Cutlass
CUDA Templates for Linear Algebra Subroutines and Solvers
fp16_sgemm_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"
32 namespace cutlass {
33 namespace gemm {
34 
36 
38 template <typename ThreadGemmShape_,
39  typename ThreadsPerWarp_>
40 struct ThreadMultiplyAdd<ThreadGemmShape_, ThreadsPerWarp_, half, half, float> {
44  typedef ThreadGemmShape_ ThreadGemmShape;
48  typedef ThreadsPerWarp_ ThreadsPerWarp;
52  typedef half ScalarA;
56  typedef half ScalarB;
60  typedef float ScalarC;
63 
65  CUTLASS_DEVICE ThreadMultiplyAdd() {}
66 
68  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
69  FragmentB const& b,
70  Accumulators const& c,
71  Accumulators& d) {
72  for (int j = 0; j < AccumulatorsPerThread::kH; ++j) {
73  for (int i = 0; i < AccumulatorsPerThread::kW; ++i) {
74  d[j * AccumulatorsPerThread::kW + i] = static_cast<ScalarC>(a[i]) * static_cast<ScalarC>(b[j]) + c[j * AccumulatorsPerThread::kW + i];
75  }
76  }
77  }
78 };
79 
81 
82 } // namespace gemm
83 } // namespace cutlass
Definition: convert.h:33
CUTLASS_DEVICE ThreadMultiplyAdd()
Ctor.
Definition: fp16_sgemm_multiply_add.h:65
Fragment< ScalarB, AccumulatorsPerThread::kH > FragmentB
The fragment for B.
Definition: fp16_sgemm_multiply_add.h:58
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: fp16_sgemm_multiply_add.h:50
Template implementing matrix multiply-add operations on fragments.
ThreadGemmShape_ ThreadGemmShape
The shape of a thread-leveel matrix multiply accumulate.
Definition: fp16_sgemm_multiply_add.h:44
CUTLASS_DEVICE void multiply_add(FragmentA const &a, FragmentB const &b, Accumulators const &c, Accumulators &d)
Multiply : d = a*b + c.
Definition: fp16_sgemm_multiply_add.h:68
half ScalarA
The type for A. specialized to half.
Definition: fp16_sgemm_multiply_add.h:52
half ScalarB
The type for B. specialized to half.
Definition: fp16_sgemm_multiply_add.h:56
ThreadsPerWarp_ ThreadsPerWarp
The number of threads per warp.
Definition: fp16_sgemm_multiply_add.h:48
Fragment< ScalarA, AccumulatorsPerThread::kW > FragmentA
The fragment for A.
Definition: fp16_sgemm_multiply_add.h:54
float ScalarC
The type for C and D. specialized to float.
Definition: fp16_sgemm_multiply_add.h:60
A Shape implementing Layout Concept describing the dimensions of a cube.
Definition: shape.h:64
Fragment< ScalarC, AccumulatorsPerThread::kH *AccumulatorsPerThread::kW, 16 > Accumulators
The accumulators.
Definition: fp16_sgemm_multiply_add.h:62
ThreadGemmShape AccumulatorsPerThread
Aliased to "AccumulatorsPerThread" for compatibility. Expect to be renamed in CUTLASS v2...
Definition: fp16_sgemm_multiply_add.h:46
Template performing matrix multiply-add operation within a thread.
Definition: thread_multiply_add.h:44
Defines Fragment, a statically-sized array for storing parts of matrices within a thread&#39;s registers...
Shape< 1, 1, 1, 1 > InstructionShape
The shape of the instruction.
Definition: fp16_sgemm_multiply_add.h:42