Cutlass
CUDA Templates for Linear Algebra Subroutines and Solvers
igemm_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  **************************************************************************************************/
29 #pragma once
30 
31 #include "cutlass/fragment.h"
32 
34 
35 namespace cutlass {
36 namespace gemm {
37 
39 
41 template <typename ThreadGemmShape_, typename ThreadsPerWarp_>
42 struct ThreadMultiplyAdd<ThreadGemmShape_, ThreadsPerWarp_, int8_t, int8_t, int> {
46  typedef ThreadGemmShape_ ThreadGemmShape;
50  typedef ThreadsPerWarp_ ThreadsPerWarp;
54  typedef int8_t ScalarA;
58  typedef int8_t ScalarB;
62  typedef int ScalarC;
65 
67  CUTLASS_DEVICE ThreadMultiplyAdd() {}
68 
70  CUTLASS_DEVICE void multiply_add(FragmentA const& a,
71  FragmentB const& b,
72  Accumulators const& c,
73  Accumulators& d) {
74 
75  #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 610)
76  // The inputs.
77  int const* a_int = reinterpret_cast<int const*>(&a[0]);
78  int const* b_int = reinterpret_cast<int const*>(&b[0]);
79 
80  for (int j = 0; j < AccumulatorsPerThread::kH; ++j) {
81  for (int i = 0; i < AccumulatorsPerThread::kW; ++i) {
82  asm volatile("dp4a.s32.s32 %0, %1, %2, %3;"
83  : "=r"(d[j * AccumulatorsPerThread::kW + i])
84  : "r"(a_int[i]), "r"(b_int[j]), "r"(c[j * AccumulatorsPerThread::kW + i]));
85  }
86  }
87  #endif
88  }
89 };
90 
92 
93 } // namespace gemm
94 } // namespace cutlass
Definition: convert.h:33
Fragment< ScalarA, AccumulatorsPerThread::kW *4 > FragmentA
The fragment for A.
Definition: igemm_multiply_add.h:56
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
Template implementing matrix multiply-add operations on fragments.
CUTLASS_DEVICE ThreadMultiplyAdd()
Ctor.
Definition: igemm_multiply_add.h:67
CUTLASS_DEVICE void multiply_add(FragmentA const &a, FragmentB const &b, Accumulators const &c, Accumulators &d)
Multiply : d = a*b + c.
Definition: igemm_multiply_add.h:70
int ScalarC
The type for C and D.
Definition: igemm_multiply_add.h:62
Shape< 4, 1, 1 > InstructionShape
The shape of the instruction.
Definition: igemm_multiply_add.h:44
ThreadsPerWarp_ ThreadsPerWarp
The number of threads per warp.
Definition: igemm_multiply_add.h:50
ThreadGemmShape_ ThreadGemmShape
Shape of the thread-level GEMM (K-by-N-by-M)
Definition: igemm_multiply_add.h:46
Fragment< ScalarC, AccumulatorsPerThread::kH *AccumulatorsPerThread::kW > Accumulators
The accumulators.
Definition: igemm_multiply_add.h:64
A Shape implementing Layout Concept describing the dimensions of a cube.
Definition: shape.h:64
ShapeMul< ThreadGemmShape, ThreadsPerWarp >::Shape AccumulatorsPerWarp
The number of accumulators per warp.
Definition: igemm_multiply_add.h:52
Template performing matrix multiply-add operation within a thread.
Definition: thread_multiply_add.h:44
ThreadGemmShape AccumulatorsPerThread
Aliased for compatibility. Will be removed in CUTLASS v2.0.
Definition: igemm_multiply_add.h:48
Fragment< ScalarB, AccumulatorsPerThread::kH *4 > FragmentB
The fragment for B.
Definition: igemm_multiply_add.h:60
Defines Fragment, a statically-sized array for storing parts of matrices within a thread&#39;s registers...