summaryrefslogtreecommitdiff
path: root/modules/raycast
diff options
context:
space:
mode:
Diffstat (limited to 'modules/raycast')
-rw-r--r--modules/raycast/SCsub3
-rw-r--r--modules/raycast/godot_update_embree.py12
-rw-r--r--modules/raycast/raycast_occlusion_cull.cpp123
-rw-r--r--modules/raycast/raycast_occlusion_cull.h17
4 files changed, 98 insertions, 57 deletions
diff --git a/modules/raycast/SCsub b/modules/raycast/SCsub
index 6e7b3e7b8d..1fdc8fe1b3 100644
--- a/modules/raycast/SCsub
+++ b/modules/raycast/SCsub
@@ -55,6 +55,9 @@ if env["builtin_embree"]:
"kernels/bvh/bvh_builder_sah_mb.cpp",
"kernels/bvh/bvh_builder_twolevel.cpp",
"kernels/bvh/bvh_intersector1_bvh4.cpp",
+ "kernels/bvh/bvh_intersector_hybrid4_bvh4.cpp",
+ "kernels/bvh/bvh_intersector_stream_bvh4.cpp",
+ "kernels/bvh/bvh_intersector_stream_filters.cpp",
]
thirdparty_sources = [thirdparty_dir + file for file in embree_src]
diff --git a/modules/raycast/godot_update_embree.py b/modules/raycast/godot_update_embree.py
index 31a25a318f..e31d88b741 100644
--- a/modules/raycast/godot_update_embree.py
+++ b/modules/raycast/godot_update_embree.py
@@ -61,6 +61,11 @@ cpp_files = [
"kernels/bvh/bvh_builder_twolevel.cpp",
"kernels/bvh/bvh_intersector1.cpp",
"kernels/bvh/bvh_intersector1_bvh4.cpp",
+ "kernels/bvh/bvh_intersector_hybrid4_bvh4.cpp",
+ "kernels/bvh/bvh_intersector_stream_bvh4.cpp",
+ "kernels/bvh/bvh_intersector_stream_filters.cpp",
+ "kernels/bvh/bvh_intersector_hybrid.cpp",
+ "kernels/bvh/bvh_intersector_stream.cpp",
]
os.chdir("../../thirdparty")
@@ -117,7 +122,7 @@ with open(os.path.join(dest_dir, "kernels/config.h"), "w") as config_file:
/* #undef EMBREE_GEOMETRY_INSTANCE */
/* #undef EMBREE_GEOMETRY_GRID */
/* #undef EMBREE_GEOMETRY_POINT */
-/* #undef EMBREE_RAY_PACKETS */
+#define EMBREE_RAY_PACKETS
/* #undef EMBREE_COMPACT_POLYS */
#define EMBREE_CURVE_SELF_INTERSECTION_AVOIDANCE_FACTOR 2.0
@@ -249,3 +254,8 @@ with open(os.path.join(dest_dir, "include/embree3/rtcore_config.h"), "w") as con
os.chdir("..")
shutil.rmtree("embree-tmp")
+
+subprocess.run(["git", "restore", "embree/patches"])
+
+for patch in os.listdir("embree/patches"):
+ subprocess.run(["git", "apply", "embree/patches/" + patch])
diff --git a/modules/raycast/raycast_occlusion_cull.cpp b/modules/raycast/raycast_occlusion_cull.cpp
index 88c0145ebc..a55b81c05d 100644
--- a/modules/raycast/raycast_occlusion_cull.cpp
+++ b/modules/raycast/raycast_occlusion_cull.cpp
@@ -66,28 +66,45 @@ void RaycastOcclusionCull::RaycastHZBuffer::resize(const Size2i &p_size) {
void RaycastOcclusionCull::RaycastHZBuffer::update_camera_rays(const Transform3D &p_cam_transform, const CameraMatrix &p_cam_projection, bool p_cam_orthogonal, ThreadWorkPool &p_thread_work_pool) {
CameraRayThreadData td;
- td.camera_matrix = p_cam_projection;
- td.camera_transform = p_cam_transform;
- td.camera_orthogonal = p_cam_orthogonal;
td.thread_count = p_thread_work_pool.get_thread_count();
+ td.z_near = p_cam_projection.get_z_near();
+ td.z_far = p_cam_projection.get_z_far() * 1.05f;
+ td.camera_pos = p_cam_transform.origin;
+ td.camera_dir = -p_cam_transform.basis.get_axis(2);
+ td.camera_orthogonal = p_cam_orthogonal;
+
+ CameraMatrix inv_camera_matrix = p_cam_projection.inverse();
+ Vector3 camera_corner_proj = Vector3(-1.0f, -1.0f, -1.0f);
+ Vector3 camera_corner_view = inv_camera_matrix.xform(camera_corner_proj);
+ td.pixel_corner = p_cam_transform.xform(camera_corner_view);
+
+ Vector3 top_corner_proj = Vector3(-1.0f, 1.0f, -1.0f);
+ Vector3 top_corner_view = inv_camera_matrix.xform(top_corner_proj);
+ Vector3 top_corner_world = p_cam_transform.xform(top_corner_view);
+
+ Vector3 left_corner_proj = Vector3(1.0f, -1.0f, -1.0f);
+ Vector3 left_corner_view = inv_camera_matrix.xform(left_corner_proj);
+ Vector3 left_corner_world = p_cam_transform.xform(left_corner_view);
+
+ td.pixel_u_interp = left_corner_world - td.pixel_corner;
+ td.pixel_v_interp = top_corner_world - td.pixel_corner;
+
+ debug_tex_range = td.z_far;
+
p_thread_work_pool.do_work(td.thread_count, this, &RaycastHZBuffer::_camera_rays_threaded, &td);
}
-void RaycastOcclusionCull::RaycastHZBuffer::_camera_rays_threaded(uint32_t p_thread, RaycastOcclusionCull::RaycastHZBuffer::CameraRayThreadData *p_data) {
+void RaycastOcclusionCull::RaycastHZBuffer::_camera_rays_threaded(uint32_t p_thread, const CameraRayThreadData *p_data) {
uint32_t packs_total = camera_rays.size();
uint32_t total_threads = p_data->thread_count;
uint32_t from = p_thread * packs_total / total_threads;
uint32_t to = (p_thread + 1 == total_threads) ? packs_total : ((p_thread + 1) * packs_total / total_threads);
- _generate_camera_rays(p_data->camera_transform, p_data->camera_matrix, p_data->camera_orthogonal, from, to);
+ _generate_camera_rays(p_data, from, to);
}
-void RaycastOcclusionCull::RaycastHZBuffer::_generate_camera_rays(const Transform3D &p_cam_transform, const CameraMatrix &p_cam_projection, bool p_cam_orthogonal, int p_from, int p_to) {
- Size2i buffer_size = sizes[0];
-
- CameraMatrix inv_camera_matrix = p_cam_projection.inverse();
- float z_far = p_cam_projection.get_z_far() * 1.05f;
- debug_tex_range = z_far;
+void RaycastOcclusionCull::RaycastHZBuffer::_generate_camera_rays(const CameraRayThreadData *p_data, int p_from, int p_to) {
+ const Size2i &buffer_size = sizes[0];
RayPacket *ray_packets = camera_rays.ptr();
uint32_t *ray_masks = camera_ray_masks.ptr();
@@ -98,56 +115,52 @@ void RaycastOcclusionCull::RaycastHZBuffer::_generate_camera_rays(const Transfor
int tile_y = (i / packs_size.x) * TILE_SIZE;
for (int j = 0; j < TILE_RAYS; j++) {
- float x = tile_x + j % TILE_SIZE;
- float y = tile_y + j / TILE_SIZE;
-
- ray_masks[i * TILE_RAYS + j] = ~0U;
+ int x = tile_x + j % TILE_SIZE;
+ int y = tile_y + j / TILE_SIZE;
if (x >= buffer_size.x || y >= buffer_size.y) {
ray_masks[i * TILE_RAYS + j] = 0U;
- } else {
- float u = x / (buffer_size.x - 1);
- float v = y / (buffer_size.y - 1);
- u = u * 2.0f - 1.0f;
- v = v * 2.0f - 1.0f;
-
- Plane pixel_proj = Plane(u, v, -1.0, 1.0);
- Plane pixel_view = inv_camera_matrix.xform4(pixel_proj);
- Vector3 pixel_world = p_cam_transform.xform(pixel_view.normal);
-
- Vector3 dir;
- if (p_cam_orthogonal) {
- dir = -p_cam_transform.basis.get_axis(2);
- } else {
- dir = (pixel_world - p_cam_transform.origin).normalized();
- }
-
- packet.ray.org_x[j] = pixel_world.x;
- packet.ray.org_y[j] = pixel_world.y;
- packet.ray.org_z[j] = pixel_world.z;
+ continue;
+ }
- packet.ray.dir_x[j] = dir.x;
- packet.ray.dir_y[j] = dir.y;
- packet.ray.dir_z[j] = dir.z;
+ ray_masks[i * TILE_RAYS + j] = ~0U;
- packet.ray.tnear[j] = 0.0f;
+ float u = (float(x) + 0.5f) / buffer_size.x;
+ float v = (float(y) + 0.5f) / buffer_size.y;
+ Vector3 pixel_pos = p_data->pixel_corner + u * p_data->pixel_u_interp + v * p_data->pixel_v_interp;
- packet.ray.time[j] = 0.0f;
+ packet.ray.tnear[j] = p_data->z_near;
- packet.ray.flags[j] = 0;
- packet.ray.mask[j] = -1;
- packet.hit.geomID[j] = RTC_INVALID_GEOMETRY_ID;
+ Vector3 dir;
+ if (p_data->camera_orthogonal) {
+ dir = -p_data->camera_dir;
+ packet.ray.org_x[j] = pixel_pos.x - dir.x * p_data->z_near;
+ packet.ray.org_y[j] = pixel_pos.y - dir.y * p_data->z_near;
+ packet.ray.org_z[j] = pixel_pos.z - dir.z * p_data->z_near;
+ } else {
+ dir = (pixel_pos - p_data->camera_pos).normalized();
+ packet.ray.org_x[j] = p_data->camera_pos.x;
+ packet.ray.org_y[j] = p_data->camera_pos.y;
+ packet.ray.org_z[j] = p_data->camera_pos.z;
+ packet.ray.tnear[j] /= dir.dot(p_data->camera_dir);
}
- packet.ray.tfar[j] = z_far;
+ packet.ray.dir_x[j] = dir.x;
+ packet.ray.dir_y[j] = dir.y;
+ packet.ray.dir_z[j] = dir.z;
+
+ packet.ray.tfar[j] = p_data->z_far;
+ packet.ray.time[j] = 0.0f;
+
+ packet.ray.flags[j] = 0;
+ packet.ray.mask[j] = -1;
+ packet.hit.geomID[j] = RTC_INVALID_GEOMETRY_ID;
}
}
}
-void RaycastOcclusionCull::RaycastHZBuffer::sort_rays() {
- if (is_empty()) {
- return;
- }
+void RaycastOcclusionCull::RaycastHZBuffer::sort_rays(const Vector3 &p_camera_dir, bool p_orthogonal) {
+ ERR_FAIL_COND(is_empty());
Size2i buffer_size = sizes[0];
for (int i = 0; i < packs_size.y; i++) {
@@ -161,7 +174,17 @@ void RaycastOcclusionCull::RaycastHZBuffer::sort_rays() {
}
int k = tile_i * TILE_SIZE + tile_j;
int packet_index = i * packs_size.x + j;
- mips[0][y * buffer_size.x + x] = camera_rays[packet_index].ray.tfar[k];
+ float d = camera_rays[packet_index].ray.tfar[k];
+
+ if (!p_orthogonal) {
+ const float &dir_x = camera_rays[packet_index].ray.dir_x[k];
+ const float &dir_y = camera_rays[packet_index].ray.dir_y[k];
+ const float &dir_z = camera_rays[packet_index].ray.dir_z[k];
+ float cos_theta = p_camera_dir.x * dir_x + p_camera_dir.y * dir_y + p_camera_dir.z * dir_z;
+ d *= cos_theta;
+ }
+
+ mips[0][y * buffer_size.x + x] = d;
}
}
}
@@ -514,7 +537,7 @@ void RaycastOcclusionCull::buffer_update(RID p_buffer, const Transform3D &p_cam_
buffer.update_camera_rays(p_cam_transform, p_cam_projection, p_cam_orthogonal, p_thread_pool);
scenario.raycast(buffer.camera_rays, buffer.camera_ray_masks, p_thread_pool);
- buffer.sort_rays();
+ buffer.sort_rays(-p_cam_transform.basis.get_axis(2), p_cam_orthogonal);
buffer.update_mips();
}
diff --git a/modules/raycast/raycast_occlusion_cull.h b/modules/raycast/raycast_occlusion_cull.h
index 85710a790c..cc87a6342c 100644
--- a/modules/raycast/raycast_occlusion_cull.h
+++ b/modules/raycast/raycast_occlusion_cull.h
@@ -51,15 +51,20 @@ public:
Size2i packs_size;
struct CameraRayThreadData {
- CameraMatrix camera_matrix;
- Transform3D camera_transform;
- bool camera_orthogonal;
int thread_count;
+ float z_near;
+ float z_far;
+ Vector3 camera_dir;
+ Vector3 camera_pos;
+ Vector3 pixel_corner;
+ Vector3 pixel_u_interp;
+ Vector3 pixel_v_interp;
+ bool camera_orthogonal;
Size2i buffer_size;
};
- void _camera_rays_threaded(uint32_t p_thread, CameraRayThreadData *p_data);
- void _generate_camera_rays(const Transform3D &p_cam_transform, const CameraMatrix &p_cam_projection, bool p_cam_orthogonal, int p_from, int p_to);
+ void _camera_rays_threaded(uint32_t p_thread, const CameraRayThreadData *p_data);
+ void _generate_camera_rays(const CameraRayThreadData *p_data, int p_from, int p_to);
public:
LocalVector<RayPacket> camera_rays;
@@ -68,7 +73,7 @@ public:
virtual void clear() override;
virtual void resize(const Size2i &p_size) override;
- void sort_rays();
+ void sort_rays(const Vector3 &p_camera_dir, bool p_orthogonal);
void update_camera_rays(const Transform3D &p_cam_transform, const CameraMatrix &p_cam_projection, bool p_cam_orthogonal, ThreadWorkPool &p_thread_work_pool);
};
generated by cgit v1.2.3 (git 2.25.1) at 2024-12-13 16:33:16 +0000