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
|
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice, this list of
* conditions and the following disclaimer.
* * 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.
* * Neither the name of the NVIDIA CORPORATION 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 NVIDIA CORPORATION 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 TOR (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 Defines structural properties of the GEMM epilogue.
*/
#pragma once
#include <cutlass/convert.h>
#include <cutlass/coord.h>
#include <cutlass/gemm/gemm_global_stream.h>
#include <cutlass/gemm/gemm_shared_stream.h>
#include <cutlass/gemm/linear_scaling.h>
#include <cutlass/reshape_tile.h>
#include <cutlass/tile_iterator.h>
namespace cutlass {
namespace gemm {
////////////////////////////////////////////////////////////////////////////////////////////////////
template <
/// The output tile.
typename OutputTile_,
/// The accumulators.
typename Accumulators_,
/// The iterator to load C from global memory.
typename GlobalLoadIteratorC_,
/// The transformer for C.
typename GlobalTransformerC_,
/// The transformer for D.
typename GlobalTransformerD_,
/// The iterator to store D to global memory.
typename GlobalStoreIteratorD_,
/// The iterator to store D to shared memory.
typename SharedStoreIteratorD_,
/// The shared store transformer for D.
typename SharedStoreTransformerD_,
/// The iterator to load D from shared memory.
typename SharedLoadIteratorD_,
/// The number of iterations in the epilogue.
typename Iterations_,
/// The iterations strides.
typename Delta_,
/// The functor to be used in the epilogue.
typename Functor_,
/// The index.
typename Index_ = int>
struct GemmEpilogueTraits {
//
/// The output tile.
typedef OutputTile_ OutputTile;
/// The number of iterations.
/// The accumulators.
typedef Accumulators_ Accumulators;
/// The iterator for C in global memory.
typedef GlobalLoadIteratorC_ GlobalLoadIteratorC;
/// The transformer for C.
typedef GlobalTransformerC_ GlobalTransformerC;
/// The transformer for D.
typedef GlobalTransformerD_ GlobalTransformerD;
/// The iterator for D in global memory.
typedef GlobalStoreIteratorD_ GlobalStoreIteratorD;
/// The iterator to store D in shared memory.
typedef SharedStoreIteratorD_ SharedStoreIteratorD;
/// The shared store transformer for D.
typedef SharedStoreTransformerD_ SharedStoreTransformerD;
/// The iterator to store D in shared memory.
typedef SharedLoadIteratorD_ SharedLoadIteratorD;
/// typedef typename GemmConfig::EpilogueIterations Iterations;
typedef Iterations_ Iterations;
/// The iterations strides.
typedef Delta_ Delta;
/// The functor in charge of the math.
typedef Functor_ Functor;
/// The index.
typedef Index_ Index;
/// We do not support 3D or 4D shapes.
static_assert(Iterations::kD == 1 && Iterations::kC == 1, "Unsupported 3D/4D shapes");
/// The scalar.
typedef typename Functor::Scalar Scalar;
/// The scalar for C.
typedef typename GlobalLoadIteratorC::Scalar ScalarC;
/// The scalar for D.
typedef typename GlobalStoreIteratorD::Scalar ScalarD;
/// The params.
struct Params {
/// The strides for H and W in the different iterations of the epilogue.
Index stride_h, stride_w;
/// The params for the C iterator.
typename GlobalLoadIteratorC::Params iterator_c;
/// The params for the D global iterator.
typename GlobalStoreIteratorD::Params iterator_d;
/// The params for the D shared store iterator.
typename SharedStoreIteratorD::Params shared_store_iterator_d;
/// The params for the D shared load iterator.
typename SharedLoadIteratorD::Params shared_load_iterator_d;
/// The functor params.
typename Functor::Params functor;
/// Setup the params.
template <typename GemmDesc_>
CUTLASS_HOST_DEVICE int initialize(GemmDesc_ const& desc) {
// The parameters for the functor.
int error_code = functor.initialize(desc);
if (error_code) {
return error_code;
}
// At the end of the H iteration, we jump over a number of columns.
this->stride_h = desc.ldd * Delta::kH;
// Nothing to do here.
this->stride_w = 0;
// Setup the params for the global memory iterator for C.
error_code = iterator_c.initialize(
reinterpret_cast<ScalarC const*>(desc.d_c), desc.ldc, desc.n, stride_w, Delta::kW);
if (error_code) {
return error_code;
}
// Setup the params for the global memory iterator for D.
return iterator_d.initialize(
reinterpret_cast<ScalarD*>(desc.d_d), desc.ldd, desc.n, stride_w, Delta::kW);
}
};
/// The shared memory storage to exchange data.
union StreamSharedStorage {
// The storage for the store iterator.
typename SharedStoreIteratorD::SharedStorage store;
// The storage for the store iterator.
typename SharedLoadIteratorD::SharedStorage load;
};
/// The shared memory to swizzle the data in the epilogue.
struct SharedStorage {
// The storage for the shared stream D.
StreamSharedStorage shared_stream;
};
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename GemmConfig_, typename EpilogueFunctor_, typename Index_ = int>
struct GemmEpilogueTraitsHelper {
/// The scalar.
typedef typename EpilogueFunctor_::Scalar Scalar;
/// The output tile.
typedef typename GemmConfig_::OutputTile OutputTile;
/// The number of iterations in the epilogue.
typedef Shape<1,
GemmConfig_::MultiplyAdd::AccumulatorsPerThread::kH /
GemmConfig_::kAccumulatorsPerLdsB,
GemmConfig_::kAccumulatorsPerLdsB>
Iterations;
// The iteration strides in the H/W dimension.
typedef Shape<0,
GemmConfig_::kAccumulatorsPerLdsB*(
GemmConfig_::Warps::kH* GemmConfig_::MultiplyAdd::ThreadsPerWarp::kH - 1),
0>
Delta;
/// The functor to do the math in the epilogue.
typedef EpilogueFunctor_ Functor;
/// The traits class to build the iterator to store to shared memory for D.
typedef GemmSharedStoreTileDTraits<
// The pointer is float.
typename Functor::Scalar,
// The output tile size.
typename GemmConfig_::OutputTile,
// The number of warps.
typename GemmConfig_::Warps,
// The number of threads per warp.
typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
// The number of scalars per STS.
GemmConfig_::kScalarsPerStsD,
// The skew -- 128 / sizeof(ScalarD) / kScalarsPerStsD is the number of threads involved in
// a single STS. We divide by 2 as our objective is to add a skew to the odd threads to
// avoid bank conflicts between odd and even threads.
128 / sizeof(typename GemmConfig_::ScalarD) / GemmConfig_::kScalarsPerStsD / 2 *
GemmConfig_::kScalarsPerStsD>
SharedStoreTileTraits;
/// The iterator to store D to shared memory.
typedef TileStoreIterator<SharedStoreTileTraits,
typename SharedStoreTileTraits::Scalar,
IteratorAdvance::kH,
MemorySpace::kShared>
SharedStoreIteratorD;
/// The shared store transformer for D.
typedef Copy<typename SharedStoreIteratorD::Fragment> SharedStoreTransformerD;
/// The traits class to build the iterator to load from shared memory for D.
typedef GemmSharedLoadTileDTraits<
// The pointer is float.
typename Functor::Scalar,
// The output tile size.
typename GemmConfig_::OutputTile,
// The number of warps.
typename GemmConfig_::Warps,
// The number of threads per warp.
typename GemmConfig_::MultiplyAdd::ThreadsPerWarp,
// The number of columns of the output tile written by iteration.
GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
// The number of scalars per LDS.
GemmConfig_::kScalarsPerLdsD,
// The skew.
SharedStoreTileTraits::kSkew>
SharedLoadTileTraits;
/// The iterator to load D from shared memory.
typedef TileLoadIterator<SharedLoadTileTraits,
typename SharedLoadTileTraits::Scalar,
IteratorAdvance::kH,
MemorySpace::kShared>
SharedLoadIteratorD;
/// The traits class to build the iterator to load data from global memory for C^N.
typedef GemmGlobalTileCdTraits<
// The pointer is float const.
typename GemmConfig_::ScalarC const,
// The tile has size (N / Iterations)xM in GEMM's terminology.
Shape<1,
GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
GemmConfig_::OutputTile::kW>,
// The threads are distributed as warps x 32 (the traits may reorganize).
Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
// How many elements do we jump over at each iteration?
Iterations::kW,
// The number of scalars per LDG (LDG.32 or LDG.128, etc).
GemmConfig_::kScalarsPerLdgC>
GlobalLoadTileTraits;
/// The iterator to load C.
typedef GemmGlobalIteratorCd<GlobalLoadTileTraits, Index_> GlobalLoadIteratorC;
/// The transformer for C.
typedef Copy<typename GlobalLoadIteratorC::Fragment> GlobalTransformerC;
/// The traits class to build the iterator to store data to global memory for D^N.
typedef GemmGlobalTileCdTraits<
// The pointer is float.
typename GemmConfig_::ScalarD,
// The tile has size (N / Iterations)xM in GEMM's terminology.
Shape<1,
GemmConfig_::OutputTile::kH / ShapeCount<Iterations>::kCount,
GemmConfig_::OutputTile::kW>,
// The threads are distributed as warps x 32 (the traits may reorganize).
Shape<1, ShapeCount<typename GemmConfig_::Warps>::kCount, GemmConfig_::kWarpSize>,
// How many elements do we jump over at each iteration?
Iterations::kW,
// The number of scalars per LDG (LDG.32 or LDG.128, etc).
GemmConfig_::kScalarsPerStgD>
GlobalStoreTileTraits;
/// The iterator to store D.
typedef GemmGlobalIteratorCd<GlobalStoreTileTraits, Index_> GlobalStoreIteratorD;
/// The transformer for D.
typedef Copy<typename GlobalStoreIteratorD::Fragment> GlobalTransformerD;
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template <
/// The GEMM config.
typename GemmConfig_,
/// The epilogue functor to do the math in the epilogue.
typename EpilogueFunctor_,
/// The index.
typename Index_ = int,
/// The helper to create the traits class.
typename Helper_ = GemmEpilogueTraitsHelper<GemmConfig_, EpilogueFunctor_, Index_> >
struct SimplifiedGemmEpilogueTraits : public GemmEpilogueTraits<
// The output tile.
typename GemmConfig_::OutputTile,
// The accumulators.
typename GemmConfig_::Accumulators,
// The global iterator for C.
typename Helper_::GlobalLoadIteratorC,
// The transformer for C.
typename Helper_::GlobalTransformerC,
// The transformer for D.
typename Helper_::GlobalTransformerD,
// The global iterator for D.
typename Helper_::GlobalStoreIteratorD,
// The iterator to store D to shared memory.
typename Helper_::SharedStoreIteratorD,
// The shared store transformer for D.
typename Helper_::SharedStoreTransformerD,
// The iterator to load D from shared memory.
typename Helper_::SharedLoadIteratorD,
// The number of iterations.
typename Helper_::Iterations,
// The strides between iterations.
typename Helper_::Delta,
// The functor to be used in the epilogue.
EpilogueFunctor_,
// The index.
Index_> {};
////////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace gemm
} // namespace cutlass
|