File size: 6,283 Bytes
7f51798
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
/*
 * SPDX-FileCopyrightText: Copyright (c) 2021-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: LicenseRef-NvidiaProprietary
 *
 * NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
 * property and proprietary rights in and to this material, related
 * documentation and any modifications thereto. Any use, reproduction,
 * disclosure or distribution of this material and related documentation
 * without an express license agreement from NVIDIA CORPORATION or
 * its affiliates is strictly prohibited.
 */

#include <c10/util/Half.h>
#include "bias_act.h"

//------------------------------------------------------------------------
// Helpers.

template <class T> struct InternalType;
template <> struct InternalType<double>     { typedef double scalar_t; };
template <> struct InternalType<float>      { typedef float  scalar_t; };
template <> struct InternalType<c10::Half>  { typedef float  scalar_t; };

//------------------------------------------------------------------------
// CUDA kernel.

template <class T, int A>
__global__ void bias_act_kernel(bias_act_kernel_params p)
{
    typedef typename InternalType<T>::scalar_t scalar_t;
    int G                 = p.grad;
    scalar_t alpha        = (scalar_t)p.alpha;
    scalar_t gain         = (scalar_t)p.gain;
    scalar_t clamp        = (scalar_t)p.clamp;
    scalar_t one          = (scalar_t)1;
    scalar_t two          = (scalar_t)2;
    scalar_t expRange     = (scalar_t)80;
    scalar_t halfExpRange = (scalar_t)40;
    scalar_t seluScale    = (scalar_t)1.0507009873554804934193349852946;
    scalar_t seluAlpha    = (scalar_t)1.6732632423543772848170429916717;

    // Loop over elements.
    int xi = blockIdx.x * p.loopX * blockDim.x + threadIdx.x;
    for (int loopIdx = 0; loopIdx < p.loopX && xi < p.sizeX; loopIdx++, xi += blockDim.x)
    {
        // Load.
        scalar_t x = (scalar_t)((const T*)p.x)[xi];
        scalar_t b = (p.b) ? (scalar_t)((const T*)p.b)[(xi / p.stepB) % p.sizeB] : 0;
        scalar_t xref = (p.xref) ? (scalar_t)((const T*)p.xref)[xi] : 0;
        scalar_t yref = (p.yref) ? (scalar_t)((const T*)p.yref)[xi] : 0;
        scalar_t dy = (p.dy) ? (scalar_t)((const T*)p.dy)[xi] : one;
        scalar_t yy = (gain != 0) ? yref / gain : 0;
        scalar_t y = 0;

        // Apply bias.
        ((G == 0) ? x : xref) += b;

        // linear
        if (A == 1)
        {
            if (G == 0) y = x;
            if (G == 1) y = x;
        }

        // relu
        if (A == 2)
        {
            if (G == 0) y = (x > 0) ? x : 0;
            if (G == 1) y = (yy > 0) ? x : 0;
        }

        // lrelu
        if (A == 3)
        {
            if (G == 0) y = (x > 0) ? x : x * alpha;
            if (G == 1) y = (yy > 0) ? x : x * alpha;
        }

        // tanh
        if (A == 4)
        {
            if (G == 0) { scalar_t c = exp(x); scalar_t d = one / c; y = (x < -expRange) ? -one : (x > expRange) ? one : (c - d) / (c + d); }
            if (G == 1) y = x * (one - yy * yy);
            if (G == 2) y = x * (one - yy * yy) * (-two * yy);
        }

        // sigmoid
        if (A == 5)
        {
            if (G == 0) y = (x < -expRange) ? 0 : one / (exp(-x) + one);
            if (G == 1) y = x * yy * (one - yy);
            if (G == 2) y = x * yy * (one - yy) * (one - two * yy);
        }

        // elu
        if (A == 6)
        {
            if (G == 0) y = (x >= 0) ? x : exp(x) - one;
            if (G == 1) y = (yy >= 0) ? x : x * (yy + one);
            if (G == 2) y = (yy >= 0) ? 0 : x * (yy + one);
        }

        // selu
        if (A == 7)
        {
            if (G == 0) y = (x >= 0) ? seluScale * x : (seluScale * seluAlpha) * (exp(x) - one);
            if (G == 1) y = (yy >= 0) ? x * seluScale : x * (yy + seluScale * seluAlpha);
            if (G == 2) y = (yy >= 0) ? 0 : x * (yy + seluScale * seluAlpha);
        }

        // softplus
        if (A == 8)
        {
            if (G == 0) y = (x > expRange) ? x : log(exp(x) + one);
            if (G == 1) y = x * (one - exp(-yy));
            if (G == 2) { scalar_t c = exp(-yy); y = x * c * (one - c); }
        }

        // swish
        if (A == 9)
        {
            if (G == 0)
                y = (x < -expRange) ? 0 : x / (exp(-x) + one);
            else
            {
                scalar_t c = exp(xref);
                scalar_t d = c + one;
                if (G == 1)
                    y = (xref > halfExpRange) ? x : x * c * (xref + d) / (d * d);
                else
                    y = (xref > halfExpRange) ? 0 : x * c * (xref * (two - d) + two * d) / (d * d * d);
                yref = (xref < -expRange) ? 0 : xref / (exp(-xref) + one) * gain;
            }
        }

        // Apply gain.
        y *= gain * dy;

        // Clamp.
        if (clamp >= 0)
        {
            if (G == 0)
                y = (y > -clamp & y < clamp) ? y : (y >= 0) ? clamp : -clamp;
            else
                y = (yref > -clamp & yref < clamp) ? y : 0;
        }

        // Store.
        ((T*)p.y)[xi] = (T)y;
    }
}

//------------------------------------------------------------------------
// CUDA kernel selection.

template <class T> void* choose_bias_act_kernel(const bias_act_kernel_params& p)
{
    if (p.act == 1) return (void*)bias_act_kernel<T, 1>;
    if (p.act == 2) return (void*)bias_act_kernel<T, 2>;
    if (p.act == 3) return (void*)bias_act_kernel<T, 3>;
    if (p.act == 4) return (void*)bias_act_kernel<T, 4>;
    if (p.act == 5) return (void*)bias_act_kernel<T, 5>;
    if (p.act == 6) return (void*)bias_act_kernel<T, 6>;
    if (p.act == 7) return (void*)bias_act_kernel<T, 7>;
    if (p.act == 8) return (void*)bias_act_kernel<T, 8>;
    if (p.act == 9) return (void*)bias_act_kernel<T, 9>;
    return NULL;
}

//------------------------------------------------------------------------
// Template specializations.

template void* choose_bias_act_kernel<double>       (const bias_act_kernel_params& p);
template void* choose_bias_act_kernel<float>        (const bias_act_kernel_params& p);
template void* choose_bias_act_kernel<c10::Half>    (const bias_act_kernel_params& p);

//------------------------------------------------------------------------