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
|
/*
This file is part of darktable,
Copyright (C) 2020 - darktable developers.
darktable is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
darktable is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with darktable. If not, see <http://www.gnu.org/licenses/>.
*/
// This is the OpenCL translation of common/noise_generator.h
typedef enum dt_noise_distribution_t
{
DT_NOISE_UNIFORM = 0, // $DESCRIPTION: "uniform"
DT_NOISE_GAUSSIAN = 1, // $DESCRIPTION: "gaussian"
DT_NOISE_POISSONIAN = 2 // $DESCRIPTION: "poissonian"
} dt_noise_distribution_t;
static inline unsigned int splitmix32(const unsigned long seed)
{
// fast random number generator
// reference : https://gist.github.com/imneme/6179748664e88ef3c34860f44309fc71
unsigned long result = (seed ^ (seed >> 33)) * 0x62a9d9ed799705f5ul;
result = (result ^ (result >> 28)) * 0xcb24d0a5c88c35b3ul;
return (unsigned int)(result >> 32);
}
static inline unsigned rol32(const unsigned int x, const int k)
{
return (x << k) | (x >> (32 - k));
}
static inline float xoshiro128plus(uint state[4])
{
// fast random number generator
// reference : http://prng.di.unimi.it/
const unsigned int result = state[0] + state[3];
const unsigned int t = state[1] << 9;
state[2] ^= state[0];
state[3] ^= state[1];
state[1] ^= state[2];
state[0] ^= state[3];
state[2] ^= t;
state[3] = rol32(state[3], 11);
return (float)(result >> 8) * 0x1.0p-24f; // take the first 24 bits and put them in mantissa
}
static inline float4 uniform_noise_simd(const float4 mu, const float4 sigma, uint state[4])
{
const float4 noise = { xoshiro128plus(state), xoshiro128plus(state), xoshiro128plus(state), 0.f };
return mu + 2.0f * (noise - 0.5f) * sigma;
}
static inline float4 gaussian_noise_simd(const float4 mu, const float4 sigma, uint state[4])
{
// Create gaussian noise centered in mu of standard deviation sigma
// state should be initialized with xoshiro256_init() before calling and private in thread
// flip needs to be flipped every next iteration
// reference : https://en.wikipedia.org/wiki/Box%E2%80%93Muller_transform
float4 u1, u2;
u1.x = xoshiro128plus(state);
u1.y = xoshiro128plus(state);
u1.z = xoshiro128plus(state);
u2.x = xoshiro128plus(state);
u2.y = xoshiro128plus(state);
u2.z = xoshiro128plus(state);
u1 = fmax(u1, FLT_MIN);
const float4 flip = { 1.0f, 0.0f, 1.0f, 0.0f };
const float4 flip_comp = { 0.0f, 1.0f, 0.0f, 0.0f };
// flip is a 4×1 boolean mask
const float4 noise = flip * native_sqrt(-2.0f * native_log(u1)) * native_cos(2.f * M_PI_F * u2) +
flip_comp * native_sqrt(-2.0f * native_log(u1)) * native_sin(2.f * M_PI_F * u2);
return noise * sigma + mu;
}
static inline float4 poisson_noise_simd(const float4 mu, const float4 sigma, uint state[4])
{
// create poissonian noise - It's just gaussian noise with Anscombe transform applied
float4 u1, u2;
// we need to generate the random numbers in this order to match CPU path.
u1.x = xoshiro128plus(state);
u2.x = xoshiro128plus(state);
u1.y = xoshiro128plus(state);
u2.y = xoshiro128plus(state);
u1.z = xoshiro128plus(state);
u2.z = xoshiro128plus(state);
u1 = fmax(u1, (float4)FLT_MIN);
const float4 flip = { 1.0f, 0.0f, 1.0f, 0.0f };
const float4 flip_comp = { 0.0f, 1.0f, 0.0f, 0.0f };
// flip is a 4×1 boolean mask
const float4 noise = flip * native_sqrt(-2.0f * native_log(u1)) * native_cos(2.f * M_PI_F * u2) +
flip_comp * native_sqrt(-2.0f * native_log(u1)) * native_sin(2.f * M_PI_F * u2);
// now we have gaussian noise, then apply Anscombe transform to get poissonian one
const float4 r = noise * sigma + 2.0f * native_sqrt(fmax(mu + (3.f / 8.f), 0.0f));
return ((r * r - sigma * sigma) / (4.f)) - (3.f / 8.f);
}
static inline float4 dt_noise_generator_simd(const dt_noise_distribution_t distribution,
const float4 mu, const float4 param,
uint state[4])
{
// vector version
switch(distribution)
{
case(DT_NOISE_UNIFORM):
default:
{
return uniform_noise_simd(mu, param, state);
}
case(DT_NOISE_GAUSSIAN):
{
return gaussian_noise_simd(mu, param, state);
}
case(DT_NOISE_POISSONIAN):
{
return poisson_noise_simd(mu, param, state);
}
}
}
|
