forked from NVIDIA/cutlass
-
Notifications
You must be signed in to change notification settings - Fork 0
/
kernel_forward.h
1323 lines (1205 loc) · 51.4 KB
/
kernel_forward.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 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* 3. Neither the name of the copyright holder nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
#pragma once
#ifdef HAS_PYTORCH
#include <ATen/cuda/CUDAGeneratorImpl.h>
#include <ATen/cuda/CUDAGraphsUtils.cuh>
#endif
#include <curand_kernel.h>
#include <cmath>
#include <vector>
#include "cutlass/fast_math.h"
#include "cutlass/gemm/gemm.h"
#include "cutlass/layout/matrix.h"
#include "cutlass/layout/vector.h"
#include "cutlass/matrix.h"
#include "cutlass/numeric_types.h"
#include "cutlass/tensor_ref.h"
#include "cutlass/epilogue/threadblock/default_epilogue_simt.h"
#include "cutlass/epilogue/threadblock/default_epilogue_tensor_op.h"
#include "cutlass/epilogue/threadblock/default_epilogue_volta_tensor_op.h"
#include "cutlass/gemm/device/default_gemm_configuration.h"
#include "cutlass/gemm/kernel/default_gemm.h"
#include "cutlass/gemm/threadblock/default_mma.h"
#include "cutlass/gemm/threadblock/default_mma_core_simt.h"
#include "cutlass/gemm/threadblock/default_mma_core_sm70.h"
#include "cutlass/gemm/threadblock/default_mma_core_sm75.h"
#include "cutlass/gemm/threadblock/default_mma_core_sm80.h"
#include "cutlass/gemm/threadblock/threadblock_swizzle.h"
#include "cutlass/matrix_shape.h"
#include "cutlass/platform/platform.h"
#include "cutlass/transform/threadblock/predicated_tile_iterator.h"
#include "debug_utils.h"
#include "epilogue/epilogue_pipelined.h"
#include "epilogue/epilogue_rescale_output.h"
#include "gemm/custom_mma.h"
#include "gemm/find_default_mma.h"
#include "gemm/mma_from_smem.h"
#include "gemm_kernel_utils.h"
#include "transform/tile_smem_loader.h"
#include <inttypes.h>
using namespace gemm_kernel_utils;
namespace {
template <typename scalar_t, typename Arch>
constexpr int getWarpsPerSmFw() {
return (
Arch::kMinComputeCapability >= 80 &&
!cutlass::platform::is_same<scalar_t, float>::value
? 16
: 12);
}
static CUTLASS_DEVICE float atomicMaxFloat(float* addr, float value) {
// source: https://stackoverflow.com/a/51549250
return (value >= 0)
? __int_as_float(atomicMax((int*)addr, __float_as_int(value)))
: __uint_as_float(atomicMin((unsigned int*)addr, __float_as_uint(value)));
}
} // namespace
// If ToBatchHookType_ is supplied other than this default (which is
// never the case in the xformers library) then the user is
// defining the logic which each block uses to find its data to work on,
// with the advance_to_batch function with the following signature.
// It should return false if there is no work to do for this block.
// In general this will not work with saving for backward due to fixed layout
// for logsumexp and incompatible rngs for dropout, so is likely only useful for
// custom inference.
struct DefaultToBatchHook {
template <typename Params>
CUTLASS_DEVICE static bool advance_to_batch(
Params&,
int64_t& /* q_start */,
int64_t& /* k_start */) {
return true;
}
};
template <
// The datatype of Q/K/V
typename scalar_t_,
// Architecture we are targeting (eg `cutlass::arch::Sm80`)
typename ArchTag,
// If Q/K/V are correctly aligned in memory and we can run a fast kernel
bool isAligned_,
int kQueriesPerBlock_,
int kKeysPerBlock_,
// upperbound on `max(value.shape[-1], query.shape[-1])`
int kMaxK_ = (int)cutlass::platform::numeric_limits<uint32_t>::max(),
// This is quite slower on V100 for some reason
// Set to false if you know at compile-time you will never need dropout
bool kSupportsDropout_ = true,
bool kSupportsBias_ = true,
typename ToBatchHookType_ = DefaultToBatchHook>
struct AttentionKernel {
enum CustomMaskType {
NoCustomMask = 0,
CausalFromTopLeft = 1,
CausalFromBottomRight = 2,
NumCustomMaskTypes,
};
using scalar_t = scalar_t_;
using accum_t = float;
using lse_scalar_t = float;
using output_t = scalar_t;
// Accumulator between 2 iterations
// Using `accum_t` improves perf on f16 at the cost of
// numerical errors
using output_accum_t = accum_t;
static constexpr bool kSupportsDropout = kSupportsDropout_;
static constexpr bool kSupportsBias = kSupportsBias_;
static constexpr int kKeysPerBlock = kKeysPerBlock_;
static constexpr int kQueriesPerBlock = kQueriesPerBlock_;
static constexpr int kMaxK = kMaxK_;
static constexpr bool kIsAligned = isAligned_;
static constexpr bool kSingleValueIteration = kMaxK <= kKeysPerBlock;
static constexpr int32_t kAlignLSE = 32; // block size of backward
static constexpr bool kIsHalf = cutlass::sizeof_bits<scalar_t>::value == 16;
static constexpr bool kPreloadV =
ArchTag::kMinComputeCapability >= 80 && kIsHalf;
static constexpr bool kKeepOutputInRF = kSingleValueIteration;
static constexpr bool kNeedsOutputAccumulatorBuffer = !kKeepOutputInRF &&
!cutlass::platform::is_same<output_accum_t, output_t>::value;
static_assert(kQueriesPerBlock % 32 == 0, "");
static_assert(kKeysPerBlock % 32 == 0, "");
static constexpr int kNumWarpsPerBlock =
kQueriesPerBlock * kKeysPerBlock / (32 * 32);
static constexpr int kWarpSize = 32;
// Launch bounds
static constexpr int kNumThreads = kWarpSize * kNumWarpsPerBlock;
static constexpr int kMinBlocksPerSm =
getWarpsPerSmFw<scalar_t, ArchTag>() / kNumWarpsPerBlock;
struct Params {
// Input tensors
scalar_t* query_ptr = nullptr; // [num_queries, num_heads, head_dim]
scalar_t* key_ptr = nullptr; // [num_keys, num_heads, head_dim]
scalar_t* value_ptr = nullptr; // [num_keys, num_heads, head_dim_value]
scalar_t* attn_bias_ptr = nullptr; // [num_heads, num_queries, num_keys]
int32_t* seqstart_q_ptr = nullptr;
int32_t* seqstart_k_ptr = nullptr;
int32_t* seqlen_k_ptr = nullptr;
uint32_t causal_diagonal_offset = 0;
// Output tensors
output_t* output_ptr = nullptr; // [num_queries, num_heads, head_dim_value]
// [num_queries, num_heads, head_dim_value]
output_accum_t* output_accum_ptr = nullptr;
// [num_heads, num_queries] - can be null
lse_scalar_t* logsumexp_ptr = nullptr;
// Scale
accum_t scale = 0.0;
// Dimensions/strides
int32_t head_dim = 0;
int32_t head_dim_value = 0;
int32_t num_queries = 0;
int32_t num_keys = 0;
int32_t num_keys_absolute = 0;
uint8_t custom_mask_type = NoCustomMask;
int32_t q_strideM = 0;
int32_t k_strideM = 0;
int32_t v_strideM = 0;
int32_t bias_strideM = 0;
int32_t o_strideM = 0;
// Everything below is only used in `advance_to_block`
// and shouldn't use registers
int32_t q_strideH = 0;
int32_t k_strideH = 0;
int32_t v_strideH = 0;
int64_t bias_strideH = 0;
int64_t q_strideB = 0;
int64_t k_strideB = 0;
int64_t v_strideB = 0;
int64_t bias_strideB = 0;
int32_t num_batches = 0;
int32_t num_heads = 0;
// dropout
bool use_dropout = false;
unsigned long long dropout_batch_head_rng_offset = 0;
float dropout_prob = 0.0f;
#ifdef HAS_PYTORCH
at::PhiloxCudaState rng_engine_inputs = at::PhiloxCudaState(0, 0);
#endif
// Moves pointers to what we should process
// Returns "false" if there is no work to do
CUTLASS_DEVICE bool advance_to_block() {
auto batch_id = blockIdx.z;
auto head_id = blockIdx.y;
auto query_start = blockIdx.x * kQueriesPerBlock;
auto lse_dim = ceil_div((int32_t)num_queries, kAlignLSE) * kAlignLSE;
if (kSupportsDropout) {
dropout_batch_head_rng_offset =
batch_id * num_heads * num_queries * num_keys +
head_id * num_queries * num_keys;
}
int64_t q_start = 0, k_start = 0;
// Advance to current batch - in case of different sequence lengths
constexpr bool kToBatchHook =
!cutlass::platform::is_same<ToBatchHookType_, DefaultToBatchHook>::
value;
if (kToBatchHook) {
// Call out to a custom implementation.
if (!ToBatchHookType_::advance_to_batch(*this, q_start, k_start)) {
return false;
}
} else if (seqstart_q_ptr != nullptr) {
assert(seqstart_k_ptr != nullptr);
seqstart_q_ptr += batch_id;
q_start = seqstart_q_ptr[0];
int64_t q_next_start = seqstart_q_ptr[1];
int64_t k_end;
seqstart_k_ptr += batch_id;
if (seqlen_k_ptr) {
k_start = seqstart_k_ptr[0];
k_end = k_start + seqlen_k_ptr[batch_id];
} else {
k_start = seqstart_k_ptr[0];
k_end = seqstart_k_ptr[1];
}
num_queries = q_next_start - q_start;
num_keys = k_end - k_start;
if (query_start >= num_queries) {
return false;
}
} else {
query_ptr += batch_id * q_strideB;
key_ptr += batch_id * k_strideB;
value_ptr += batch_id * v_strideB;
output_ptr += int64_t(batch_id * num_queries) * o_strideM;
if (output_accum_ptr != nullptr) {
output_accum_ptr +=
int64_t(batch_id * num_queries) * (head_dim_value * num_heads);
}
q_start = 0;
k_start = 0;
}
// Advance to the current batch / head / query_start
query_ptr += (q_start + query_start) * q_strideM + head_id * q_strideH;
key_ptr += k_start * k_strideM + head_id * k_strideH;
value_ptr += k_start * v_strideM + head_id * v_strideH;
output_ptr +=
int64_t(q_start + query_start) * o_strideM + head_id * head_dim_value;
if (kSupportsBias && attn_bias_ptr != nullptr) {
attn_bias_ptr += (batch_id * bias_strideB) + (head_id * bias_strideH);
}
if (output_accum_ptr != nullptr) {
output_accum_ptr +=
int64_t(q_start + query_start) * (head_dim_value * num_heads) +
head_id * head_dim_value;
} else {
// Accumulate directly in the destination buffer (eg for f32)
output_accum_ptr = (accum_t*)output_ptr;
}
if (logsumexp_ptr != nullptr) {
// lse[batch_id, head_id, query_start]
logsumexp_ptr +=
batch_id * lse_dim * num_heads + head_id * lse_dim + query_start;
}
// Custom masking
if (custom_mask_type == CausalFromBottomRight) {
causal_diagonal_offset = num_keys - num_queries;
}
// We use num_keys_absolute to index into the rng_state
// We need this index to match between forward and backwards
num_keys_absolute = num_keys;
if (custom_mask_type == CausalFromTopLeft ||
custom_mask_type == CausalFromBottomRight) {
// the bottom row of the current block is query_start + kQueriesPerBlock
// the last active key is then query_start + causal_diagonal_offset +
// kQueriesPerBlock so num_keys is the min between actual num_keys and
// this to avoid extra computations
num_keys = cutlass::fast_min(
int32_t(query_start + causal_diagonal_offset + kQueriesPerBlock),
num_keys);
}
num_queries -= query_start;
num_batches = 0; // no longer used after
// If num_queries == 1, and there is only one key head we're wasting
// 15/16th of tensor core compute In that case :
// - we only launch kernels for head_id % kQueriesPerBlock == 0
// - we iterate over heads instead of queries (strideM = strideH)
if (num_queries == 1 && k_strideH == 0 && v_strideH == 0) {
if (head_id % kQueriesPerBlock != 0)
return false;
q_strideM = q_strideH;
num_queries = num_heads;
num_heads = 1; // unused but here for intent
// remove causal since n_query = 1
// otherwise, offset would change with head !
custom_mask_type = NoCustomMask;
o_strideM = head_dim_value;
}
// Make sure the compiler knows these variables are the same on all
// the threads of the warp.
// Only worth doing if they could have been modified above.
query_ptr = warp_uniform(query_ptr);
key_ptr = warp_uniform(key_ptr);
value_ptr = warp_uniform(value_ptr);
if (kSupportsBias) {
attn_bias_ptr = warp_uniform(attn_bias_ptr);
}
output_ptr = warp_uniform(output_ptr);
output_accum_ptr = warp_uniform(output_accum_ptr);
logsumexp_ptr = warp_uniform(logsumexp_ptr);
num_queries = warp_uniform(num_queries);
num_keys = warp_uniform(num_keys);
num_heads = warp_uniform(num_heads);
o_strideM = warp_uniform(o_strideM);
custom_mask_type = warp_uniform(custom_mask_type);
return true;
}
__host__ dim3 getBlocksGrid() const {
return dim3(
ceil_div(num_queries, (int32_t)kQueriesPerBlock),
num_heads,
num_batches);
}
__host__ dim3 getThreadsGrid() const {
return dim3(kWarpSize, kNumWarpsPerBlock, 1);
}
};
struct MM0 {
/*
In this first matmul, we compute a block of `Q @ K.T`.
While the calculation result is still hot in registers, we update
`mi`, `m_prime`, `s_prime` in shared-memory, and then store this value
into a shared-memory ("AccumulatorSharedStorage") that is used later as
operand A for the second matmul (see MM1)
*/
using GemmType = DefaultGemmType<ArchTag, scalar_t>;
using OpClass = typename GemmType::OpClass;
using DefaultConfig =
typename cutlass::gemm::device::DefaultGemmConfiguration<
OpClass,
ArchTag,
scalar_t,
scalar_t,
scalar_t, // ElementC
accum_t // ElementAccumulator
>;
static constexpr int kAlignmentA =
kIsAligned ? DefaultConfig::kAlignmentA : GemmType::kMinimumAlignment;
static constexpr int kAlignmentB =
kIsAligned ? DefaultConfig::kAlignmentB : GemmType::kMinimumAlignment;
using ThreadblockShape = cutlass::gemm::
GemmShape<kQueriesPerBlock, kKeysPerBlock, GemmType::ThreadK>;
using WarpShape = cutlass::gemm::GemmShape<32, 32, GemmType::WarpK>;
using DefaultMma = typename cutlass::gemm::threadblock::FindDefaultMma<
scalar_t, // ElementA,
cutlass::layout::RowMajor, // LayoutA,
kAlignmentA,
scalar_t, // ElementB,
cutlass::layout::ColumnMajor, // LayoutB,
kAlignmentB,
accum_t,
cutlass::layout::RowMajor, // LayoutC,
OpClass,
ArchTag, // ArchTag
ThreadblockShape, // ThreadblockShape
WarpShape, // WarpShape
typename GemmType::InstructionShape, // InstructionShape
ArchTag::kMinComputeCapability >= 80 && kIsHalf
? 4
: DefaultConfig::kStages,
typename GemmType::Operator // Operator
>::DefaultMma;
using MmaCore = typename DefaultMma::MmaCore;
using IteratorA = typename DefaultMma::IteratorA;
using IteratorB = typename DefaultMma::IteratorB;
using DefaultThreadblockMma = typename DefaultMma::ThreadblockMma;
using Mma = typename cutlass::platform::conditional<
kSingleValueIteration,
typename MakeCustomMma<DefaultThreadblockMma, kMaxK>::Mma,
DefaultThreadblockMma>::type;
using AccumLambdaIterator = typename DefaultMmaAccumLambdaIterator<
typename Mma::Operator::IteratorC,
accum_t,
kWarpSize>::Iterator;
static_assert(
MmaCore::WarpCount::kM * MmaCore::WarpCount::kN *
MmaCore::WarpCount::kK ==
kNumWarpsPerBlock,
"");
// used for efficient load of bias tile Bij from global to shared memory
using BiasLoader = TileSmemLoader<
scalar_t,
cutlass::MatrixShape<kQueriesPerBlock, kKeysPerBlock>,
MmaCore::kThreads,
// input restriction: kv_len has to be a multiple of this value
128 / cutlass::sizeof_bits<scalar_t>::value>;
// Epilogue to store to shared-memory in a format that we can use later for
// the second matmul
using B2bGemm = typename cutlass::gemm::threadblock::B2bGemm<
typename Mma::Operator::IteratorC,
typename Mma::Operator,
scalar_t,
WarpShape,
ThreadblockShape>;
using AccumulatorSharedStorage = typename B2bGemm::AccumulatorSharedStorage;
};
struct MM1 {
/**
Second matmul: perform `attn @ V` where `attn` is the attention (not
normalized) and stored in shared memory
*/
using GemmType = DefaultGemmType<ArchTag, scalar_t>;
using OpClass = typename GemmType::OpClass;
using DefaultConfig =
typename cutlass::gemm::device::DefaultGemmConfiguration<
OpClass,
ArchTag,
scalar_t,
scalar_t,
output_accum_t, // ElementC
accum_t // ElementAccumulator
>;
static constexpr int kAlignmentA = DefaultConfig::kAlignmentA; // from smem
static constexpr int kAlignmentB =
kIsAligned ? DefaultConfig::kAlignmentB : GemmType::kMinimumAlignment;
using ThreadblockShape = cutlass::gemm::
GemmShape<kQueriesPerBlock, kKeysPerBlock, GemmType::ThreadK>;
using WarpShape = cutlass::gemm::GemmShape<32, 32, GemmType::WarpK>;
using InstructionShape = typename GemmType::InstructionShape;
using LayoutB = cutlass::layout::RowMajor;
using DefaultGemm = cutlass::gemm::kernel::DefaultGemm<
scalar_t, // ElementA,
cutlass::layout::RowMajor, // LayoutA,
kAlignmentA,
scalar_t, // ElementB,
LayoutB, // LayoutB,
kAlignmentB,
output_accum_t,
cutlass::layout::RowMajor, // LayoutC,
accum_t,
OpClass,
ArchTag,
ThreadblockShape,
WarpShape,
typename GemmType::InstructionShape,
typename DefaultConfig::EpilogueOutputOp,
void, // ThreadblockSwizzle - not used
ArchTag::kMinComputeCapability >= 80 && kIsHalf
? 4
: DefaultConfig::kStages,
false, // SplitKSerial
typename GemmType::Operator>;
using WarpIteratorA = typename cutlass::gemm::threadblock::
DefaultWarpIteratorAFromSharedMemory<
typename DefaultGemm::Mma::Policy::Operator::Shape, // WarpShape
typename DefaultGemm::Mma::Policy::Operator::InstructionShape,
typename DefaultGemm::Mma::Policy::Operator::IteratorA,
typename DefaultGemm::Mma::Policy>::WarpIterator;
using DefaultMmaFromSmem =
typename cutlass::gemm::threadblock::DefaultMmaFromSharedMemory<
typename DefaultGemm::Mma,
MM0::AccumulatorSharedStorage::Shape::kN, // kMaxK
WarpIteratorA,
false>; // kScaleOperandA
using Mma = typename DefaultMmaFromSmem::Mma;
using IteratorB = typename Mma::IteratorB;
using WarpCount = typename Mma::WarpCount;
static_assert(
WarpCount::kM * WarpCount::kN * WarpCount::kK == kNumWarpsPerBlock,
"");
using DefaultEpilogue = typename DefaultGemm::Epilogue;
using OutputTileIterator =
typename cutlass::epilogue::threadblock::PredicatedTileIterator<
typename DefaultEpilogue::OutputTileIterator::ThreadMap,
output_t>;
using OutputTileIteratorAccum =
typename cutlass::epilogue::threadblock::PredicatedTileIterator<
typename DefaultEpilogue::OutputTileIterator::ThreadMap,
output_accum_t>;
};
static constexpr int64_t kAlignmentQ = MM0::kAlignmentA;
static constexpr int64_t kAlignmentK = MM0::kAlignmentB;
static constexpr int64_t kAlignmentV = 1;
// Shared storage - depends on kernel params
struct ScalingCoefs {
cutlass::Array<accum_t, kQueriesPerBlock> m_prime;
cutlass::Array<accum_t, kQueriesPerBlock> s_prime;
cutlass::Array<accum_t, kQueriesPerBlock> mi;
cutlass::Array<accum_t, kQueriesPerBlock> out_rescale;
cutlass::Array<accum_t, kQueriesPerBlock * MM0::MmaCore::WarpCount::kN>
addition_storage;
};
struct SharedStorageEpilogueAtEnd : ScalingCoefs {
struct SharedStorageAfterMM0 {
// Everything here might be overwritten during MM0
union {
typename MM0::BiasLoader::SmemTile bias;
typename MM0::AccumulatorSharedStorage si;
};
typename MM1::Mma::SharedStorage mm1;
};
union {
typename MM0::Mma::SharedStorage mm0;
SharedStorageAfterMM0 after_mm0;
typename MM1::DefaultEpilogue::SharedStorage epilogue;
};
CUTLASS_DEVICE typename MM1::DefaultEpilogue::SharedStorage&
epilogue_shared_storage() {
return epilogue;
}
};
struct SharedStorageEpilogueInLoop : ScalingCoefs {
struct SharedStorageAfterMM0 {
// Everything here might be overwritten during MM0
union {
typename MM0::BiasLoader::SmemTile bias;
typename MM0::AccumulatorSharedStorage si;
};
typename MM1::Mma::SharedStorage mm1;
typename MM1::DefaultEpilogue::SharedStorage epilogue;
};
union {
typename MM0::Mma::SharedStorage mm0;
SharedStorageAfterMM0 after_mm0;
};
CUTLASS_DEVICE typename MM1::DefaultEpilogue::SharedStorage&
epilogue_shared_storage() {
return after_mm0.epilogue;
}
};
using SharedStorage = typename cutlass::platform::conditional<
kSingleValueIteration || kKeepOutputInRF,
SharedStorageEpilogueAtEnd,
SharedStorageEpilogueInLoop>::type;
static bool __host__ check_supported(Params const& p) {
CHECK_ALIGNED_PTR(p.query_ptr, kAlignmentQ);
CHECK_ALIGNED_PTR(p.key_ptr, kAlignmentK);
CHECK_ALIGNED_PTR(p.value_ptr, kAlignmentV);
if (kSupportsBias) {
CHECK_ALIGNED_PTR(p.attn_bias_ptr, kAlignmentQ);
XFORMERS_CHECK(
p.num_batches <= 1 || p.bias_strideB % kAlignmentQ == 0,
"attn_bias is not correctly aligned (strideB)");
XFORMERS_CHECK(
p.num_heads <= 1 || p.bias_strideH % kAlignmentQ == 0,
"attn_bias is not correctly aligned (strideH)");
XFORMERS_CHECK(
p.bias_strideM % kAlignmentQ == 0,
"attn_bias is not correctly aligned");
}
XFORMERS_CHECK(
p.q_strideM % kAlignmentQ == 0,
"query is not correctly aligned (strideM)");
XFORMERS_CHECK(
p.k_strideM % kAlignmentK == 0,
"key is not correctly aligned (strideM)");
XFORMERS_CHECK(
p.v_strideM % kAlignmentV == 0,
"value is not correctly aligned (strideM)");
XFORMERS_CHECK(
p.num_heads <= 1 || p.q_strideH % kAlignmentQ == 0,
"query is not correctly aligned (strideH)");
XFORMERS_CHECK(
p.num_heads <= 1 || p.k_strideH % kAlignmentK == 0,
"key is not correctly aligned (strideH)");
XFORMERS_CHECK(
p.num_heads <= 1 || p.v_strideH % kAlignmentV == 0,
"value is not correctly aligned (strideH)");
XFORMERS_CHECK(
p.custom_mask_type < NumCustomMaskTypes,
"invalid value for `custom_mask_type`");
return true;
}
static void CUTLASS_DEVICE attention_kernel(Params& p) {
// In this block, we will only ever:
// - read query[query_start:query_end, :]
// - write to output[query_start:query_end, :]
extern __shared__ char smem_buffer[];
SharedStorage& shared_storage = *((SharedStorage*)smem_buffer);
auto& m_prime = shared_storage.m_prime;
auto& s_prime = shared_storage.s_prime;
auto& mi = shared_storage.mi;
auto& out_rescale = shared_storage.out_rescale;
const uint32_t query_start = blockIdx.x * kQueriesPerBlock;
static_assert(kQueriesPerBlock < kNumWarpsPerBlock * kWarpSize, "");
if (thread_id() < kQueriesPerBlock) {
s_prime[thread_id()] = accum_t(0);
out_rescale[thread_id()] = accum_t(1.0);
m_prime[thread_id()] =
-cutlass::platform::numeric_limits<accum_t>::infinity();
mi[thread_id()] = -cutlass::platform::numeric_limits<accum_t>::infinity();
}
typename MM1::Mma::FragmentC accum_o;
accum_o.clear();
auto createOutputIter = [&](int col) -> typename MM1::OutputTileIterator {
using OutputTileIterator = typename MM1::OutputTileIterator;
return OutputTileIterator(
typename OutputTileIterator::Params{(int32_t)p.o_strideM},
p.output_ptr,
typename OutputTileIterator::TensorCoord{
p.num_queries, p.head_dim_value},
thread_id(),
{0, col});
};
auto createOutputAccumIter = [&](int col) ->
typename MM1::OutputTileIteratorAccum {
using OutputTileIteratorAccum = typename MM1::OutputTileIteratorAccum;
return OutputTileIteratorAccum(
typename OutputTileIteratorAccum::Params{
(int32_t)(p.head_dim_value * p.num_heads)},
p.output_accum_ptr,
typename OutputTileIteratorAccum::TensorCoord{
p.num_queries, p.head_dim_value},
thread_id(),
{0, col});
};
#ifdef HAS_PYTORCH
curandStatePhilox4_32_10_t curand_state_init;
if (kSupportsDropout && p.use_dropout) {
const auto seeds = at::cuda::philox::unpack(p.rng_engine_inputs);
// each element of the attention matrix P with shape
// (batch_sz, n_heads, n_queries, n_keys) is associated with a single
// offset in RNG sequence. we initialize the RNG state with offset that
// starts at the beginning of a (n_queries, n_keys) matrix for this
// block's batch_id and head_id
// initializing rng state is very expensive, so we run once per kernel,
// rather than once per iteration. each iteration takes a copy of the
// initialized RNG state and offsets it as needed.
curand_init(
std::get<0>(seeds),
0,
std::get<1>(seeds) + p.dropout_batch_head_rng_offset,
&curand_state_init);
}
#endif
// Iterate through keys
for (int32_t iter_key_start = 0; iter_key_start < p.num_keys;
iter_key_start += kKeysPerBlock) {
int32_t problem_size_0_m =
cutlass::fast_min((int32_t)kQueriesPerBlock, p.num_queries);
int32_t problem_size_0_n = cutlass::fast_min(
int32_t(kKeysPerBlock), p.num_keys - iter_key_start);
int32_t const& problem_size_0_k = p.head_dim;
int32_t const& problem_size_1_n = p.head_dim_value;
int32_t const& problem_size_1_k = problem_size_0_n;
auto prologueV = [&](int blockN) {
typename MM1::Mma::IteratorB iterator_V(
typename MM1::IteratorB::Params{typename MM1::LayoutB(p.v_strideM)},
p.value_ptr + iter_key_start * p.v_strideM,
{problem_size_1_k, problem_size_1_n},
thread_id(),
cutlass::MatrixCoord{0, blockN * MM1::Mma::Shape::kN});
MM1::Mma::prologue(
shared_storage.after_mm0.mm1,
iterator_V,
thread_id(),
problem_size_1_k);
};
__syncthreads(); // Need to have shared memory initialized, and `m_prime`
// updated from end of prev iter
//
// MATMUL: Q.K_t
//
// Computes the block-matrix product of:
// (a) query[query_start:query_end, :]
// with
// (b) key[iter_key_start:iter_key_start + kKeysPerBlock]
// and stores that into `shared_storage.si`
//
// Compute threadblock location
cutlass::gemm::GemmCoord tb_tile_offset = {0, 0, 0};
cutlass::MatrixCoord tb_offset_A{
tb_tile_offset.m() * MM0::Mma::Shape::kM, tb_tile_offset.k()};
cutlass::MatrixCoord tb_offset_B{
tb_tile_offset.k(), tb_tile_offset.n() * MM0::Mma::Shape::kN};
// Construct iterators to A and B operands
typename MM0::IteratorA iterator_A(
typename MM0::IteratorA::Params(
typename MM0::MmaCore::LayoutA(p.q_strideM)),
p.query_ptr,
{problem_size_0_m, problem_size_0_k},
thread_id(),
tb_offset_A);
typename MM0::IteratorB iterator_B(
typename MM0::IteratorB::Params(
typename MM0::MmaCore::LayoutB(p.k_strideM)),
p.key_ptr + iter_key_start * p.k_strideM,
{problem_size_0_k, problem_size_0_n},
thread_id(),
tb_offset_B);
auto my_warp_id = warp_uniform(warp_id());
auto my_lane_id = lane_id();
// Construct thread-scoped matrix multiply
typename MM0::Mma mma(
shared_storage.mm0, thread_id(), my_warp_id, my_lane_id);
typename MM0::Mma::FragmentC accum;
accum.clear();
auto gemm_k_iterations =
(problem_size_0_k + MM0::Mma::Shape::kK - 1) / MM0::Mma::Shape::kK;
// Compute threadblock-scoped matrix multiply-add
mma(gemm_k_iterations, accum, iterator_A, iterator_B, accum);
__syncthreads();
if (kPreloadV) {
prologueV(0);
} else {
MM1::Mma::drain_cp_asyncs();
}
typename MM0::Mma::Operator::IteratorC::TensorCoord
iteratorC_tile_offset = {
(tb_tile_offset.m() * MM0::Mma::WarpCount::kM) +
(my_warp_id % MM0::Mma::WarpCount::kM),
(tb_tile_offset.n() * MM0::Mma::WarpCount::kN) +
(my_warp_id / MM0::Mma::WarpCount::kM)};
// multiply by scaling factor
if (kSupportsBias) {
accum =
cutlass::multiplies<typename MM0::Mma::FragmentC>()(p.scale, accum);
}
// apply attention bias if applicable
if (kSupportsBias && p.attn_bias_ptr != nullptr) {
// load bias tile Bij into shared memory
typename MM0::BiasLoader::GmemTileIterator bias_iter(
{cutlass::layout::RowMajor(p.bias_strideM)},
// attn_bias_pointer points to matrix of size (n_queries, n_keys)
// for the relevant batch_id and head_id
p.attn_bias_ptr + query_start * p.bias_strideM + iter_key_start,
{problem_size_0_m, problem_size_0_n},
thread_id());
cutlass::TensorRef<scalar_t, cutlass::layout::RowMajor> bias_tensor_ref(
shared_storage.after_mm0.bias.data(),
cutlass::layout::RowMajor(MM0::ThreadblockShape::kN));
typename MM0::BiasLoader::SmemTileIterator smem_tile_iter(
bias_tensor_ref, thread_id());
MM0::BiasLoader::load(bias_iter, smem_tile_iter);
// Pij += Bij, Pij is in register fragment and Bij is in shared memory
auto lane_offset = MM0::AccumLambdaIterator::get_lane_offset(
my_lane_id, my_warp_id, iteratorC_tile_offset);
MM0::AccumLambdaIterator::iterateRows(
lane_offset,
[&](int accum_m) {},
[&](int accum_m, int accum_n, int idx) {
if (accum_m < problem_size_0_m && accum_n < problem_size_0_n) {
accum[idx] += bias_tensor_ref.at({accum_m, accum_n});
}
},
[&](int accum_m) {});
}
// Mask out last if causal
// This is only needed if upper-right corner of current query / key block
// intersects the mask Coordinates of upper-right corner of current block
// is y=query_start x=min(iter_key_start + kKeysPerBlock, num_keys)) The
// first masked element is x = y + offset -> query_start + offset There is
// intersection (and we need to mask) if min(iter_key_start +
// kKeysPerBlock, num_keys)) >= query_start + offset
if (p.custom_mask_type &&
cutlass::fast_min(iter_key_start + kKeysPerBlock, p.num_keys) >=
(query_start + p.causal_diagonal_offset)) {
auto query_start = blockIdx.x * kQueriesPerBlock;
auto lane_offset = MM0::AccumLambdaIterator::get_lane_offset(
my_lane_id, my_warp_id, iteratorC_tile_offset);
int32_t last_col;
MM0::AccumLambdaIterator::iterateRows(
lane_offset,
[&](int accum_m) {
// last absolute col is (last absolute query + offset)
// last local col is (last absolute query + offset -
// iter_key_start)
last_col = query_start + accum_m + p.causal_diagonal_offset -
iter_key_start;
},
[&](int accum_m, int accum_n, int idx) {
if (accum_n > last_col) {
accum[idx] =
-cutlass::platform::numeric_limits<accum_t>::infinity();
}
},
[&](int accum_m) {});
}
// Update `mi` from accum stored in registers
// Also does accum[i] <- exp(accum[i] - mi)
iterative_softmax<typename MM0::Mma::Operator::IteratorC>(
accum_o,
accum,
mi,
m_prime,
s_prime,
out_rescale,
shared_storage.addition_storage,
my_lane_id,
thread_id(),
my_warp_id,
p.num_keys - iter_key_start,
iter_key_start == 0,
iteratorC_tile_offset,
kSupportsBias ? 1.0f : p.scale);
// Output results to shared-memory
int warp_idx_mn_0 = my_warp_id %
(MM0::Mma::Base::WarpCount::kM * MM0::Mma::Base::WarpCount::kN);
auto output_tile_coords = cutlass::MatrixCoord{
warp_idx_mn_0 % MM0::Mma::Base::WarpCount::kM,
warp_idx_mn_0 / MM0::Mma::Base::WarpCount::kM};
MM0::B2bGemm::accumToSmem(
shared_storage.after_mm0.si, accum, my_lane_id, output_tile_coords);
__syncthreads();
#ifdef HAS_PYTORCH
// apply dropout (if applicable) after we've written Pij to smem.
// dropout is applied by multiplying each element of Pij by:
// - 0 with probability dropout_p
// - 1 / (1 - dropout_p) with probability 1 - dropout_p
//
// for backward purposes we want to be able to map each element of the
// attention matrix to the same random uniform number as the one we used
// in forward, without needing to use the same iteration order or having
// to store the dropout matrix. its possible to do this in registers but
// it ends up being very slow because each thread having noncontiguous
// strips of the Pij tile means we have to skip around a lot, and also
// have to generate a single random number at a time
if (kSupportsDropout && p.use_dropout) {
auto si = shared_storage.after_mm0.si.accum_ref();
// each thread handles a contiguous sequence of elements from Sij, all
// coming from the same row. the reason they have to come from the same
// row is that the sampling random numbers from a contiguous random
// number sequence is much more efficient than jumping around, and the
// linear offset of each element of S (the global matrix) maps to an
// offset in a random number sequence. for S, the end of a row and the
// beginning of the next have adjacent offsets, but for Sij, this is not
// necessarily the case.
const int num_threads = blockDim.x * blockDim.y * blockDim.z;
const int threads_per_row =
cutlass::fast_min(num_threads / problem_size_0_m, problem_size_0_n);
const int elts_per_thread = cutlass::round_nearest(
cutlass::ceil_div(problem_size_0_n, threads_per_row), 4);
const int thread_i = thread_id() / threads_per_row;
const int thread_start_j =
(thread_id() % threads_per_row) * elts_per_thread;
if (thread_i < problem_size_0_m && thread_start_j < problem_size_0_n) {
curandStatePhilox4_32_10_t curand_state = curand_state_init;
skipahead(
static_cast<unsigned long long>(
(query_start + thread_i) * p.num_keys_absolute +
(iter_key_start + thread_start_j)),
&curand_state);
const float dropout_scale = 1.0 / (1.0 - p.dropout_prob);
// apply dropout scaling to elements this thread is responsible for,
// in chunks of 4
for (int sij_start_col_idx = thread_start_j; sij_start_col_idx <
cutlass::fast_min(thread_start_j + elts_per_thread,
problem_size_0_n);
sij_start_col_idx += 4) {
const float4 rand_uniform_quad = curand_uniform4(&curand_state);
CUTLASS_PRAGMA_UNROLL
for (int quad_idx = 0; quad_idx < 4; ++quad_idx) {
si.at({thread_i, sij_start_col_idx + quad_idx}) *=
static_cast<scalar_t>(
dropout_scale *
((&rand_uniform_quad.x)[quad_idx] > p.dropout_prob));
}
}
}
__syncthreads(); // p.use_dropout should have same value kernel-wide
}
#endif
//
// MATMUL: Attn . V
// Run the matmul `attn @ V` for a block of attn and V.
// `attn` is read from shared memory (in `shared_storage_si`)
// `V` is read from global memory (with iterator_B)
//
const int64_t nBlockN = kSingleValueIteration
? 1
: ceil_div(
(int64_t)problem_size_1_n, int64_t(MM1::ThreadblockShape::kN));
for (int blockN = 0; blockN < nBlockN; ++blockN) {
int gemm_k_iterations =
(problem_size_1_k + MM1::Mma::Shape::kK - 1) / MM1::Mma::Shape::kK;
// Compute threadblock-scoped matrix multiply-add and store it in accum
// (in registers)
if (!kPreloadV) {
__syncthreads(); // we share shmem between mma and epilogue
}
typename MM1::Mma::IteratorB iterator_V(
typename MM1::IteratorB::Params{typename MM1::LayoutB(p.v_strideM)},
p.value_ptr + iter_key_start * p.v_strideM,