Cutlass
CUDA Templates for Linear Algebra Subroutines and Solvers
thread_multiply_add.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/fragment.h"
31 
32 namespace cutlass {
33 namespace gemm {
34 
36 
38 template <typename ThreadGemmShape_,
39  typename ThreadsPerWarp_,
40  typename ScalarA_,
41  typename ScalarB_,
42  typename ScalarC_,
48  typedef ThreadGemmShape_ ThreadGemmShape;
52  typedef ThreadsPerWarp_ ThreadsPerWarp;
56  typedef ScalarA_ ScalarA;
60  typedef ScalarB_ ScalarB;
64  typedef ScalarC_ ScalarC;
67 
69  CUTLASS_DEVICE ThreadMultiplyAdd() {}
70 
72  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
73  FragmentB const& b,
74  Accumulators const& c,
75  Accumulators& d) {
76  if(kLayout_ == MatrixLayout::kColumnMajor) {
77  for (int j = 0; j < AccumulatorsPerThread::kH; ++j) {
78  for (int i = 0; i < AccumulatorsPerThread::kW; ++i) {
79  d[j * AccumulatorsPerThread::kW + i] = a[i] * b[j] + c[j * AccumulatorsPerThread::kW + i];
80  }
81  }
82  }
83  else {
84  for(int i = 0; i < AccumulatorsPerThread::kW; ++i) {
85  for(int j = 0; j < AccumulatorsPerThread::kH; ++j) {
86  d[i * AccumulatorsPerThread::kH + j] = a[i] * b[j] + c[i * AccumulatorsPerThread::kH + j];
87  }
88  }
89  }
90  }
91 };
92 
94 
95 } // namespace gemm
96 } // namespace cutlass
Definition: convert.h:33
Fragment< ScalarA, AccumulatorsPerThread::kW > FragmentA
The fragment for A.
Definition: thread_multiply_add.h:58
Kind
Enumeration defining fundamental contiguous layouts.
Definition: matrix_traits.h:159
Shape< A_::kD *B_::kD, A_::kH *B_::kH, A_::kW *B_::kW, A_::kC *B_::kC > Shape
Definition: shape.h:119
A template defining Fragment Concept.
Definition: fragment.h:99
ShapeMul< ThreadGemmShape, ThreadsPerWarp >::Shape AccumulatorsPerWarp
The number of accumulators per warp.
Definition: thread_multiply_add.h:54
ScalarB_ ScalarB
The type for B.
Definition: thread_multiply_add.h:60
CUTLASS_DEVICE void multiply_add(FragmentA const &a, FragmentB const &b, Accumulators const &c, Accumulators &d)
Multiply : d = a*b + c.
Definition: thread_multiply_add.h:72
CUTLASS_DEVICE ThreadMultiplyAdd()
Ctor.
Definition: thread_multiply_add.h:69
ThreadGemmShape_ ThreadGemmShape
The shape of a thread-leveel matrix multiply accumulate.
Definition: thread_multiply_add.h:48
A Shape implementing Layout Concept describing the dimensions of a cube.
Definition: shape.h:64
Definition: matrix_traits.h:159
Fragment< ScalarC, AccumulatorsPerThread::kH *AccumulatorsPerThread::kW, 16 > Accumulators
The accumulators.
Definition: thread_multiply_add.h:66
ScalarA_ ScalarA
The type for A.
Definition: thread_multiply_add.h:56
Template performing matrix multiply-add operation within a thread.
Definition: thread_multiply_add.h:44
ThreadsPerWarp_ ThreadsPerWarp
The number of threads per warp.
Definition: thread_multiply_add.h:52
ScalarC_ ScalarC
The type for C and D.
Definition: thread_multiply_add.h:64
ThreadGemmShape AccumulatorsPerThread
Aliased to "AccumulatorsPerThread" for compatibility. Expect to be renamed in CUTLASS v2...
Definition: thread_multiply_add.h:50
Shape< 1, 1, 1, 1 > InstructionShape
The shape of the instruction.
Definition: thread_multiply_add.h:46
Defines Fragment, a statically-sized array for storing parts of matrices within a thread&#39;s registers...
Fragment< ScalarB, AccumulatorsPerThread::kH > FragmentB
The fragment for B.
Definition: thread_multiply_add.h:62