures runtime
Result profile() {
std::cout << "Batched GEMM:\n"
<< "====================================================" << std::endl;
Result result;
result.passed = false;
// Initialize the problem
this->allocate();
this->initialize();
if (this->options.verbose) {
print_problem_sizes();
}
//
// Prepare batched GEMM environment
//
int32_t effective_streams = (this->options.cuda_streams ? this->options.cuda_streams : 1);
// Array of leading dimensions used by batched GEMM calls
std::vector<cutlass::gemm::GemmCoord> bin_problem_sizes;
std::vector<int32_t> bin_count;
std::vector<int32_t> bin_ldm_A;
std::vector<int32_t> bin_ldm_B;
std::vector<int32_t> bin_ldm_C;
std::vector<int32_t> bin_start;
std::vector<void const *> ptr_A_batched_host;
std::vector<void const *> ptr_B_batched_host;
std::vector<void *> ptr_C_batched_host;
for (auto const & bin : this->options.problem_bins) {
int first_idx = bin.second.front();
bin_problem_sizes.push_back(this->options.problem_sizes.at(first_idx));
bin_count.push_back(int32_t(bin.second.size()));
bin_ldm_A.push_back(static_cast<int32_t>(this->lda_host.at(first_idx)));
bin_ldm_B.push_back(static_cast<int32_t>(this->ldb_host.at(first_idx)));
bin_ldm_C.push_back(static_cast<int32_t>(this->ldc_host.at(first_idx)));
if (ptr_A_batched_host.size() % 2) {
ptr_A_batched_host.push_back(nullptr);
ptr_B_batched_host.push_back(nullptr);
ptr_C_batched_host.push_back(nullptr);
}
bin_start.push_back(int32_t(ptr_A_batched_host.size()));
for (int idx : bin.second) {
if (bin_problem_sizes.back() != this->options.problem_sizes.at(idx)) {
std::cerr << "Error - failed to group problems.\n";
return result;
}
if (bin_ldm_A.back() != this->lda_host.at(idx)) {
std::cerr << "Error - failed to group problems.\n";
return result;
}
if (bin_ldm_B.back() != this->ldb_host.at(idx)) {
std::cerr << "Error - failed to group problems.\n";
return result;
}
if (bin_ldm_C.back() != this->ldc_host.at(idx)) {
std::cerr << "Error - failed to group problems.\n";
return result;
}
ptr_A_batched_host.push_back(this->block_A.get() + this->offset_A.at(idx));
ptr_B_batched_host.push_back(this->block_B.get() + this->offset_B.at(idx));
ptr_C_batched_host.push_back(this->block_D.get() + this->offset_C.at(idx));
}
}
// Array of GMEM pointers used by batched array GEMM calls
cutlass::DeviceAllocation<void const *> ptr_A_batched;
cutlass::DeviceAllocation<void const *> ptr_B_batched;
cutlass::DeviceAllocation<void *> ptr_C_batched;
ptr_A_batched.reset(ptr_A_batched_host.size());
ptr_B_batched.reset(ptr_A_batched_host.size());
ptr_C_batched.reset(ptr_A_batched_host.size());
ptr_A_batched.copy_from_host(ptr_A_batched_host.data());
ptr_B_batched.copy_from_host(ptr_B_batched_host.data());
ptr_C_batched.copy_from_host(ptr_C_batched_host.data());
//
// Create CUDA streams to maximize concurrency of batched-array GEMM kernels
//
std::vector<cudaStream_t> cuda_streams;
//
// Warmup run
//
if (this->options.cuda_streams) {
for (int i = 0; i < this->options.cuda_streams; ++i) {
cudaStream_t stream;
result.error = cudaStreamCreate(&stream);
if (result.error != cudaSuccess) {
std::cerr << "Failed to create CUDA stream." << std::endl;
return result;
}
cuda_streams.push_back(stream);
}
}
else {
cuda_streams.push_back(nullptr);
}
// Use 'D' for the in/out workspace
this->block_D.copy_from_device(this->block_C.get());
for (int bin_idx = 0; bin_idx < int32_t(bin_problem_sizes.size()); ++bin_idx) {
cutlass::gemm::GemmCoord const & problem = bin_problem_sizes[bin_idx];
int32_t batch_count = bin_count[bin_idx];
int32_t bin_start_idx = bin_start[bin_idx];
int32_t lda = bin_ldm_A[bin_idx];
int32_t ldb = bin_ldm_B[bin_idx];
int32_t ldc = bin_ldm_C[bin_idx];
void const ** ptr_A_array = ptr_A_batched.get() + bin_start[bin_idx];
void const ** ptr_B_array = ptr_B_batched.get() + bin_start[bin_idx];
void ** ptr_C_array = ptr_C_batched.get() + bin_start[bin_idx];
//
// Initialize the CUTLASS GEMM operator
//
// Configure the GEMM arguments
typename Gemm::EpilogueOutputOp::Params epilogue_op(this->options.alpha, this->options.beta);
typename Gemm::Arguments arguments{
cutlass::gemm::GemmUniversalMode::kArray,
problem,
batch_count,
epilogue_op,
(void const *)ptr_A_array,
(void const *)ptr_B_array,
(void const *)ptr_C_array,
(void *)ptr_C_array,
int64_t(),
int64_t(),
int64_t(),
int64_t(),
int64_t(lda),
int64_t(ldb),
int64_t(ldc),
int64_t(ldc)
};
Gemm gemm_op;
cutlass::Status status = gemm_op.initialize(arguments);
if (status != cutlass::Status::kSuccess) {
std::cerr << "CUTLASS error on line " << __LINE__ << std::endl;
return result;
}
status = gemm_op();
if (status != cutlass::Status::kSuccess) {
std::cerr << "CUTLASS error on line " << __LINE__ << std::endl;
return result;
}
}
//
// Wait for completion
//
result.error = cudaDeviceSynchronize();
if (result.error != cudaSuccess) {
std::cerr << "Kernel execution error: " << cudaGetErrorString(result.error);
return result;
}
//
// Construct events
//
cudaEvent_t events[2];
for (auto & event : events) {
result.error = cudaEventCreate(&event);
if (result.error != cudaSuccess) {
std::cerr << "cudaEventCreate() failed: " << cudaGetErrorString(result.error) << std::endl;