blender
/
blender
136
406
Fork 590
Code Issues 6.2k Pull Requests 764 Projects 19 Wiki Activity

Cycles: use more PMJ patterns and make their size adaptive. 

This resolves some issues with correlation artifacts at higher sample counts.

Fix T101356, correlation issues in new PMJ pattern.

Differential Revision: https://developer.blender.org/D16561
This commit is contained in:
Nathan Vegdahl 2022-11-21 18:16:27 +01:00 committed by Brecht Van Lommel
parent 41a3de878f
commit 03b5be4e3c
Notes: blender-bot 2023-02-13 14:24:32 +01:00 
Referenced by issue #101356, New PMJ does not converge to the same result as Sobol-Burley
4 changed files with 45 additions and 39 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
intern/cycles/kernel/data_template.h
View File

@ -183,6 +183,7 @@ KERNEL_STRUCT_MEMBER(integrator, int, use_lamp_mis)
KERNEL_STRUCT_MEMBER(integrator, int, use_caustics)
/* Sampling pattern. */
KERNEL_STRUCT_MEMBER(integrator, int, sampling_pattern)
KERNEL_STRUCT_MEMBER(integrator, int, pmj_sequence_size)
KERNEL_STRUCT_MEMBER(integrator, float, scrambling_distance)
/* Volume render. */
KERNEL_STRUCT_MEMBER(integrator, int, use_volumes)
@ -205,6 +206,11 @@ KERNEL_STRUCT_MEMBER(integrator, int, use_surface_guiding)
KERNEL_STRUCT_MEMBER(integrator, int, use_volume_guiding)
KERNEL_STRUCT_MEMBER(integrator, int, use_guiding_direct_light)
KERNEL_STRUCT_MEMBER(integrator, int, use_guiding_mis_weights)
/* Padding. */
KERNEL_STRUCT_MEMBER(integrator, int, pad1)
KERNEL_STRUCT_MEMBER(integrator, int, pad2)
KERNEL_STRUCT_MEMBER(integrator, int, pad3)
KERNEL_STRUCT_END(KernelIntegrator)
/* SVM. For shader specialization. */

55
intern/cycles/kernel/sample/jitter.h
View File

@ -7,6 +7,25 @@
#pragma once
CCL_NAMESPACE_BEGIN
ccl_device uint pmj_shuffled_sample_index(KernelGlobals kg, uint sample, uint dimension, uint seed)
{
const uint sample_count = kernel_data.integrator.pmj_sequence_size;
/* Shuffle the pattern order and sample index to better decorrelate
* dimensions and make the most of the finite patterns we have.
* The funky sample mask stuff is to ensure that we only shuffle
* *within* the current sample pattern, which is necessary to avoid
* early repeat pattern use. */
const uint pattern_i = hash_shuffle_uint(dimension, NUM_PMJ_PATTERNS, seed);
/* sample_count should always be a power of two, so this results in a mask. */
const uint sample_mask = sample_count - 1;
const uint sample_shuffled = nested_uniform_scramble(sample,
hash_wang_seeded_uint(dimension, seed));
sample = (sample & ~sample_mask) | (sample_shuffled & sample_mask);
return ((pattern_i * sample_count) + sample) % (sample_count * NUM_PMJ_PATTERNS);
}
ccl_device float pmj_sample_1D(KernelGlobals kg,
uint sample,
const uint rng_hash,
@ -20,22 +39,9 @@ ccl_device float pmj_sample_1D(KernelGlobals kg,
seed = kernel_data.integrator.seed;
}
/* Shuffle the pattern order and sample index to better decorrelate
* dimensions and make the most of the finite patterns we have.
* The funky sample mask stuff is to ensure that we only shuffle
* *within* the current sample pattern, which is necessary to avoid
* early repeat pattern use. */
const uint pattern_i = hash_shuffle_uint(dimension, NUM_PMJ_PATTERNS, seed);
/* NUM_PMJ_SAMPLES should be a power of two, so this results in a mask. */
const uint sample_mask = NUM_PMJ_SAMPLES - 1;
const uint sample_shuffled = nested_uniform_scramble(sample,
hash_wang_seeded_uint(dimension, seed));
sample = (sample & ~sample_mask) | (sample_shuffled & sample_mask);
/* Fetch the sample. */
const uint index = ((pattern_i * NUM_PMJ_SAMPLES) + sample) %
(NUM_PMJ_SAMPLES * NUM_PMJ_PATTERNS);
float x = kernel_data_fetch(sample_pattern_lut, index * 2);
const uint index = pmj_shuffled_sample_index(kg, sample, dimension, seed);
float x = kernel_data_fetch(sample_pattern_lut, index * NUM_PMJ_DIMENSIONS);
/* Do limited Cranley-Patterson rotation when using scrambling distance. */
if (kernel_data.integrator.scrambling_distance < 1.0f) {
@ -61,23 +67,10 @@ ccl_device float2 pmj_sample_2D(KernelGlobals kg,
seed = kernel_data.integrator.seed;
}
/* Shuffle the pattern order and sample index to better decorrelate
* dimensions and make the most of the finite patterns we have.
* The funky sample mask stuff is to ensure that we only shuffle
* *within* the current sample pattern, which is necessary to avoid
* early repeat pattern use. */
const uint pattern_i = hash_shuffle_uint(dimension, NUM_PMJ_PATTERNS, seed);
/* NUM_PMJ_SAMPLES should be a power of two, so this results in a mask. */
const uint sample_mask = NUM_PMJ_SAMPLES - 1;
const uint sample_shuffled = nested_uniform_scramble(sample,
hash_wang_seeded_uint(dimension, seed));
sample = (sample & ~sample_mask) | (sample_shuffled & sample_mask);
/* Fetch the sample. */
const uint index = ((pattern_i * NUM_PMJ_SAMPLES) + sample) %
(NUM_PMJ_SAMPLES * NUM_PMJ_PATTERNS);
float x = kernel_data_fetch(sample_pattern_lut, index * 2);
float y = kernel_data_fetch(sample_pattern_lut, index * 2 + 1);
const uint index = pmj_shuffled_sample_index(kg, sample, dimension, seed);
float x = kernel_data_fetch(sample_pattern_lut, index * NUM_PMJ_DIMENSIONS);
float y = kernel_data_fetch(sample_pattern_lut, index * NUM_PMJ_DIMENSIONS + 1);
/* Do limited Cranley-Patterson rotation when using scrambling distance. */
if (kernel_data.integrator.scrambling_distance < 1.0f) {

9
intern/cycles/kernel/types.h
View File

@ -1382,12 +1382,13 @@ static_assert_align(KernelShaderEvalInput, 16);
/* Pre-computed sample table sizes for PMJ02 sampler.
*
* NOTE: divisions *must* be a power of two, and patterns
* NOTE: min and max samples *must* be a power of two, and patterns
* ideally should be as well.
*/
#define NUM_PMJ_DIVISIONS 32
#define NUM_PMJ_SAMPLES ((NUM_PMJ_DIVISIONS) * (NUM_PMJ_DIVISIONS))
#define NUM_PMJ_PATTERNS 64
#define MIN_PMJ_SAMPLES 256
#define MAX_PMJ_SAMPLES 8192
#define NUM_PMJ_DIMENSIONS 2
#define NUM_PMJ_PATTERNS 256
/* Device kernels.
*

14
intern/cycles/scene/integrator.cpp
View File

@ -257,12 +257,18 @@ void Integrator::device_update(Device *device, DeviceScene *dscene, Scene *scene
kintegrator->light_inv_rr_threshold = 0.0f;
}
constexpr int num_sequences = NUM_PMJ_PATTERNS;
int sequence_size = clamp(next_power_of_two(aa_samples - 1), MIN_PMJ_SAMPLES, MAX_PMJ_SAMPLES);
if (kintegrator->sampling_pattern == SAMPLING_PATTERN_PMJ &&
dscene->sample_pattern_lut.size() == 0) {
constexpr int sequence_size = NUM_PMJ_SAMPLES;
constexpr int num_sequences = NUM_PMJ_PATTERNS;
dscene->sample_pattern_lut.size() !=
(sequence_size * NUM_PMJ_DIMENSIONS * NUM_PMJ_PATTERNS)) {
kintegrator->pmj_sequence_size = sequence_size;
if (dscene->sample_pattern_lut.size() != 0) {
dscene->sample_pattern_lut.free();
}
float2 *directions = (float2 *)dscene->sample_pattern_lut.alloc(sequence_size * num_sequences *
2);
NUM_PMJ_DIMENSIONS);
TaskPool pool;
for (int j = 0; j < num_sequences; ++j) {
float2 *sequence = directions + j * sequence_size;