LIVE / DiffVG /scene.cpp
Xu Ma
update
be11144
raw
history blame
51.5 kB
#include "scene.h"
#include "aabb.h"
#include "cuda_utils.h"
#include "filter.h"
#include "shape.h"
#include <numeric>
#include <algorithm>
#include <cstring>
#include <chrono>
#include <cstddef>
size_t align(size_t s) {
auto a = alignof(std::max_align_t);
return ((s + a - 1) / a) * a;
}
template <typename T>
void allocate(bool use_gpu, T **p) {
if (use_gpu) {
#ifdef __NVCC__
checkCuda(cudaMallocManaged(p, sizeof(T)));
#else
throw std::runtime_error("diffvg not compiled with GPU");
assert(false);
#endif
} else {
*p = (T*)malloc(sizeof(T));
}
}
template <typename T>
void allocate(bool use_gpu, size_t size, T **p) {
if (use_gpu) {
#ifdef __NVCC__
checkCuda(cudaMallocManaged(p, size * sizeof(T)));
#else
throw std::runtime_error("diffvg not compiled with GPU");
assert(false);
#endif
} else {
*p = (T*)malloc(size * sizeof(T));
}
}
void copy_and_init_shapes(Scene &scene,
const std::vector<const Shape *> &shape_list) {
for (int shape_id = 0; shape_id < scene.num_shapes; shape_id++) {
switch (shape_list[shape_id]->type) {
case ShapeType::Circle: {
Circle *p = (Circle *)scene.shapes[shape_id].ptr;
const Circle *p_ = (const Circle*)(shape_list[shape_id]->ptr);
*p = *p_;
Circle *d_p = (Circle *)scene.d_shapes[shape_id].ptr;
d_p->radius = 0;
d_p->center = Vector2f{0, 0};
break;
} case ShapeType::Ellipse: {
Ellipse *p = (Ellipse *)scene.shapes[shape_id].ptr;
const Ellipse *p_ = (const Ellipse*)(shape_list[shape_id]->ptr);
*p = *p_;
Ellipse *d_p = (Ellipse *)scene.d_shapes[shape_id].ptr;
d_p->radius = Vector2f{0, 0};
d_p->center = Vector2f{0, 0};
break;
} case ShapeType::Path: {
Path *p = (Path *)scene.shapes[shape_id].ptr;
const Path *p_ = (const Path*)(shape_list[shape_id]->ptr);
p->num_points = p_->num_points;
p->num_base_points = p_->num_base_points;
for (int i = 0; i < p_->num_base_points; i++) {
p->num_control_points[i] = p_->num_control_points[i];
}
for (int i = 0; i < 2 * p_->num_points; i++) {
p->points[i] = p_->points[i];
}
p->is_closed = p_->is_closed;
p->use_distance_approx = p_->use_distance_approx;
Path *d_p = (Path *)scene.d_shapes[shape_id].ptr;
d_p->num_points = p_->num_points;
d_p->num_base_points = p_->num_base_points;
for (int i = 0; i < 2 * p_->num_points; i++) {
d_p->points[i] = 0;
}
d_p->is_closed = p_->is_closed;
if (p_->thickness != nullptr) {
for (int i = 0; i < p_->num_points; i++) {
p->thickness[i] = p_->thickness[i];
d_p->thickness[i] = 0;
}
}
d_p->use_distance_approx = p_->use_distance_approx;
break;
} case ShapeType::Rect: {
Rect *p = (Rect *)scene.shapes[shape_id].ptr;
const Rect *p_ = (const Rect*)(shape_list[shape_id]->ptr);
*p = *p_;
Rect *d_p = (Rect *)scene.d_shapes[shape_id].ptr;
d_p->p_min = Vector2f{0, 0};
d_p->p_max = Vector2f{0, 0};
break;
} default: {
assert(false);
break;
}
}
scene.shapes[shape_id].type = shape_list[shape_id]->type;
scene.shapes[shape_id].stroke_width = shape_list[shape_id]->stroke_width;
scene.d_shapes[shape_id].type = shape_list[shape_id]->type;
scene.d_shapes[shape_id].stroke_width = 0;
}
}
std::vector<float>
compute_shape_length(const std::vector<const Shape *> &shape_list) {
int num_shapes = (int)shape_list.size();
std::vector<float> shape_length_list(num_shapes, 0.f);
for (int shape_id = 0; shape_id < num_shapes; shape_id++) {
auto shape_length = 0.f;
switch (shape_list[shape_id]->type) {
case ShapeType::Circle: {
const Circle *p_ = (const Circle*)(shape_list[shape_id]->ptr);
shape_length += float(2.f * M_PI) * p_->radius;
break;
} case ShapeType::Ellipse: {
const Ellipse *p_ = (const Ellipse*)(shape_list[shape_id]->ptr);
// https://en.wikipedia.org/wiki/Ellipse#Circumference
// Ramanujan's ellipse circumference approximation
auto a = p_->radius.x;
auto b = p_->radius.y;
shape_length += float(M_PI) * (3 * (a + b) - sqrt((3 * a + b) * (a + 3 * b)));
break;
} case ShapeType::Path: {
const Path *p_ = (const Path*)(shape_list[shape_id]->ptr);
auto length = 0.f;
auto point_id = 0;
for (int i = 0; i < p_->num_base_points; i++) {
if (p_->num_control_points[i] == 0) {
// Straight line
auto i0 = point_id;
assert(i0 < p_->num_points);
auto i1 = (i0 + 1) % p_->num_points;
point_id += 1;
auto p0 = Vector2f{p_->points[2 * i0], p_->points[2 * i0 + 1]};
auto p1 = Vector2f{p_->points[2 * i1], p_->points[2 * i1 + 1]};
length += distance(p1, p0);
} else if (p_->num_control_points[i] == 1) {
// Quadratic Bezier curve
auto i0 = point_id;
auto i1 = i0 + 1;
auto i2 = (i0 + 2) % p_->num_points;
point_id += 2;
auto p0 = Vector2f{p_->points[2 * i0], p_->points[2 * i0 + 1]};
auto p1 = Vector2f{p_->points[2 * i1], p_->points[2 * i1 + 1]};
auto p2 = Vector2f{p_->points[2 * i2], p_->points[2 * i2 + 1]};
auto eval = [&](float t) -> Vector2f {
auto tt = 1 - t;
return (tt*tt)*p0 + (2*tt*t)*p1 + (t*t)*p2;
};
// We use 3-point samples to approximate the length
auto v0 = p0;
auto v1 = eval(0.5f);
auto v2 = p2;
length += distance(v1, v0) + distance(v1, v2);
} else if (p_->num_control_points[i] == 2) {
// Cubic Bezier curve
auto i0 = point_id;
auto i1 = i0 + 1;
auto i2 = i0 + 2;
auto i3 = (i0 + 3) % p_->num_points;
point_id += 3;
auto p0 = Vector2f{p_->points[2 * i0], p_->points[2 * i0 + 1]};
auto p1 = Vector2f{p_->points[2 * i1], p_->points[2 * i1 + 1]};
auto p2 = Vector2f{p_->points[2 * i2], p_->points[2 * i2 + 1]};
auto p3 = Vector2f{p_->points[2 * i3], p_->points[2 * i3 + 1]};
auto eval = [&](float t) -> Vector2f {
auto tt = 1 - t;
return (tt*tt*tt)*p0 + (3*tt*tt*t)*p1 + (3*tt*t*t)*p2 + (t*t*t)*p3;
};
// We use 4-point samples to approximate the length
auto v0 = p0;
auto v1 = eval(1.f/3.f);
auto v2 = eval(2.f/3.f);
auto v3 = p3;
length += distance(v1, v0) + distance(v1, v2) + distance(v2, v3);
} else {
assert(false);
}
}
assert(isfinite(length));
shape_length += length;
break;
} case ShapeType::Rect: {
const Rect *p_ = (const Rect*)(shape_list[shape_id]->ptr);
shape_length += 2 * (p_->p_max.x - p_->p_min.x + p_->p_max.y - p_->p_min.y);
break;
} default: {
assert(false);
break;
}
}
assert(isfinite(shape_length));
shape_length_list[shape_id] = shape_length;
}
return shape_length_list;
}
void build_shape_cdfs(Scene &scene,
const std::vector<const ShapeGroup *> &shape_group_list,
const std::vector<float> &shape_length_list) {
int sample_id = 0;
for (int shape_group_id = 0; shape_group_id < (int)shape_group_list.size(); shape_group_id++) {
const ShapeGroup *shape_group = shape_group_list[shape_group_id];
for (int i = 0; i < shape_group->num_shapes; i++) {
int shape_id = shape_group->shape_ids[i];
float length = shape_length_list[shape_id];
scene.sample_shape_id[sample_id] = shape_id;
if (sample_id == 0) {
scene.sample_shapes_cdf[sample_id] = length;
} else {
scene.sample_shapes_cdf[sample_id] = length +
scene.sample_shapes_cdf[sample_id - 1];
}
assert(isfinite(length));
scene.sample_shapes_pmf[sample_id] = length;
scene.sample_group_id[sample_id] = shape_group_id;
sample_id++;
}
}
assert(sample_id == scene.num_total_shapes);
auto normalization = scene.sample_shapes_cdf[scene.num_total_shapes - 1];
if (normalization <= 0) {
char buf[256];
sprintf(buf, "The total length of the shape boundaries in the scene is equal or less than 0. Length = %f", normalization);
throw std::runtime_error(buf);
}
if (!isfinite(normalization)) {
char buf[256];
sprintf(buf, "The total length of the shape boundaries in the scene is not a number. Length = %f", normalization);
throw std::runtime_error(buf);
}
assert(normalization > 0);
for (int sample_id = 0; sample_id < scene.num_total_shapes; sample_id++) {
scene.sample_shapes_cdf[sample_id] /= normalization;
scene.sample_shapes_pmf[sample_id] /= normalization;
}
}
void build_path_cdfs(Scene &scene,
const std::vector<const Shape *> &shape_list,
const std::vector<float> &shape_length_list) {
for (int shape_id = 0; shape_id < scene.num_shapes; shape_id++) {
if (shape_list[shape_id]->type == ShapeType::Path) {
const Path &path = shape_list[shape_id]->as_path();
float *pmf = scene.path_length_pmf[shape_id];
float *cdf = scene.path_length_cdf[shape_id];
int *point_id_map = scene.path_point_id_map[shape_id];
auto path_length = shape_length_list[shape_id];
auto inv_length = 1.f / path_length;
auto point_id = 0;
for (int i = 0; i < path.num_base_points; i++) {
point_id_map[i] = point_id;
if (path.num_control_points[i] == 0) {
// Straight line
auto i0 = point_id;
auto i1 = (i0 + 1) % path.num_points;
point_id += 1;
auto p0 = Vector2f{path.points[2 * i0], path.points[2 * i0 + 1]};
auto p1 = Vector2f{path.points[2 * i1], path.points[2 * i1 + 1]};
auto d = distance(p0, p1) * inv_length;
pmf[i] = d;
if (i == 0) {
cdf[i] = d;
} else {
cdf[i] = d + cdf[i - 1];
}
} else if (path.num_control_points[i] == 1) {
// Quadratic Bezier curve
auto i0 = point_id;
auto i1 = i0 + 1;
auto i2 = (i0 + 2) % path.num_points;
point_id += 2;
auto p0 = Vector2f{path.points[2 * i0], path.points[2 * i0 + 1]};
auto p1 = Vector2f{path.points[2 * i1], path.points[2 * i1 + 1]};
auto p2 = Vector2f{path.points[2 * i2], path.points[2 * i2 + 1]};
auto eval = [&](float t) -> Vector2f {
auto tt = 1 - t;
return (tt*tt)*p0 + (2*tt*t)*p1 + (t*t)*p2;
};
// We use 3-point samples to approximate the length
auto v0 = p0;
auto v1 = eval(0.5f);
auto v2 = p2;
auto d = (distance(v0, v1) + distance(v1, v2)) * inv_length;
pmf[i] = d;
if (i == 0) {
cdf[i] = d;
} else {
cdf[i] = d + cdf[i - 1];
}
} else if (path.num_control_points[i] == 2) {
// Cubic Bezier curve
auto i0 = point_id;
auto i1 = point_id + 1;
auto i2 = point_id + 2;
auto i3 = (point_id + 3) % path.num_points;
point_id += 3;
auto p0 = Vector2f{path.points[2 * i0], path.points[2 * i0 + 1]};
auto p1 = Vector2f{path.points[2 * i1], path.points[2 * i1 + 1]};
auto p2 = Vector2f{path.points[2 * i2], path.points[2 * i2 + 1]};
auto p3 = Vector2f{path.points[2 * i3], path.points[2 * i3 + 1]};
auto eval = [&](float t) -> Vector2f {
auto tt = 1 - t;
return (tt*tt*tt)*p0 + (3*tt*tt*t)*p1 + (3*tt*t*t)*p2 + (t*t*t)*p3;
};
// We use 4-point samples to approximate the length
auto v0 = p0;
auto v1 = eval(1.f/3.f);
auto v2 = eval(2.f/3.f);
auto v3 = p3;
auto d = (distance(v1, v0) + distance(v1, v2) + distance(v2, v3)) * inv_length;
pmf[i] = d;
if (i == 0) {
cdf[i] = d;
} else {
cdf[i] = d + cdf[i - 1];
}
} else {
assert(false);
}
}
}
}
}
void copy_and_init_shape_groups(Scene &scene,
const std::vector<const ShapeGroup *> &shape_group_list) {
for (int group_id = 0; group_id < scene.num_shape_groups; group_id++) {
const ShapeGroup *shape_group = shape_group_list[group_id];
auto copy_and_init_color = [&](const ColorType &color_type, void *color_ptr, void *target_ptr, void *d_target_ptr) {
switch (color_type) {
case ColorType::Constant: {
Constant *c = (Constant*)target_ptr;
Constant *d_c = (Constant*)d_target_ptr;
const Constant *c_ = (const Constant*)color_ptr;
*c = *c_;
d_c->color = Vector4{0, 0, 0, 0};
break;
} case ColorType::LinearGradient: {
LinearGradient *c = (LinearGradient*)target_ptr;
LinearGradient *d_c = (LinearGradient*)d_target_ptr;
const LinearGradient *c_ = (const LinearGradient*)color_ptr;
c->begin = c_->begin;
c->end = c_->end;
c->num_stops = c_->num_stops;
for (int i = 0; i < c_->num_stops; i++) {
c->stop_offsets[i] = c_->stop_offsets[i];
}
for (int i = 0; i < 4 * c_->num_stops; i++) {
c->stop_colors[i] = c_->stop_colors[i];
}
d_c->begin = Vector2f{0, 0};
d_c->end = Vector2f{0, 0};
d_c->num_stops = c_->num_stops;
for (int i = 0; i < c_->num_stops; i++) {
d_c->stop_offsets[i] = 0;
}
for (int i = 0; i < 4 * c_->num_stops; i++) {
d_c->stop_colors[i] = 0;
}
break;
} case ColorType::RadialGradient: {
RadialGradient *c = (RadialGradient*)target_ptr;
RadialGradient *d_c = (RadialGradient*)d_target_ptr;
const RadialGradient *c_ = (const RadialGradient*)color_ptr;
c->center = c_->center;
c->radius = c_->radius;
c->num_stops = c_->num_stops;
for (int i = 0; i < c_->num_stops; i++) {
c->stop_offsets[i] = c_->stop_offsets[i];
}
for (int i = 0; i < 4 * c_->num_stops; i++) {
c->stop_colors[i] = c_->stop_colors[i];
}
d_c->center = Vector2f{0, 0};
d_c->radius = Vector2f{0, 0};
d_c->num_stops = c_->num_stops;
for (int i = 0; i < c_->num_stops; i++) {
d_c->stop_offsets[i] = 0;
}
for (int i = 0; i < 4 * c_->num_stops; i++) {
d_c->stop_colors[i] = 0;
}
break;
} default: {
assert(false);
}
}
};
for (int i = 0; i < shape_group->num_shapes; i++) {
scene.shape_groups[group_id].shape_ids[i] = shape_group->shape_ids[i];
}
scene.shape_groups[group_id].num_shapes = shape_group->num_shapes;
scene.shape_groups[group_id].use_even_odd_rule = shape_group->use_even_odd_rule;
scene.shape_groups[group_id].canvas_to_shape = shape_group->canvas_to_shape;
scene.shape_groups[group_id].shape_to_canvas = shape_group->shape_to_canvas;
scene.d_shape_groups[group_id].shape_ids = nullptr;
scene.d_shape_groups[group_id].num_shapes = shape_group->num_shapes;
scene.d_shape_groups[group_id].use_even_odd_rule = shape_group->use_even_odd_rule;
scene.d_shape_groups[group_id].canvas_to_shape = Matrix3x3f{};
scene.d_shape_groups[group_id].shape_to_canvas = Matrix3x3f{};
scene.shape_groups[group_id].fill_color_type = shape_group->fill_color_type;
scene.d_shape_groups[group_id].fill_color_type = shape_group->fill_color_type;
if (shape_group->fill_color != nullptr) {
copy_and_init_color(shape_group->fill_color_type,
shape_group->fill_color,
scene.shape_groups[group_id].fill_color,
scene.d_shape_groups[group_id].fill_color);
}
scene.shape_groups[group_id].stroke_color_type = shape_group->stroke_color_type;
scene.d_shape_groups[group_id].stroke_color_type = shape_group->stroke_color_type;
if (shape_group->stroke_color != nullptr) {
copy_and_init_color(shape_group->stroke_color_type,
shape_group->stroke_color,
scene.shape_groups[group_id].stroke_color,
scene.d_shape_groups[group_id].stroke_color);
}
}
}
DEVICE uint32_t morton2D(const Vector2f &p, int canvas_width, int canvas_height) {
auto scene_bounds = Vector2f{canvas_width, canvas_height};
auto pp = p / scene_bounds;
TVector2<uint32_t> pp_i{pp.x * 1023, pp.y * 1023};
return (expand_bits(pp_i.x) << 1u) |
(expand_bits(pp_i.y) << 0u);
}
template <bool sort>
void build_bvh(const Scene &scene, BVHNode *nodes, int num_primitives) {
auto bvh_size = 2 * num_primitives - 1;
if (bvh_size > 1) {
if (sort) {
// Sort by Morton code
std::sort(nodes, nodes + num_primitives,
[&] (const BVHNode &n0, const BVHNode &n1) {
auto p0 = 0.5f * (n0.box.p_min + n0.box.p_max);
auto p1 = 0.5f * (n1.box.p_min + n1.box.p_max);
auto m0 = morton2D(p0, scene.canvas_width, scene.canvas_height);
auto m1 = morton2D(p1, scene.canvas_width, scene.canvas_height);
return m0 < m1;
});
}
for (int i = num_primitives; i < bvh_size; i++) {
nodes[i] = BVHNode{-1, -1, AABB{}, 0.f};
}
int prev_beg = 0;
int prev_end = num_primitives;
// For handling odd number of nodes at a level
int leftover = prev_end % 2 == 0 ? -1 : prev_end - 1;
while (prev_end - prev_beg >= 1 || leftover != -1) {
int length = (prev_end - prev_beg) / 2;
if ((prev_end - prev_beg) % 2 == 1 && leftover != -1 &&
leftover != prev_end - 1) {
length += 1;
}
for (int i = 0; i < length; i++) {
BVHNode node;
node.child0 = prev_beg + 2 * i;
node.child1 = prev_beg + 2 * i + 1;
if (node.child1 >= prev_end) {
assert(leftover != -1);
node.child1 = leftover;
leftover = -1;
}
AABB child0_box = nodes[node.child0].box;
AABB child1_box = nodes[node.child1].box;
node.box = merge(child0_box, child1_box);
node.max_radius = std::max(nodes[node.child0].max_radius,
nodes[node.child1].max_radius);
nodes[prev_end + i] = node;
}
if (length == 1 && leftover == -1) {
break;
}
prev_beg = prev_end;
prev_end = prev_beg + length;
if (length % 2 == 1 && leftover == -1) {
leftover = prev_end - 1;
}
}
}
assert(nodes[2 * num_primitives - 2].child0 != -1);
}
void compute_bounding_boxes(Scene &scene,
const std::vector<const Shape *> &shape_list,
const std::vector<const ShapeGroup *> &shape_group_list) {
for (int shape_id = 0; shape_id < scene.num_shapes; shape_id++) {
switch (shape_list[shape_id]->type) {
case ShapeType::Circle: {
const Circle *p = (const Circle*)(shape_list[shape_id]->ptr);
scene.shapes_bbox[shape_id] = AABB{p->center - p->radius,
p->center + p->radius};
break;
} case ShapeType::Ellipse: {
const Ellipse *p = (const Ellipse*)(shape_list[shape_id]->ptr);
scene.shapes_bbox[shape_id] = AABB{p->center - p->radius,
p->center + p->radius};
break;
} case ShapeType::Path: {
const Path *p = (const Path*)(shape_list[shape_id]->ptr);
AABB box;
if (p->num_points > 0) {
box = AABB{Vector2f{p->points[0], p->points[1]},
Vector2f{p->points[0], p->points[1]}};
}
for (int i = 1; i < p->num_points; i++) {
box = merge(box, Vector2f{p->points[2 * i], p->points[2 * i + 1]});
}
scene.shapes_bbox[shape_id] = box;
std::vector<AABB> boxes(p->num_base_points);
std::vector<float> thickness(p->num_base_points);
std::vector<int> first_point_id(p->num_base_points);
auto r = shape_list[shape_id]->stroke_width;
auto point_id = 0;
for (int i = 0; i < p->num_base_points; i++) {
first_point_id[i] = point_id;
if (p->num_control_points[i] == 0) {
// Straight line
auto i0 = point_id;
auto i1 = (i0 + 1) % p->num_points;
point_id += 1;
auto p0 = Vector2f{p->points[2 * i0], p->points[2 * i0 + 1]};
auto p1 = Vector2f{p->points[2 * i1], p->points[2 * i1 + 1]};
boxes[i] = AABB();
boxes[i] = merge(boxes[i], p0);
boxes[i] = merge(boxes[i], p1);
auto r0 = r;
auto r1 = r;
// override radius if path has thickness
if (p->thickness != nullptr) {
r0 = p->thickness[i0];
r1 = p->thickness[i1];
}
thickness[i] = max(r0, r1);
} else if (p->num_control_points[i] == 1) {
// Quadratic Bezier curve
auto i0 = point_id;
auto i1 = i0 + 1;
auto i2 = (i0 + 2) % p->num_points;
point_id += 2;
auto p0 = Vector2f{p->points[2 * i0], p->points[2 * i0 + 1]};
auto p1 = Vector2f{p->points[2 * i1], p->points[2 * i1 + 1]};
auto p2 = Vector2f{p->points[2 * i2], p->points[2 * i2 + 1]};
boxes[i] = AABB();
boxes[i] = merge(boxes[i], p0);
boxes[i] = merge(boxes[i], p1);
boxes[i] = merge(boxes[i], p2);
auto r0 = r;
auto r1 = r;
auto r2 = r;
// override radius if path has thickness
if (p->thickness != nullptr) {
r0 = p->thickness[i0];
r1 = p->thickness[i1];
r2 = p->thickness[i2];
}
thickness[i] = max(max(r0, r1), r2);
} else if (p->num_control_points[i] == 2) {
// Cubic Bezier curve
auto i0 = point_id;
auto i1 = i0 + 1;
auto i2 = i0 + 2;
auto i3 = (i0 + 3) % p->num_points;
point_id += 3;
auto p0 = Vector2f{p->points[2 * i0], p->points[2 * i0 + 1]};
auto p1 = Vector2f{p->points[2 * i1], p->points[2 * i1 + 1]};
auto p2 = Vector2f{p->points[2 * i2], p->points[2 * i2 + 1]};
auto p3 = Vector2f{p->points[2 * i3], p->points[2 * i3 + 1]};
boxes[i] = AABB();
boxes[i] = merge(boxes[i], p0);
boxes[i] = merge(boxes[i], p1);
boxes[i] = merge(boxes[i], p2);
boxes[i] = merge(boxes[i], p3);
auto r0 = r;
auto r1 = r;
auto r2 = r;
auto r3 = r;
// override radius if path has thickness
if (p->thickness != nullptr) {
r0 = p->thickness[i0];
r1 = p->thickness[i1];
r2 = p->thickness[i2];
r3 = p->thickness[i3];
}
thickness[i] = max(max(max(r0, r1), r2), r3);
} else {
assert(false);
}
}
// Sort the boxes by y
std::vector<int> idx(boxes.size());
std::iota(idx.begin(), idx.end(), 0);
std::sort(idx.begin(), idx.end(), [&](int i0, int i1) {
const AABB &b0 = boxes[i0];
const AABB &b1 = boxes[i1];
auto b0y = 0.5f * (b0.p_min.y + b0.p_max.y);
auto b1y = 0.5f * (b1.p_min.y + b1.p_max.y);
return b0y < b1y;
});
BVHNode *nodes = scene.path_bvhs[shape_id];
for (int i = 0; i < (int)idx.size(); i++) {
nodes[i] = BVHNode{idx[i],
-(first_point_id[idx[i]]+1),
boxes[idx[i]],
thickness[idx[i]]};
}
build_bvh<false /*sort*/>(scene, nodes, boxes.size());
break;
} case ShapeType::Rect: {
const Rect *p = (const Rect*)(shape_list[shape_id]->ptr);
scene.shapes_bbox[shape_id] = AABB{p->p_min, p->p_max};
break;
} default: {
assert(false);
break;
}
}
}
for (int shape_group_id = 0; shape_group_id < (int)shape_group_list.size(); shape_group_id++) {
const ShapeGroup *shape_group = shape_group_list[shape_group_id];
// Build a BVH for each shape group
BVHNode *nodes = scene.shape_groups_bvh_nodes[shape_group_id];
for (int i = 0; i < shape_group->num_shapes; i++) {
auto shape_id = shape_group->shape_ids[i];
auto r = shape_group->stroke_color == nullptr ? 0 : shape_list[shape_id]->stroke_width;
nodes[i] = BVHNode{shape_id,
-1,
scene.shapes_bbox[shape_id],
r};
}
build_bvh<true /*sort*/>(scene, nodes, shape_group->num_shapes);
}
BVHNode *nodes = scene.bvh_nodes;
for (int shape_group_id = 0; shape_group_id < (int)shape_group_list.size(); shape_group_id++) {
const ShapeGroup *shape_group = shape_group_list[shape_group_id];
auto max_radius = shape_list[shape_group->shape_ids[0]]->stroke_width;
if (shape_list[shape_group->shape_ids[0]]->type == ShapeType::Path) {
const Path *p = (const Path*)(shape_list[shape_group->shape_ids[0]]->ptr);
if (p->thickness != nullptr) {
const BVHNode *nodes = scene.path_bvhs[shape_group->shape_ids[0]];
max_radius = nodes[0].max_radius;
}
}
for (int i = 1; i < shape_group->num_shapes; i++) {
auto shape_id = shape_group->shape_ids[i];
auto shape = shape_list[shape_id];
auto r = shape->stroke_width;
if (shape->type == ShapeType::Path) {
const Path *p = (const Path*)(shape_list[shape_id]->ptr);
if (p->thickness != nullptr) {
const BVHNode *nodes = scene.path_bvhs[shape_id];
r = nodes[0].max_radius;
}
}
max_radius = std::max(max_radius, r);
}
// Fetch group bbox from BVH
auto bbox = scene.shape_groups_bvh_nodes[shape_group_id][2 * shape_group->num_shapes - 2].box;
// Transform box from local to world space
nodes[shape_group_id].child0 = shape_group_id;
nodes[shape_group_id].child1 = -1;
nodes[shape_group_id].box = transform(shape_group->shape_to_canvas, bbox);
if (shape_group->stroke_color == nullptr) {
nodes[shape_group_id].max_radius = 0;
} else {
nodes[shape_group_id].max_radius = max_radius;
}
}
build_bvh<true /*sort*/>(scene, nodes, shape_group_list.size());
}
template <bool alloc_mode>
size_t allocate_buffers(Scene &scene,
const std::vector<const Shape *> &shape_list,
const std::vector<const ShapeGroup *> &shape_group_list) {
auto num_shapes = shape_list.size();
auto num_shape_groups = shape_group_list.size();
size_t buffer_size = 0;
if (alloc_mode) scene.shapes = (Shape*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Shape) * num_shapes);
if (alloc_mode) scene.d_shapes = (Shape*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Shape) * num_shapes);
if (alloc_mode) scene.shape_groups = (ShapeGroup*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(ShapeGroup) * num_shape_groups);
if (alloc_mode) scene.d_shape_groups = (ShapeGroup*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(ShapeGroup) * num_shape_groups);
if (alloc_mode) scene.sample_shapes_cdf = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * scene.num_total_shapes);
if (alloc_mode) scene.sample_shapes_pmf = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * scene.num_total_shapes);
if (alloc_mode) scene.sample_shape_id = (int*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(int) * scene.num_total_shapes);
if (alloc_mode) scene.sample_group_id = (int*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(int) * scene.num_total_shapes);
if (alloc_mode) scene.shapes_length = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * num_shapes);
if (alloc_mode) scene.path_length_cdf = (float**)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float*) * num_shapes);
if (alloc_mode) scene.path_length_pmf = (float**)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float*) * num_shapes);
if (alloc_mode) scene.path_point_id_map = (int**)&scene.buffer[buffer_size];
buffer_size += align(sizeof(int*) * num_shapes);
if (alloc_mode) scene.filter = (Filter*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Filter));
if (alloc_mode) scene.d_filter = (DFilter*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(DFilter));
if (alloc_mode) scene.shapes_bbox = (AABB*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(AABB) * num_shapes);
if (alloc_mode) scene.path_bvhs = (BVHNode**)&scene.buffer[buffer_size];
buffer_size += align(sizeof(BVHNode*) * num_shapes);
if (alloc_mode) scene.shape_groups_bvh_nodes = (BVHNode**)&scene.buffer[buffer_size];
buffer_size += align(sizeof(BVHNode*) * num_shape_groups);
if (alloc_mode) scene.bvh_nodes = (BVHNode*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(BVHNode) * (2 * num_shape_groups - 1));
if (alloc_mode) {
for (int i = 0; i < num_shapes; i++) {
scene.path_length_cdf[i] = nullptr;
scene.path_length_pmf[i] = nullptr;
scene.path_point_id_map[i] = nullptr;
scene.path_bvhs[i] = nullptr;
}
}
for (int shape_id = 0; shape_id < scene.num_shapes; shape_id++) {
switch (shape_list[shape_id]->type) {
case ShapeType::Circle: {
if (alloc_mode) scene.shapes[shape_id].ptr = (Circle*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Circle)); // scene.shapes[shape_id].ptr
if (alloc_mode) scene.d_shapes[shape_id].ptr = (Circle*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Circle)); // scene.d_shapes[shape_id].ptr
break;
} case ShapeType::Ellipse: {
if (alloc_mode) scene.shapes[shape_id].ptr = (Ellipse*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Ellipse)); // scene.shapes[shape_id].ptr
if (alloc_mode) scene.d_shapes[shape_id].ptr = (Ellipse*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Ellipse)); // scene.d_shapes[shape_id].ptr
break;
} case ShapeType::Path: {
if (alloc_mode) scene.shapes[shape_id].ptr = (Path*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Path)); // scene.shapes[shape_id].ptr
if (alloc_mode) scene.d_shapes[shape_id].ptr = (Path*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Path)); // scene.d_shapes[shape_id].ptr
const Path *p_ = (const Path*)(shape_list[shape_id]->ptr);
Path *p = nullptr, *d_p = nullptr;
if (alloc_mode) p = (Path*)scene.shapes[shape_id].ptr;
if (alloc_mode) d_p = (Path*)scene.d_shapes[shape_id].ptr;
if (alloc_mode) p->num_control_points = (int*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(int) * p_->num_base_points); // p->num_control_points
if (alloc_mode) p->points = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * (2 * p_->num_points)); // p->points
if (alloc_mode) d_p->points = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * (2 * p_->num_points)); // d_p->points
if (p_->thickness != nullptr) {
if (alloc_mode) p->thickness = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * p_->num_points); // p->thickness
if (alloc_mode) d_p->thickness = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * p_->num_points); // d_p->thickness
} else {
if (alloc_mode) p->thickness = nullptr;
if (alloc_mode) d_p->thickness = nullptr;
}
if (alloc_mode) scene.path_length_pmf[shape_id] = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * p_->num_base_points); // scene.path_length_pmf
if (alloc_mode) scene.path_length_cdf[shape_id] = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * p_->num_base_points); // scene.path_length_cdf
if (alloc_mode) scene.path_point_id_map[shape_id] = (int*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(int) * p_->num_base_points); // scene.path_point_id_map
if (alloc_mode) scene.path_bvhs[shape_id] = (BVHNode*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(BVHNode) * (2 * p_->num_base_points - 1));
break;
} case ShapeType::Rect: {
if (alloc_mode) scene.shapes[shape_id].ptr = (Ellipse*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Rect)); // scene.shapes[shape_id].ptr
if (alloc_mode) scene.d_shapes[shape_id].ptr = (Ellipse*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Rect)); // scene.d_shapes[shape_id].ptr
break;
} default: {
assert(false);
break;
}
}
}
for (int group_id = 0; group_id < scene.num_shape_groups; group_id++) {
const ShapeGroup *shape_group = shape_group_list[group_id];
if (shape_group->fill_color != nullptr) {
switch (shape_group->fill_color_type) {
case ColorType::Constant: {
if (alloc_mode) scene.shape_groups[group_id].fill_color = (Constant*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Constant)); // color
if (alloc_mode) scene.d_shape_groups[group_id].fill_color = (Constant*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Constant)); // d_color
break;
} case ColorType::LinearGradient: {
if (alloc_mode) scene.shape_groups[group_id].fill_color = (LinearGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(LinearGradient)); // color
if (alloc_mode) scene.d_shape_groups[group_id].fill_color = (LinearGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(LinearGradient)); // d_color
const LinearGradient *c_ = (const LinearGradient *)shape_group->fill_color;
LinearGradient *c = nullptr, *d_c = nullptr;
if (alloc_mode) c = (LinearGradient *)scene.shape_groups[group_id].fill_color;
if (alloc_mode) d_c = (LinearGradient *)scene.d_shape_groups[group_id].fill_color;
if (alloc_mode) c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // c->stop_offsets
if (alloc_mode) c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // c->stop_colors
if (alloc_mode) d_c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // d_c->stop_offsets
if (alloc_mode) d_c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // d_c->stop_colors
break;
} case ColorType::RadialGradient: {
if (alloc_mode) scene.shape_groups[group_id].fill_color = (RadialGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(RadialGradient)); // color
if (alloc_mode) scene.d_shape_groups[group_id].fill_color = (RadialGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(RadialGradient)); // d_color
const RadialGradient *c_ = (const RadialGradient *)shape_group->fill_color;
RadialGradient *c = nullptr, *d_c = nullptr;
if (alloc_mode) c = (RadialGradient *)scene.shape_groups[group_id].fill_color;
if (alloc_mode) d_c = (RadialGradient *)scene.d_shape_groups[group_id].fill_color;
if (alloc_mode) c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // c->stop_offsets
if (alloc_mode) c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // c->stop_colors
if (alloc_mode) d_c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // d_c->stop_offsets
if (alloc_mode) d_c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // d_c->stop_colors
break;
} default: {
assert(false);
}
}
} else {
if (alloc_mode) scene.shape_groups[group_id].fill_color = nullptr;
if (alloc_mode) scene.d_shape_groups[group_id].fill_color = nullptr;
}
if (shape_group->stroke_color != nullptr) {
switch (shape_group->stroke_color_type) {
case ColorType::Constant: {
if (alloc_mode) scene.shape_groups[group_id].stroke_color = (Constant*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Constant)); // color
if (alloc_mode) scene.d_shape_groups[group_id].stroke_color = (Constant*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(Constant)); // d_color
break;
} case ColorType::LinearGradient: {
if (alloc_mode) scene.shape_groups[group_id].stroke_color = (LinearGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(LinearGradient)); // color
if (alloc_mode) scene.shape_groups[group_id].stroke_color = (LinearGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(LinearGradient)); // d_color
const LinearGradient *c_ = (const LinearGradient *)shape_group->stroke_color;
LinearGradient *c = nullptr, *d_c = nullptr;
if (alloc_mode) c = (LinearGradient *)scene.shape_groups[group_id].stroke_color;
if (alloc_mode) d_c = (LinearGradient *)scene.d_shape_groups[group_id].stroke_color;
if (alloc_mode) c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // c->stop_offsets
if (alloc_mode) c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // c->stop_colors
if (alloc_mode) d_c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // d_c->stop_offsets
if (alloc_mode) d_c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // d_c->stop_colors
break;
} case ColorType::RadialGradient: {
if (alloc_mode) scene.shape_groups[group_id].stroke_color = (RadialGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(RadialGradient)); // color
if (alloc_mode) scene.shape_groups[group_id].stroke_color = (RadialGradient*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(RadialGradient)); // d_color
const RadialGradient *c_ = (const RadialGradient *)shape_group->stroke_color;
RadialGradient *c = nullptr, *d_c = nullptr;
if (alloc_mode) c = (RadialGradient *)scene.shape_groups[group_id].stroke_color;
if (alloc_mode) d_c = (RadialGradient *)scene.d_shape_groups[group_id].stroke_color;
if (alloc_mode) c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // c->stop_offsets
if (alloc_mode) c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // c->stop_colors
if (alloc_mode) d_c->stop_offsets = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * c_->num_stops); // d_c->stop_offsets
if (alloc_mode) d_c->stop_colors = (float*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(float) * 4 * c_->num_stops); // d_c->stop_colors
break;
} default: {
assert(false);
}
}
} else {
if (alloc_mode) scene.shape_groups[group_id].stroke_color = nullptr;
if (alloc_mode) scene.d_shape_groups[group_id].stroke_color = nullptr;
}
if (alloc_mode) scene.shape_groups[group_id].shape_ids = (int*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(int) * shape_group->num_shapes); // shape_group->shape_ids
if (alloc_mode) scene.shape_groups_bvh_nodes[group_id] = (BVHNode*)&scene.buffer[buffer_size];
buffer_size += align(sizeof(BVHNode) * (2 * shape_group->num_shapes - 1)); // scene.shape_groups_bvh_nodes[group_id]
}
return buffer_size;
}
Scene::Scene(int canvas_width,
int canvas_height,
const std::vector<const Shape *> &shape_list,
const std::vector<const ShapeGroup *> &shape_group_list,
const Filter &filter,
bool use_gpu,
int gpu_index)
: canvas_width(canvas_width),
canvas_height(canvas_height),
num_shapes(shape_list.size()),
num_shape_groups(shape_group_list.size()),
use_gpu(use_gpu),
gpu_index(gpu_index) {
if (num_shapes == 0) {
return;
}
// Shape group may reuse some of the shapes,
// record the total number of shapes.
int num_total_shapes = 0;
for (const ShapeGroup *sg : shape_group_list) {
num_total_shapes += sg->num_shapes;
}
this->num_total_shapes = num_total_shapes;
// Memory initialization
#ifdef __NVCC__
int old_device_id = -1;
#endif
if (use_gpu) {
#ifdef __NVCC__
checkCuda(cudaGetDevice(&old_device_id));
if (gpu_index != -1) {
checkCuda(cudaSetDevice(gpu_index));
}
#else
throw std::runtime_error("diffvg not compiled with GPU");
assert(false);
#endif
}
size_t buffer_size = allocate_buffers<false /*alloc_mode*/>(*this, shape_list, shape_group_list);
// Allocate a huge buffer for everything
allocate<uint8_t>(use_gpu, buffer_size, &buffer);
// memset(buffer, 111, buffer_size);
// Actually distribute the buffer
allocate_buffers<true /*alloc_mode*/>(*this, shape_list, shape_group_list);
copy_and_init_shapes(*this, shape_list);
copy_and_init_shape_groups(*this, shape_group_list);
std::vector<float> shape_length_list = compute_shape_length(shape_list);
// Copy shape_length
if (use_gpu) {
#ifdef __NVCC__
checkCuda(cudaMemcpy(this->shapes_length, &shape_length_list[0], num_shapes * sizeof(float), cudaMemcpyHostToDevice));
#else
throw std::runtime_error("diffvg not compiled with GPU");
assert(false);
#endif
} else {
memcpy(this->shapes_length, &shape_length_list[0], num_shapes * sizeof(float));
}
build_shape_cdfs(*this, shape_group_list, shape_length_list);
build_path_cdfs(*this, shape_list, shape_length_list);
compute_bounding_boxes(*this, shape_list, shape_group_list);
// Filter initialization
*(this->filter) = filter;
this->d_filter->radius = 0;
if (use_gpu) {
#ifdef __NVCC__
if (old_device_id != -1) {
checkCuda(cudaSetDevice(old_device_id));
}
#else
throw std::runtime_error("diffvg not compiled with GPU");
assert(false);
#endif
}
}
Scene::~Scene() {
if (num_shapes == 0) {
return;
}
if (use_gpu) {
#ifdef __NVCC__
int old_device_id = -1;
checkCuda(cudaGetDevice(&old_device_id));
if (gpu_index != -1) {
checkCuda(cudaSetDevice(gpu_index));
}
checkCuda(cudaFree(buffer));
checkCuda(cudaSetDevice(old_device_id));
#else
// Don't throw because C++ don't want a destructor to throw.
std::cerr << "diffvg not compiled with GPU";
exit(1);
#endif
} else {
free(buffer);
}
}
Shape Scene::get_d_shape(int shape_id) const {
return d_shapes[shape_id];
}
ShapeGroup Scene::get_d_shape_group(int group_id) const {
return d_shape_groups[group_id];
}
float Scene::get_d_filter_radius() const {
return d_filter->radius;
}