forked from NVIDIA/cutlass
-
Notifications
You must be signed in to change notification settings - Fork 28
/
Copy pathgemm_streamk_with_fused_epilogue.h
2396 lines (1962 loc) · 78.3 KB
/
gemm_streamk_with_fused_epilogue.h
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 - 2025 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 Stream-K Gemm kernel compatible with fused epilogues
that broadcast a bias vector over the MMA output.
*/
#pragma once
#include "cutlass/cutlass.h"
#include "cutlass/fast_math.h"
#include "cutlass/layout/layout.h"
#include "cutlass/gemm/gemm.h"
#include "cutlass/matrix_coord.h"
#include "cutlass/complex.h"
#include "cutlass/barrier.h"
#include "cutlass/block_striped.h"
#include "cutlass/semaphore.h"
#include "cutlass/trace.h"
/////////////////////////////////////////////////////////////////////////////////////////////////
namespace cutlass {
namespace gemm {
namespace kernel {
/////////////////////////////////////////////////////////////////////////////////////////////////
template <
typename Mma_, ///! Threadblock-scoped matrix multiply-accumulate
typename Epilogue_, ///! Epilogue
typename ThreadblockSwizzle_, ///! Threadblock swizzling function
bool IsSingleSource = Epilogue_::kIsSingleSource
>
struct GemmStreamkWithFusedEpilogue;
// GemmStreamkWithFusedEpilogue with two sources
template <
typename Mma_, ///! Threadblock-scoped matrix multiply-accumulate
typename Epilogue_, ///! Epilogue
typename ThreadblockSwizzle_ ///! Threadblock swizzling function
>
struct GemmStreamkWithFusedEpilogue<Mma_, Epilogue_, ThreadblockSwizzle_, false> {
using Mma = Mma_;
using Epilogue = Epilogue_;
using EpilogueOutputOp = typename Epilogue::OutputOp;
using ThreadblockSwizzle = ThreadblockSwizzle_;
using ElementA = typename Mma::IteratorA::Element;
using LayoutA = typename Mma::IteratorA::Layout;
using ElementB = typename Mma::IteratorB::Element;
using LayoutB = typename Mma::IteratorB::Layout;
using ElementC = typename Epilogue::OutputTileIterator::Element;
using LayoutC = typename Epilogue::OutputTileIterator::Layout;
/// The per-thread tile of raw accumulators
using AccumulatorTile = typename Mma::FragmentC;
static ComplexTransform const kTransformA = Mma::kTransformA;
static ComplexTransform const kTransformB = Mma::kTransformB;
using Operator = typename Mma::Operator;
using OperatorClass = typename Mma::Operator::OperatorClass;
using ThreadblockShape = typename Mma::Shape;
using WarpShape = typename Mma::Operator::Shape;
using InstructionShape = typename Mma::Policy::Operator::InstructionShape;
using ArchTag = typename Mma::ArchTag;
static int const kStages = Mma::kStages;
static int const kAlignmentA = Mma::IteratorA::AccessType::kElements;
static int const kAlignmentB = Mma::IteratorB::AccessType::kElements;
static int const kAlignmentC = Epilogue::OutputTileIterator::kElementsPerAccess;
/// Warp count (concept: GemmShape)
using WarpCount = typename Mma::WarpCount;
static int const kThreadCount = 32 * WarpCount::kCount;
/// Workspace bytes per thread block
static size_t const kWorkspaceBytesPerBlock =
__NV_STD_MAX(
kThreadCount * sizeof(AccumulatorTile),
Epilogue::kWorkspaceBytesPerBlock);
/// Block-striped reduction utility
using BlockStripedReduceT = BlockStripedReduce<kThreadCount, AccumulatorTile>;
//
// Structures
//
/// Argument structure
struct Arguments {
//
// Data members
//
GemmUniversalMode mode{GemmUniversalMode::kGemm};
GemmCoord problem_size{};
int batch_count{1}; // Either (mode == GemmUniversalMode::kBatched) the batch count, or (mode == GemmUniversalMode::kGemm) the tile-splitting factor
typename EpilogueOutputOp::Params epilogue{};
void const * ptr_A{nullptr};
void const * ptr_B{nullptr};
void const * ptr_C1{nullptr};
void const * ptr_C2{nullptr};
void * ptr_D{nullptr};
void * ptr_Vector;
void * ptr_Tensor;
int64_t batch_stride_A{0};
int64_t batch_stride_B{0};
int64_t batch_stride_C1{0};
int64_t batch_stride_C2{0};
int64_t batch_stride_D{0};
int64_t batch_stride_Vector{0};
int64_t batch_stride_Tensor{0};
typename LayoutA::Stride::Index lda{};
typename LayoutB::Stride::Index ldb{};
typename LayoutC::Stride::Index ldc1{};
typename LayoutC::Stride::Index ldc2{};
typename LayoutC::Stride::Index ldd{};
typename LayoutC::Stride::Index ldr{};
typename LayoutC::Stride::Index ldt{};
int avail_sms{-1}; /// The number of SMs that StreamK dispatch heuristics will attempt to load-balance across (-1 defaults to device width, 1 implies classic data-parallel scheduling)
//
// Methods
//
/// Default Constructor
Arguments() = default;
/// constructs an arguments structure
Arguments(
GemmUniversalMode mode,
GemmCoord problem_size,
int batch_split, /// Either (mode == GemmUniversalMode::kBatched) the batch count, or (mode == GemmUniversalMode::kGemm) the tile-splitting factor (1 defaults to StreamK, >1 emulates Split-K)
typename EpilogueOutputOp::Params epilogue,
void const * ptr_A,
void const * ptr_B,
void const * ptr_C1,
void const * ptr_C2,
void * ptr_D,
void * ptr_Vector,
void * ptr_Tensor,
int64_t batch_stride_A,
int64_t batch_stride_B,
int64_t batch_stride_C1,
int64_t batch_stride_C2,
int64_t batch_stride_D,
int64_t batch_stride_Vector,
int64_t batch_stride_Tensor,
typename LayoutA::Stride::Index lda,
typename LayoutB::Stride::Index ldb,
typename LayoutC::Stride::Index ldc1,
typename LayoutC::Stride::Index ldc2,
typename LayoutC::Stride::Index ldd,
typename LayoutC::Stride::Index ldr,
typename LayoutC::Stride::Index ldt,
int avail_sms = -1) /// The number of SMs that StreamK dispatch heuristics will attempt to load-balance across (-1 defaults to device width, 1 implies classic data-parallel scheduling)
:
mode(mode),
problem_size(problem_size),
batch_count(batch_split),
epilogue(epilogue),
ptr_A(ptr_A), ptr_B(ptr_B), ptr_C1(ptr_C1), ptr_C2(ptr_C2), ptr_D(ptr_D),
ptr_Vector(ptr_Vector),
ptr_Tensor(ptr_Tensor),
batch_stride_A(batch_stride_A),
batch_stride_B(batch_stride_B),
batch_stride_C1(batch_stride_C1),
batch_stride_C2(batch_stride_C2),
batch_stride_Vector(batch_stride_Vector),
batch_stride_Tensor(batch_stride_Tensor),
lda(lda), ldb(ldb), ldc1(ldc1), ldc2(ldc2), ldd(ldd), ldr(ldr), ldt(ldt), avail_sms(avail_sms)
{
CUTLASS_TRACE_HOST("GemmStreamkWithFusedEpilogue::Arguments::Arguments() - problem_size: " << problem_size);
CUTLASS_TRACE_HOST(" ptr_Vector: " << (void *)this->ptr_Vector);
CUTLASS_TRACE_HOST(" ptr_Tensor: " << (void *)this->ptr_Tensor);
CUTLASS_TRACE_HOST(" ldr: " << this->ldr);
CUTLASS_TRACE_HOST(" ldt: " << this->ldt);
CUTLASS_TRACE_HOST(" avail_sms: " << this->avail_sms);
}
/// Returns arguments for the transposed problem
Arguments transposed_problem() const {
Arguments args(*this);
std::swap(args.problem_size.m(), args.problem_size.n());
std::swap(args.ptr_A, args.ptr_B);
std::swap(args.lda, args.ldb);
std::swap(args.batch_stride_A, args.batch_stride_B);
return args;
}
};
/// Parameters structure
struct Params
{
public:
//
// Data members
//
void * ptr_A{nullptr};
void * ptr_B{nullptr};
typename Mma::IteratorA::Params params_A{};
typename Mma::IteratorB::Params params_B{};
int64_t batch_stride_A{0};
int64_t batch_stride_B{0};
GemmUniversalMode mode{GemmUniversalMode::kGemm};
ThreadblockSwizzle block_mapping{};
void *barrier_workspace{nullptr};
void *partials_workspace{nullptr};
typename EpilogueOutputOp::Params output_op{};
void * ptr_C1{nullptr};
void * ptr_C2{nullptr};
void * ptr_D{nullptr};
void * ptr_Tensor{nullptr};
void * ptr_Vector{nullptr};
typename Epilogue::OutputTileIterator::Params params_C1{};
typename Epilogue::OutputTileIterator::Params params_C2{};
typename Epilogue::OutputTileIterator::Params params_D{};
typename Epilogue::TensorTileIterator::Params params_Tensor{};
int64_t batch_stride_C1{0};
int64_t batch_stride_C2{0};
int64_t batch_stride_D{0};
int64_t batch_stride_Vector{0};
int64_t batch_stride_Tensor{0};
typename LayoutC::Stride::Index ldr{};
protected:
//
// Host-only dispatch-utilities
//
/// Pad the given allocation size up to the nearest cache line
static size_t cacheline_align_up(size_t size)
{
static const int CACHELINE_SIZE = 128;
return (size + CACHELINE_SIZE - 1) / CACHELINE_SIZE * CACHELINE_SIZE;
}
/// Get the workspace size needed for barrier
size_t get_barrier_workspace_size() const
{
// For atomic reduction, each SK-block needs a synchronization flag. For parallel reduction,
// each reduction block needs its own synchronization flag.
int sk_blocks = block_mapping.sk_regions() * block_mapping.sk_blocks_per_region();
int num_flags = fast_max(sk_blocks, block_mapping.reduction_blocks);
return cacheline_align_up(sizeof(typename Barrier::T) * num_flags);
}
/// Get the workspace size needed for intermediate partial sums
size_t get_partials_workspace_size() const
{
int sk_blocks = block_mapping.sk_regions() * block_mapping.sk_blocks_per_region();
return cacheline_align_up(kWorkspaceBytesPerBlock * sk_blocks);
}
public:
//
// Host dispatch API
//
/// Default constructor
Params() = default;
/// Constructor
Params(
Arguments const &args, /// GEMM application arguments
int device_sms, /// Number of SMs on the device
int sm_occupancy) /// Kernel SM occupancy (in thread blocks)
:
params_A(args.lda),
params_B(args.ldb),
params_C1(args.ldc1),
params_C2(args.ldc2),
params_D(args.ldd),
params_Tensor(args.ldt),
output_op(args.epilogue),
mode(args.mode),
ptr_A(const_cast<void *>(args.ptr_A)),
ptr_B(const_cast<void *>(args.ptr_B)),
ptr_C1(const_cast<void *>(args.ptr_C1)),
ptr_C2(const_cast<void *>(args.ptr_C2)),
ptr_D(args.ptr_D),
ptr_Vector(args.ptr_Vector),
ldr(args.ldr),
ptr_Tensor(args.ptr_Tensor),
batch_stride_A(args.batch_stride_A),
batch_stride_B(args.batch_stride_B),
batch_stride_C1(args.batch_stride_C1),
batch_stride_C2(args.batch_stride_C2),
batch_stride_D(args.batch_stride_D),
batch_stride_Vector(args.batch_stride_Vector),
batch_stride_Tensor(args.batch_stride_Tensor),
barrier_workspace(nullptr),
partials_workspace(nullptr)
{
CUTLASS_TRACE_HOST("GemmStreamkWithFusedEpilogue::Params::Params()");
CUTLASS_TRACE_HOST(" ptr_Vector: " << (void *)this->ptr_Vector);
CUTLASS_TRACE_HOST(" ptr_Tensor: " << (void *)this->ptr_Tensor);
CUTLASS_TRACE_HOST(" ldr: " << this->ldr);
CUTLASS_TRACE_HOST(" ldt: " << args.ldt);
// Number of SMs to make available for StreamK decomposition
int avail_sms = (args.avail_sms == -1) ?
device_sms :
fast_min(args.avail_sms, device_sms);
CUTLASS_TRACE_HOST(" avail_sms: " << avail_sms);
// Initialize the block mapping structure
block_mapping = ThreadblockSwizzle(
args.mode,
args.problem_size,
{ThreadblockShape::kM, ThreadblockShape::kN, ThreadblockShape::kK},
args.batch_count,
sm_occupancy,
device_sms,
avail_sms,
sizeof(ElementA),
sizeof(ElementB),
sizeof(ElementC),
Epilogue::kAccumulatorFragments);
}
/// Returns the workspace size (in bytes) needed for these parameters
size_t get_workspace_size() const
{
return
get_barrier_workspace_size() +
get_partials_workspace_size();
}
/// Assign and initialize the specified workspace buffer. Assumes
/// the memory allocated to workspace is at least as large as get_workspace_size().
Status init_workspace(
void *workspace,
cudaStream_t stream = nullptr)
{
uint8_t *ptr = static_cast<uint8_t*>(workspace);
// Establish partials workspace
partials_workspace = nullptr;
size_t partials_workspace_bytes = get_partials_workspace_size();
if (partials_workspace_bytes > 0)
{
if (!workspace) {
return Status::kErrorWorkspaceNull;
}
partials_workspace = ptr;
ptr += partials_workspace_bytes;
}
// Establish barrier workspace
barrier_workspace = nullptr;
size_t barrier_workspace_bytes = get_barrier_workspace_size();
if (barrier_workspace_bytes > 0)
{
if (!workspace) {
return Status::kErrorWorkspaceNull;
}
barrier_workspace = ptr;
ptr += barrier_workspace_bytes;
}
// Zero-initialize barrier workspace
if (barrier_workspace)
{
size_t barrier_workspace_bytes = get_barrier_workspace_size();
CUTLASS_TRACE_HOST(" Initialize " << barrier_workspace_bytes << " barrier bytes");
cudaError_t result = cudaMemsetAsync(
barrier_workspace,
0,
barrier_workspace_bytes,
stream);
if (result != cudaSuccess) {
CUTLASS_TRACE_HOST(" cudaMemsetAsync() returned error " << cudaGetErrorString(result));
return Status::kErrorInternal;
}
}
return Status::kSuccess;
}
/// Returns the GEMM volume in thread block tiles
cutlass::gemm::GemmCoord get_tiled_shape() const
{
return block_mapping.tiled_shape();
}
/// Returns the total number of thread blocks to launch
int get_grid_blocks() const
{
dim3 grid_dims = get_grid_dims();
return grid_dims.x * grid_dims.y * grid_dims.z;
}
/// Returns the grid extents in thread blocks to launch
dim3 get_grid_dims() const
{
return block_mapping.get_grid_dims();
}
/// Lightweight update given a subset of arguments. Problem geometry is assumed
/// to remain the same.
CUTLASS_HOST_DEVICE
void update(Arguments const &args)
{
ptr_A = const_cast<void *>(args.ptr_A);
ptr_B = const_cast<void *>(args.ptr_B);
ptr_C1 = const_cast<void *>(args.ptr_C1);
ptr_C2 = const_cast<void *>(args.ptr_C2);
ptr_D = args.ptr_D;
ptr_Vector = args.ptr_Vector;
ldr = args.ldr;
ptr_Tensor = args.ptr_Tensor;
batch_stride_A = args.batch_stride_A;
batch_stride_B = args.batch_stride_B;
batch_stride_C1 = args.batch_stride_C1;
batch_stride_C2 = args.batch_stride_C2;
batch_stride_D = args.batch_stride_D;
batch_stride_Vector = args.batch_stride_Vector;
batch_stride_Tensor = args.batch_stride_Tensor;
output_op = args.epilogue;
CUTLASS_TRACE_HOST("GemmStreamkWithFusedEpilogue::Params::update()");
CUTLASS_TRACE_HOST(" ptr_Vector: " << (void *)this->ptr_Vector);
CUTLASS_TRACE_HOST(" ptr_Tensor: " << (void *)this->ptr_Tensor);
CUTLASS_TRACE_HOST(" ldr: " << this->ldr);
}
};
/// Tile work descriptor
struct TileWorkDesc
{
/// The linear tile index
int tile_idx;
/// The location of this tile (in threadblock-tile coordinates) in the output matrix
cutlass::gemm::GemmCoord tiled_coord;
// The first global-scoped MAC-iteration this threadblock will perform for this tile
int iter_begin;
// The starting index in the k-domain for MAC-iterations this threadblock will perform for this tile
int k_begin;
// The ending index (one-past) in the k-domain for MAC-iterations this threadblock will perform for this tile
int k_end;
/// The number of remaining MAC-iterations this threadblock will perform for this tile
int k_iters_remaining;
// Whether this block will perform the first iteration of this tile
CUTLASS_DEVICE
bool tile_started()
{
return (k_begin == 0);
}
// Whether this block will perform the last iteration of this tile
CUTLASS_DEVICE
bool tile_finished(Params const ¶ms)
{
return (k_end == params.block_mapping.problem_size.k());
}
};
/// Shared memory storage structure
union SharedStorage {
typename Mma::SharedStorage main_loop;
typename Epilogue::SharedStorage epilogue;
};
protected:
//
// Data members
//
/// GEMM problem parameters
Params const ¶ms;
/// Shared storage reference
SharedStorage &shared_storage;
/// ID within the threadblock
int thread_idx;
/// ID of warp
int warp_idx;
/// ID of each thread within a warp
int lane_idx;
/// Threadblock scoped epilogue
Epilogue epilogue;
public:
//
// Host dispatch API
//
/// Determines whether kernel satisfies alignment
static Status can_implement(
cutlass::gemm::GemmCoord const & problem_size) {
CUTLASS_TRACE_HOST("GemmStreamkWithFusedEpilogue::can_implement()");
static int const kAlignmentA = Mma::IteratorA::AccessType::kElements;
static int const kAlignmentB = Mma::IteratorB::AccessType::kElements;
static int const kAlignmentC = Epilogue::OutputTileIterator::kElementsPerAccess;
bool isAMisaligned = false;
bool isBMisaligned = false;
bool isCMisaligned = false;
if (platform::is_same<LayoutA, layout::RowMajor>::value) {
isAMisaligned = problem_size.k() % kAlignmentA;
} else if (platform::is_same<LayoutA, layout::ColumnMajor>::value) {
isAMisaligned = problem_size.m() % kAlignmentA;
} else if (platform::is_same<LayoutA, layout::ColumnMajorInterleaved<32>>::value
|| platform::is_same<LayoutA, layout::ColumnMajorInterleaved<64>>::value) {
isAMisaligned = problem_size.k() % kAlignmentA;
}
if (platform::is_same<LayoutB, layout::RowMajor>::value) {
isBMisaligned = problem_size.n() % kAlignmentB;
} else if (platform::is_same<LayoutB, layout::ColumnMajor>::value) {
isBMisaligned = problem_size.k() % kAlignmentB;
} else if (platform::is_same<LayoutB, layout::RowMajorInterleaved<32>>::value
|| platform::is_same<LayoutB, layout::RowMajorInterleaved<64>>::value) {
isBMisaligned = problem_size.k() % kAlignmentB;
}
if (platform::is_same<LayoutC, layout::RowMajor>::value) {
isCMisaligned = problem_size.n() % kAlignmentC;
} else if (platform::is_same<LayoutC, layout::ColumnMajor>::value) {
isCMisaligned = problem_size.m() % kAlignmentC;
} else if (platform::is_same<LayoutC, layout::ColumnMajorInterleaved<32>>::value
|| platform::is_same<LayoutC, layout::ColumnMajorInterleaved<64>>::value) {
isCMisaligned = problem_size.n() % kAlignmentC;
}
if (isAMisaligned) {
CUTLASS_TRACE_HOST(" returning kErrorMisalignedOperand for A operand");
return Status::kErrorMisalignedOperand;
}
if (isBMisaligned) {
CUTLASS_TRACE_HOST(" returning kErrorMisalignedOperand for B operand");
return Status::kErrorMisalignedOperand;
}
if (isCMisaligned) {
CUTLASS_TRACE_HOST(" returning kErrorMisalignedOperand for C operand");
return Status::kErrorMisalignedOperand;
}
CUTLASS_TRACE_HOST(" returning kSuccess");
return Status::kSuccess;
}
static Status can_implement(Arguments const &args) {
return can_implement(args.problem_size);
}
protected:
//
// Device-only utility methods
//
/// Iterator for fetching tile fragments from A
CUTLASS_DEVICE
typename Mma::IteratorA init_iterator_A(
TileWorkDesc &tile_work,
GemmUniversalMode mode)
{
// The input A matrix
ElementA *ptr_A = static_cast<ElementA *>(params.ptr_A);
// Update input pointers based on batched/array mode
if (mode == GemmUniversalMode::kBatched) {
ptr_A += tile_work.tiled_coord.k() * params.batch_stride_A;
}
if (mode == GemmUniversalMode::kArray) {
ptr_A = static_cast<ElementA * const *>(params.ptr_A)[tile_work.tiled_coord.k()];
}
int m_begin = tile_work.tiled_coord.m() * Mma::Shape::kM;
int m_end = params.block_mapping.problem_size.m();
return Mma::IteratorA(
params.params_A,
ptr_A,
{ m_end, tile_work.k_end },
threadIdx.x,
{ m_begin, tile_work.k_begin });
}
/// Iterator for fetching tile fragments from B
CUTLASS_DEVICE
typename Mma::IteratorB init_iterator_B(
TileWorkDesc &tile_work,
GemmUniversalMode mode)
{
// The input B matrix
ElementB *ptr_B = static_cast<ElementB *>(params.ptr_B);
// Update input pointers based on batched/array mode
if (mode == GemmUniversalMode::kBatched) {
ptr_B += tile_work.tiled_coord.k() * params.batch_stride_B;
}
if (mode == GemmUniversalMode::kArray) {
ptr_B = static_cast<ElementB * const *>(params.ptr_B)[tile_work.tiled_coord.k()];
}
int n_begin = tile_work.tiled_coord.n() * Mma::Shape::kN;
int n_end = params.block_mapping.problem_size.n();
return Mma::IteratorB(
params.params_B,
ptr_B,
{ tile_work.k_end, n_end },
threadIdx.x,
{ tile_work.k_begin, n_begin });
}
CUTLASS_DEVICE
void init_dp_tile_work(
TileWorkDesc &tile_work,
int tile_idx)
{
// The linear tile index
tile_work.tile_idx = tile_idx;
// The first global-scoped MAC-iteration this threadblock will perform for this tile
tile_work.iter_begin = tile_idx * params.block_mapping.iters_per_tile();
// The number of MAC-iterations this threadblock will perform for this tile
tile_work.k_iters_remaining = params.block_mapping.iters_per_tile();
// The starting index in the k-domain for MAC-iterations this threadblock will perform for this tile
tile_work.k_begin = 0;
// The ending index (one-past) in the k-domain for MAC-iterations this threadblock will perform for this tile
tile_work.k_end = params.block_mapping.problem_size.k();
// The location of this tile (in threadblock-tile coordinates) in the output matrix
tile_work.tiled_coord = params.block_mapping.get_tile_offset(tile_work.tile_idx);
}
CUTLASS_DEVICE
void init_sk_tile_work(
TileWorkDesc &tile_work,
int tile_idx,
int block_iter_begin,
int block_iter_end)
{
// The linear tile index
tile_work.tile_idx = tile_idx;
// The first global-scoped MAC-iteration for this tile
int tile_iter_begin = tile_idx * params.block_mapping.iters_per_tile();
// The first global-scoped MAC-iteration this threadblock will perform for this tile
tile_work.iter_begin = max(block_iter_begin, tile_iter_begin);
// The first tile-scoped MAC-iteration this threadblock will perform for this tile
int k_iter_begin = tile_work.iter_begin - tile_iter_begin;
// The last (one past) tile-scoped MAC-iteration this threadblock will perform for this tile
int k_iter_end = block_iter_end - tile_iter_begin;
// The number of MAC-iterations this threadblock will perform for this tile
tile_work.k_iters_remaining = k_iter_end - k_iter_begin;
// The starting index in the k-domain for MAC-iterations this threadblock will perform for this tile
tile_work.k_begin = k_iter_begin * Mma::Shape::kK;
// The ending index (one-past) in the k-domain for MAC-iterations this threadblock will perform for this tile
tile_work.k_end = min(
params.block_mapping.problem_size.k(), // extent of k domain
(k_iter_end * Mma::Shape::kK)); // extent of the threadblock's global iteration assignment
// The location of this tile (in threadblock-tile coordinates) in the output matrix
tile_work.tiled_coord = params.block_mapping.get_tile_offset(tile_work.tile_idx);
}
/// Share accumulators with peers
CUTLASS_DEVICE
void share_accumulators(
AccumulatorTile const &accumulator_tile,
int block_idx,
int first_block_idx)
{
AccumulatorTile *accum_tile_workspace = reinterpret_cast<AccumulatorTile *>(params.partials_workspace);
int accum_tile_offset = first_block_idx * kThreadCount;
if (block_idx == first_block_idx)
{
// First peer initializes the workspace partials
BlockStripedReduceT::store(accum_tile_workspace + accum_tile_offset, accumulator_tile, thread_idx);
}
else
{
// Subsequent peers atomically accumulate into the workspace partials
if (ThreadblockSwizzle::kReductionStrategy == ThreadblockSwizzle::kAtomic)
{
// Non-deterministic reduction order: wait for the first peer to have initialized the partials before we add to them
Barrier::wait_lt(params.barrier_workspace, thread_idx, first_block_idx, 1);
}
else
{
// Turnstile reduction order: wait until the previous peer has written
int wait_count = block_idx - first_block_idx;
Barrier::wait_eq(params.barrier_workspace, thread_idx, first_block_idx, wait_count);
}
// Perform reduction in workspace
BlockStripedReduceT::reduce(accum_tile_workspace + accum_tile_offset, accumulator_tile, thread_idx);
}
// Signal our arrival
Barrier::arrive_inc(params.barrier_workspace, thread_idx, first_block_idx);
}
/// Acquire accumulators from peers
CUTLASS_DEVICE
void acquire_accumulators(
AccumulatorTile &accumulator_tile,
int block_idx,
int first_block_idx)
{
AccumulatorTile *accum_tile_workspace = reinterpret_cast<AccumulatorTile *>(params.partials_workspace);
// Wait for arrival
int num_carry_in = block_idx - first_block_idx;
Barrier::wait_eq_reset(params.barrier_workspace, thread_idx, first_block_idx, num_carry_in);
// Load and add peer-partials accumulator tile to local accumulator tile
int accum_tile_offset = first_block_idx * kThreadCount;
BlockStripedReduceT::load_add(accumulator_tile, accum_tile_workspace + accum_tile_offset, thread_idx);
}
/// Perform epilogue computations and output
CUTLASS_DEVICE
void do_epilogue(
TileWorkDesc &tile_work,
AccumulatorTile &accumulator_tile)
{
ElementC *ptr_C1 = static_cast<ElementC *>(params.ptr_C1);
ElementC *ptr_C2 = static_cast<ElementC *>(params.ptr_C2);
ElementC *ptr_D = static_cast<ElementC *>(params.ptr_D);
typename Epilogue::ElementTensor *ptr_Tensor = static_cast<typename Epilogue::ElementTensor *>(params.ptr_Tensor);
// Define the reduction output pointer and move to the appropriate place
typename Epilogue::ElementVector *ptr_Vector =
static_cast<typename Epilogue::ElementVector *>(params.ptr_Vector);
// Update pointers for batched/array mode(s)
if (params.mode == GemmUniversalMode::kBatched) {
ptr_C1 += tile_work.tiled_coord.k() * params.batch_stride_C1;
if (ptr_C2) {
ptr_C2 += tile_work.tiled_coord.k() * params.batch_stride_C2;
}
ptr_D += tile_work.tiled_coord.k() * params.batch_stride_D;
if (ptr_Tensor) {
ptr_Tensor = ReferenceFactory<typename Epilogue::ElementTensor>::add_pointer_offset(
ptr_Tensor,
tile_work.tiled_coord.k() * params.batch_stride_Tensor);
}
if (ptr_Vector) {
ptr_Vector += tile_work.tiled_coord.k() * params.batch_stride_Vector;
}
}
if (params.mode == GemmUniversalMode::kArray) {
ptr_C1 = static_cast<ElementC * const *>(params.ptr_C1)[tile_work.tiled_coord.k()];
if (ptr_C2) {
ptr_C2 = static_cast<ElementC * const *>(params.ptr_C2)[tile_work.tiled_coord.k()];
}
ptr_D = static_cast<ElementC * const *>(params.ptr_D)[tile_work.tiled_coord.k()];
if (ptr_Tensor) {
ptr_Tensor = static_cast<typename Epilogue::ElementTensor * const *>(params.ptr_Tensor)[tile_work.tiled_coord.k()];
}
if (ptr_Vector) {
ptr_Vector = static_cast<typename Epilogue::ElementVector * const *>(params.ptr_Vector)[tile_work.tiled_coord.k()];
}
}
// Location of this tile in item-coords
MatrixCoord threadblock_item_begin(
tile_work.tiled_coord.m() * Mma::Shape::kM,
tile_work.tiled_coord.n() * Mma::Shape::kN
);
// Tile iterator loading from residual1.
typename Epilogue::OutputTileIterator iterator_C1(
params.params_C1,
ptr_C1,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Tile iterator loading from residual2.
typename Epilogue::OutputTileIterator iterator_C2(
params.params_C2,
ptr_C2,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Tile iterator writing to destination tensor.
typename Epilogue::OutputTileIterator iterator_D(
params.params_D,
ptr_D,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Additional tensor to load from
typename Epilogue::TensorTileIterator tensor_iterator(
params.params_Tensor,
ptr_Tensor,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Move to appropriate location for this output tile
if (ptr_Vector) {
ptr_Vector += threadblock_item_begin.column() + tile_work.tiled_coord.m() * params.ldr;
}
// Execute the epilogue operator to update the destination tensor.
epilogue(
EpilogueOutputOp(params.output_op),
ptr_Vector,
iterator_D,
accumulator_tile,
iterator_C1,
iterator_C2,
tensor_iterator,
params.block_mapping.problem_size.mn(),
threadblock_item_begin);
}
CUTLASS_DEVICE
void separate_reduction(int reduce_idx)
{
int peer_idx_begin, peer_idx_last, reduce_tile_idx, reduce_fragment_idx;
// Reduce by sk-tile (every tile contributed to by one or more blocks)
reduce_tile_idx = reduce_idx / Epilogue::kAccumulatorFragments;
reduce_fragment_idx = reduce_idx % Epilogue::kAccumulatorFragments;
int iter_tile_first = reduce_tile_idx * params.block_mapping.iters_per_tile();
int iter_tile_last = iter_tile_first + params.block_mapping.iters_per_tile() - 1;
peer_idx_begin = params.block_mapping.get_sk_block_idx(iter_tile_first);
peer_idx_last = params.block_mapping.get_sk_block_idx(iter_tile_last);
// Wait for peers to complete
int peer_idx_end = peer_idx_last + 1;
int num_peers = peer_idx_end - peer_idx_begin;
Barrier::wait_eq_reset(
params.barrier_workspace,
thread_idx,
(reduce_tile_idx * Epilogue::kAccumulatorFragments) + reduce_fragment_idx,
num_peers);
/// The location of this tile (in threadblock-tile coordinates) in the output matrix
GemmCoord tiled_coord = params.block_mapping.get_tile_offset(reduce_tile_idx);
// Location of this tile in item-coords
MatrixCoord threadblock_item_begin(
tiled_coord.m() * Mma::Shape::kM,
tiled_coord.n() * Mma::Shape::kN
);
ElementC *ptr_C1 = static_cast<ElementC *>(params.ptr_C1);
ElementC *ptr_C2 = static_cast<ElementC *>(params.ptr_C2);
ElementC *ptr_D = static_cast<ElementC *>(params.ptr_D);
typename Epilogue::ElementTensor *ptr_Tensor = static_cast<typename Epilogue::ElementTensor *>(params.ptr_Tensor);
// Define the reduction output pointer and move to the appropriate place
typename Epilogue::ElementVector *ptr_Vector =
static_cast<typename Epilogue::ElementVector *>(params.ptr_Vector);
// Tile iterator loading from residual1.
typename Epilogue::OutputTileIterator iterator_C1(
params.params_C1,
ptr_C1,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Tile iterator loading from residual2.
typename Epilogue::OutputTileIterator iterator_C2(
params.params_C2,
ptr_C2,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Tile iterator writing to destination tensor.
typename Epilogue::OutputTileIterator iterator_D(
params.params_D,
ptr_D,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Additional tensor to load from
typename Epilogue::TensorTileIterator tensor_iterator(
params.params_Tensor,
ptr_Tensor,
params.block_mapping.problem_size.mn(),
thread_idx,
threadblock_item_begin);
// Move to appropriate location for this output tile
if (ptr_Vector) {
ptr_Vector += threadblock_item_begin.column() + tiled_coord.m() * params.ldr;
}