Reverts max_mip_dimension to control the short side.

Author: beicause

crates/bevy_post_process/src/bloom/bloom.wgsl

@@ -52,35 +52,35 @@ fn karis_average(color: vec3<f32>) -> f32 {
 // https://www.shadertoy.com/view/mdsyDf
 #ifdef FAST_BLUR
 fn bloom_down_kernel4(uv: vec2<f32>) -> vec3<f32> {
-	let ps = uniforms.scale / vec2<f32>(textureDimensions(input_texture));
-	let o = 0.5 + 1.0 / 4.0;
+    let ps = uniforms.scale / vec2<f32>(textureDimensions(input_texture));
+    let o = 0.5 + 1.0 / 4.0;
 
-	let a = textureSample(input_texture, s, uv + vec2<f32>(-1.0, -1.0) * o * ps).rgb * 0.25;
-	let b = textureSample(input_texture, s, uv + vec2<f32>(1.0, -1.0) * o * ps).rgb * 0.25;
-	let c = textureSample(input_texture, s, uv + vec2<f32>(-1.0, 1.0) * o * ps).rgb * 0.25;
-	let d = textureSample(input_texture, s, uv + vec2<f32>(1.0, 1.0) * o * ps).rgb * 0.25;
+    let a = textureSample(input_texture, s, uv + vec2<f32>(-1.0, -1.0) * o * ps).rgb * 0.25;
+    let b = textureSample(input_texture, s, uv + vec2<f32>(1.0, -1.0) * o * ps).rgb * 0.25;
+    let c = textureSample(input_texture, s, uv + vec2<f32>(-1.0, 1.0) * o * ps).rgb * 0.25;
+    let d = textureSample(input_texture, s, uv + vec2<f32>(1.0, 1.0) * o * ps).rgb * 0.25;
 
 #ifdef FIRST_DOWNSAMPLE
-	return (a + b + c + d) * karis_average(a + b + c + d);
+    return (a + b + c + d) * karis_average(a + b + c + d);
 #else
-	return a + b + c + d;
+    return a + b + c + d;
 #endif
 }
 
 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 ps = uniforms.scale / vec2<f32>(textureDimensions(input_texture));
     let w = vec4<f32>(0.211029, 0.288971, 0.288971, 0.211029);
     // 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;
-	let c = textureSample(input_texture, s, uv + (vec2<f32>( -0.5, 0.5) + l01) * ps).rgb * w.z;
-	let d = textureSample(input_texture, s, uv + (vec2<f32>( -1.5,-0.5) + l11) * ps).rgb * w.w;
+    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;
+    let c = textureSample(input_texture, s, uv + (vec2<f32>(-0.5, 0.5) + l01) * ps).rgb * w.z;
+    let d = textureSample(input_texture, s, uv + (vec2<f32>(-1.5, -0.5) + l11) * ps).rgb * w.w;
 
     return a + b + c + d;
 }
@@ -199,7 +199,7 @@ fn sample_input_3x3_tent(uv: vec2<f32>) -> vec3<f32> {
 fn downsample_first(@location(0) output_uv: vec2<f32>) -> @location(0) vec4<f32> {
     let sample_uv = uniforms.viewport.xy + output_uv * uniforms.viewport.zw;
 #ifdef FAST_BLUR
-	var sample = bloom_down_kernel4(sample_uv);
+    var sample = bloom_down_kernel4(sample_uv);
 #else
     var sample = sample_input_13_tap(sample_uv);
 #endif
@@ -220,7 +220,7 @@ fn downsample_first(@location(0) output_uv: vec2<f32>) -> @location(0) vec4<f32>
 @fragment
 fn downsample(@location(0) uv: vec2<f32>) -> @location(0) vec4<f32> {
 #ifdef FAST_BLUR
-	return vec4<f32>(bloom_down_kernel4(uv), 1.0);
+    return vec4<f32>(bloom_down_kernel4(uv), 1.0);
 #else
     return vec4<f32>(sample_input_13_tap(uv), 1.0);
 #endif
@@ -229,7 +229,7 @@ fn downsample(@location(0) uv: vec2<f32>) -> @location(0) vec4<f32> {
 @fragment
 fn upsample(@location(0) uv: vec2<f32>) -> @location(0) vec4<f32> {
 #ifdef FAST_BLUR
-	return vec4<f32>(bloom_up_kernel4(uv), 1.0);
+    return vec4<f32>(bloom_up_kernel4(uv), 1.0);
 #else
     return vec4<f32>(sample_input_3x3_tent(uv), 1.0);
 #endif

crates/bevy_post_process/src/bloom/mod.rs

@@ -366,8 +366,9 @@ fn prepare_bloom_textures(
                 .ilog2()
                 .clamp(2, bloom.max_mip_count)
                 - 1;
+            // To better predict the mip count, `max_mip_dimension` actually controls the maximum short-side size of the bloom texture.
             let mip_dim_ratio = if viewport.y != 0 && viewport.x != 0 {
-                (bloom.max_mip_dimension as f32 / viewport.as_vec2()).max_element()
+                (bloom.max_mip_dimension as f32 / viewport.as_vec2()).min_element()
             } else {
                 0.
             };

crates/bevy_post_process/src/bloom/settings.rs

@@ -115,7 +115,7 @@ pub struct Bloom {
     /// otherwise set to [`BloomCompositeMode::EnergyConserving`].
     pub composite_mode: BloomCompositeMode,
 
-    /// Maximum size of each dimension for the largest mipchain texture used in downscaling/upscaling.
+    /// Maximum size of the short side for the largest mipchain texture used in downscaling/upscaling.
     /// Lower values can improve performance but result in more aliasing.
     pub max_mip_dimension: u32,
 

release-content/release-notes/bloom_performance_options.md

@@ -0,0 +1,11 @@
+---
+title: Add performance options to Bloom.
+authors: ["@beicause"]
+pull_requests: [21340]
+---
+
+Bloom is a relatively expensive post-processing for low-end devices, as it requires multiple render passes for downsampling and upsampling. For more performance configurability, we added the `high_quality` (default: true) and `max_mip_count` (default: unlimited) options to Bloom, in addition to the existing `max_mip_dimension`.
+
+If `high_quality` is false, Bloom will use a faster but lower quality implementation, which significantly reduces texture sampling but still maintains reasonable visual quality. For low-end devices, this could potentially reduce frame time by a few milliseconds.
+
+You can also set `max_mip_count` and/or `max_mip_dimension` to a lower value for a significant performance gain. By default the bloom texture has a maximum short-side size of 512 and uses all 8 mipmaps. You may be able to cut the Bloom frame time in half by reducing the mipmap count to a smaller value (such as 3 or 4). However, please note that these two options impact the bloom quality and need to be balanced for your needs.