1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
|
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice, this list of
* conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright notice, this list of
* conditions and the following disclaimer in the documentation and/or other materials
* provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
* to endorse or promote products derived from this software without specific prior written
* permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
* FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
* STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
/*! \file
\brief Defines multiply-add operations on fragments within a thread.
*/
#pragma once
#include <cutlass/fragment.h>
namespace cutlass {
namespace gemm {
////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename Scalar_>
struct FragmentMultiplyAdd {
/// The shape of the instruction.
typedef Shape<1, 1, 1, 1> InstructionShape;
/// The type for A.
typedef Scalar_ ScalarA;
/// The type for B.
typedef Scalar_ ScalarB;
/// The type for C and D.
typedef Scalar_ ScalarC;
/// Ctor.
CUTLASS_DEVICE FragmentMultiplyAdd() {}
/// Multiply : d = a*b.
template <typename FragmentB_, typename FragmentCd_>
CUTLASS_DEVICE void multiply(Scalar_ a, FragmentB_ const& b, FragmentCd_& d) {
int const kReduction = FragmentB_::kElements / FragmentCd_::kElements;
for (int j = 0; j < FragmentCd_::kElements; ++j) {
d[j] = a * b[j * kReduction + 0];
for (int k = 1; k < kReduction; ++k) {
d[j] += a * b[j * kReduction + k];
}
}
}
/// Multiply : d = a*b + c.
template <typename FragmentB_, typename FragmentCd_>
CUTLASS_DEVICE void multiply_add(Scalar_ a,
FragmentB_ const& b,
FragmentCd_ const& c,
FragmentCd_& d) {
int const kReduction = FragmentB_::kElements / FragmentCd_::kElements;
for (int j = 0; j < FragmentCd_::kElements; ++j) {
d[j] = a * b[j * kReduction + 0] + c[j];
for (int k = 1; k < kReduction; ++k) {
d[j] += a * b[j * kReduction + k];
}
}
}
};
////////////////////////////////////////////////////////////////////////////////////////////////////
#if !defined(__CUDACC_RTC__) || defined(CUTLASS_NVRTC_HAS_FP16)
template <>
struct FragmentMultiplyAdd<half> {
/// The shape of the instruction.
typedef Shape<1, 1, 2, 1> InstructionShape;
/// The type for A.
typedef half ScalarA;
/// The type for B.
typedef half ScalarB;
/// The type for C and D.
typedef half ScalarC;
/// Ctor.
CUTLASS_DEVICE FragmentMultiplyAdd() {}
/// Multiply : d = a*b.
template <typename FragmentB_, typename FragmentCd_>
CUTLASS_DEVICE void multiply(half a, FragmentB_ const& b, FragmentCd_& d) {
#if defined(__CUDACC__) && __CUDA_ARCH__ >= 530
// Assemble a half2 from a.
__half2 const a_half2 = __half2half2(a);
// The input.
__half2 const* b_half2 = reinterpret_cast<__half2 const*>(&b[0]);
// The output.
__half2* d_half2 = reinterpret_cast<__half2*>(&d[0]);
int const kReduction = FragmentB_::kElements / FragmentCd_::kElements;
for (int j = 0; j < FragmentCd_::kElements / 2; ++j) {
d_half2[j] = __hmul2(a_half2, b_half2[j * kReduction + 0]);
for (int k = 1; k < kReduction; ++k) {
d_half2[j] = __hfma2(a_half2, b_half2[j * kReduction + k], d_half2[j]);
}
}
#endif
}
/// Multiply : d = a*b + c.
template <typename FragmentB_, typename FragmentCd_>
CUTLASS_DEVICE void multiply_add(half a,
FragmentB_ const& b,
FragmentCd_ const& c,
FragmentCd_& d) {
#if defined(__CUDACC__) && __CUDA_ARCH__ >= 530
// Assemble a half2 from a.
__half2 const a_half2 = __half2half2(a);
// The inputs.
__half2 const* b_half2 = reinterpret_cast<__half2 const*>(&b[0]);
__half2 const* c_half2 = reinterpret_cast<__half2 const*>(&c[0]);
// The output.
__half2* d_half2 = reinterpret_cast<__half2*>(&d[0]);
int const kReduction = (FragmentB_::kElements / FragmentCd_::kElements);
for (int j = 0; j < FragmentCd_::kElements / 2; ++j) {
d_half2[j] = __hfma2(a_half2, b_half2[j * kReduction + 0], c_half2[j]);
for (int k = 1; k < kReduction; ++k) {
d_half2[j] = __hfma2(a_half2, b_half2[j * kReduction + k], d_half2[j]);
}
}
#endif
}
};
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace gemm
} // namespace cutlass
|
