perf(wire): pack instance color as Unorm8x4 (88 -> 76 B per instance)

Step 4 (final) of the wire pipeline cold-open optimization plan. Color
shrinks from `[f32; 4]` (16 B) to `[u8; 4]` (4 B) and the vertex attribute
format flips to `Unorm8x4`; the GPU still hands the shader a `vec4<f32>`
in [0, 1], so wire.wgsl needs no change. 8-bit per channel is below
display precision.

Net: 76 B per instance, ~14 % smaller than the Step-3 instance layout and
~7.6× smaller than the pre-Step-3 per-vertex format. Wire upload memory
and PCIe bandwidth on cold open scale down accordingly.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
HakanandClaude Opus 4.7 committed 2026-05-23 09:53:49 +03:00
1 parent 4b4f5c94cd
commit a3c982c4f3
1 file changed
+31 -18
+31 -18
View File
@@ -9,18 +9,18 @@
// WireModel, separated by [NaN, NaN, NaN] points. Segments where either // WireModel, separated by [NaN, NaN, NaN] points. Segments where either
// endpoint contains NaN are silently skipped during emission. // endpoint contains NaN are silently skipped during emission.
// //
// Instance layout (88 bytes, stride = 88, step_mode = Instance): // Instance layout (76 bytes, stride = 76, step_mode = Instance):
// pos_a [f32; 3] offset 0 12 B — segment start (world) // pos_a [f32; 3] offset 0 12 B — segment start (world)
// pos_b [f32; 3] offset 12 12 B — segment end (world) // pos_b [f32; 3] offset 12 12 B — segment end (world)
// color [f32; 4] offset 24 16 B — RGBA [0,1] // color [u8; 4] offset 24 4 B — RGBA, Unorm8x4 → vec4<f32> in shader
// distance_a f32 offset 40 4 B — arc-length at endpoint A // distance_a f32 offset 28 4 B — arc-length at endpoint A
// distance_b f32 offset 44 4 B — arc-length at endpoint B // distance_b f32 offset 32 4 B — arc-length at endpoint B
// half_width f32 offset 48 4 B — half line width in pixels // half_width f32 offset 36 4 B — half line width in pixels
// pattern_length f32 offset 52 4 B — dash pattern total length // pattern_length f32 offset 40 4 B — dash pattern total length
// pat0 [f32; 4] offset 56 16 B — pattern elements 0-3 // pat0 [f32; 4] offset 44 16 B — pattern elements 0-3
// pat1 [f32; 4] offset 72 16 B — pattern elements 4-7 // pat1 [f32; 4] offset 60 16 B — pattern elements 4-7
// ------ // ------
// 88 B / instance // 76 B / instance
use crate::scene::wire_model::WireModel; use crate::scene::wire_model::WireModel;
use iced::wgpu; use iced::wgpu;
@@ -61,7 +61,10 @@ fn instance_buffer_mapped(
pub struct WireInstance { pub struct WireInstance {
pub pos_a: [f32; 3], pub pos_a: [f32; 3],
pub pos_b: [f32; 3], pub pos_b: [f32; 3],
pub color: [f32; 4], /// RGBA packed as `Unorm8x4` — the vertex shader receives a `vec4<f32>`
/// in [0, 1] after the GPU does the conversion. 8 bits per channel is
/// indistinguishable from f32 at 8-bit display output.
pub color: [u8; 4],
pub distance_a: f32, pub distance_a: f32,
pub distance_b: f32, pub distance_b: f32,
pub half_width: f32, pub half_width: f32,
@@ -89,35 +92,35 @@ impl WireInstance {
wgpu::VertexAttribute { wgpu::VertexAttribute {
offset: 24, offset: 24,
shader_location: 2, shader_location: 2,
format: wgpu::VertexFormat::Float32x4, format: wgpu::VertexFormat::Unorm8x4,
}, // color }, // color
wgpu::VertexAttribute { wgpu::VertexAttribute {
offset: 40, offset: 28,
shader_location: 3, shader_location: 3,
format: wgpu::VertexFormat::Float32, format: wgpu::VertexFormat::Float32,
}, // distance_a }, // distance_a
wgpu::VertexAttribute { wgpu::VertexAttribute {
offset: 44, offset: 32,
shader_location: 4, shader_location: 4,
format: wgpu::VertexFormat::Float32, format: wgpu::VertexFormat::Float32,
}, // distance_b }, // distance_b
wgpu::VertexAttribute { wgpu::VertexAttribute {
offset: 48, offset: 36,
shader_location: 5, shader_location: 5,
format: wgpu::VertexFormat::Float32, format: wgpu::VertexFormat::Float32,
}, // half_width }, // half_width
wgpu::VertexAttribute { wgpu::VertexAttribute {
offset: 52, offset: 40,
shader_location: 6, shader_location: 6,
format: wgpu::VertexFormat::Float32, format: wgpu::VertexFormat::Float32,
}, // pattern_length }, // pattern_length
wgpu::VertexAttribute { wgpu::VertexAttribute {
offset: 56, offset: 44,
shader_location: 7, shader_location: 7,
format: wgpu::VertexFormat::Float32x4, format: wgpu::VertexFormat::Float32x4,
}, // pat0 }, // pat0
wgpu::VertexAttribute { wgpu::VertexAttribute {
offset: 72, offset: 60,
shader_location: 8, shader_location: 8,
format: wgpu::VertexFormat::Float32x4, format: wgpu::VertexFormat::Float32x4,
}, // pat1 }, // pat1
@@ -140,7 +143,17 @@ pub struct WireGpu {
/// finite segment). Pulled out so both the single-wire and batched paths share /// finite segment). Pulled out so both the single-wire and batched paths share
/// the same emission logic, and so the batched path can `par_iter` across /// the same emission logic, and so the batched path can `par_iter` across
/// wires on cold open. /// wires on cold open.
fn pack_color(color: [f32; 4]) -> [u8; 4] {
[
(color[0].clamp(0.0, 1.0) * 255.0 + 0.5) as u8,
(color[1].clamp(0.0, 1.0) * 255.0 + 0.5) as u8,
(color[2].clamp(0.0, 1.0) * 255.0 + 0.5) as u8,
(color[3].clamp(0.0, 1.0) * 255.0 + 0.5) as u8,
]
}
fn emit_wire_instances(wire: &WireModel, color: [f32; 4]) -> Vec<WireInstance> { fn emit_wire_instances(wire: &WireModel, color: [f32; 4]) -> Vec<WireInstance> {
let color_u8 = pack_color(color);
let pat0 = [ let pat0 = [
wire.pattern[0], wire.pattern[0],
wire.pattern[1], wire.pattern[1],
@@ -196,7 +209,7 @@ fn emit_wire_instances(wire: &WireModel, color: [f32; 4]) -> Vec<WireInstance> {
instances.push(WireInstance { instances.push(WireInstance {
pos_a: a, pos_a: a,
pos_b: b, pos_b: b,
color, color: color_u8,
distance_a: dists[i], distance_a: dists[i],
distance_b: dists[i + 1], distance_b: dists[i + 1],
half_width, half_width,