forked from NVIDIA/cutlass
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlayout.hpp
1687 lines (1455 loc) · 50.7 KB
/
layout.hpp
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) 2023 - 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.
*
**************************************************************************************************/
#pragma once
#include <cute/config.hpp>
#include <cute/underscore.hpp>
#include <cute/int_tuple.hpp>
#include <cute/stride.hpp>
#include <cute/numeric/arithmetic_tuple.hpp>
namespace cute
{
// Aliases
template <class... Shapes>
using Shape = IntTuple<Shapes...>;
template <class... Strides>
using Stride = IntTuple<Strides...>;
template <class... Strides>
using Step = IntTuple<Strides...>;
template <class... Coords>
using Coord = IntTuple<Coords...>;
template <class... Ts>
CUTE_HOST_DEVICE constexpr
Shape<Ts...>
make_shape(Ts const&... t) {
return {t...};
}
template <class... Ts>
CUTE_HOST_DEVICE constexpr
Stride<Ts...>
make_stride(Ts const&... t) {
return {t...};
}
template <class... Ts>
CUTE_HOST_DEVICE constexpr
Step<Ts...>
make_step(Ts const&... t) {
return {t...};
}
template <class... Ts>
CUTE_HOST_DEVICE constexpr
Coord<Ts...>
make_coord(Ts const&... t) {
return {t...};
}
template <class Shape, class Stride = LayoutLeft::Apply<Shape> >
struct Layout
: private cute::tuple<Shape, Stride> // EBO for static layouts
{
// Expensive in compilation time...
//static_assert(is_congruent<Shape, Stride>::value, "Shape and Stride must be congruent");
// NOTE: This defaults static Shapes/Strides correctly, but not dynamic
CUTE_HOST_DEVICE constexpr
Layout(Shape const& shape = {}, Stride const& stride = {})
: cute::tuple<Shape, Stride>(shape, stride)
{}
//
// Accessors
//
static constexpr int rank = rank_v<Shape>;
CUTE_HOST_DEVICE constexpr
decltype(auto)
layout() {
return *this;
}
CUTE_HOST_DEVICE constexpr
decltype(auto)
layout() const {
return *this;
}
template <int... I>
CUTE_HOST_DEVICE constexpr
decltype(auto)
shape() {
return get<0,I...>(static_cast<cute::tuple<Shape, Stride>&>(*this));
}
template <int... I>
CUTE_HOST_DEVICE constexpr
decltype(auto)
shape() const {
return get<0,I...>(static_cast<cute::tuple<Shape, Stride> const&>(*this));
}
template <int... I>
CUTE_HOST_DEVICE constexpr
decltype(auto)
stride() {
return get<1,I...>(static_cast<cute::tuple<Shape, Stride>&>(*this));
}
template <int... I>
CUTE_HOST_DEVICE constexpr
decltype(auto)
stride() const {
return get<1,I...>(static_cast<cute::tuple<Shape, Stride> const&>(*this));
}
//
// Mappings
//
// Map a logical coordinate to a linear index (Coord has no Underscore slice operators)
// OR
// Slice the layout and return the sublayout (Coord has an Underscore slice op)
template <class Coord>
CUTE_HOST_DEVICE constexpr
auto
operator()(Coord const& coord) const {
if constexpr (has_underscore<Coord>::value) {
return slice(coord, *this);
} else {
return crd2idx(coord, shape(), stride());
}
CUTE_GCC_UNREACHABLE;
}
// Convenience function for multi-dimensional coordinates
template <class Coord0, class Coord1, class... Coords>
CUTE_HOST_DEVICE constexpr
auto
operator()(Coord0 const& c0, Coord1 const& c1, Coords const&... cs) const {
return operator()(make_coord(c0,c1,cs...));
}
// Map a linear index to a hier ND logical coordinate
// NOTE: Dangerous and error-prone
template <class Int>
CUTE_HOST_DEVICE constexpr
auto
operator[](Int const& linear_idx) const {
static_assert(is_integral<Int>::value);
return get_hier_coord(linear_idx);
}
//
// Compose
//
template <class OtherLayout>
CUTE_HOST_DEVICE constexpr
auto
compose(OtherLayout const& other) const {
return composition(*this, other);
}
template <class... Layouts>
CUTE_HOST_DEVICE constexpr
auto
compose(Layouts const&... layouts) const {
return composition(*this, make_tile(layouts...));
}
template <class OtherShape>
CUTE_HOST_DEVICE constexpr
auto
with_shape(OtherShape const& shape) const {
return composition(*this, make_layout(shape));
}
template <class... Shapes>
CUTE_HOST_DEVICE constexpr
auto
with_shape(Shapes const&... shapes) const {
return composition(*this, make_layout(make_shape(shapes...)));
}
//
// Tile
//
template <class OtherLayout>
CUTE_HOST_DEVICE constexpr
auto
tile(OtherLayout const& other) const {
return tiled_divide(*this, other);
}
template <class... Layouts>
CUTE_HOST_DEVICE constexpr
auto
tile(Layouts const&... layouts) const {
return tiled_divide(*this, make_tile(layouts...));
}
//
// Utility
//
//
// Index to Coordinate
//
// NOTE: Only valid for compact layouts
// Return the (hierarchical) ND logical coordinate corresponding to the linear index
// @post crd2idx(@a result, shape(), stride()) == idx
// @post congruent(@a result, shape())
template <class IInt,
__CUTE_REQUIRES(is_integral<IInt>::value)>
CUTE_HOST_DEVICE constexpr
auto
get_hier_coord(IInt const& idx) const {
return cute::idx2crd(idx, shape(), stride());
}
// Return the (flat) ND logical coordinate corresponding to the linear index
// @post crd2idx(@a result, shape(), stride()) == idx
// @post rank(@a result) == rank(shape()) && depth(@a result) == 1
template <class IInt,
__CUTE_REQUIRES(is_integral<IInt>::value)>
CUTE_HOST_DEVICE constexpr
auto
get_flat_coord(IInt const& idx) const {
return cute::crd2crd(this->get_hier_coord(idx), shape(), repeat<rank>(Int<1>{}));
}
// Return the generalized column-major 1D logical coordinate corresponding to the linear index
// @post crd2idx(@a result, shape(), stride()) == idx
// @post is_integral<decltype(@a result)>::value
template <class IInt,
__CUTE_REQUIRES(is_integral<IInt>::value)>
CUTE_HOST_DEVICE constexpr
auto
get_1d_coord(IInt const& idx) const {
return cute::crd2idx(this->get_hier_coord(idx), shape());
}
//
// Coordinate to Coordinate
//
#if 0
// Return the (hierarchical) ND logical coordinate corresponding to the linear index
// @post congruent(@a result, shape())
template <class Coord>
CUTE_HOST_DEVICE constexpr
auto
crd_2_hier_coord(Coord const& crd) const {
return cute::crd2crd(crd, shape(), shape());
}
// Return the (flat) ND logical coordinate corresponding to the linear index
// @post rank(@a result) == rank(shape()) && depth(@a result) == 1
template <class Coord>
CUTE_HOST_DEVICE constexpr
auto
crd_2_flat_coord(Coord const& crd) const {
return cute::crd2crd(crd, shape(), product_each(shape()));
}
// Return the generalized column-major 1D logical coordinate corresponding to the linear index
// @post is_integral<decltype(@a result)>::value
template <class Coord>
CUTE_HOST_DEVICE constexpr
auto
crd_2_1d_coord(Coord const& crd) const {
//return cute::crd2crd(crd, shape(), product(shape()));
return cute::crd2idx(crd, shape());
}
#endif
};
template <class Layout>
struct is_layout : false_type {};
template <class Shape, class Stride>
struct is_layout<Layout<Shape,Stride>> : true_type {};
template <class Shape, class Stride,
__CUTE_REQUIRES((is_tuple<Shape >::value || is_integral<Shape >::value) &&
(is_tuple<Stride>::value || is_integral<Stride>::value))>
CUTE_HOST_DEVICE constexpr
auto
make_layout(Shape const& shape, Stride const& stride)
{
return Layout<Shape,Stride>(shape, stride);
}
template <class Shape,
__CUTE_REQUIRES(is_tuple<Shape>::value || is_integral<Shape>::value)>
CUTE_HOST_DEVICE constexpr
auto
make_layout(Shape const& shape)
{
return make_layout(shape, compact_col_major(shape));
}
// Construct a layout from multiple layouts by
// concatenating each layout as an independent mode
template <class... Shapes, class... Strides>
CUTE_HOST_DEVICE constexpr
auto
make_layout(Layout<Shapes,Strides> const&... layouts)
{
return make_layout(make_shape (layouts.shape()...),
make_stride(layouts.stride()...));
}
//
// Convenience tags for common layouts
//
template <class Shape>
CUTE_HOST_DEVICE constexpr
auto
make_layout(Shape const& shape, GenColMajor)
{
return make_layout(shape, compact_col_major(shape));
}
template <class Shape>
CUTE_HOST_DEVICE constexpr
auto
make_layout(Shape const& shape, GenRowMajor)
{
return make_layout(shape, compact_row_major(shape));
}
// Follow the same ordering induced by the strides, but make the layout compact
template <class Shape, class Order>
CUTE_HOST_DEVICE constexpr
auto
make_ordered_layout(Shape const& shape, Order const& order)
{
static_assert(is_static<Shape>::value && is_static<Order>::value);
return make_layout(shape, compact_order(shape, order));
}
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
make_ordered_layout(Layout<Shape,Stride> const& layout)
{
return make_ordered_layout(layout.shape(), layout.stride());
}
// Make a layout of the same shape that is either ordered or colmajor depending on staticness
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
make_layout_like(Layout<Shape,Stride> const& layout)
{
auto any_zero = any_of(layout.stride(), [](auto d) { return is_constant<0, decltype(d)>{}; });
if constexpr (any_zero) {
// If there are static-0 strides, then make a col-major layout that keeps those 0s
return make_layout(layout.shape(),
compact_col_major(filter_zeros(layout.stride(), layout.shape())));
} else
if constexpr (is_static<Shape>::value && is_static<Stride>::value) {
// If the layout is fully static, then make a layout that follows the same order as the strides
// Assumes the strides are unique
return make_ordered_layout(layout.shape(), layout.stride());
} else {
return make_layout(layout.shape());
}
CUTE_GCC_UNREACHABLE;
}
//
// Make a layout of the same shape,
// with mode-0 being colmajor then following the mode order in layout
//
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
make_fragment_like(Layout<Shape,Stride> const& layout)
{
constexpr int R = Layout<Shape,Stride>::rank;
if constexpr (R > 1 && is_static<Shape>::value && is_static<Stride>::value) {
return tiled_product(make_layout(shape<0>(layout)), make_ordered_layout(take<1,R>(layout)));
} else {
return make_layout(layout.shape());
}
CUTE_GCC_UNREACHABLE;
}
template <class Shape,
__CUTE_REQUIRES(is_tuple<Shape>::value || is_integral<Shape>::value)>
CUTE_HOST_DEVICE constexpr
auto
make_fragment_like(Shape const& shape)
{
return make_layout(shape);
}
//
// Make an identity layout that maps a coordinate to itself
//
template <class Shape>
CUTE_HOST_DEVICE constexpr
auto
make_identity_layout(Shape const& shape)
{
return make_layout(shape, make_basis_like(shape));
}
//
// Operations to manipulate Layouts like a tuple of pairs
//
template <size_t... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
get(Layout<Shape,Stride> const& layout)
{
// Let the static_asserts in get<I>(shape|stride) catch problems
return make_layout(get<Is...>(layout.shape()), get<Is...>(layout.stride()));
}
template <int B, int E, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
take(Layout<Shape,Stride> const& layout)
{
// Let the static_asserts in take<B,E>(shape|stride) catch problems
return make_layout(take<B,E>(layout.shape()), take<B,E>(layout.stride()));
}
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
flatten(Layout<Shape,Stride> const& layout)
{
return make_layout(flatten(layout.shape()), flatten(layout.stride()));
}
//
// Utilities
//
// Return the layout of a mode
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
decltype(auto)
layout(Layout<Shape,Stride> const& layout)
{
if constexpr (sizeof...(Is) == 0) {
return layout;
} else {
return get<Is...>(layout);
}
CUTE_GCC_UNREACHABLE;
}
// Return the shape of a mode
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
decltype(auto)
shape(Layout<Shape,Stride>& layout)
{
return layout.template shape<Is...>();
}
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
decltype(auto)
shape(Layout<Shape,Stride> const& layout)
{
return layout.template shape<Is...>();
}
// Return the stride of a mode
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
decltype(auto)
stride(Layout<Shape,Stride>& layout)
{
return layout.template stride<Is...>();
}
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
decltype(auto)
stride(Layout<Shape,Stride> const& layout)
{
return layout.template stride<Is...>();
}
// Return the number of elements in a mode
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
size(Layout<Shape,Stride> const& layout)
{
return size(shape<Is...>(layout));
}
// Return the number of modes
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
rank(Layout<Shape,Stride> const& layout)
{
return rank(shape<Is...>(layout));
}
// Return the depth of the layout
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
depth(Layout<Shape,Stride> const& layout)
{
return depth(shape<Is...>(layout));
}
// Return the codomain size of a mode
// @return M smallest integer such that @a sub_layout(c) < M for all c < size(@a sub_layout)
// where sub_layout = get<Is...>(layout).
template <int... Is, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
cosize(Layout<Shape,Stride> const& layout)
{
// Protect against negative strides
auto abs_sub_layout = make_layout(shape<Is...>(layout),
transform_leaf(stride<Is...>(layout), abs_fn{}));
return abs_sub_layout(size(abs_sub_layout) - Int<1>{}) + Int<1>{};
}
template <class Layout>
using cosize_t = decltype(cosize(declval<Layout>()));
template <class Layout>
static constexpr int cosize_v = cosize_t<Layout>::value;
// Equality
// Return a static or dynamic boolean
template <class ShapeA, class StrideA,
class ShapeB, class StrideB>
CUTE_HOST_DEVICE constexpr
auto
operator==(Layout<ShapeA,StrideA> const& layoutA, Layout<ShapeB,StrideB> const& layoutB)
{
return layoutA.shape() == layoutB.shape() && layoutA.stride() == layoutB.stride();
}
// With crd2idx(coord, shape), makes sense to have crd2idx(coord, Layout) as well
template <class Coord, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
crd2idx(Coord const& c, Layout<Shape,Stride> const& layout)
{
return crd2idx(c, layout.shape(), layout.stride());
}
//
// Slice and Dice a layout
//
template <class Coord, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
slice(Coord const& c, Layout<Shape,Stride> const& layout)
{
return make_layout(slice(c, layout.shape()),
slice(c, layout.stride()));
}
template <class Coord, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
slice_and_offset(Coord const& c, Layout<Shape,Stride> const& layout)
{
return cute::make_tuple(slice(c, layout), crd2idx(c, layout));
}
template <class Coord, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
dice(Coord const& c, Layout<Shape,Stride> const& layout)
{
return make_layout(dice(c, layout.shape()),
dice(c, layout.stride()));
}
//
// Transform the modes of a layout
//
namespace detail {
template <class Tuple, class F, int... I>
CUTE_HOST_DEVICE constexpr
auto
transform_layout(Tuple const& t, F&& f, seq<I...>)
{
return make_layout(f(get<I>(t))...);
}
template <class Tuple0, class Tuple1, class F, int... I, int... I0, int... I1>
CUTE_HOST_DEVICE constexpr
auto
transform_layout(Tuple0 const& t0, Tuple1 const& t1, F&& f, seq<I...>, seq<I0...>, seq<I1...>)
{
return make_layout(f(get<I>(t0),get<I>(t1))..., get<I0>(t0)..., get<I1>(t1)...);
}
} // end namespace detail
template <class Tuple, class F>
CUTE_HOST_DEVICE constexpr
auto
transform_layout(Tuple const& t, F&& f)
{
return detail::transform_layout(t, f, make_seq<decltype(rank(t))::value>{});
}
template <class Tuple0, class Tuple1, class F>
CUTE_HOST_DEVICE constexpr
auto
transform_layout(Tuple0 const& t0, Tuple1 const& t1, F&& f)
{
constexpr int R0 = decltype(rank(t0))::value;
constexpr int R1 = decltype(rank(t1))::value;
constexpr int R = (R0 < R1) ? R0 : R1;
return detail::transform_layout(t0, t1, f, make_seq<R>{}, make_range<R,R0>{}, make_range<R,R1>{});
}
//
// Coalesce and Filter
//
namespace detail {
// Look at each element and the front of the stack (in order of priority)
// front(NewLayout) get<I>(Layout)
// s0:d0 _1:d1 => continue
// _1:d0 s1:d1 => replace_front s1:d1
// s0:s1*d1 s1:d1 => replace_front s0*s1:d1
// s0:d0 s1:d1 => prepend s1:d1
//
// @pre OldShape and OldStride are flat
template <int I, class OldShape, class OldStride, class NewShape, class NewStride>
CUTE_HOST_DEVICE constexpr
auto
bw_coalesce(OldShape const& old_shape, OldStride const& old_stride,
NewShape const& new_shape, NewStride const& new_stride)
{
if constexpr (I == -1) {
// Base case, we're done
if constexpr (is_constant<1, NewShape>::value) {
return Layout<_1,_0>{};
} else {
return Layout<NewShape,NewStride>{new_shape,new_stride};
}
} else if constexpr (is_constant<1, decltype(get<I>(old_shape))>::value) {
// shape<I>(layout) == _1, skip it and continue
return bw_coalesce<I-1>(old_shape, old_stride, new_shape, new_stride);
} else if constexpr (is_constant<1, NewShape>::value) {
// Replace our shape-1 with anything (Can only happen on input new_shape/new_stride)
return bw_coalesce<I-1>(old_shape, old_stride, get<I>(old_shape), get<I>(old_stride));
} else if constexpr (is_constant<true, decltype(get<I>(old_shape) * get<I>(old_stride) == get<0>(new_stride))>::value) {
// Merge modes because the shapes and strides match
return bw_coalesce<I-1>(old_shape, old_stride,
replace_front(new_shape, get<I>(old_shape) * get<0>(new_shape)),
replace_front(new_stride, get<I>(old_stride)));
} else {
// Can't replace or merge, so prepend a new mode
return bw_coalesce<I-1>(old_shape, old_stride,
prepend(new_shape, get<I>(old_shape)),
prepend(new_stride, get<I>(old_stride)));
}
CUTE_GCC_UNREACHABLE;
}
} // end namespace detail
// Combine all the modes that are possible to combine
// Does not respect the profile of the layout, but does preserve total size
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
coalesce(Layout<Shape,Stride> const& layout)
{
auto flat_shape = flatten(layout.shape());
auto flat_stride = flatten(layout.stride());
constexpr int R = decltype(rank(flat_shape))::value;
return detail::bw_coalesce<R-2>(flat_shape, flat_stride, get<R-1>(flat_shape), get<R-1>(flat_stride));
}
// Apply coalesce at the terminals of trg_profile
template <class Shape, class Stride, class IntTuple>
CUTE_HOST_DEVICE constexpr
auto
coalesce(Layout<Shape,Stride> const& layout, IntTuple const& trg_profile)
{
if constexpr (is_tuple<IntTuple>::value) {
static_assert(tuple_size<IntTuple>::value <= Layout<Shape,Stride>::rank);
return transform_layout(layout, trg_profile, [](auto const& l, auto const& t) { return coalesce(l,t); });
} else {
return coalesce(layout);
}
CUTE_GCC_UNREACHABLE;
}
// Replace the modes in layout that have a 0-stride with a 1-size
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
filter_zeros(Layout<Shape,Stride> const& layout)
{
return make_layout(filter_zeros(layout.stride(), layout.shape()), layout.stride());
}
// Remove all of the 0-strides and 1-sizes
// Return 1-shape if empty
template <class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
filter(Layout<Shape,Stride> const& layout)
{
return coalesce(filter_zeros(layout));
}
// Apply filter at the terminals of trg_profile
template <class Shape, class Stride, class IntTuple>
CUTE_HOST_DEVICE constexpr
auto
filter(Layout<Shape,Stride> const& layout, IntTuple const& trg_profile)
{
if constexpr (is_tuple<IntTuple>::value) {
static_assert(tuple_size<IntTuple>::value <= Layout<Shape,Stride>::rank);
return transform_layout(layout, trg_profile, [](auto const& l, auto const& t) { return filter(l,t); });
} else {
return filter(layout);
}
CUTE_GCC_UNREACHABLE;
}
//
// Append, Prepend, Replace
//
template <int N, class ShapeA, class StrideA, class ShapeX = _1, class StrideX = _0>
CUTE_HOST_DEVICE constexpr
auto
append(Layout<ShapeA,StrideA> const& layout,
Layout<ShapeX,StrideX> const& x = {})
{
return make_layout(append<N>(layout.shape(), x.shape()),
append<N>(layout.stride(), x.stride()));
}
template <int N, class ShapeA, class StrideA, class ShapeX = _1, class StrideX = _0>
CUTE_HOST_DEVICE constexpr
auto
prepend(Layout<ShapeA,StrideA> const& layout,
Layout<ShapeX,StrideX> const& x = {})
{
return make_layout(prepend<N>(layout.shape(), x.shape()),
prepend<N>(layout.stride(), x.stride()));
}
template <int N, class ShapeA, class StrideA, class ShapeX, class StrideX>
CUTE_HOST_DEVICE constexpr
auto
replace(Layout<ShapeA,StrideA> const& layout,
Layout<ShapeX,StrideX> const& x)
{
return make_layout(replace<N>(layout.shape(), x.shape()),
replace<N>(layout.stride(), x.stride()));
}
template <int B, int E, class Shape, class Stride>
CUTE_HOST_DEVICE constexpr
auto
group(Layout<Shape,Stride> const& layout)
{
return make_layout(group<B,E>(layout.shape()),
group<B,E>(layout.stride()));
}
//
// Composition of two layouts: lhs o rhs
// @post compatible(rhs, result)
// @post result(c) = lhs(rhs(c))
// for all c in the domain of result
//
namespace detail {
template <class LShape, class LStride,
class RShape, class RStride>
CUTE_HOST_DEVICE constexpr
auto
composition(Layout<LShape,LStride> const& lhs,
RShape const& rhs_shape, RStride const& rhs_stride)
{
if constexpr (is_tuple<RShape>::value) {
// Apply the right-distributivity of Layout composition
return transform_layout(rhs_shape, rhs_stride, [&](auto const& s, auto const& d) { return composition(lhs, s, d); });
} else
if constexpr (is_scaled_basis<RStride>::value) {
// Special case for a ScaledBasis stride
return composition(get<rhs_stride.mode()>(lhs), rhs_shape, rhs_stride.value());
} else
if constexpr (is_integral<RStride>::value) {
// Integral Rstride (and RShape)
// NOTE: Should only flatten once for efficiency
auto flat_shape = flatten(lhs.shape());
[[maybe_unused]] auto flat_stride = flatten(lhs.stride());
[[maybe_unused]] constexpr int R = rank(flat_shape);
if constexpr (is_constant<0, RStride>::value) {
// Special case shortcut for any static stride-0
return Layout<RShape, RStride>{rhs_shape, rhs_stride};
} else
if constexpr (is_integral<decltype(flat_shape)>::value) {
// Special case shortcut for any integral LShape
auto result_stride = rhs_stride * flat_stride;
return Layout<RShape, decltype(result_stride)>{rhs_shape, result_stride};
} else
if constexpr (is_constant<1, RStride>::value) {
// Special case shortcut for any static stride-1
auto result_shape_0 = take<0,R-1>(flat_shape);
// Mod out the rhs_shape from the lhs.shape()
auto const [result_shape_1, rest_shape] = fold(result_shape_0, cute::make_tuple(cute::make_tuple(), rhs_shape),
[] (auto const& init, auto const& si) {
return cute::make_tuple(append(get<0>(init), cute::min(abs(si), get<1>(init))), shape_div(get<1>(init), abs(si)));
});
// Jump into coalesce and append (rest_shape, get<R-1>(lhs.stride())
return detail::bw_coalesce<R-2>(result_shape_1, flat_stride, rest_shape, get<R-1>(flat_stride));
} else
{
// General case
auto result_shape_0 = take<0,R-1>(flat_shape);
auto result_stride_0 = take<0,R-1>(flat_stride);
// Divide out the rhs_stride from the lhs.shape()
auto const [result_shape_1, rest_stride] = fold(result_shape_0, cute::make_tuple(cute::make_tuple(), rhs_stride),
[] (auto const& init, auto const& di) {
return cute::make_tuple(append(get<0>(init), shape_div(di, get<1>(init))), shape_div(get<1>(init), di));
});
// Apply any lhs.shape() changes to the stride
auto result_stride_1 = elem_scale(result_stride_0, shape_div(result_shape_0, result_shape_1));
// Mod out the rhs_shape from the lhs.shape()
auto const [result_shape_2, rest_shape] = fold(result_shape_1, cute::make_tuple(cute::make_tuple(), rhs_shape),
[] (auto const& init, auto const& si) {
return cute::make_tuple(append(get<0>(init), cute::min(abs(si), get<1>(init))), shape_div(get<1>(init), abs(si)));
});
// Jump into coalesce and append (rest_shape, rest_stride * get<R-1>(lhs.stride())
return detail::bw_coalesce<R-2>(result_shape_2, result_stride_1, rest_shape, rest_stride * get<R-1>(flat_stride));
}
}
CUTE_GCC_UNREACHABLE;
}
} // end namespace detail
template <class LShape, class LStride,
class RShape, class RStride>
CUTE_HOST_DEVICE constexpr
auto
composition(Layout<LShape,LStride> const& lhs,
Layout<RShape,RStride> const& rhs)
{
//return detail::composition(flatten(lhs), rhs.shape(), rhs.stride());
return detail::composition(lhs, rhs.shape(), rhs.stride());
}
template <class LShape, class LStride, class IntTuple>
CUTE_HOST_DEVICE constexpr
auto
composition(Layout<LShape,LStride> const& lhs,
IntTuple const& rhs)
{
if constexpr (is_tuple<IntTuple>::value) {
static_assert(tuple_size<IntTuple>::value <= Layout<LShape,LStride>::rank);
// Drop any modes of lhs that aren't hit by rhs
return detail::transform_layout(lhs, rhs, [](auto const& l, auto const& r) { return composition(l,r); }, make_seq<tuple_size<IntTuple>::value>{}, seq<>{}, seq<>{});
} else if constexpr (is_underscore<IntTuple>::value) {
return lhs;
} else {
return composition(lhs, make_layout(rhs));
}
CUTE_GCC_UNREACHABLE;
}
//
// Complement
//
// Build the complement of a layout.
// @post size(@a result) >= @a cosize_hi / size(filter(@a layout)));
// @post For all i in [1,size(@a result)),
// @a result(i) < @a result(i-1)
// For all j in [0, size(@a layout)),
// @a result(i) != @a layout(j)
//
namespace detail {
// @pre @a layout has been filtered (flattened and no stride-0 or size-1 modes).
template <class Shape, class Stride, class CoSizeHi>
CUTE_HOST_DEVICE constexpr
auto
complement(Shape const& shape, Stride const& stride, CoSizeHi const& cosize_hi)
{
if constexpr (is_constant<0, Stride>::value) {
// Special case for irreducible rank-1 stride-0 layout
return make_layout(cosize_hi);
} else {
// General case
constexpr int R = rank_v<Shape>;
static_assert(R == 1 || is_static<Stride>::value,
"Dynamic-stride complement only for rank-1 layouts");
// Should just be a sort and a fold...
// Then we could even handle dynamic strides (but they would destroy all static strides)
auto result = fold(make_seq<R-1>{},
cute::make_tuple(shape,
stride,
cute::make_tuple(),
cute::make_tuple(Int<1>{})),
[](auto const& init, auto i)
{
auto curr_stride = cute::min(get<1>(init));
auto curr_idx = find(get<1>(init), curr_stride);
auto curr_shape = get<curr_idx>(get<0>(init));
return cute::make_tuple(remove<curr_idx>(get<0>(init)), // Remove the curr shape
remove<curr_idx>(get<1>(init)), // Remove the curr stride
append(get<2>(init), curr_stride / get<3,i>(init)), // new shape = curr_stride / last_stride
append(get<3>(init), curr_shape * curr_stride)); // new stride = curr_shape * curr_stride
});
// Append the last shape mode
auto result_stride = get<3>(result);
auto result_shape = append(get<2>(result), get<1,0>(result) / back(result_stride)); // new shape = curr_stride / last_stride
// Compute the rest_stride
auto rest_stride = get<0,0>(result) * get<1,0>(result);
//return make_layout(append(result_shape, ceil_div(cosize_hi, rest_stride)), append(result_stride, rest_stride));
// Jump into coalesce and append (ceil_div(cosize_hi, rest_stride), rest_stride)
return detail::bw_coalesce<R-1>(result_shape, result_stride, ceil_div(cosize_hi, rest_stride), rest_stride);
}
CUTE_GCC_UNREACHABLE;
}