Cutlass
CUDA Templates for Linear Algebra Subroutines and Solvers
sgemm_traits.h
Go to the documentation of this file.
1 /***************************************************************************************************
2  * Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without modification, are permitted
5  * provided that the following conditions are met:
6  * * Redistributions of source code must retain the above copyright notice, this list of
7  * conditions and the following disclaimer.
8  * * Redistributions in binary form must reproduce the above copyright notice, this list of
9  * conditions and the following disclaimer in the documentation and/or other materials
10  * provided with the distribution.
11  * * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
12  * to endorse or promote products derived from this software without specific prior written
13  * permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
17  * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
19  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
20  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
21  * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
22  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
23  *
24  **************************************************************************************************/
28 #pragma once
29 
30 #include "cutlass/gemm/gemm.h"
37 
38 namespace cutlass {
39 namespace gemm {
40 
42 
43 template <
45  typename OutputTile_,
47  typename ThreadGemmShape_,
49  int kScalarsPerLdgA_ = 1,
51  int kScalarsPerLdgB_ = 1,
53  bool kLaunchBounds = true>
54 struct SgemmConfig : public GemmConfig<
56  float,
58  float,
60  float,
62  float,
64  OutputTile_,
66  ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, float, float, float>,
68  kScalarsPerLdgA_,
70  kScalarsPerLdgA_,
72  4,
74  kScalarsPerLdgB_,
76  kScalarsPerLdgB_,
78  4,
80  1,
82  4,
84  1,
86  2,
88  false,
90  true,
92  kLaunchBounds> {};
93 
95 
96 template <
98  MatrixLayout::Kind kLayoutA_,
100  MatrixLayout::Kind kLayoutB_,
102  typename OutputTile_ = Shape<8, 128, 128>,
104  typename EpilogueFunctor_ = LinearScaling<float>,
106  typename ThreadGemmShape_ = Shape<8, 8, 8>,
108  int kScalarsPerLdgA_ = 1,
110  int kScalarsPerLdgB_ = 1,
112  typename Index_ = int,
114  typename GemmConfig_ =
117  typename GemmEpilogueTraits_ =
120  // The layout for A.
121  kLayoutA_,
122  // The layout for B.
123  kLayoutB_,
124  // The config.
125  GemmConfig_,
126  // The epilogue.
127  GemmEpilogue<GemmEpilogueTraits_>,
128  // The index.
129  Index_> {};
130 
132 
134 template <
136  MatrixLayout::Kind kLayoutA_,
138  MatrixLayout::Kind kLayoutB_,
140  typename OutputTile_ = Shape<8, 128, 128>,
142  typename EpilogueFunctor_ = LinearScaling<float>,
144  typename ThreadGemmShape_ = Shape<8, 8, 8>,
146  int kScalarsPerLdgA_ = 1,
148  int kScalarsPerLdgB_ = 1,
150  typename Index_ = int,
152  typename GemmConfig_ =
155  typename GemmEpilogueTraits_ =
158  // The layout for A.
159  kLayoutA_,
160  // The layout for B.
161  kLayoutB_,
162  // The config.
163  GemmConfig_,
164  // The epilogue.
165  GemmEpilogue<GemmEpilogueTraits_>,
166  // The index.
167  Index_> {};
168 
170 
171 } // namespace gemm
172 } // namespace cutlass
Definition: convert.h:33
Defines iterators for efficiently loading and storing to global memory.
Defines structural properties of complete GEMM computation.
Kind
Enumeration defining fundamental contiguous layouts.
Definition: matrix_traits.h:159
Definition: sgemm_traits.h:54
Template implementing matrix multiply-add operations on fragments.
Helper to define SGEMM traits using Launch Bounds.
Definition: sgemm_traits.h:157
Implements the epilogue phase of the GEMM kernel that efficiently updates global memory with the comp...
Defines iterators for efficiently loading and storing tiles to and from shared memory.
Definition: gemm_config.h:76
A Shape implementing Layout Concept describing the dimensions of a cube.
Definition: shape.h:64
Definition: gemm_epilogue_traits.h:340
Definition: sgemm_traits.h:119
Functor to compute linear combination of fragments.
Definition: linear_scaling.h:51
Implements a software-pipelined efficient GEMM.
Defines structural properties of the GEMM epilogue.
Definition: gemm_traits.h:773