diff options
author | Rémi Verschelde <rverschelde@gmail.com> | 2023-01-24 09:15:26 +0100 |
---|---|---|
committer | Rémi Verschelde <rverschelde@gmail.com> | 2023-01-24 09:15:26 +0100 |
commit | a6042b649a3626488550a822b7d9469aa2bd3f92 (patch) | |
tree | aa085871bd57af6b432cc957183e2cb66f1e1d8a /servers | |
parent | 514b3486196ac4779d7a9dbe84e7223d8d365a6a (diff) | |
parent | 4bd01a93dc82b4ac5734ac2a4af70d49b4559496 (diff) |
Merge pull request #71832 from Geometror/fix-spotlight-artifacts
Fix some `SpotLight3D` issues (clustering artifacts, leaking light, AABB)
Diffstat (limited to 'servers')
3 files changed, 89 insertions, 61 deletions
diff --git a/servers/rendering/renderer_rd/cluster_builder_rd.cpp b/servers/rendering/renderer_rd/cluster_builder_rd.cpp index 73a0c652a4..959a752fba 100644 --- a/servers/rendering/renderer_rd/cluster_builder_rd.cpp +++ b/servers/rendering/renderer_rd/cluster_builder_rd.cpp @@ -74,7 +74,7 @@ ClusterBuilderSharedDataRD::ClusterBuilderSharedDataRD() { cluster_debug.shader_pipeline = RD::get_singleton()->compute_pipeline_create(cluster_debug.shader); } - { // SPHERE + { // Sphere mesh data. static const uint32_t icosphere_vertex_count = 42; static const float icosphere_vertices[icosphere_vertex_count * 3] = { 0, 0, -1, 0.7236073, -0.5257253, -0.4472195, -0.276388, -0.8506492, -0.4472199, -0.8944262, 0, -0.4472156, -0.276388, 0.8506492, -0.4472199, 0.7236073, 0.5257253, -0.4472195, 0.276388, -0.8506492, 0.4472199, -0.7236073, -0.5257253, 0.4472195, -0.7236073, 0.5257253, 0.4472195, 0.276388, 0.8506492, 0.4472199, 0.8944262, 0, 0.4472156, 0, 0, 1, -0.1624555, -0.4999952, -0.8506544, 0.4253227, -0.3090114, -0.8506542, 0.2628688, -0.8090116, -0.5257377, 0.8506479, 0, -0.5257359, 0.4253227, 0.3090114, -0.8506542, -0.5257298, 0, -0.8506517, -0.6881894, -0.4999969, -0.5257362, -0.1624555, 0.4999952, -0.8506544, -0.6881894, 0.4999969, -0.5257362, 0.2628688, 0.8090116, -0.5257377, 0.9510579, -0.3090126, 0, 0.9510579, 0.3090126, 0, 0, -1, 0, 0.5877856, -0.8090167, 0, -0.9510579, -0.3090126, 0, -0.5877856, -0.8090167, 0, -0.5877856, 0.8090167, 0, -0.9510579, 0.3090126, 0, 0.5877856, 0.8090167, 0, 0, 1, 0, 0.6881894, -0.4999969, 0.5257362, -0.2628688, -0.8090116, 0.5257377, -0.8506479, 0, 0.5257359, -0.2628688, 0.8090116, 0.5257377, 0.6881894, 0.4999969, 0.5257362, 0.1624555, -0.4999952, 0.8506544, 0.5257298, 0, 0.8506517, -0.4253227, -0.3090114, 0.8506542, -0.4253227, 0.3090114, 0.8506542, 0.1624555, 0.4999952, 0.8506544 @@ -118,7 +118,7 @@ ClusterBuilderSharedDataRD::ClusterBuilderSharedDataRD() { sphere_overfit = 1.0 / min_d; } - { // CONE + { // Cone mesh data. static const uint32_t cone_vertex_count = 99; static const float cone_vertices[cone_vertex_count * 3] = { 0, 1, -1, 0.1950903, 0.9807853, -1, 0.3826835, 0.9238795, -1, 0.5555703, 0.8314696, -1, 0.7071068, 0.7071068, -1, 0.8314697, 0.5555702, -1, 0.9238795, 0.3826834, -1, 0.9807853, 0.1950903, -1, 1, 0, -1, 0.9807853, -0.1950902, -1, 0.9238796, -0.3826833, -1, 0.8314697, -0.5555702, -1, 0.7071068, -0.7071068, -1, 0.5555702, -0.8314697, -1, 0.3826833, -0.9238796, -1, 0.1950901, -0.9807853, -1, -3.25841e-7, -1, -1, -0.1950907, -0.9807852, -1, -0.3826839, -0.9238793, -1, -0.5555707, -0.8314693, -1, -0.7071073, -0.7071063, -1, -0.83147, -0.5555697, -1, -0.9238799, -0.3826827, -1, 0, 0, 0, -0.9807854, -0.1950894, -1, -1, 9.65599e-7, -1, -0.9807851, 0.1950913, -1, -0.9238791, 0.3826845, -1, -0.8314689, 0.5555713, -1, -0.7071059, 0.7071077, -1, -0.5555691, 0.8314704, -1, -0.3826821, 0.9238801, -1, -0.1950888, 0.9807856, -1 @@ -172,7 +172,7 @@ ClusterBuilderSharedDataRD::ClusterBuilderSharedDataRD() { cone_overfit = 1.0 / min_d; } - { // BOX + { // Box mesh data. static const uint32_t box_vertex_count = 8; static const float box_vertices[box_vertex_count * 3] = { -1, -1, -1, -1, -1, 1, -1, 1, -1, -1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, -1, 1, 1, 1 @@ -219,8 +219,9 @@ ClusterBuilderSharedDataRD::~ClusterBuilderSharedDataRD() { void ClusterBuilderRD::_clear() { if (cluster_buffer.is_null()) { - return; //nothing to clear + return; } + RD::get_singleton()->free(cluster_buffer); RD::get_singleton()->free(cluster_render_buffer); RD::get_singleton()->free(element_buffer); @@ -254,7 +255,7 @@ void ClusterBuilderRD::setup(Size2i p_screen_size, uint32_t p_max_elements, RID cluster_screen_size.height = (p_screen_size.height - 1) / cluster_size + 1; max_elements_by_type = p_max_elements; - if (max_elements_by_type % 32) { //need to be 32 aligned + if (max_elements_by_type % 32) { // Needs to be aligned to 32. max_elements_by_type += 32 - (max_elements_by_type % 32); } @@ -264,7 +265,8 @@ void ClusterBuilderRD::setup(Size2i p_screen_size, uint32_t p_max_elements, RID uint32_t element_tag_bits_size = render_element_max / 32; uint32_t element_tag_depth_bits_size = render_element_max; - cluster_render_buffer_size = cluster_screen_size.x * cluster_screen_size.y * (element_tag_bits_size + element_tag_depth_bits_size) * 4; // tag bits (element was used) and tag depth (depth range in which it was used) + + cluster_render_buffer_size = cluster_screen_size.x * cluster_screen_size.y * (element_tag_bits_size + element_tag_depth_bits_size) * 4; // Tag bits (element was used) and tag depth (depth range in which it was used). cluster_render_buffer = RD::get_singleton()->storage_buffer_create(cluster_render_buffer_size); cluster_buffer = RD::get_singleton()->storage_buffer_create(cluster_buffer_size); @@ -379,9 +381,9 @@ void ClusterBuilderRD::begin(const Transform3D &p_view_transform, const Projecti projection = p_cam_projection; z_near = projection.get_z_near(); z_far = projection.get_z_far(); - orthogonal = p_cam_projection.is_orthogonal(); + camera_orthogonal = p_cam_projection.is_orthogonal(); adjusted_projection = projection; - if (!orthogonal) { + if (!camera_orthogonal) { adjusted_projection.adjust_perspective_znear(0.0001); } @@ -390,7 +392,7 @@ void ClusterBuilderRD::begin(const Transform3D &p_view_transform, const Projecti projection = correction * projection; adjusted_projection = correction * adjusted_projection; - //reset counts + // Reset counts. render_element_count = 0; for (uint32_t i = 0; i < ELEMENT_TYPE_MAX; i++) { cluster_count_by_type[i] = 0; @@ -402,14 +404,14 @@ void ClusterBuilderRD::bake_cluster() { RD::get_singleton()->draw_command_begin_label("Bake Light Cluster"); - //clear cluster buffer + // Clear cluster buffer. RD::get_singleton()->buffer_clear(cluster_buffer, 0, cluster_buffer_size, RD::BARRIER_MASK_RASTER | RD::BARRIER_MASK_COMPUTE); if (render_element_count > 0) { - //clear render buffer + // Clear render buffer. RD::get_singleton()->buffer_clear(cluster_render_buffer, 0, cluster_render_buffer_size, RD::BARRIER_MASK_RASTER); - { //fill state uniform + { // Fill state uniform. StateUniform state; @@ -425,13 +427,13 @@ void ClusterBuilderRD::bake_cluster() { RD::get_singleton()->buffer_update(state_uniform, 0, sizeof(StateUniform), &state, RD::BARRIER_MASK_RASTER | RD::BARRIER_MASK_COMPUTE); } - //update instances + // Update instances. RD::get_singleton()->buffer_update(element_buffer, 0, sizeof(RenderElementData) * render_element_count, render_elements, RD::BARRIER_MASK_RASTER | RD::BARRIER_MASK_COMPUTE); RENDER_TIMESTAMP("Render 3D Cluster Elements"); - //render elements + // Render elements. { RD::DrawListID draw_list = RD::get_singleton()->draw_list_begin(framebuffer, RD::INITIAL_ACTION_DROP, RD::FINAL_ACTION_DISCARD, RD::INITIAL_ACTION_DROP, RD::FINAL_ACTION_DISCARD); ClusterBuilderSharedDataRD::ClusterRender::PushConstant push_constant = {}; @@ -447,8 +449,16 @@ void ClusterBuilderRD::bake_cluster() { RD::get_singleton()->draw_list_bind_index_array(draw_list, shared->sphere_index_array); } break; case ELEMENT_TYPE_SPOT_LIGHT: { - RD::get_singleton()->draw_list_bind_vertex_array(draw_list, shared->cone_vertex_array); - RD::get_singleton()->draw_list_bind_index_array(draw_list, shared->cone_index_array); + // If the spot angle is above a certain threshold, use a sphere instead of a cone for building the clusters + // since the cone gets too flat/large (spot angle close to 90 degrees) or + // can't even cover the affected area of the light (spot angle above 90 degrees). + if (render_elements[i].has_wide_spot_angle) { + RD::get_singleton()->draw_list_bind_vertex_array(draw_list, shared->sphere_vertex_array); + RD::get_singleton()->draw_list_bind_index_array(draw_list, shared->sphere_index_array); + } else { + RD::get_singleton()->draw_list_bind_vertex_array(draw_list, shared->cone_vertex_array); + RD::get_singleton()->draw_list_bind_index_array(draw_list, shared->cone_index_array); + } } break; case ELEMENT_TYPE_DECAL: case ELEMENT_TYPE_REFLECTION_PROBE: { @@ -465,7 +475,7 @@ void ClusterBuilderRD::bake_cluster() { } RD::get_singleton()->draw_list_end(RD::BARRIER_MASK_COMPUTE); } - //store elements + // Store elements. RENDER_TIMESTAMP("Pack 3D Cluster Elements"); { @@ -509,7 +519,7 @@ void ClusterBuilderRD::debug(ElementType p_element) { push_constant.cluster_screen_size[1] = cluster_screen_size.y; push_constant.cluster_shift = get_shift_from_power_of_2(cluster_size); push_constant.cluster_type = p_element; - push_constant.orthogonal = orthogonal; + push_constant.orthogonal = camera_orthogonal; push_constant.z_far = z_far; push_constant.z_near = z_near; push_constant.max_cluster_element_count_div_32 = max_elements_by_type / 32; diff --git a/servers/rendering/renderer_rd/cluster_builder_rd.h b/servers/rendering/renderer_rd/cluster_builder_rd.h index 0b20a5d7ee..a13e6c8172 100644 --- a/servers/rendering/renderer_rd/cluster_builder_rd.h +++ b/servers/rendering/renderer_rd/cluster_builder_rd.h @@ -43,13 +43,13 @@ class ClusterBuilderSharedDataRD { RID sphere_vertex_array; RID sphere_index_buffer; RID sphere_index_array; - float sphere_overfit = 0.0; //because an icosphere is not a perfect sphere, we need to enlarge it to cover the sphere area + float sphere_overfit = 0.0; // Because an icosphere is not a perfect sphere, we need to enlarge it to cover the sphere area. RID cone_vertex_buffer; RID cone_vertex_array; RID cone_index_buffer; RID cone_index_array; - float cone_overfit = 0.0; //because an cone mesh is not a perfect sphere, we need to enlarge it to cover the actual cone area + float cone_overfit = 0.0; // Because an cone mesh is not a perfect cone, we need to enlarge it to cover the actual cone area. RID box_vertex_buffer; RID box_vertex_array; @@ -73,6 +73,7 @@ class ClusterBuilderSharedDataRD { ClusterRenderShaderRD cluster_render_shader; RID shader_version; RID shader; + enum PipelineVersion { PIPELINE_NORMAL, PIPELINE_MSAA, @@ -85,10 +86,11 @@ class ClusterBuilderSharedDataRD { struct ClusterStore { struct PushConstant { uint32_t cluster_render_data_size; // how much data for a single cluster takes - uint32_t max_render_element_count_div_32; //divided by 32 + uint32_t max_render_element_count_div_32; // divided by 32 uint32_t cluster_screen_size[2]; - uint32_t render_element_count_div_32; //divided by 32 - uint32_t max_cluster_element_count_div_32; //divided by 32 + uint32_t render_element_count_div_32; // divided by 32 + uint32_t max_cluster_element_count_div_32; // divided by 32 + uint32_t pad1; uint32_t pad2; }; @@ -111,6 +113,7 @@ class ClusterBuilderSharedDataRD { uint32_t orthogonal; uint32_t max_cluster_element_count_div_32; + uint32_t pad1; uint32_t pad2; }; @@ -128,6 +131,8 @@ public: class ClusterBuilderRD { public: + static constexpr float WIDE_SPOT_ANGLE_THRESHOLD_DEG = 60.0f; + enum LightType { LIGHT_TYPE_OMNI, LIGHT_TYPE_SPOT @@ -144,21 +149,20 @@ public: ELEMENT_TYPE_DECAL, ELEMENT_TYPE_REFLECTION_PROBE, ELEMENT_TYPE_MAX, - }; private: ClusterBuilderSharedDataRD *shared = nullptr; struct RenderElementData { - uint32_t type; //0-4 + uint32_t type; // 0-4 uint32_t touches_near; uint32_t touches_far; uint32_t original_index; - float transform_inv[12]; //transposed transform for less space + float transform_inv[12]; // Transposed transform for less space. float scale[3]; - uint32_t pad; - }; + uint32_t has_wide_spot_angle; + }; // Keep aligned to 32 bytes. uint32_t cluster_count_by_type[ELEMENT_TYPE_MAX] = {}; uint32_t max_elements_by_type = 0; @@ -172,7 +176,7 @@ private: Projection projection; float z_far = 0; float z_near = 0; - bool orthogonal = false; + bool camera_orthogonal = false; enum Divisor { DIVISOR_1, @@ -188,26 +192,27 @@ private: Size2i cluster_screen_size; RID framebuffer; - RID cluster_render_buffer; //used for creating - RID cluster_buffer; //used for rendering - RID element_buffer; //used for storing, to hint element touches far plane or near plane + RID cluster_render_buffer; // Used for creating. + RID cluster_buffer; // Used for rendering. + RID element_buffer; // Used for storing, to hint element touches far plane or near plane. uint32_t cluster_render_buffer_size = 0; uint32_t cluster_buffer_size = 0; RID cluster_render_uniform_set; RID cluster_store_uniform_set; - //persistent data + // Persistent data. void _clear(); struct StateUniform { float projection[16]; float inv_z_far; - uint32_t screen_to_clusters_shift; // shift to obtain coordinates in block indices - uint32_t cluster_screen_width; // - uint32_t cluster_data_size; // how much data for a single cluster takes + uint32_t screen_to_clusters_shift; // Shift to obtain coordinates in block indices. + uint32_t cluster_screen_width; + uint32_t cluster_data_size; // How much data is needed for a single cluster. uint32_t cluster_depth_offset; + uint32_t pad0; uint32_t pad1; uint32_t pad2; @@ -224,10 +229,10 @@ public: _FORCE_INLINE_ void add_light(LightType p_type, const Transform3D &p_transform, float p_radius, float p_spot_aperture) { if (p_type == LIGHT_TYPE_OMNI && cluster_count_by_type[ELEMENT_TYPE_OMNI_LIGHT] == max_elements_by_type) { - return; //max number elements reached + return; // Max number elements reached. } if (p_type == LIGHT_TYPE_SPOT && cluster_count_by_type[ELEMENT_TYPE_SPOT_LIGHT] == max_elements_by_type) { - return; //max number elements reached + return; // Max number elements reached. } RenderElementData &e = render_elements[render_element_count]; @@ -242,15 +247,14 @@ public: radius *= p_radius; if (p_type == LIGHT_TYPE_OMNI) { - radius *= shared->sphere_overfit; // overfit icosphere + radius *= shared->sphere_overfit; // Overfit icosphere. - //omni float depth = -xform.origin.z; - if (orthogonal) { + if (camera_orthogonal) { e.touches_near = (depth - radius) < z_near; } else { - //contains camera inside light - float radius2 = radius * shared->sphere_overfit; // overfit again for outer size (camera may be outside actual sphere but behind an icosphere vertex) + // Contains camera inside light. + float radius2 = radius * shared->sphere_overfit; // Overfit again for outer size (camera may be outside actual sphere but behind an icosphere vertex) e.touches_near = xform.origin.length_squared() < radius2 * radius2; } @@ -265,12 +269,11 @@ public: cluster_count_by_type[ELEMENT_TYPE_OMNI_LIGHT]++; - } else { - //spot - radius *= shared->cone_overfit; // overfit icosphere + } else /*LIGHT_TYPE_SPOT */ { + radius *= shared->cone_overfit; // Overfit icosphere real_t len = Math::tan(Math::deg_to_rad(p_spot_aperture)) * radius; - //approximate, probably better to use a cone support function + // Approximate, probably better to use a cone support function. float max_d = -1e20; float min_d = 1e20; #define CONE_MINMAX(m_x, m_y) \ @@ -285,14 +288,13 @@ public: CONE_MINMAX(-1, -1); CONE_MINMAX(1, -1); - if (orthogonal) { + if (camera_orthogonal) { e.touches_near = min_d < z_near; } else { - //contains camera inside light Plane base_plane(-xform.basis.get_column(Vector3::AXIS_Z), xform.origin); float dist = base_plane.distance_to(Vector3()); if (dist >= 0 && dist < radius) { - //inside, check angle + // Contains camera inside light, check angle. float angle = Math::rad_to_deg(Math::acos((-xform.origin.normalized()).dot(-xform.basis.get_column(Vector3::AXIS_Z)))); e.touches_near = angle < p_spot_aperture * 1.05; //overfit aperture a little due to cone overfit } else { @@ -302,12 +304,23 @@ public: e.touches_far = max_d > z_far; - e.scale[0] = len * shared->cone_overfit; - e.scale[1] = len * shared->cone_overfit; - e.scale[2] = radius; + // If the spot angle is above the threshold, use a sphere instead of a cone for building the clusters + // since the cone gets too flat/large (spot angle close to 90 degrees) or + // can't even cover the affected area of the light (spot angle above 90 degrees). + if (p_spot_aperture > WIDE_SPOT_ANGLE_THRESHOLD_DEG) { + e.scale[0] = radius; + e.scale[1] = radius; + e.scale[2] = radius; + e.has_wide_spot_angle = true; + } else { + e.scale[0] = len * shared->cone_overfit; + e.scale[1] = len * shared->cone_overfit; + e.scale[2] = radius; + e.has_wide_spot_angle = false; + } e.type = ELEMENT_TYPE_SPOT_LIGHT; - e.original_index = cluster_count_by_type[ELEMENT_TYPE_SPOT_LIGHT]; //use omni since they share index + e.original_index = cluster_count_by_type[ELEMENT_TYPE_SPOT_LIGHT]; // Use omni light since they share index. RendererRD::MaterialStorage::store_transform_transposed_3x4(xform, e.transform_inv); @@ -319,16 +332,16 @@ public: _FORCE_INLINE_ void add_box(BoxType p_box_type, const Transform3D &p_transform, const Vector3 &p_half_extents) { if (p_box_type == BOX_TYPE_DECAL && cluster_count_by_type[ELEMENT_TYPE_DECAL] == max_elements_by_type) { - return; //max number elements reached + return; // Max number elements reached. } if (p_box_type == BOX_TYPE_REFLECTION_PROBE && cluster_count_by_type[ELEMENT_TYPE_REFLECTION_PROBE] == max_elements_by_type) { - return; //max number elements reached + return; // Max number elements reached. } RenderElementData &e = render_elements[render_element_count]; Transform3D xform = view_xform * p_transform; - //extract scale and scale the matrix by it, makes things simpler + // Extract scale and scale the matrix by it, makes things simpler. Vector3 scale = p_half_extents; for (uint32_t i = 0; i < 3; i++) { float s = xform.basis.rows[i].length(); @@ -339,10 +352,10 @@ public: float box_depth = Math::abs(xform.basis.xform_inv(Vector3(0, 0, -1)).dot(scale)); float depth = -xform.origin.z; - if (orthogonal) { + if (camera_orthogonal) { e.touches_near = depth - box_depth < z_near; } else { - //contains camera inside box + // Contains camera inside box. Vector3 inside = xform.xform_inv(Vector3(0, 0, 0)).abs(); e.touches_near = inside.x < scale.x && inside.y < scale.y && inside.z < scale.z; } diff --git a/servers/rendering/renderer_rd/shaders/scene_forward_lights_inc.glsl b/servers/rendering/renderer_rd/shaders/scene_forward_lights_inc.glsl index b30b0c8169..9dda62c28d 100644 --- a/servers/rendering/renderer_rd/shaders/scene_forward_lights_inc.glsl +++ b/servers/rendering/renderer_rd/shaders/scene_forward_lights_inc.glsl @@ -794,8 +794,13 @@ void light_process_spot(uint idx, vec3 vertex, vec3 eye_vec, vec3 normal, vec3 v float light_length = length(light_rel_vec); float spot_attenuation = get_omni_attenuation(light_length, spot_lights.data[idx].inv_radius, spot_lights.data[idx].attenuation); vec3 spot_dir = spot_lights.data[idx].direction; - float scos = max(dot(-normalize(light_rel_vec), spot_dir), spot_lights.data[idx].cone_angle); - float spot_rim = max(0.0001, (1.0 - scos) / (1.0 - spot_lights.data[idx].cone_angle)); + + // This conversion to a highp float is crucial to prevent light leaking + // due to precision errors in the following calculations (cone angle is mediump). + highp float cone_angle = spot_lights.data[idx].cone_angle; + float scos = max(dot(-normalize(light_rel_vec), spot_dir), cone_angle); + float spot_rim = max(0.0001, (1.0 - scos) / (1.0 - cone_angle)); + spot_attenuation *= 1.0 - pow(spot_rim, spot_lights.data[idx].cone_attenuation); float light_attenuation = spot_attenuation; vec3 color = spot_lights.data[idx].color; |