Cutlass
CUDA Templates for Linear Algebra Subroutines and Solvers
fp16_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  **************************************************************************************************/
29 #pragma once
30 
31 #include "cutlass/gemm/gemm.h"
38 
39 namespace cutlass {
40 namespace gemm {
41 
43 
44 template <
46  typename OutputTile_,
48  typename ThreadGemmShape_,
50  typename ScalarA_,
52  typename ScalarB_,
54  typename ScalarC_,
56  typename ScalarD_,
58  int kScalarsPerLdgA_ = 1,
60  int kScalarsPerLdgB_ = 1>
61 struct Fp16SgemmConfig : public GemmConfig<
63  ScalarA_,
65  ScalarB_,
67  ScalarC_,
69  ScalarD_,
71  OutputTile_,
73  ThreadMultiplyAdd<ThreadGemmShape_, Shape<1, 4, 8>, ScalarA_, ScalarB_, float /*for sgemm accum is float*/>,
75  kScalarsPerLdgA_,
77  kScalarsPerLdgA_,
79  4,
81  kScalarsPerLdgB_,
83  kScalarsPerLdgB_,
85  4,
87  1,
89  4,
91  1,
93  2> {};
94 
96 
97 template <
99  MatrixLayout::Kind kLayoutA_,
101  MatrixLayout::Kind kLayoutB_,
103  typename OutputTile_ = Shape<8, 128, 128>,
105  typename ScalarA_ = half,
107  typename ScalarB_ = half,
109  typename ScalarC_ = half,
111  typename ScalarD_ = half,
113  typename Scalar_ = half,
115  typename EpilogueFunctor_ = LinearScaling<Scalar_, FragmentMultiplyAdd<Scalar_, float/*accumulator type*/> >,
117  typename ThreadGemmShape_ = Shape<8, 8, 8>,
119  int kScalarsPerLdgA_ = 1,
121  int kScalarsPerLdgB_ = 1,
123  typename Index_ = int,
125  typename GemmConfig_ =
126  Fp16SgemmConfig<OutputTile_,
127  ThreadGemmShape_,
128  ScalarA_,
129  ScalarB_,
130  ScalarC_,
131  ScalarD_,
132  kScalarsPerLdgA_,
133  kScalarsPerLdgB_>,
135  typename GemmEpilogueTraits_ =
138  // The layout for A.
139  kLayoutA_,
140  // The layout for B.
141  kLayoutB_,
142  // The config.
143  GemmConfig_,
144  // The epilogue.
145  GemmEpilogue<GemmEpilogueTraits_>,
146  // The index.
147  Index_> {};
148 
150 
151 } // namespace gemm
152 } // 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
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: fp16_sgemm_traits.h:61
Template implementing matrix multiply-add operations on fragments.
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: fp16_sgemm_traits.h:137
Definition: gemm_traits.h:773
Definition: fragment_multiply_add.h:41