Merge pull request #1885 from bungoboingo/gradient-packing-optimization

Small gradient optimization

graphics/Cargo.toml

@@ -18,6 +18,7 @@ web-colors = []
 
 [dependencies]
 glam = "0.24"
+half = "2.2.1"
 log = "0.4"
 raw-window-handle = "0.5"
 thiserror = "1.0"

graphics/src/gradient.rs

@@ -7,6 +7,7 @@ use crate::color;
 use crate::core::gradient::ColorStop;
 use crate::core::{self, Color, Point, Rectangle};
 
+use half::f16;
 use std::cmp::Ordering;
 
 #[derive(Debug, Clone, PartialEq)]
@@ -99,61 +100,96 @@ impl Linear {
 
     /// Packs the [`Gradient`] for use in shader code.
     pub fn pack(&self) -> Packed {
-        let mut data: [f32; 44] = [0.0; 44];
+        let mut colors = [[0u32; 2]; 8];
+        let mut offsets = [f16::from(0u8); 8];
 
         for (index, stop) in self.stops.iter().enumerate() {
             let [r, g, b, a] =
                 color::pack(stop.map_or(Color::default(), |s| s.color))
                     .components();
 
-            data[index * 4] = r;
+            colors[index] = [
-            data[(index * 4) + 1] = g;
+                pack_f16s([f16::from_f32(r), f16::from_f32(g)]),
-            data[(index * 4) + 2] = b;
+                pack_f16s([f16::from_f32(b), f16::from_f32(a)]),
-            data[(index * 4) + 3] = a;
+            ];
 
-            data[32 + index] = stop.map_or(2.0, |s| s.offset);
+            offsets[index] =
+                stop.map_or(f16::from_f32(2.0), |s| f16::from_f32(s.offset));
         }
 
-        data[40] = self.start.x;
+        let offsets = [
-        data[41] = self.start.y;
+            pack_f16s([offsets[0], offsets[1]]),
-        data[42] = self.end.x;
+            pack_f16s([offsets[2], offsets[3]]),
-        data[43] = self.end.y;
+            pack_f16s([offsets[4], offsets[5]]),
+            pack_f16s([offsets[6], offsets[7]]),
+        ];
 
-        Packed(data)
+        let direction = [self.start.x, self.start.y, self.end.x, self.end.y];
+
+        Packed {
+            colors,
+            offsets,
+            direction,
+        }
     }
 }
 
 /// Packed [`Gradient`] data for use in shader code.
 #[derive(Debug, Copy, Clone, PartialEq)]
 #[repr(C)]
-pub struct Packed([f32; 44]);
+pub struct Packed {
+    // 8 colors, each channel = 16 bit float, 2 colors packed into 1 u32
+    colors: [[u32; 2]; 8],
+    // 8 offsets, 8x 16 bit floats packed into 4 u32s
+    offsets: [u32; 4],
+    direction: [f32; 4],
+}
 
 /// Creates a new [`Packed`] gradient for use in shader code.
 pub fn pack(gradient: &core::Gradient, bounds: Rectangle) -> Packed {
     match gradient {
         core::Gradient::Linear(linear) => {
-            let mut data: [f32; 44] = [0.0; 44];
+            let mut colors = [[0u32; 2]; 8];
+            let mut offsets = [f16::from(0u8); 8];
 
             for (index, stop) in linear.stops.iter().enumerate() {
                 let [r, g, b, a] =
                     color::pack(stop.map_or(Color::default(), |s| s.color))
                         .components();
 
-                data[index * 4] = r;
+                colors[index] = [
-                data[(index * 4) + 1] = g;
+                    pack_f16s([f16::from_f32(r), f16::from_f32(g)]),
-                data[(index * 4) + 2] = b;
+                    pack_f16s([f16::from_f32(b), f16::from_f32(a)]),
-                data[(index * 4) + 3] = a;
+                ];
-                data[32 + index] = stop.map_or(2.0, |s| s.offset);
+
+                offsets[index] = stop
+                    .map_or(f16::from_f32(2.0), |s| f16::from_f32(s.offset));
             }
 
+            let offsets = [
+                pack_f16s([offsets[0], offsets[1]]),
+                pack_f16s([offsets[2], offsets[3]]),
+                pack_f16s([offsets[4], offsets[5]]),
+                pack_f16s([offsets[6], offsets[7]]),
+            ];
+
             let (start, end) = linear.angle.to_distance(&bounds);
 
-            data[40] = start.x;
+            let direction = [start.x, start.y, end.x, end.y];
-            data[41] = start.y;
-            data[42] = end.x;
-            data[43] = end.y;
 
-            Packed(data)
+            Packed {
+                colors,
+                offsets,
+                direction,
+            }
         }
     }
 }
+
+/// Packs two f16s into one u32.
+fn pack_f16s(f: [f16; 2]) -> u32 {
+    let one = (f[0].to_bits() as u32) << 16;
+    let two = f[1].to_bits() as u32;
+
+    one | two
+}

wgpu/src/quad/gradient.rs

@@ -96,36 +96,26 @@ impl Pipeline {
                                 as u64,
                             step_mode: wgpu::VertexStepMode::Instance,
                             attributes: &wgpu::vertex_attr_array!(
-                                // Color 1
+                                // Colors 1-2
-                                1 => Float32x4,
+                                1 => Uint32x4,
-                                // Color 2
+                                // Colors 3-4
-                                2 => Float32x4,
+                                2 => Uint32x4,
-                                // Color 3
+                                // Colors 5-6
-                                3 => Float32x4,
+                                3 => Uint32x4,
-                                // Color 4
+                                // Colors 7-8
-                                4 => Float32x4,
+                                4 => Uint32x4,
-                                // Color 5
+                                // Offsets 1-8
-                                5 => Float32x4,
+                                5 => Uint32x4,
-                                // Color 6
-                                6 => Float32x4,
-                                // Color 7
-                                7 => Float32x4,
-                                // Color 8
-                                8 => Float32x4,
-                                // Offsets 1-4
-                                9 => Float32x4,
-                                // Offsets 5-8
-                                10 => Float32x4,
                                 // Direction
-                                11 => Float32x4,
+                                6 => Float32x4,
                                 // Position & Scale
-                                12 => Float32x4,
+                                7 => Float32x4,
                                 // Border color
-                                13 => Float32x4,
+                                8 => Float32x4,
                                 // Border radius
-                                14 => Float32x4,
+                                9 => Float32x4,
                                 // Border width
-                                15 => Float32
+                                10 => Float32
                             ),
                         },
                     ],

wgpu/src/shader/quad.wgsl

@@ -38,6 +38,13 @@ fn select_border_radius(radi: vec4<f32>, position: vec2<f32>, center: vec2<f32>)
     return rx;
 }
 
+fn unpack_u32(color: vec2<u32>) -> vec4<f32> {
+    let rg: vec2<f32> = unpack2x16float(color.x);
+    let ba: vec2<f32> = unpack2x16float(color.y);
+
+    return vec4<f32>(rg.y, rg.x, ba.y, ba.x);
+}
+
 struct SolidVertexInput {
     @location(0) v_pos: vec2<f32>,
     @location(1) color: vec4<f32>,
@@ -140,40 +147,30 @@ fn solid_fs_main(
 
 struct GradientVertexInput {
     @location(0) v_pos: vec2<f32>,
-    @location(1) color_1: vec4<f32>,
+    @location(1) colors_1: vec4<u32>,
-    @location(2) color_2: vec4<f32>,
+    @location(2) colors_2: vec4<u32>,
-    @location(3) color_3: vec4<f32>,
+    @location(3) colors_3: vec4<u32>,
-    @location(4) color_4: vec4<f32>,
+    @location(4) colors_4: vec4<u32>,
-    @location(5) color_5: vec4<f32>,
+    @location(5) offsets: vec4<u32>,
-    @location(6) color_6: vec4<f32>,
+    @location(6) direction: vec4<f32>,
-    @location(7) color_7: vec4<f32>,
+    @location(7) position_and_scale: vec4<f32>,
-    @location(8) color_8: vec4<f32>,
+    @location(8) border_color: vec4<f32>,
-    @location(9) offsets_1: vec4<f32>,
+    @location(9) border_radius: vec4<f32>,
-    @location(10) offsets_2: vec4<f32>,
+    @location(10) border_width: f32,
-    @location(11) direction: vec4<f32>,
-    @location(12) position_and_scale: vec4<f32>,
-    @location(13) border_color: vec4<f32>,
-    @location(14) border_radius: vec4<f32>,
-    @location(15) border_width: f32
 }
 
 struct GradientVertexOutput {
     @builtin(position) position: vec4<f32>,
-    @location(1) color_1: vec4<f32>,
+    @location(1) colors_1: vec4<u32>,
-    @location(2) color_2: vec4<f32>,
+    @location(2) colors_2: vec4<u32>,
-    @location(3) color_3: vec4<f32>,
+    @location(3) colors_3: vec4<u32>,
-    @location(4) color_4: vec4<f32>,
+    @location(4) colors_4: vec4<u32>,
-    @location(5) color_5: vec4<f32>,
+    @location(5) offsets: vec4<u32>,
-    @location(6) color_6: vec4<f32>,
+    @location(6) direction: vec4<f32>,
-    @location(7) color_7: vec4<f32>,
+    @location(7) position_and_scale: vec4<f32>,
-    @location(8) color_8: vec4<f32>,
+    @location(8) border_color: vec4<f32>,
-    @location(9) offsets_1: vec4<f32>,
+    @location(9) border_radius: vec4<f32>,
-    @location(10) offsets_2: vec4<f32>,
+    @location(10) border_width: f32,
-    @location(11) direction: vec4<f32>,
-    @location(12) position_and_scale: vec4<f32>,
-    @location(13) border_color: vec4<f32>,
-    @location(14) border_radius: vec4<f32>,
-    @location(15) border_width: f32
 }
 
 @vertex
@@ -199,16 +196,11 @@ fn gradient_vs_main(input: GradientVertexInput) -> GradientVertexOutput {
     );
 
     out.position = globals.transform * transform * vec4<f32>(input.v_pos, 0.0, 1.0);
-    out.color_1 = input.color_1;
+    out.colors_1 = input.colors_1;
-    out.color_2 = input.color_2;
+    out.colors_2 = input.colors_2;
-    out.color_3 = input.color_3;
+    out.colors_3 = input.colors_3;
-    out.color_4 = input.color_4;
+    out.colors_4 = input.colors_4;
-    out.color_5 = input.color_5;
+    out.offsets = input.offsets;
-    out.color_6 = input.color_6;
-    out.color_7 = input.color_7;
-    out.color_8 = input.color_8;
-    out.offsets_1 = input.offsets_1;
-    out.offsets_2 = input.offsets_2;
     out.direction = input.direction * globals.scale;
     out.position_and_scale = vec4<f32>(pos, scale);
     out.border_color = input.border_color;
@@ -274,25 +266,28 @@ fn gradient(
 @fragment
 fn gradient_fs_main(input: GradientVertexOutput) -> @location(0) vec4<f32> {
     let colors = array<vec4<f32>, 8>(
-        input.color_1,
+        unpack_u32(input.colors_1.xy),
-        input.color_2,
+        unpack_u32(input.colors_1.zw),
-        input.color_3,
+        unpack_u32(input.colors_2.xy),
-        input.color_4,
+        unpack_u32(input.colors_2.zw),
-        input.color_5,
+        unpack_u32(input.colors_3.xy),
-        input.color_6,
+        unpack_u32(input.colors_3.zw),
-        input.color_7,
+        unpack_u32(input.colors_4.xy),
-        input.color_8,
+        unpack_u32(input.colors_4.zw),
     );
 
+    let offsets_1: vec4<f32> = unpack_u32(input.offsets.xy);
+    let offsets_2: vec4<f32> = unpack_u32(input.offsets.zw);
+
     var offsets = array<f32, 8>(
-        input.offsets_1.x,
+        offsets_1.x,
-        input.offsets_1.y,
+        offsets_1.y,
-        input.offsets_1.z,
+        offsets_1.z,
-        input.offsets_1.w,
+        offsets_1.w,
-        input.offsets_2.x,
+        offsets_2.x,
-        input.offsets_2.y,
+        offsets_2.y,
-        input.offsets_2.z,
+        offsets_2.z,
-        input.offsets_2.w,
+        offsets_2.w,
     );
 
     //TODO could just pass this in to the shader but is probably more performant to just check it here

wgpu/src/shader/triangle.wgsl

@@ -4,6 +4,13 @@ struct Globals {
 
 @group(0) @binding(0) var<uniform> globals: Globals;
 
+fn unpack_u32(color: vec2<u32>) -> vec4<f32> {
+    let rg: vec2<f32> = unpack2x16float(color.x);
+    let ba: vec2<f32> = unpack2x16float(color.y);
+
+    return vec4<f32>(rg.y, rg.x, ba.y, ba.x);
+}
+
 struct SolidVertexInput {
     @location(0) position: vec2<f32>,
     @location(1) color: vec4<f32>,
@@ -29,52 +36,39 @@ fn solid_fs_main(input: SolidVertexOutput) -> @location(0) vec4<f32> {
     return input.color;
 }
 
+struct GradientVertexInput {
+    @location(0) v_pos: vec2<f32>,
+    @location(1) colors_1: vec4<u32>,
+    @location(2) colors_2: vec4<u32>,
+    @location(3) colors_3: vec4<u32>,
+    @location(4) colors_4: vec4<u32>,
+    @location(5) offsets: vec4<u32>,
+    @location(6) direction: vec4<f32>,
+}
+
 struct GradientVertexOutput {
     @builtin(position) position: vec4<f32>,
     @location(0) raw_position: vec2<f32>,
-    @location(1) color_1: vec4<f32>,
+    @location(1) colors_1: vec4<u32>,
-    @location(2) color_2: vec4<f32>,
+    @location(2) colors_2: vec4<u32>,
-    @location(3) color_3: vec4<f32>,
+    @location(3) colors_3: vec4<u32>,
-    @location(4) color_4: vec4<f32>,
+    @location(4) colors_4: vec4<u32>,
-    @location(5) color_5: vec4<f32>,
+    @location(5) offsets: vec4<u32>,
-    @location(6) color_6: vec4<f32>,
+    @location(6) direction: vec4<f32>,
-    @location(7) color_7: vec4<f32>,
-    @location(8) color_8: vec4<f32>,
-    @location(9) offsets_1: vec4<f32>,
-    @location(10) offsets_2: vec4<f32>,
-    @location(11) direction: vec4<f32>,
 }
 
 @vertex
-fn gradient_vs_main(
+fn gradient_vs_main(input: GradientVertexInput) -> GradientVertexOutput {
-    @location(0) input: vec2<f32>,
-    @location(1) color_1: vec4<f32>,
-    @location(2) color_2: vec4<f32>,
-    @location(3) color_3: vec4<f32>,
-    @location(4) color_4: vec4<f32>,
-    @location(5) color_5: vec4<f32>,
-    @location(6) color_6: vec4<f32>,
-    @location(7) color_7: vec4<f32>,
-    @location(8) color_8: vec4<f32>,
-    @location(9) offsets_1: vec4<f32>,
-    @location(10) offsets_2: vec4<f32>,
-    @location(11) direction: vec4<f32>,
-) -> GradientVertexOutput {
     var output: GradientVertexOutput;
 
-    output.position = globals.transform * vec4<f32>(input.xy, 0.0, 1.0);
+    output.position = globals.transform * vec4<f32>(input.v_pos, 0.0, 1.0);
-    output.raw_position = input;
+    output.raw_position = input.v_pos;
-    output.color_1 = color_1;
+    output.colors_1 = input.colors_1;
-    output.color_2 = color_2;
+    output.colors_2 = input.colors_2;
-    output.color_3 = color_3;
+    output.colors_3 = input.colors_3;
-    output.color_4 = color_4;
+    output.colors_4 = input.colors_4;
-    output.color_5 = color_5;
+    output.offsets = input.offsets;
-    output.color_6 = color_6;
+    output.direction = input.direction;
-    output.color_7 = color_7;
-    output.color_8 = color_8;
-    output.offsets_1 = offsets_1;
-    output.offsets_2 = offsets_2;
-    output.direction = direction;
 
     return output;
 }
@@ -135,25 +129,28 @@ fn gradient(
 @fragment
 fn gradient_fs_main(input: GradientVertexOutput) -> @location(0) vec4<f32> {
     let colors = array<vec4<f32>, 8>(
-        input.color_1,
+        unpack_u32(input.colors_1.xy),
-        input.color_2,
+        unpack_u32(input.colors_1.zw),
-        input.color_3,
+        unpack_u32(input.colors_2.xy),
-        input.color_4,
+        unpack_u32(input.colors_2.zw),
-        input.color_5,
+        unpack_u32(input.colors_3.xy),
-        input.color_6,
+        unpack_u32(input.colors_3.zw),
-        input.color_7,
+        unpack_u32(input.colors_4.xy),
-        input.color_8,
+        unpack_u32(input.colors_4.zw),
     );
 
+    let offsets_1: vec4<f32> = unpack_u32(input.offsets.xy);
+    let offsets_2: vec4<f32> = unpack_u32(input.offsets.zw);
+
     var offsets = array<f32, 8>(
-        input.offsets_1.x,
+        offsets_1.x,
-        input.offsets_1.y,
+        offsets_1.y,
-        input.offsets_1.z,
+        offsets_1.z,
-        input.offsets_1.w,
+        offsets_1.w,
-        input.offsets_2.x,
+        offsets_2.x,
-        input.offsets_2.y,
+        offsets_2.y,
-        input.offsets_2.z,
+        offsets_2.z,
-        input.offsets_2.w,
+        offsets_2.w,
     );
 
     var last_index = 7;

wgpu/src/triangle.rs

@@ -652,28 +652,18 @@ mod gradient {
                             attributes: &wgpu::vertex_attr_array!(
                                 // Position
                                 0 => Float32x2,
-                                // Color 1
+                                // Colors 1-2
-                                1 => Float32x4,
+                                1 => Uint32x4,
-                                // Color 2
+                                // Colors 3-4
-                                2 => Float32x4,
+                                2 => Uint32x4,
-                                // Color 3
+                                // Colors 5-6
-                                3 => Float32x4,
+                                3 => Uint32x4,
-                                // Color 4
+                                // Colors 7-8
-                                4 => Float32x4,
+                                4 => Uint32x4,
-                                // Color 5
+                                // Offsets
-                                5 => Float32x4,
+                                5 => Uint32x4,
-                                // Color 6
-                                6 => Float32x4,
-                                // Color 7
-                                7 => Float32x4,
-                                // Color 8
-                                8 => Float32x4,
-                                // Offsets 1-4
-                                9 => Float32x4,
-                                // Offsets 5-8
-                                10 => Float32x4,
                                 // Direction
-                                11 => Float32x4
+                                6 => Float32x4
                             ),
                         }],
                     },