griefith/test2
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Compositor: Redesign Sun Beams node for CPU

Browse Source 
This patch ports the new GPU implementation of the Sun Beams node to the
CPU compositor. Introduced in 9e358fcd44.
This commit is contained in:
Omar Emara
2024-01-10 19:36:52 +02:00
parent c08ba9b0bd
commit cbb738191e
2 changed files with 107 additions and 335 deletions
Download Patch File Download Diff File Expand all files Collapse all files

428
source/blender/compositor/operations/COM_SunBeamsOperation.cc
View File

@@ -2,6 +2,11 @@
*
* SPDX-License-Identifier: GPL-2.0-or-later */
#include "BLI_math_base.hh"
#include "BLI_math_vector.h"
#include "BLI_math_vector.hh"
#include "BLI_math_vector_types.hh"
#include "MEM_guardedalloc.h"
#include "COM_SunBeamsOperation.h"
@@ -17,331 +22,61 @@ SunBeamsOperation::SunBeamsOperation()
flags_.complex = true;
}
void SunBeamsOperation::calc_rays_common_data()
{
/* convert to pixels */
source_px_[0] = data_.source[0] * this->get_width();
source_px_[1] = data_.source[1] * this->get_height();
ray_length_px_ = data_.ray_length * std::max(this->get_width(), this->get_height());
}
void SunBeamsOperation::init_execution()
{
calc_rays_common_data();
input_program_ = this->get_input_socket_reader(0);
}
/**
* Defines a line accumulator for a specific sector,
* given by the four matrix entries that rotate from buffer space into the sector
*
* (x,y) is used to designate buffer space coordinates
* (u,v) is used to designate sector space coordinates
*
* For a target point (x,y) the sector should be chosen such that
* `u >= v >= 0`
* This removes the need to handle all sorts of special cases.
*
* Template parameters:
* \param fxu: buffer increment in x for sector `u + 1`.
* \param fxv: buffer increment in x for sector `v + 1`.
* \param fyu: buffer increment in y for sector `u + 1`.
* \param fyv: buffer increment in y for sector `v + 1`.
*/
template<int fxu, int fxv, int fyu, int fyv> struct BufferLineAccumulator {
void SunBeamsOperation::execute_pixel(float output[4], int x, int y, void * /* data */)
{
const float2 input_size = float2(input_program_->get_width(), input_program_->get_height());
const int max_steps = int(data_.ray_length * math::length(input_size));
const float2 source = float2(data_.source);
/* utility functions implementing the matrix transform to/from sector space */
const float2 texel = float2(x, y);
static inline void buffer_to_sector(const float source[2], float x, float y, float &u, float &v)
{
int x0 = int(source[0]);
int y0 = int(source[1]);
x -= float(x0);
y -= float(y0);
u = x * fxu + y * fyu;
v = x * fxv + y * fyv;
/* The number of steps is the distance in pixels from the source to the current texel. With
* at least a single step and at most the user specified maximum ray length, which is
* proportional to the diagonal pixel count. */
const float unbounded_steps = math::max(1.0f, math::distance(texel, source * input_size));
const int steps = math::min(max_steps, int(unbounded_steps));
/* We integrate from the current pixel to the source pixel, so compute the start coordinates
* and step vector in the direction to source. Notice that the step vector is still computed
* from the unbounded steps, such that the total integration length becomes limited by the
* bounded steps, and thus by the maximum ray length. */
const float2 coordinates = (texel + float2(0.5f)) / input_size;
const float2 vector_to_source = source - coordinates;
const float2 step_vector = vector_to_source / unbounded_steps;
float accumulated_weight = 0.0f;
float4 accumulated_color = float4(0.0f);
for (int i = 0; i <= steps; i++) {
float2 position = coordinates + i * step_vector;
/* We are already past the image boundaries, and any future steps are also past the image
* boundaries, so break. */
if (position.x < 0.0f || position.y < 0.0f || position.x > 1.0f || position.y > 1.0f) {
break;
}
const float2 coordinates = position * input_size;
float4 sample_color;
input_program_->read_sampled(
sample_color, coordinates.x, coordinates.y, PixelSampler::Bilinear);
/* Attenuate the contributions of pixels that are further away from the source using a
* quadratic falloff. Also weight by the alpha to give more significance to opaque pixels.
*/
const float weight = (math::square(1.0f - i / float(steps))) * sample_color.w;
accumulated_weight += weight;
accumulated_color += sample_color * weight;
}
static inline void sector_to_buffer(const float source[2], int u, int v, int &x, int &y)
{
int x0 = int(source[0]);
int y0 = int(source[1]);
x = x0 + u * fxu + v * fxv;
y = y0 + u * fyu + v * fyv;
}
/**
* Set up the initial buffer pointer and calculate necessary variables for looping.
*
* Note that sector space is centered around the "source" point while the loop starts
* at dist_min from the target pt. This way the loop can be canceled as soon as it runs
* out of the buffer rect, because no pixels further along the line can contribute.
*
* \param x, y: Start location in the buffer
* \param num: Total steps in the loop
* \param v, dv: Vertical offset in sector space, for line offset perpendicular to the loop axis
*/
static float *init_buffer_iterator(MemoryBuffer *input,
const float source[2],
const float co[2],
float dist_min,
float dist_max,
int &x,
int &y,
int &num,
float &v,
float &dv,
float &falloff_factor)
{
float pu, pv;
buffer_to_sector(source, co[0], co[1], pu, pv);
/* line angle */
double tan_phi = pv / double(pu);
double dr = sqrt(tan_phi * tan_phi + 1.0);
double cos_phi = 1.0 / dr;
/* clamp u range to avoid influence of pixels "behind" the source */
float umin = max_ff(pu - cos_phi * dist_min, 0.0f);
float umax = max_ff(pu - cos_phi * dist_max, 0.0f);
v = umin * tan_phi;
dv = tan_phi;
int start = int(floorf(umax));
int end = int(ceilf(umin));
num = end - start;
sector_to_buffer(source, end, int(ceilf(v)), x, y);
falloff_factor = dist_max > dist_min ? dr / double(dist_max - dist_min) : 0.0f;
float *iter = input->get_buffer() + input->get_coords_offset(x, y);
return iter;
}
/**
* Perform the actual accumulation along a ray segment from source to pt.
* Only pixels within dist_min..dist_max contribute.
*
* The loop runs backwards(!) over the primary sector space axis u, i.e. increasing distance to
* pt. After each step it decrements v by dv < 1, adding a buffer shift when necessary.
*/
static void eval(MemoryBuffer *input,
float output[4],
const float co[2],
const float source[2],
float dist_min,
float dist_max)
{
const rcti &rect = input->get_rect();
int x, y, num;
float v, dv;
float falloff_factor;
float border[4];
zero_v4(output);
if (int(co[0] - source[0]) == 0 && int(co[1] - source[1]) == 0) {
copy_v4_v4(output, input->get_elem(source[0], source[1]));
return;
}
/* Initialize the iteration variables. */
float *buffer = init_buffer_iterator(
input, source, co, dist_min, dist_max, x, y, num, v, dv, falloff_factor);
zero_v3(border);
border[3] = 1.0f;
/* v_local keeps track of when to decrement v (see below) */
float v_local = v - floorf(v);
for (int i = 0; i < num; i++) {
float weight = 1.0f - float(i) * falloff_factor;
weight *= weight;
/* range check, use last valid color when running beyond the image border */
if (x >= rect.xmin && x < rect.xmax && y >= rect.ymin && y < rect.ymax) {
madd_v4_v4fl(output, buffer, buffer[3] * weight);
/* use as border color in case subsequent pixels are out of bounds */
copy_v4_v4(border, buffer);
}
else {
madd_v4_v4fl(output, border, border[3] * weight);
}
/* TODO: implement proper filtering here, see
* https://en.wikipedia.org/wiki/Lanczos_resampling
* https://en.wikipedia.org/wiki/Sinc_function
*
* using lanczos with x = distance from the line segment,
* normalized to a == 0.5f, could give a good result
*
* for now just divide equally at the end ...
*/
/* decrement u */
x -= fxu;
y -= fyu;
buffer -= fxu * input->elem_stride + fyu * input->row_stride;
/* decrement v (in steps of dv < 1) */
v_local -= dv;
if (v_local < 0.0f) {
v_local += 1.0f;
x -= fxv;
y -= fyv;
buffer -= fxv * input->elem_stride + fyv * input->row_stride;
}
}
/* normalize */
if (num > 0) {
mul_v4_fl(output, 1.0f / float(num));
}
}
};
/**
* Dispatch function which selects an appropriate accumulator based on the sector of the target
* point, relative to the source.
*
* The BufferLineAccumulator defines the actual loop over the buffer, with an efficient inner loop
* due to using compile time constants instead of a local matrix variable defining the sector
* space.
*/
static void accumulate_line(MemoryBuffer *input,
float output[4],
const float co[2],
const float source[2],
float dist_min,
float dist_max)
{
/* coordinates relative to source */
float pt_ofs[2] = {co[0] - source[0], co[1] - source[1]};
/* The source sectors are defined like so:
*
* \ 3 | 2 /
* \ | /
* 4 \ | / 1
* \|/
* -----------
* /|\
* 5 / | \ 8
* / | \
* / 6 | 7 \
*
* The template arguments encode the transformation into "sector space",
* by means of rotation/mirroring matrix elements.
*/
if (fabsf(pt_ofs[1]) > fabsf(pt_ofs[0])) {
if (pt_ofs[0] > 0.0f) {
if (pt_ofs[1] > 0.0f) {
/* 2 */
BufferLineAccumulator<0, 1, 1, 0>::eval(input, output, co, source, dist_min, dist_max);
}
else {
/* 7 */
BufferLineAccumulator<0, 1, -1, 0>::eval(input, output, co, source, dist_min, dist_max);
}
}
else {
if (pt_ofs[1] > 0.0f) {
/* 3 */
BufferLineAccumulator<0, -1, 1, 0>::eval(input, output, co, source, dist_min, dist_max);
}
else {
/* 6 */
BufferLineAccumulator<0, -1, -1, 0>::eval(input, output, co, source, dist_min, dist_max);
}
}
}
else {
if (pt_ofs[0] > 0.0f) {
if (pt_ofs[1] > 0.0f) {
/* 1 */
BufferLineAccumulator<1, 0, 0, 1>::eval(input, output, co, source, dist_min, dist_max);
}
else {
/* 8 */
BufferLineAccumulator<1, 0, 0, -1>::eval(input, output, co, source, dist_min, dist_max);
}
}
else {
if (pt_ofs[1] > 0.0f) {
/* 4 */
BufferLineAccumulator<-1, 0, 0, 1>::eval(input, output, co, source, dist_min, dist_max);
}
else {
/* 5 */
BufferLineAccumulator<-1, 0, 0, -1>::eval(input, output, co, source, dist_min, dist_max);
}
}
}
}
void *SunBeamsOperation::initialize_tile_data(rcti * /*rect*/)
{
void *buffer = get_input_operation(0)->initialize_tile_data(nullptr);
return buffer;
}
void SunBeamsOperation::execute_pixel(float output[4], int x, int y, void *data)
{
const float co[2] = {float(x), float(y)};
accumulate_line((MemoryBuffer *)data, output, co, source_px_, 0.0f, ray_length_px_);
}
static void calc_ray_shift(rcti *rect, float x, float y, const float source[2], float ray_length)
{
float co[2] = {float(x), float(y)};
float dir[2], dist;
/* move (x,y) vector toward the source by ray_length distance */
sub_v2_v2v2(dir, co, source);
dist = normalize_v2(dir);
mul_v2_fl(dir, min_ff(dist, ray_length));
sub_v2_v2(co, dir);
int ico[2] = {int(co[0]), int(co[1])};
BLI_rcti_do_minmax_v(rect, ico);
}
bool SunBeamsOperation::determine_depending_area_of_interest(rcti *input,
ReadBufferOperation *read_operation,
rcti *output)
{
/* Enlarges the rect by moving each corner toward the source.
* This is the maximum distance that pixels can influence each other
* and gives a rect that contains all possible accumulated pixels.
*/
rcti rect = *input;
calc_ray_shift(&rect, input->xmin, input->ymin, source_px_, ray_length_px_);
calc_ray_shift(&rect, input->xmin, input->ymax, source_px_, ray_length_px_);
calc_ray_shift(&rect, input->xmax, input->ymin, source_px_, ray_length_px_);
calc_ray_shift(&rect, input->xmax, input->ymax, source_px_, ray_length_px_);
return NodeOperation::determine_depending_area_of_interest(&rect, read_operation, output);
}
void SunBeamsOperation::get_area_of_interest(const int input_idx,
const rcti &output_area,
rcti &r_input_area)
{
BLI_assert(input_idx == 0);
UNUSED_VARS(input_idx);
calc_rays_common_data();
r_input_area = output_area;
/* Enlarges the rect by moving each corner toward the source.
* This is the maximum distance that pixels can influence each other
* and gives a rect that contains all possible accumulated pixels. */
calc_ray_shift(&r_input_area, output_area.xmin, output_area.ymin, source_px_, ray_length_px_);
calc_ray_shift(&r_input_area, output_area.xmin, output_area.ymax, source_px_, ray_length_px_);
calc_ray_shift(&r_input_area, output_area.xmax, output_area.ymin, source_px_, ray_length_px_);
calc_ray_shift(&r_input_area, output_area.xmax, output_area.ymax, source_px_, ray_length_px_);
accumulated_color /= accumulated_weight != 0.0f ? accumulated_weight : 1.0f;
copy_v4_v4(output, accumulated_color);
}
void SunBeamsOperation::update_memory_buffer_partial(MemoryBuffer *output,
@@ -349,16 +84,63 @@ void SunBeamsOperation::update_memory_buffer_partial(MemoryBuffer *output,
Span<MemoryBuffer *> inputs)
{
MemoryBuffer *input = inputs[0];
float coords[2];
const float2 input_size = float2(input->get_width(), input->get_height());
const int max_steps = int(data_.ray_length * math::length(input_size));
const float2 source = float2(data_.source);
for (int y = area.ymin; y < area.ymax; y++) {
coords[1] = y;
float *out_elem = output->get_elem(area.xmin, y);
for (int x = area.xmin; x < area.xmax; x++) {
coords[0] = x;
accumulate_line(input, out_elem, coords, source_px_, 0.0f, ray_length_px_);
out_elem += output->elem_stride;
const float2 texel = float2(x, y);
/* The number of steps is the distance in pixels from the source to the current texel. With
* at least a single step and at most the user specified maximum ray length, which is
* proportional to the diagonal pixel count. */
const float unbounded_steps = math::max(1.0f, math::distance(texel, source * input_size));
const int steps = math::min(max_steps, int(unbounded_steps));
/* We integrate from the current pixel to the source pixel, so compute the start coordinates
* and step vector in the direction to source. Notice that the step vector is still computed
* from the unbounded steps, such that the total integration length becomes limited by the
* bounded steps, and thus by the maximum ray length. */
const float2 coordinates = (texel + float2(0.5f)) / input_size;
const float2 vector_to_source = source - coordinates;
const float2 step_vector = vector_to_source / unbounded_steps;
float accumulated_weight = 0.0f;
float4 accumulated_color = float4(0.0f);
for (int i = 0; i <= steps; i++) {
float2 position = coordinates + i * step_vector;
/* We are already past the image boundaries, and any future steps are also past the image
* boundaries, so break. */
if (position.x < 0.0f || position.y < 0.0f || position.x > 1.0f || position.y > 1.0f) {
break;
}
const float2 coordinates = position * input_size;
float4 sample_color;
input->read_elem_bilinear(coordinates.x, coordinates.y, sample_color);
/* Attenuate the contributions of pixels that are further away from the source using a
* quadratic falloff. Also weight by the alpha to give more significance to opaque pixels.
*/
const float weight = (math::square(1.0f - i / float(steps))) * sample_color.w;
accumulated_weight += weight;
accumulated_color += sample_color * weight;
}
accumulated_color /= accumulated_weight != 0.0f ? accumulated_weight : 1.0f;
copy_v4_v4(output->get_elem(x, y), accumulated_color);
}
}
}
void SunBeamsOperation::deinit_execution()
{
input_program_ = nullptr;
}
} // namespace blender::compositor

14
source/blender/compositor/operations/COM_SunBeamsOperation.h
View File

@@ -16,12 +16,6 @@ class SunBeamsOperation : public MultiThreadedOperation {
void init_execution() override;
void *initialize_tile_data(rcti *rect) override;
bool determine_depending_area_of_interest(rcti *input,
ReadBufferOperation *read_operation,
rcti *output) override;
void set_data(const NodeSunBeams &data)
{
data_ = data;
@@ -30,16 +24,12 @@ class SunBeamsOperation : public MultiThreadedOperation {
void update_memory_buffer_partial(MemoryBuffer *output,
const rcti &area,
Span<MemoryBuffer *> inputs) override;
void get_area_of_interest(int input_idx, const rcti &output_area, rcti &r_input_area) override;
private:
void calc_rays_common_data();
void deinit_execution() override;
private:
NodeSunBeams data_;
float source_px_[2];
float ray_length_px_;
SocketReader *input_program_;
};
} // namespace blender::compositor