forked from NVIDIA/cutlass
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpredicated_tile_iterator.h
1341 lines (1035 loc) · 39.6 KB
/
predicated_tile_iterator.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 - 2023 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 Epilogue for threadblock scoped GEMMs using Tensor Ops.
The epilogue rearranges the result of a matrix product through shared memory to match canonical
tensor layouts in global memory. Epilogues support conversion and reduction operations.
*/
#pragma once
#include "cutlass/cutlass.h"
#include "cutlass/numeric_types.h"
#include "cutlass/array.h"
#include "cutlass/layout/matrix.h"
#include "cutlass/layout/tensor.h"
#include "cutlass/layout/permute.h"
#include "cutlass/matrix_shape.h"
#include "cutlass/tensor_ref.h"
#include "cutlass/transform/pitch_linear_thread_map.h"
#include "cutlass/epilogue/threadblock/output_tile_thread_map.h"
#include "cutlass/arch/arch.h"
#include "cutlass/arch/memory.h"
#include "cutlass/epilogue/threadblock/predicated_tile_iterator_params.h"
////////////////////////////////////////////////////////////////////////////////
namespace cutlass {
////////////////////////////////////////////////////////////////////////////////
namespace epilogue {
namespace threadblock {
////////////////////////////////////////////////////////////////////////////////
/// Tile iterator used to load and store output tile from global memory in epilogue.
///
/// Satisfies: ReadableTileIterator | PredicatedTileIterator | ForwardTileIterator
///
template <
typename ThreadMap_, ///< Thread map (conept: OutputTileThreadMap)
typename Element_, ///< Element data type
bool ScatterD = false, ///< Scatter D operand or not
typename PermuteDLayout = layout::NoPermute, ///< Permute D operand or not
bool UseCUDAStore = false
>
class PredicatedTileIterator {
public:
using ThreadMap = ThreadMap_;
using Shape = typename ThreadMap::Shape;
using Element = Element_;
using Layout = layout::RowMajor;
using TensorRef = TensorRef<Element, Layout>;
using ConstTensorRef = typename TensorRef::ConstTensorRef;
using Index = typename Layout::Index;
using LongIndex = typename Layout::LongIndex;
using TensorCoord = MatrixCoord;
static int const kElementsPerAccess = ThreadMap::kElementsPerAccess;
static int const kThreads = ThreadMap::kThreads;
static int const kIterations = ThreadMap::Count::kTile;
static bool constexpr PermuteD = !layout::is_trivial_permute<PermuteDLayout>;
static_assert( ThreadMap::Iterations::kRow > 0,"ThreadMap::Iterations::kRow must be > 0");
static_assert( ThreadMap::Iterations::kGroup > 0,"ThreadMap::Iterations::kGroup must be > 0");
static_assert( ThreadMap::Iterations::kCluster > 0,"ThreadMap::Iterations::kCluster must be > 0");
static_assert( ThreadMap::Iterations::kColumn > 0,"ThreadMap::Iterations::kColumn must be > 0");
/// Fragment object
using Fragment = Array<
Element,
ThreadMap::Iterations::kColumn *
ThreadMap::Iterations::kRow *
ThreadMap::Iterations::kGroup *
ThreadMap::Iterations::kCluster * ThreadMap::kElementsPerAccess>;
/// Memory access size
using AccessType = AlignedArray<Element, ThreadMap::kElementsPerAccess>;
//
// Parameters struct
//
/// Uses a non-template class
struct Params : PredicatedTileIteratorParams {
using Base = PredicatedTileIteratorParams;
CUTLASS_HOST_DEVICE
Params() { }
CUTLASS_HOST_DEVICE
Params(Layout const &layout):
PredicatedTileIteratorParams(
layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess,
make_OutputTileThreadMapDesc<ThreadMap>()
)
{ }
CUTLASS_HOST_DEVICE
Params(Base const &base) :
Base(base) { }
};
/// Mask object
struct Mask {
static int const kCount = ThreadMap::Iterations::kColumn;
/// Predicate state
bool predicates[kCount];
//
// Mask
//
CUTLASS_HOST_DEVICE
Mask() {
enable();
}
///< Efficiently disables all accesses guarded by mask
CUTLASS_HOST_DEVICE void clear() {
CUTLASS_PRAGMA_UNROLL
for (int i = 0; i < kCount; ++i) {
predicates[i] = false;
}
}
///< CUTLASS_HOST_DEVICE enables all accesses guarded by mask
CUTLASS_DEVICE void enable() {
CUTLASS_PRAGMA_UNROLL
for (int i = 0; i < kCount; ++i) {
predicates[i] = true;
}
}
};
private:
//
// Data members
//
/// Parameters structure containing reference and precomputed state.
PredicatedTileIteratorParams params_;
/// Byte-level pointer. This pointer is usually for both load() and store(), unless PermuteD is performed. When having PermuteD, byte_pointer_ is only for load().
uint8_t *byte_pointer_;
/// Byte-level pointer for store(). Due to PermuteD Op, store_byte_pointer_ may be with different address computation compared to byte_pointer_.
uint8_t *store_byte_pointer_;
/// Array of boolean values to contain steady-state predicates
Mask mask_;
/// Extent of the matrix tile in rows
Index extent_row_;
/// Extent of the matrix tile in rows
Index extent_column_;
/// A thread's starting row position (assuming steady-state predicates have been computed)
Index thread_start_row_;
/// A thread's starting column
Index thread_start_column_;
/// Internal state counter
int state_[3];
/// Scatter indices
int const *indices_;
/// PermuteDLayout
PermuteDLayout permute_layout_;
//
// Static asserts about internal strides
//
static_assert(sizeof(extent_row_) == 4, "Expected 32b extents");
static_assert(sizeof(thread_start_row_) == 4, "Expected 32b extents");
static_assert(sizeof(PredicatedTileIteratorParams::stride) == 8, "Expected 64b strides");
private:
//
// Methods
//
public:
//
// Methods
//
/// Constructor
CUTLASS_DEVICE
PredicatedTileIterator(
PredicatedTileIteratorParams const & params,
Element *pointer,
TensorCoord extent,
int thread_idx,
TensorCoord threadblock_offset = TensorCoord(),
int const *indices = nullptr
):
params_(params), indices_(indices),
permute_layout_(PitchLinearCoord(extent.column(), extent.row()), params_.stride * kElementsPerAccess / sizeof(AccessType))
{
TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset;
extent_row_ = extent.row();
extent_column_ = extent.column();
thread_start_row_ = thread_offset.row();
thread_start_column_ = thread_offset.column();
// Initialize predicates
CUTLASS_PRAGMA_UNROLL
for (int c = 0; c < ThreadMap::Iterations::kColumn; ++c) {
mask_.predicates[c] = ((thread_offset.column()
+ ThreadMap::Delta::kColumn * c) < extent.column());
}
// Null pointer performs no accesses
if (!pointer) {
mask_.clear();
}
if (ScatterD && !indices) {
mask_.clear();
}
// Initialize byte_pointer_
byte_pointer_ = reinterpret_cast<uint8_t *>(pointer) +
LongIndex(thread_offset.row()) * LongIndex(params_.stride) +
LongIndex(thread_offset.column()) * sizeof(AccessType) / kElementsPerAccess;
if (ScatterD) {
byte_pointer_ = reinterpret_cast<uint8_t *>(pointer) +
LongIndex(thread_offset.column()) * sizeof(AccessType) / kElementsPerAccess;
}
// store_byte_pointer_ is set to be the same with byte_pointer_ unless PermuteD is used.
store_byte_pointer_ = PermuteD ? reinterpret_cast<uint8_t *>(pointer) : byte_pointer_;
// Initialize internal state counter
state_[0] = state_[1] = state_[2] = 0;
}
/// Adds a pointer offset in units of Element
CUTLASS_HOST_DEVICE
void add_pointer_offset(LongIndex pointer_offset) {
store_byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8;
byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8;
}
/// Loads a fragment from memory
CUTLASS_DEVICE
void load_with_byte_offset(Fragment &frag, int64_t byte_offset) const {
uint8_t *byte_pointer = byte_pointer_;
AccessType *frag_ptr = reinterpret_cast<AccessType *>(&frag);
CUTLASS_PRAGMA_UNROLL
for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {
CUTLASS_PRAGMA_UNROLL
for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {
CUTLASS_PRAGMA_UNROLL
for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {
int frag_row_idx =
(row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));
int row_offset = row * ThreadMap::Delta::kRow
+ group * ThreadMap::Delta::kGroup
+ cluster * ThreadMap::Delta::kCluster;
bool row_guard = ((row_offset + thread_start_row_) < extent_row_);
AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);
if (ScatterD && row_guard) {
assert(indices_);
memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset +
LongIndex(indices_[row_offset + thread_start_row_]) * LongIndex(params_.stride));
}
CUTLASS_PRAGMA_UNROLL
for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {
bool guard = row_guard && mask_.predicates[column];
cutlass::arch::global_load<
AccessType,
sizeof(AccessType)
>(
frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn +
column],
(void *)&memory_pointer[column * ThreadMap::Delta::kColumn /
kElementsPerAccess],
guard);
}
if (row + 1 < ThreadMap::Iterations::kRow) {
if (!ScatterD) {
byte_pointer += params_.increment_row;
}
}
}
if (group + 1 < ThreadMap::Iterations::kGroup) {
byte_pointer += params_.increment_group;
}
}
if (cluster + 1 < ThreadMap::Iterations::kCluster) {
byte_pointer += params_.increment_cluster;
}
}
}
/// Loads a fragment from memory
CUTLASS_DEVICE
void load(Fragment &frag) const {
load_with_byte_offset(frag, 0);
}
/// Stores a fragment to memory
CUTLASS_DEVICE
void store_with_byte_offset(Fragment const &frag, int64_t byte_offset) const {
uint8_t *byte_pointer = store_byte_pointer_;
AccessType const *frag_ptr = reinterpret_cast<AccessType const *>(&frag);
CUTLASS_PRAGMA_UNROLL
for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {
CUTLASS_PRAGMA_UNROLL
for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {
CUTLASS_PRAGMA_UNROLL
for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {
int frag_row_idx =
(row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));
int row_offset = row * ThreadMap::Delta::kRow
+ group * ThreadMap::Delta::kGroup
+ cluster * ThreadMap::Delta::kCluster;
bool row_guard = ((row_offset + thread_start_row_) < extent_row_);
AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);
if (ScatterD && row_guard) {
assert(indices_);
memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset +
LongIndex(indices_[row_offset + thread_start_row_]) * LongIndex(params_.stride));
}
CUTLASS_PRAGMA_UNROLL
for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {
bool guard = row_guard && mask_.predicates[column];
if (PermuteD) {
int col_offset = column * ThreadMap::Delta::kColumn;
int col = col_offset + thread_start_column_;
int row = row_offset + thread_start_row_;
// Locate memory_pointer
memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset
+ permute_layout_(PitchLinearCoord(col, row)) * sizeof(AccessType) / kElementsPerAccess);
}
if (UseCUDAStore) {
if (guard) {
memory_pointer[0] =
frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column];
}
} else {
cutlass::arch::global_store<AccessType, sizeof(AccessType)>(
frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column],
(void *)&memory_pointer[0],
guard);
}
if (!PermuteD) {
memory_pointer += (ThreadMap::Delta::kColumn / kElementsPerAccess);
}
}
if (row + 1 < ThreadMap::Iterations::kRow) {
if (!ScatterD && !PermuteD) {
byte_pointer += params_.increment_row;
}
}
}
if (group + 1 < ThreadMap::Iterations::kGroup) {
byte_pointer += params_.increment_group;
}
}
if (cluster + 1 < ThreadMap::Iterations::kCluster) {
byte_pointer += params_.increment_cluster;
}
}
}
/// Stores a fragment to memory
CUTLASS_DEVICE
void store(Fragment const &frag) const {
store_with_byte_offset(frag, 0);
}
/// Loads a fragment from memory
CUTLASS_DEVICE
void downsample_load_with_byte_offset(Fragment &frag, int64_t byte_offset, int convolution_P, int convolution_Q, int add_P, int add_Q, int problem_N) const {
uint8_t *byte_pointer = byte_pointer_;
AccessType *frag_ptr = reinterpret_cast<AccessType *>(&frag);
CUTLASS_PRAGMA_UNROLL
for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {
CUTLASS_PRAGMA_UNROLL
for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {
CUTLASS_PRAGMA_UNROLL
for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {
int frag_row_idx =
(row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));
int row_offset = row * ThreadMap::Delta::kRow
+ group * ThreadMap::Delta::kGroup
+ cluster * ThreadMap::Delta::kCluster;
bool row_guard = ((row_offset + thread_start_row_) < extent_row_);
int output_row = row_offset + thread_start_row_;
int output_N = output_row / (convolution_P * convolution_Q);
int output_PQ = output_row % (convolution_P * convolution_Q);
int output_P = output_PQ / convolution_Q;
int output_Q = output_PQ % convolution_Q;
int input_row = output_N * 2 * convolution_P * 2 * convolution_Q +
(2 * output_P + add_P) * 2 * convolution_Q + 2 * output_Q + add_Q;
int64_t byte_offset = (input_row-output_row)*problem_N*sizeof(float);
AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);
CUTLASS_PRAGMA_UNROLL
for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {
bool guard = row_guard && mask_.predicates[column];
cutlass::arch::global_load<
AccessType,
sizeof(AccessType)
>(
frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn +
column],
(void *)&memory_pointer[column * ThreadMap::Delta::kColumn /
kElementsPerAccess],
guard);
}
if (row + 1 < ThreadMap::Iterations::kRow) {
byte_pointer += params_.increment_row;
}
}
if (group + 1 < ThreadMap::Iterations::kGroup) {
byte_pointer += params_.increment_group;
}
}
if (cluster + 1 < ThreadMap::Iterations::kCluster) {
byte_pointer += params_.increment_cluster;
}
}
}
/// Loads a fragment from memory
CUTLASS_DEVICE
void upsample_load_with_byte_offset(Fragment &frag, int64_t byte_offset, int convolution_P, int convolution_Q, int add_P, int add_Q, int problem_N) const {
uint8_t *byte_pointer = byte_pointer_;
AccessType *frag_ptr = reinterpret_cast<AccessType *>(&frag);
CUTLASS_PRAGMA_UNROLL
for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) {
CUTLASS_PRAGMA_UNROLL
for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) {
CUTLASS_PRAGMA_UNROLL
for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) {
int frag_row_idx =
(row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster));
int row_offset = row * ThreadMap::Delta::kRow
+ group * ThreadMap::Delta::kGroup
+ cluster * ThreadMap::Delta::kCluster;
bool row_guard = ((row_offset + thread_start_row_) < extent_row_);
int output_row = row_offset + thread_start_row_;
int output_N = output_row / (convolution_P * convolution_Q);
int output_PQ = output_row % (convolution_P * convolution_Q);
int output_P = output_PQ / convolution_Q;
int output_Q = output_PQ % convolution_Q;
int row_add_P = add_P;
int row_add_Q = add_Q;
if (output_P > convolution_P - 2) row_add_P = 0;
if (output_Q > convolution_Q - 2) row_add_Q = 0;
int input_row = output_N * (convolution_P/2) * (convolution_Q/2) +
((output_P + row_add_P)/2) * (convolution_Q/2) + (output_Q + row_add_Q)/2;
int64_t byte_offset = (input_row-output_row)*problem_N*sizeof(float);
AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer + byte_offset);
CUTLASS_PRAGMA_UNROLL
for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) {
bool guard = row_guard && mask_.predicates[column];
cutlass::arch::global_load<
AccessType,
sizeof(AccessType)
>(
frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn +
column],
(void *)&memory_pointer[column * ThreadMap::Delta::kColumn /
kElementsPerAccess],
guard);
}
if (row + 1 < ThreadMap::Iterations::kRow) {
byte_pointer += params_.increment_row;
}
}
if (group + 1 < ThreadMap::Iterations::kGroup) {
byte_pointer += params_.increment_group;
}
}
if (cluster + 1 < ThreadMap::Iterations::kCluster) {
byte_pointer += params_.increment_cluster;
}
}
}
CUTLASS_DEVICE
MatrixCoord thread_start() const {
return MatrixCoord(thread_start_row_, thread_start_column_);
}
/// Need to get the thread start row from the tile iterator
CUTLASS_DEVICE
int32_t thread_start_row() const {
return thread_start_row_;
}
/// Need to get the thread start row from the tile iterator
CUTLASS_DEVICE
int32_t thread_start_column() const {
return thread_start_column_;
}
/// Extent of the matrix in rows
CUTLASS_DEVICE
Index extent_row() const {
return extent_row_;
}
/// Extent of the matrix in columns
CUTLASS_DEVICE
Index extent_column() const {
return extent_column_;
}
/// Advances to the next position to load or store
CUTLASS_HOST_DEVICE
PredicatedTileIterator &operator++() {
++state_[0];
if (!ScatterD && !PermuteD) {
store_byte_pointer_ += params_.advance_row;
}
if (!ScatterD) {
byte_pointer_ += params_.advance_row;
}
thread_start_row_ += ThreadMap::Shape::kRow;
if (state_[0] == ThreadMap::Count::kRow) {
state_[0] = 0;
++state_[1];
byte_pointer_ += params_.advance_group;
store_byte_pointer_ += params_.advance_group;
thread_start_row_ += (ThreadMap::Shape::kGroup - 1) *
ThreadMap::Shape::kRow * ThreadMap::Count::kRow;
if (state_[1] == ThreadMap::Count::kGroup) {
state_[1] = 0;
++state_[2];
byte_pointer_ += params_.advance_cluster;
store_byte_pointer_ += params_.advance_cluster;
thread_start_row_ += ThreadMap::Count::kGroup *
ThreadMap::Shape::kGroup * ThreadMap::Count::kRow * ThreadMap::Shape::kRow;
if (state_[2] == ThreadMap::Count::kCluster) {
state_[2] = 0;
byte_pointer_ += params_.advance_tile;
store_byte_pointer_ += params_.advance_tile;
thread_start_row_ += ThreadMap::Shape::kGroup * ThreadMap::Shape::kRow
* ThreadMap::Shape::kCluster * ThreadMap::Shape::kTile;
}
}
}
return *this;
}
/// Advances a number of positions to load or store
CUTLASS_HOST_DEVICE
PredicatedTileIterator &operator+=(int increment)
{
// Row
state_[0] += increment;
int increment_row = state_[0] / ThreadMap::Count::kRow;
state_[0] = state_[0] % ThreadMap::Count::kRow;
byte_pointer_ += (params_.advance_row * increment);
store_byte_pointer_ += (params_.advance_row * increment);
thread_start_row_ += (ThreadMap::Shape::kRow * increment);
// Group
state_[1] += increment_row;
int increment_group = state_[1] / ThreadMap::Count::kGroup;
state_[1] = state_[1] % ThreadMap::Count::kGroup;
byte_pointer_ += (params_.advance_group * increment_row);
store_byte_pointer_ += (params_.advance_group * increment_row);
thread_start_row_ +=
(ThreadMap::Shape::kGroup - 1) *
ThreadMap::Shape::kRow *
ThreadMap::Count::kRow *
increment_row;
// Cluster
state_[2] += increment_group;
int increment_cluster = state_[2] / ThreadMap::Count::kCluster;
state_[2] = state_[2] % ThreadMap::Count::kCluster;
byte_pointer_ += (params_.advance_cluster * increment_group);
store_byte_pointer_ += (params_.advance_cluster * increment_group);
thread_start_row_ +=
ThreadMap::Count::kGroup *
ThreadMap::Shape::kGroup *
ThreadMap::Count::kRow *
ThreadMap::Shape::kRow *
increment_group;
// Tile
byte_pointer_ += (params_.advance_tile * increment_cluster);
store_byte_pointer_ += (params_.advance_tile * increment_cluster);
thread_start_row_ +=
ThreadMap::Shape::kGroup *
ThreadMap::Shape::kRow *
ThreadMap::Shape::kCluster *
ThreadMap::Shape::kTile *
increment_cluster;
return *this;
}
///< Efficiently disables all accesses guarded by mask
CUTLASS_DEVICE void clear_mask() {
mask_.clear();
}
///< Efficiently enables all accesses guarded by mask
CUTLASS_DEVICE void enable_mask() {
mask_.enable();
}
///< Sets the mask
CUTLASS_DEVICE void get_mask(Mask &mask) const {
mask = mask_;
}
///< Sets the mask
CUTLASS_DEVICE void set_mask(Mask const &mask) {
mask_ = mask;
}
};
////////////////////////////////////////////////////////////////////////////////
/// Tile iterator used to load output tile from global memory in epilogue.
///
/// Satisfies: ReadableTileIterator | InterleavedPredicatedTileIterator | ForwardTileIterator
///
template <
typename ThreadMap_, ///< Thread map (conept: OutputTileThreadMap)
typename Element_, ///< Element data type
int InterleavedN ///< Number of Interleaved N
>
class InterleavedPredicatedTileIterator {
public:
using ThreadMap = ThreadMap_;
using Element = Element_;
using Layout = layout::ColumnMajorInterleaved<InterleavedN>;
using TensorRef = TensorRef<Element, Layout>;
using ConstTensorRef = typename TensorRef::ConstTensorRef;
using Index = typename Layout::Index;
using LongIndex = typename Layout::LongIndex;
using TensorCoord = layout::PitchLinearCoord;
static int const kElementsPerAccess = ThreadMap::kElementsPerAccess;
static int const kThreads = ThreadMap::kThreads;
static int const kIterations = ThreadMap::Iterations::kCount;
/// Fragment object
using Fragment = Array<Element, ThreadMap::kElementsPerAccess>;
/// Memory access size
using AccessType = AlignedArray<Element, ThreadMap::kElementsPerAccess>;
/// Uses a non-template class
struct Params : InterleavedPredicatedTileIteratorParams {
using Base = InterleavedPredicatedTileIteratorParams;
CUTLASS_HOST_DEVICE
Params() { }
CUTLASS_HOST_DEVICE
Params(Layout const &layout):
Base(
layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess,
make_InterleavedPredicatedTileIteratorDesc<Element, ThreadMap>()
) { }
CUTLASS_HOST_DEVICE
Params(Base const &base) :
Base(base) { }
};
/// Mask object
struct Mask {
static int const kCount = (ThreadMap::Iterations::kContiguous < 8)
? 8
: ThreadMap::Iterations::kContiguous;
/// Predicate state
bool predicates[kCount];
//
// Mask
//
CUTLASS_HOST_DEVICE
Mask() {
enable();
}
///< Efficiently disables all accesses guarded by mask
CUTLASS_HOST_DEVICE void clear() {
CUTLASS_PRAGMA_UNROLL
for (int i = 0; i < kCount; ++i) {
predicates[i] = false;
}
}
///< CUTLASS_HOST_DEVICE enables all accesses guarded by mask
CUTLASS_DEVICE void enable() {
CUTLASS_PRAGMA_UNROLL
for (int i = 0; i < kCount; ++i) {
predicates[i] = true;
}
}
};
private:
//
// Data members
//
/// Parameters structure containing reference and precomputed state.
Params params_;
/// Byte-level pointer
uint8_t *byte_pointer_;
/// Array of boolean values to contain steady-state predicates
Mask mask_;
/// Extent of the matrix tile in columns
Index extent_col_;
/// A thread's starting column position (assuming steady-state predicates have
/// been computed)
Index thread_start_col_;
/// Internal iteration counter
int iteration_contiguous_;
int iteration_strided_;
private:
//
// Methods
//
public:
//
// Methods
//
/// Constructor
CUTLASS_DEVICE
InterleavedPredicatedTileIterator(
Params const & params,
Element *pointer,
TensorCoord extent,
int thread_idx,
TensorCoord threadblock_offset,
int const *indices = nullptr ///< gather/scatter indices, note no support for gather/scatter at this specialization
):
params_(params) {
TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) +
TensorCoord(threadblock_offset.contiguous() * InterleavedN,
threadblock_offset.strided() / InterleavedN);
extent_col_ = extent.strided() / InterleavedN;
thread_start_col_ = thread_offset.strided();
// Initialize predicates
CUTLASS_PRAGMA_UNROLL
for (int c = 0; c < ThreadMap::Iterations::kContiguous; ++c) {
mask_.predicates[c] =
((thread_offset.contiguous() + ThreadMap::Delta::kContiguous * c) <
(extent.contiguous() * InterleavedN));
}
// Initialize pointer
byte_pointer_ = reinterpret_cast<uint8_t *>(pointer) +
LongIndex(thread_offset.strided()) * LongIndex(params_.stride) +
LongIndex(thread_offset.contiguous()) * sizeof(AccessType) / kElementsPerAccess;
// Initialize internal state counter
iteration_contiguous_ = iteration_strided_ = 0;
}
/// Adds a pointer offset in units of Element
CUTLASS_HOST_DEVICE
void add_pointer_offset(LongIndex pointer_offset) {
byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8;
}
/// Loads a fragment from memory
CUTLASS_DEVICE
void load(Fragment &frag) {
uint8_t *byte_pointer = byte_pointer_;
AccessType *frag_ptr = reinterpret_cast<AccessType *>(&frag);
AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer);
int col_offset = iteration_strided_ * ThreadMap::Delta::kStrided;
bool col_guard = ((thread_start_col_ + col_offset) < extent_col_);
bool guard = col_guard && mask_.predicates[iteration_contiguous_];
cutlass::arch::global_load<
AccessType,
sizeof(AccessType)
>(
*frag_ptr,
(void *)memory_pointer,
guard);
}
/// Stores a fragment to memory
CUTLASS_DEVICE
void store(Fragment const &frag) {
uint8_t *byte_pointer = byte_pointer_;
AccessType const *frag_ptr = reinterpret_cast<AccessType const *>(&frag);
AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer);
int col_offset = iteration_strided_ * ThreadMap::Delta::kStrided;
bool col_guard = ((thread_start_col_ + col_offset) < extent_col_);
bool guard = col_guard && mask_.predicates[iteration_contiguous_];
cutlass::arch::global_store<AccessType, sizeof(AccessType)>(
*frag_ptr, (void *)memory_pointer, guard);
}
/// Overrides the internal iteration index
CUTLASS_HOST_DEVICE
void set_iteration_index(int iteration) {
iteration_contiguous_ = iteration % ThreadMap::Iterations::kContiguous;
iteration_strided_ = iteration / ThreadMap::Iterations::kContiguous;
}
/// Advances to the next position to load or store
CUTLASS_HOST_DEVICE
InterleavedPredicatedTileIterator &operator++() {
++iteration_contiguous_;
byte_pointer_ += params_.advance_row;
if (iteration_contiguous_ == ThreadMap::Iterations::kContiguous) {
iteration_contiguous_ = 0;
++iteration_strided_;
byte_pointer_ += params_.advance_column;
if (iteration_strided_ == ThreadMap::Iterations::kStrided) {
iteration_strided_ = 0;
}
}
return *this;
}
/// Advances a number of positions to load or store
CUTLASS_HOST_DEVICE
InterleavedPredicatedTileIterator &operator+=(int increment)
{
// Contiguous
iteration_contiguous_ += increment;
int increment_strided = iteration_contiguous_ / ThreadMap::Iterations::kContiguous;