Tweak solari world cache (#21776)

* Make the world cache transition LODs faster. On big scenes like
bistro, we had way too many small cells before.
* Limit world cache GI rays to 4m, we save a ton of time and improve
stability by not sampling super far away cells.
* Switch world cache active cell update workgroup size from 1024 -> 64.
Using large workgroup sizes for RT work kills occupancy.

World cache sample_radiance time goes from 1.42ms to 0.09ms.

Author: JMS55

crates/bevy_solari/src/realtime/world_cache_compact.wgsl

@@ -65,6 +65,6 @@ fn compact_world_cache_write_active_cells(
 
     if thread_index == 1023u && workgroup_id.x == 1023u {
         world_cache_active_cells_count = compacted_index + 1u; // TODO: This is 1 even when there are zero active entries in the cache
-        world_cache_active_cells_dispatch = vec3((world_cache_active_cells_count + 1023u) / 1024u, 1u, 1u);
+        world_cache_active_cells_dispatch = vec3((world_cache_active_cells_count + 63u) / 64u, 1u, 1u);
     }
 }

crates/bevy_solari/src/realtime/world_cache_query.wgsl

@@ -13,7 +13,7 @@ const WORLD_CACHE_MAX_SEARCH_STEPS: u32 = 3u;
 /// The size of a cache cell at the lowest LOD in meters
 const WORLD_CACHE_POSITION_BASE_CELL_SIZE: f32 = 0.25;
 /// How fast the world cache transitions between LODs as a function of distance to the camera
-const WORLD_CACHE_POSITION_LOD_SCALE: f32 = 30.0;
+const WORLD_CACHE_POSITION_LOD_SCALE: f32 = 8.0;
 
 /// Marker value for an empty cell
 const WORLD_CACHE_EMPTY_CELL: u32 = 0u;

crates/bevy_solari/src/realtime/world_cache_update.wgsl

@@ -3,7 +3,7 @@
 #import bevy_render::view::View
 #import bevy_solari::presample_light_tiles::{ResolvedLightSamplePacked, unpack_resolved_light_sample}
 #import bevy_solari::sampling::{calculate_resolved_light_contribution, trace_light_visibility}
-#import bevy_solari::scene_bindings::{trace_ray, resolve_ray_hit_full, RAY_T_MIN, RAY_T_MAX}
+#import bevy_solari::scene_bindings::{trace_ray, resolve_ray_hit_full, RAY_T_MIN}
 #import bevy_solari::world_cache::{
     WORLD_CACHE_MAX_TEMPORAL_SAMPLES,
     query_world_cache,
@@ -20,8 +20,9 @@ struct PushConstants { frame_index: u32, reset: u32 }
 var<push_constant> constants: PushConstants;
 
 const DIRECT_LIGHT_SAMPLE_COUNT: u32 = 32u;
+const MAX_GI_RAY_DISTANCE: f32 = 4.0;
 
-@compute @workgroup_size(1024, 1, 1)
+@compute @workgroup_size(64, 1, 1)
 fn sample_radiance(@builtin(workgroup_id) workgroup_id: vec3<u32>, @builtin(global_invocation_id) active_cell_id: vec3<u32>) {
     if active_cell_id.x < world_cache_active_cells_count {
         let cell_index = world_cache_active_cell_indices[active_cell_id.x];
@@ -34,7 +35,7 @@ fn sample_radiance(@builtin(workgroup_id) workgroup_id: vec3<u32>, @builtin(glob
 
 #ifndef NO_MULTIBOUNCE
         let ray_direction = sample_cosine_hemisphere(geometry_data.world_normal, &rng);
-        let ray_hit = trace_ray(geometry_data.world_position, ray_direction, RAY_T_MIN, RAY_T_MAX, RAY_FLAG_NONE);
+        let ray_hit = trace_ray(geometry_data.world_position, ray_direction, RAY_T_MIN, MAX_GI_RAY_DISTANCE, RAY_FLAG_NONE);
         if ray_hit.kind != RAY_QUERY_INTERSECTION_NONE {
             let ray_hit = resolve_ray_hit_full(ray_hit);
             new_radiance += ray_hit.material.base_color * query_world_cache(ray_hit.world_position, ray_hit.geometric_world_normal, view.world_position, &rng);
@@ -45,7 +46,7 @@ fn sample_radiance(@builtin(workgroup_id) workgroup_id: vec3<u32>, @builtin(glob
     }
 }
 
-@compute @workgroup_size(1024, 1, 1)
+@compute @workgroup_size(64, 1, 1)
 fn blend_new_samples(@builtin(global_invocation_id) active_cell_id: vec3<u32>) {
     if active_cell_id.x < world_cache_active_cells_count {
         let cell_index = world_cache_active_cell_indices[active_cell_id.x];