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
|
/***************************************************************************************************
* Copyright (c) 2017 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* 3. Neither the name of the copyright holder nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
/*! \file
\brief
Default kernel-level Depthwise implicit GEMM convolution definitions combine threadblock-scoped
matrix multiply-add with the appropriate threadblock-scoped epilogue.
*/
#pragma once
#include "cutlass/cutlass.h"
#include "cutlass/conv/kernel/default_conv2d.h"
#include "cutlass/conv/kernel/direct_convolution.h"
#include "cutlass/conv/threadblock/depthwise_mma_core_with_lane_access_size.h"
#include "cutlass/conv/threadblock/conv2d_fprop_activation_tile_access_iterator_analytic.h"
#include "cutlass/conv/threadblock/conv2d_fprop_filter_tile_access_iterator_analytic.h"
#include "cutlass/conv/threadblock/depthwise_fprop_pipelined.h"
// Direct Conv Related Header files
#include "cutlass/conv/threadblock/depthwise_fprop_activation_tile_access_iterator_direct_conv_optimized.h"
#include "cutlass/conv/threadblock/depthwise_fprop_activation_tile_access_iterator_direct_conv_fixed_stride_dilation.h"
#include "cutlass/conv/threadblock/depthwise_fprop_filter_tile_access_iterator_direct_conv_optimized.h"
#include "cutlass/conv/threadblock/depthwise_fprop_direct_conv_multistage.h"
/////////////////////////////////////////////////////////////////////////////////////////////////
namespace cutlass {
namespace conv {
namespace kernel {
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines a kernel for DepthwiseFprop
template <
typename ElementA,
typename LayoutA,
typename ElementB,
typename LayoutB,
typename ElementC,
typename LayoutC,
typename ElementAccumulator,
typename OperatorClass,
typename ArchTag,
typename ThreadblockShape,
typename WarpShape,
typename InstructionShape,
typename EpilogueOutputOp,
typename ThreadblockSwizzle,
int Stages,
typename MathOperatorTag,
conv::IteratorAlgorithm IteratorAlgorithm = IteratorAlgorithm::kAnalytic,
conv::StrideSupport StrideSupport = StrideSupport::kStrided,
/// Access granularity of A matrix in units of elements
int AlignmentA = 128 / cutlass::sizeof_bits<ElementA>::value,
/// Access granularity of B matrix in units of elements
int AlignmentB = cutlass::sizeof_bits<ElementB>::value / cutlass::sizeof_bits<ElementB>::value
> struct DefaultDepthwiseFprop;
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines a kernel for DepthwiseFprop with direct convolution algorithm
template <
typename ElementA,
typename LayoutA,
typename ElementB,
typename LayoutB,
typename ElementC,
typename LayoutC,
typename ElementAccumulator,
typename OperatorClass,
typename ArchTag,
typename ThreadblockShape,
typename ThreadBlockOutputShape,
typename FilterShape,
typename WarpShape,
typename InstructionShape,
typename EpilogueOutputOp,
typename ThreadblockSwizzle,
int Stages,
typename MathOperatorTag,
conv::IteratorAlgorithm IteratorAlgorithm = IteratorAlgorithm::kAnalytic,
conv::StrideSupport StrideSupport = StrideSupport::kStrided,
// MatrixShape<Height, Width>
typename StrideShape = cutlass::MatrixShape<-1, -1>,
// MatrixShape< Height, Width>
typename DilationShape = cutlass::MatrixShape<-1, -1>,
/// Access granularity of A matrix in units of elements
int AlignmentA = 128 / cutlass::sizeof_bits<ElementA>::value,
/// Access granularity of B matrix in units of elements
int AlignmentB = 128 / cutlass::sizeof_bits<ElementB>::value
> struct DefaultDepthwiseDirect2dConvFprop;
/////////////////////////////////////////////////////////////////////////////////////////////////
// OpClassSimt convolutions
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines a kernel for Depthwise specialization for Analytic IteratorAlgorithm
template <
typename ElementA,
typename LayoutA,
typename ElementB,
typename LayoutB,
typename ElementC,
typename LayoutC,
typename ElementAccumulator,
typename ArchTag,
typename ThreadblockShape,
typename WarpShape,
typename InstructionShape,
typename EpilogueOutputOp,
typename ThreadblockSwizzle,
typename MathOperatorTag,
conv::StrideSupport StrideSupport,
int AlignmentA,
int AlignmentB
>
struct DefaultDepthwiseFprop <
ElementA,
LayoutA,
ElementB,
LayoutB,
ElementC,
LayoutC,
ElementAccumulator,
arch::OpClassSimt,
ArchTag,
ThreadblockShape,
WarpShape,
InstructionShape,
EpilogueOutputOp,
ThreadblockSwizzle,
2,
MathOperatorTag, // cutlass::arch::OpMultiplyAdd
IteratorAlgorithm::kAnalytic,
StrideSupport,
AlignmentA,
AlignmentB
> {
// Define the core components from GEMM
using MmaCore = typename cutlass::conv::threadblock::DepthwiseMmaCoreWithLaneAccessSize<
ThreadblockShape,
WarpShape,
InstructionShape,
ElementA,
layout::RowMajor,
ElementB,
layout::ColumnMajor,
ElementAccumulator,
layout::RowMajor,
arch::OpClassSimt,
128,
sizeof_bits<ElementB>::value,
2,
MathOperatorTag>;
// Define iterators over tiles from the A operand
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
using IteratorA =
cutlass::conv::threadblock::TileIterator<
cutlass::conv::threadblock::Conv2dFpropActivationTileAccessIteratorAnalytic<
cutlass::MatrixShape<ThreadblockShape::kM, ThreadblockShape::kK>,
ElementA, LayoutA,
ThreadMapA
>
>;
using SmemIteratorA = typename MmaCore::SmemIteratorA;
// Define iterators over tiles from the B operand
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
using IteratorB =
cutlass::conv::threadblock::TileIterator<
cutlass::conv::threadblock::Conv2dFpropFilterTileAccessIteratorAnalytic<
cutlass::MatrixShape<ThreadblockShape::kK, ThreadblockShape::kN>,
ElementB, LayoutB,
ThreadMapB,
AccessTypeB,
cutlass::conv::GroupMode::kDepthwise
>
>;
using SmemIteratorB = typename MmaCore::SmemIteratorB;
// Warp-level GEMM components
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
using MmaPolicy = typename MmaCore::MmaPolicy;
// Define the Mma
using Mma = threadblock::DepthwiseFpropPipelined<
ThreadblockShape,
IteratorA,
SmemIteratorA,
IteratorB,
SmemIteratorB,
ElementC,
LayoutC,
MmaPolicy
>;
// Define the epilogue
using Epilogue = typename epilogue::threadblock::DefaultEpilogueSimt<
ThreadblockShape,
WarpMmaSimtOp,
EpilogueOutputOp,
EpilogueOutputOp::kCount
>::Epilogue;
// Define the kernel
using Kernel = cutlass::conv::kernel::ImplicitGemmConvolution<
Mma,
Epilogue,
ThreadblockSwizzle,
conv::Operator::kFprop,
Conv2dProblemSize,
cutlass::conv::GroupMode::kDepthwise
>;
};
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines a kernel for Depthwise specialization for direct 2d conv implementation,
/// multiple stage pipeline, and SIMT-based mainloop
template <
typename ElementA,
typename LayoutA,
typename ElementB,
typename LayoutB,
typename ElementC,
typename LayoutC,
typename ElementAccumulator,
typename ArchTag,
typename ThreadblockShape,
typename ThreadBlockOutputShape,
typename FilterShape,
typename WarpShape,
typename InstructionShape,
typename EpilogueOutputOp,
typename ThreadblockSwizzle,
int Stages,
typename MathOperatorTag,
conv::StrideSupport StrideSupport,
typename StrideShape,
typename DilationShape,
int AlignmentA,
int AlignmentB
>
struct DefaultDepthwiseDirect2dConvFprop <
ElementA,
LayoutA,
ElementB,
LayoutB,
ElementC,
LayoutC,
ElementAccumulator,
arch::OpClassSimt,
ArchTag,
ThreadblockShape,
ThreadBlockOutputShape,
FilterShape,
WarpShape,
InstructionShape,
EpilogueOutputOp,
ThreadblockSwizzle,
Stages,
MathOperatorTag,
IteratorAlgorithm::kOptimized,
StrideSupport,
StrideShape,
DilationShape,
AlignmentA,
AlignmentB
> {
// One warp handles the entrie groups per cta.
static_assert(ThreadblockShape::kN == WarpShape::kN,
"ThreadblockShape::kN should be same as WarpShape::kN ");
static_assert(ThreadblockShape::kK == FilterShape::kCount && WarpShape::kK == FilterShape::kCount,
"ThreadblockShape::kK and WarpShape::kK should be same as filter size");
static_assert(ThreadblockShape::kM % WarpShape::kM == 0,
"ThreadblockShape::kM must be divisible by WarpShape shape::kM");
static_assert(ThreadBlockOutputShape::kN, "ThreadBlockOutputShape::kN should be 1");
// Define the core components from GEMM
using MmaCore = typename cutlass::conv::threadblock::DepthwiseDirectConvMmaCoreWithLaneAccessSize<
ThreadblockShape,
ThreadBlockOutputShape,
FilterShape,
WarpShape,
InstructionShape,
ElementA,
layout::RowMajor,
ElementB,
layout::ColumnMajor,
ElementAccumulator,
layout::RowMajor,
arch::OpClassSimt,
128,
128,
Stages,
MathOperatorTag>;
// Define iterators over tiles from the A operand
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
using IteratorA =
cutlass::conv::threadblock::DepthwiseFpropActivationDirect2dConvTileAccessIteratorOptimized<
cutlass::MatrixShape<ThreadblockShape::kM,ThreadblockShape::kN>, // < outputShape:KMNK, groups per cta>
ThreadBlockOutputShape,
ElementA, LayoutA,
ThreadMapA
>;
using SmemIteratorA = typename MmaCore::SmemIteratorA;
// Define iterators over tiles from the B operand
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
using IteratorB =
cutlass::conv::threadblock::DepthwiseFpropFilterDirectConvTileAccessIteratorOptimized<
cutlass::MatrixShape<ThreadblockShape::kN, FilterShape::kCount>,
ElementB, LayoutB,
ThreadMapB
>;
using SmemIteratorB = typename MmaCore::SmemIteratorB;
// Warp-level GEMM components
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
using MmaPolicy = typename MmaCore::MmaPolicy;
using ThreadOutputShape = typename MmaCore::ThreadOutputShape;
static cutlass::arch::CacheOperation::Kind const CacheOpA =
((sizeof_bits<ElementA>::value * AlignmentA) == 128)
? cutlass::arch::CacheOperation::Global
: cutlass::arch::CacheOperation::Always;
static cutlass::arch::CacheOperation::Kind const CacheOpB =
((sizeof_bits<ElementB>::value * AlignmentB) == 128)
? cutlass::arch::CacheOperation::Global
: cutlass::arch::CacheOperation::Always;
// Define the epilogue
using Epilogue = typename epilogue::threadblock::DefaultDirectConvEpilogueSimt<
ThreadblockShape, // < outputShape:KMNK, groups per cta>
WarpMmaSimtOp,
EpilogueOutputOp,
EpilogueOutputOp::kCount,
ThreadOutputShape,
ThreadBlockOutputShape
>::Epilogue;
// Define the Mma
using Mma = threadblock::DepthwiseFpropDirectConvMultipleStage<
ThreadblockShape,
IteratorA,
SmemIteratorA,
CacheOpA,
IteratorB,
SmemIteratorB,
CacheOpB,
MmaPolicy,
Stages,
Epilogue
>;
// Define the kernel
using Kernel = cutlass::conv::kernel::DirectConvolution<
Mma,
Epilogue,
ThreadblockSwizzle,
conv::Operator::kFprop,
Conv2dProblemSize,
cutlass::conv::GroupMode::kDepthwise,
ThreadBlockOutputShape
>;
};
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines a kernel for Depthwise specialization for direct 2d conv implementation,
/// multiple stage pipeline, and SIMT-based mainloop
template <
typename ElementA,
typename LayoutA,
typename ElementB,
typename LayoutB,
typename ElementC,
typename LayoutC,
typename ElementAccumulator,
typename ArchTag,
typename ThreadblockShape,
typename ThreadBlockOutputShape,
typename FilterShape,
typename WarpShape,
typename InstructionShape,
typename EpilogueOutputOp,
typename ThreadblockSwizzle,
int Stages,
typename MathOperatorTag,
conv::StrideSupport StrideSupport,
typename StrideShape,
typename DilationShape,
int AlignmentA,
int AlignmentB
>
struct DefaultDepthwiseDirect2dConvFprop <
ElementA,
LayoutA,
ElementB,
LayoutB,
ElementC,
LayoutC,
ElementAccumulator,
arch::OpClassSimt,
ArchTag,
ThreadblockShape,
ThreadBlockOutputShape,
FilterShape,
WarpShape,
InstructionShape,
EpilogueOutputOp,
ThreadblockSwizzle,
Stages,
MathOperatorTag,
IteratorAlgorithm::kFixedStrideDilation,
StrideSupport,
StrideShape,
DilationShape,
AlignmentA,
AlignmentB
> {
// One warp handles the entrie groups per cta.
static_assert(ThreadblockShape::kN == WarpShape::kN,
"ThreadblockShape::kN should be same as WarpShape::kN ");
static_assert(ThreadblockShape::kK == FilterShape::kCount && WarpShape::kK == FilterShape::kCount,
"ThreadblockShape::kK and WarpShape::kK should be same as filter size");
static_assert(ThreadblockShape::kM % WarpShape::kM == 0,
"ThreadblockShape::kM must be divisible by WarpShape shape::kM");
static_assert(ThreadBlockOutputShape::kN, "ThreadBlockOutputShape::kN should be 1");
static_assert(StrideShape::kRow >= 0 && StrideShape::kColumn >= 0, "Stride should be fixed");
static_assert(DilationShape::kRow >= 0 && DilationShape::kColumn >= 0, "Stride should be fixed");
// Activations loaded by threadblock
static int const ActivationShapeH = (ThreadBlockOutputShape::kH - 1) * StrideShape::kRow +
(FilterShape::kRow - 1) * DilationShape::kRow + 1;
static int const ActivationShapeW = (ThreadBlockOutputShape::kW - 1) * StrideShape::kColumn +
(FilterShape::kColumn - 1) * DilationShape::kColumn + 1;
using ActivationShape =
cutlass::conv::TensorNHWCShape<1, ActivationShapeH, ActivationShapeW, ThreadblockShape::kN >;
// Define the core components from GEMM
using MmaCore = typename cutlass::conv::threadblock::DepthwiseDirectConvMmaCoreWithLaneAccessSize<
ThreadblockShape,
ThreadBlockOutputShape,
FilterShape,
WarpShape,
InstructionShape,
ElementA,
layout::RowMajor,
ElementB,
layout::ColumnMajor,
ElementAccumulator,
layout::RowMajor,
arch::OpClassSimt,
128,
128,
Stages,
MathOperatorTag,
IteratorAlgorithm::kFixedStrideDilation,
StrideShape,
DilationShape,
ActivationShape>;
// Define iterators over tiles from the A operand
using ThreadMapA = typename MmaCore::IteratorThreadMapA;
using IteratorA =
cutlass::conv::threadblock::DepthwiseFpropActivationDirect2dConvTileAccessIteratorFixedStrideDilation<
cutlass::MatrixShape<ThreadblockShape::kM,ThreadblockShape::kN>, // < outputShape:KMNK, groups per cta>
ThreadBlockOutputShape,
StrideShape,
DilationShape,
ActivationShape,
ElementA, LayoutA,
ThreadMapA
>;
using SmemIteratorA = typename MmaCore::SmemIteratorA;
// Define iterators over tiles from the B operand
using ThreadMapB = typename MmaCore::IteratorThreadMapB;
using AccessTypeB = cutlass::AlignedArray<ElementB, AlignmentB>;
using IteratorB =
cutlass::conv::threadblock::DepthwiseFpropFilterDirectConvTileAccessIteratorOptimized<
cutlass::MatrixShape<ThreadblockShape::kN, FilterShape::kCount>,
ElementB, LayoutB,
ThreadMapB
>;
using SmemIteratorB = typename MmaCore::SmemIteratorB;
// Warp-level GEMM components
using WarpMmaSimtOp = typename MmaCore::MmaWarpSimt;
using MmaPolicy = typename MmaCore::MmaPolicy;
using ThreadOutputShape = typename MmaCore::ThreadOutputShape;
static cutlass::arch::CacheOperation::Kind const CacheOpA =
((sizeof_bits<ElementA>::value * AlignmentA) == 128)
? cutlass::arch::CacheOperation::Global
: cutlass::arch::CacheOperation::Always;
static cutlass::arch::CacheOperation::Kind const CacheOpB =
((sizeof_bits<ElementB>::value * AlignmentB) == 128)
? cutlass::arch::CacheOperation::Global
: cutlass::arch::CacheOperation::Always;
// Define the epilogue
using Epilogue = typename epilogue::threadblock::DefaultDirectConvEpilogueSimt<
ThreadblockShape, // < outputShape:KMNK, groups per cta>
WarpMmaSimtOp,
EpilogueOutputOp,
EpilogueOutputOp::kCount,
ThreadOutputShape,
ThreadBlockOutputShape
>::Epilogue;
// Define the Mma
using Mma = threadblock::DepthwiseFpropDirectConvMultipleStage<
ThreadblockShape,
IteratorA,
SmemIteratorA,
CacheOpA,
IteratorB,
SmemIteratorB,
CacheOpB,
MmaPolicy,
Stages,
Epilogue,
IteratorAlgorithm::kFixedStrideDilation
>;
// Define the kernel
using Kernel = cutlass::conv::kernel::DirectConvolution<
Mma,
Epilogue,
ThreadblockSwizzle,
conv::Operator::kFprop,
Conv2dProblemSize,
cutlass::conv::GroupMode::kDepthwise,
ThreadBlockOutputShape
>;
};
} // namespace kernel
} // namespace conv
} // namespace cutlass
/////////////////////////////////////////////////////////////////////////////////////////////////
|
