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
|
/* SPDX-License-Identifier: Apache-2.0
* Copyright 2011-2022 Blender Foundation */
#pragma once
#include "kernel/sample/jitter.h"
#include "kernel/sample/sobol_burley.h"
#include "util/hash.h"
CCL_NAMESPACE_BEGIN
/* Pseudo random numbers, uncomment this for debugging correlations. Only run
* this single threaded on a CPU for repeatable results. */
//#define __DEBUG_CORRELATION__
/* High Dimensional Sobol.
*
* Multidimensional sobol with generator matrices. Dimension 0 and 1 are equal
* to classic Van der Corput and Sobol sequences. */
#ifdef __SOBOL__
/* Skip initial numbers that for some dimensions have clear patterns that
* don't cover the entire sample space. Ideally we would have a better
* progressive pattern that doesn't suffer from this problem, because even
* with this offset some dimensions are quite poor.
*/
# define SOBOL_SKIP 64
ccl_device uint sobol_dimension(KernelGlobals kg, int index, int dimension)
{
uint result = 0;
uint i = index + SOBOL_SKIP;
for (int j = 0, x; (x = find_first_set(i)); i >>= x) {
j += x;
result ^= __float_as_uint(kernel_data_fetch(sample_pattern_lut, 32 * dimension + j - 1));
}
return result;
}
#endif /* __SOBOL__ */
ccl_device_forceinline float path_rng_1D(KernelGlobals kg,
uint rng_hash,
int sample,
int dimension)
{
#ifdef __DEBUG_CORRELATION__
return (float)drand48();
#endif
if (kernel_data.integrator.sampling_pattern == SAMPLING_PATTERN_SOBOL_BURLEY) {
return sobol_burley_sample_1D(sample, dimension, rng_hash);
}
#ifdef __SOBOL__
if (kernel_data.integrator.sampling_pattern == SAMPLING_PATTERN_PMJ)
#endif
{
return pmj_sample_1D(kg, sample, rng_hash, dimension);
}
#ifdef __SOBOL__
/* Sobol sequence value using direction vectors. */
uint result = sobol_dimension(kg, sample, dimension);
float r = (float)result * (1.0f / (float)0xFFFFFFFF);
/* Cranly-Patterson rotation using rng seed */
float shift;
/* Hash rng with dimension to solve correlation issues.
* See T38710, T50116.
*/
uint tmp_rng = hash_wang_seeded_uint(dimension, rng_hash);
shift = tmp_rng * (kernel_data.integrator.scrambling_distance / (float)0xFFFFFFFF);
return r + shift - floorf(r + shift);
#endif
}
ccl_device_forceinline void path_rng_2D(KernelGlobals kg,
uint rng_hash,
int sample,
int dimension,
ccl_private float *fx,
ccl_private float *fy)
{
#ifdef __DEBUG_CORRELATION__
*fx = (float)drand48();
*fy = (float)drand48();
return;
#endif
if (kernel_data.integrator.sampling_pattern == SAMPLING_PATTERN_SOBOL_BURLEY) {
sobol_burley_sample_2D(sample, dimension, rng_hash, fx, fy);
return;
}
#ifdef __SOBOL__
if (kernel_data.integrator.sampling_pattern == SAMPLING_PATTERN_PMJ)
#endif
{
pmj_sample_2D(kg, sample, rng_hash, dimension, fx, fy);
return;
}
#ifdef __SOBOL__
/* Sobol. */
*fx = path_rng_1D(kg, rng_hash, sample, dimension);
*fy = path_rng_1D(kg, rng_hash, sample, dimension + 1);
#endif
}
/**
* 1D hash recommended from "Hash Functions for GPU Rendering" JCGT Vol. 9, No. 3, 2020
* See https://www.shadertoy.com/view/4tXyWN and https://www.shadertoy.com/view/XlGcRh
* http://www.jcgt.org/published/0009/03/02/paper.pdf
*/
ccl_device_inline uint hash_iqint1(uint n)
{
n = (n << 13U) ^ n;
n = n * (n * n * 15731U + 789221U) + 1376312589U;
return n;
}
/**
* 2D hash recommended from "Hash Functions for GPU Rendering" JCGT Vol. 9, No. 3, 2020
* See https://www.shadertoy.com/view/4tXyWN and https://www.shadertoy.com/view/XlGcRh
* http://www.jcgt.org/published/0009/03/02/paper.pdf
*/
ccl_device_inline uint hash_iqnt2d(const uint x, const uint y)
{
const uint qx = 1103515245U * ((x >> 1U) ^ (y));
const uint qy = 1103515245U * ((y >> 1U) ^ (x));
const uint n = 1103515245U * ((qx) ^ (qy >> 3U));
return n;
}
ccl_device_inline uint path_rng_hash_init(KernelGlobals kg,
const int sample,
const int x,
const int y)
{
const uint rng_hash = hash_iqnt2d(x, y) ^ kernel_data.integrator.seed;
#ifdef __DEBUG_CORRELATION__
srand48(rng_hash + sample);
#else
(void)sample;
#endif
return rng_hash;
}
ccl_device_inline bool sample_is_even(int pattern, int sample)
{
if (pattern == SAMPLING_PATTERN_PMJ) {
/* See Section 10.2.1, "Progressive Multi-Jittered Sample Sequences", Christensen et al.
* We can use this to get divide sample sequence into two classes for easier variance
* estimation. */
return popcount(uint(sample) & 0xaaaaaaaa) & 1;
}
else {
/* TODO(Stefan): Are there reliable ways of dividing CMJ and Sobol into two classes? */
return sample & 0x1;
}
}
CCL_NAMESPACE_END
|