tweak up kernel, mix ClampToBorder and ClampToEdge

Author: beicause

crates/bevy_post_process/src/bloom/bloom.wgsl

@@ -68,12 +68,14 @@ fn bloom_down_kernel4(uv: vec2<f32>) -> vec3<f32> {
 }
 
 fn bloom_up_kernel4(uv: vec2<f32>) -> vec3<f32> {
+	// Modified version of BloomUpKernel4B https://www.shadertoy.com/view/mdsyDf. I couldn't get a good result with the original version.
 	let ps = uniforms.scale / vec2<f32>(textureDimensions(input_texture));
     let w = vec4<f32>(0.211029, 0.288971, 0.288971, 0.211029);
-    let l00 = vec2<f32>(0.347209, 0.526425);
-    let l10 = vec2<f32>(0.109840, 0.334045);
-    let l01 = vec2<f32>(0.334045, 0.109840);
-    let l11 = vec2<f32>(0.526425, 0.347209);
+    // Add a small offset for better radial symmetry.
+    let l00 = vec2<f32>(0.347209, 0.526425) + 0.1;
+    let l10 = vec2<f32>(0.109840, 0.334045) + 0.1;
+    let l01 = vec2<f32>(0.334045, 0.109840) + 0.1;
+    let l11 = vec2<f32>(0.526425, 0.347209) + 0.1;
 
 	let a = textureSample(input_texture, s, uv + (vec2<f32>( -0.5, -1.5) + l00) * ps).rgb * w.x;
 	let b = textureSample(input_texture, s, uv + (vec2<f32>(0.5, -0.5) + l10) * ps).rgb * w.y;

crates/bevy_post_process/src/bloom/mod.rs

@@ -181,7 +181,7 @@ impl ViewNode for BloomNode {
                     &BindGroupEntries::sequential((
                         // Read from main texture directly
                         view_texture,
-                        &bind_groups.sampler,
+                        &bind_groups.downsample_sampler,
                         uniforms.clone(),
                     )),
                 );
@@ -418,7 +418,7 @@ fn prepare_bloom_textures(
 struct BloomBindGroups {
     downsampling_bind_groups: Box<[BindGroup]>,
     upsampling_bind_groups: Box<[BindGroup]>,
-    sampler: Sampler,
+    downsample_sampler: Sampler,
 }
 
 fn prepare_bloom_bind_groups(
@@ -429,8 +429,6 @@ fn prepare_bloom_bind_groups(
     views: Query<(Entity, &BloomTexture)>,
     uniforms: Res<ComponentUniforms<BloomUniforms>>,
 ) {
-    let sampler = &downsampling_pipeline.sampler;
-
     for (entity, bloom_texture) in &views {
         let bind_group_count = bloom_texture.mip_count as usize - 1;
 
@@ -441,7 +439,7 @@ fn prepare_bloom_bind_groups(
                 &downsampling_pipeline.bind_group_layout,
                 &BindGroupEntries::sequential((
                     &bloom_texture.view(mip - 1),
-                    sampler,
+                    &downsampling_pipeline.sampler,
                     uniforms.binding().unwrap(),
                 )),
             ));
@@ -454,7 +452,7 @@ fn prepare_bloom_bind_groups(
                 &upsampling_pipeline.bind_group_layout,
                 &BindGroupEntries::sequential((
                     &bloom_texture.view(mip),
-                    sampler,
+                    &upsampling_pipeline.sampler,
                     uniforms.binding().unwrap(),
                 )),
             ));
@@ -463,7 +461,7 @@ fn prepare_bloom_bind_groups(
         commands.entity(entity).insert(BloomBindGroups {
             downsampling_bind_groups: downsampling_bind_groups.into_boxed_slice(),
             upsampling_bind_groups: upsampling_bind_groups.into_boxed_slice(),
-            sampler: sampler.clone(),
+            downsample_sampler: downsampling_pipeline.sampler.clone(),
         });
     }
 }

crates/bevy_post_process/src/bloom/upsampling_pipeline.rs

@@ -29,6 +29,7 @@ pub struct UpsamplingPipelineIds {
 #[derive(Resource)]
 pub struct BloomUpsamplingPipeline {
     pub bind_group_layout: BindGroupLayout,
+    pub sampler: Sampler,
     /// The asset handle for the fullscreen vertex shader.
     pub fullscreen_shader: FullscreenShader,
     /// The fragment shader asset handle.
@@ -63,8 +64,18 @@ pub fn init_bloom_upscaling_pipeline(
         ),
     );
 
+    // Sampler
+    let sampler = render_device.create_sampler(&SamplerDescriptor {
+        min_filter: FilterMode::Linear,
+        mag_filter: FilterMode::Linear,
+        address_mode_u: AddressMode::ClampToEdge,
+        address_mode_v: AddressMode::ClampToEdge,
+        ..Default::default()
+    });
+
     commands.insert_resource(BloomUpsamplingPipeline {
         bind_group_layout,
+        sampler,
         fullscreen_shader: fullscreen_shader.clone(),
         fragment_shader: load_embedded_asset!(asset_server.as_ref(), "bloom.wgsl"),
     });

release-content/release-notes/faster_bloom.md

@@ -0,0 +1,7 @@
+---
+title: Add a faster bloom implementation.
+authors: ["@beicause"]
+pull_requests: [21340]
+---
+
+Bloom now has a `high_quality` (default: true) option to control whether to use a high quality implementation, or a faster but lower quality implementation. The lower quality bloom still maintains reasonable visual quality while significantly reducing texture sampling. For low-end devices, this could potentially reduce frame time by a few milliseconds.