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
|
/***************************************************************************************************
* 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 iterator traits for efficiently loading and storing fragment to and from shared
memory, specialized for WMMA GEMM.
*/
#pragma once
#include <cutlass/wmma_matrix.h>
#ifdef CUTLASS_USE_WMMA_API
#include <cutlass/gemm/gemm_operand.h>
#include <cutlass/reshape_tile.h>
namespace cutlass {
namespace gemm {
template <class>
struct Debug {};
////////////////////////////////////////////////////////////////////////////////////////////////////
template <MatrixLayout::Kind kLayout_,
typename Scalar_,
typename Tile_,
typename Warps_,
int kWarpStride_,
typename Iterations_,
typename Delta_,
typename WmmaShape_>
struct WmmaGemmSharedLoadTileATraits {
/// The operand.
static GemmOperand::Kind const kOperand = GemmOperand::kA;
/// The layout.
static MatrixLayout::Kind const kLayout = kLayout_;
/// The scalar.
typedef Scalar_ Scalar;
/// The pointer.
typedef Scalar const* Pointer;
/// The access size
static int const kAccessSize = 1;
/// The tile with skew.
typedef Tile_ Tile;
/// The number of warps.
typedef Warps_ Warps;
/// The warps strides.
static int const kWarpStride = kWarpStride_;
/// The number of iterations.
typedef Iterations_ Iterations;
/// The strides between iterations.
typedef Delta_ Delta;
/// The strides between iterations.
typedef Delta_ ImmediateOffsetStrides;
/// The shape of the WMMA instruction.
typedef WmmaShape_ WmmaShape;
/// The memory space.
static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
/// ThreadOffset
struct ThreadOffset {
CUTLASS_HOST_DEVICE
Coord<4> operator()() const {
// The warp id.
int const warp = threadIdx.x / kWarpSize;
// The offset.
int const offset = warp % Warps::kW * kWarpStride;
return make_Coord(0, 0, offset, 0);
}
};
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template <MatrixLayout::Kind kLayout_,
typename Scalar_,
typename Tile_,
typename Warps_,
int kWarpStride_,
typename Iterations_,
typename Delta_,
typename WmmaShape_>
struct WmmaGemmSharedLoadTileBTraits {
/// The operand.
static GemmOperand::Kind const kOperand = GemmOperand::kB;
/// The layout.
static MatrixLayout::Kind const kLayout = kLayout_;
/// The scalar.
typedef Scalar_ Scalar;
/// The pointer.
typedef Scalar const* Pointer;
/// The access size
static int const kAccessSize = 1;
/// The tile with skew.
typedef Tile_ Tile;
/// The number of warps.
typedef Warps_ Warps;
/// The warps strides.
static int const kWarpStride = kWarpStride_;
/// The number of iterations.
typedef Iterations_ Iterations;
/// The strides between iterations.
typedef Delta_ Delta;
/// The strides between iterations.
typedef Delta_ ImmediateOffsetStrides;
/// The shape of the WMMA instruction.
typedef WmmaShape_ WmmaShape;
/// The memory space.
static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
/// ThreadOffset
struct ThreadOffset {
CUTLASS_HOST_DEVICE
Coord<4> operator()() const {
// The warp id.
int const warp = threadIdx.x / kWarpSize;
// The offset.
int const offset = warp / Warps::kW * kWarpStride;
return make_Coord(0, 0, offset, 0);
}
};
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template <MatrixLayout::Kind kLayout_,
typename Scalar_,
typename OutputTile_,
typename Warps_,
typename WmmaShape_,
int kSkew_ = 0>
struct WmmaGemmSharedStoreTileDTraits {
/// The operand.
static GemmOperand::Kind const kOperand = GemmOperand::kC;
/// The layout.
static MatrixLayout::Kind const kLayout = kLayout_;
/// The scalar.
typedef Scalar_ Scalar;
// The access size
static int const kAccessSize = 1;
/// The pointer.
typedef Scalar* Pointer;
/// The number of warps.
typedef Warps_ Warps;
/// The shape of the WMMA instruction.
typedef WmmaShape_ WmmaShape;
/// The skew.
static int const kSkew = kSkew_;
/// The memory space.
static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
/// The tile with skew.
typedef Shape<1, Warps_::kH * WmmaShape_::kH, OutputTile_::kW + kSkew_> Tile;
/// The number of iterations needed to store the tile.
typedef Shape<1, 1, OutputTile_::kW / Warps::kW / WmmaShape_::kW> Iterations;
/// The strides in each dimension between different loads/stores.
typedef Shape<0, 0, Warps::kW * WmmaShape_::kW, 0> Delta;
/// The strides in each dimension between different loads/stores.
typedef Shape<0, 0, Warps::kW * WmmaShape_::kW, 0> ImmediateOffsetStrides;
/// ThreadOffset
struct ThreadOffset {
CUTLASS_HOST_DEVICE
Coord<4> operator()() const {
// The warp id.
int const warp = threadIdx.x / kWarpSize;
// The starting column.
int const h = warp / Warps::kW * WmmaShape::kH;
// The w.
int const w = warp % Warps::kW * WmmaShape::kW;
// The offset.
int const offset = h * Tile::kW + w;
return make_Coord(0, 0, offset, 0);
}
};
};
////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename Scalar_, typename Tile_, typename Threads_, int kScalarsPerLds_>
struct WmmaGemmSharedLoadTileDTraits {
/// The scalar.
typedef Scalar_ Scalar;
/// The pointer.
typedef Scalar const* Pointer;
/// The access size
static int const kAccessSize = kScalarsPerLds_;
/// The tile.
typedef typename ReshapeTile<Tile_, kScalarsPerLds_>::Tile Tile;
/// The threads.
typedef typename ReshapeThreads<Tile, Threads_>::Threads Threads;
/// The threads strides.
typedef Shape<1, Tile::kW * Tile::kC, Tile::kC> ThreadsStrides;
/// The memory space.
static MemorySpace::Kind const kMemorySpace = MemorySpace::kShared;
/// The strides in each dimension between different loads/stores.
typedef Shape<0, Threads::kH * ShapeCount<Tile>::kWc, Threads::kW * kScalarsPerLds_> Delta;
/// The strides in each dimension between different loads/stores.
typedef Shape<0, Threads::kH * ShapeCount<Tile>::kWc, Threads::kW * kScalarsPerLds_>
ImmediateOffsetStrides;
/// The number of iterations needed to load/store the tile.
typedef Shape<1, Tile::kH / Threads::kH, Tile::kW / Threads::kW, Tile::kC / kScalarsPerLds_>
Iterations;
/// ThreadOffset
struct ThreadOffset {
CUTLASS_HOST_DEVICE
Coord<4> operator()() const {
// The offset.
int const offset = ComputeThreadOffsetFromStrides<Threads, ThreadsStrides>::get();
return make_Coord(0, 0, offset, 0);
}
};
};
////////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace gemm
} // namespace cutlass
#endif // defined CUTLASS_USE_WMMA_API
|
