Merge pull request #2463 from HEOJUNFO/fix/classic3d-shadow-diffusion

Classic 3D Shadow Diffusion blurs instead of stamping copies (#2444)

crates/gpu/src/shaders/classic3d.wgsl

@@ -199,6 +199,12 @@ fn attenuation(l: u32, p: vec3<f32>) -> f32 {
     return k;
 }
 
+// compose::shadow_taps: samples per side of the Shadow Diffusion kernel.
+fn shadow_taps(r_px: f32) -> u32 {
+    let n = clamp(u32(ceil(2.0 * r_px / 1.5)), 3u, 15u);
+    return n | 1u;
+}
+
 // compose::occlusion: RGB transmittance of light `l` reaching `p` past caster `c`.
 fn occlusion(c: u32, p: vec3<f32>, l: u32) -> vec3<f32> {
     let pl = planes[c];
@@ -243,9 +249,14 @@ fn occlusion(c: u32, p: vec3<f32>, l: u32) -> vec3<f32> {
     if (r * pl.tex.z < 0.5) {
         px = texel(c, hit.x, hit.y);
     } else {
-        for (var j = -1; j <= 1; j++) {
-            for (var i = -1; i <= 1; i++) {
-                px += texel(c, hit.x + f32(i) * r * 0.66, hit.y + f32(j) * r * 0.66) / 9.0;
+        // compose::shadow_taps: n × n box filter over the penumbra square.
+        let n = shadow_taps(r * pl.tex.z);
+        let step = 2.0 * r / f32(n);
+        let h = i32(n / 2u);
+        let w = 1.0 / f32(n * n);
+        for (var j = -h; j <= h; j++) {
+            for (var i = -h; i <= h; i++) {
+                px += texel(c, hit.x + f32(i) * step, hit.y + f32(j) * step) * w;
             }
         }
     }

crates/gpu/src/tests/three_d.rs

@@ -133,6 +133,22 @@ fn lights_specular_transmission_and_shadows() {
     }
 }
 
+#[test]
+fn wide_shadow_diffusion() {
+    // #2444: a wide penumbra takes more than the 3 × 3 kernel's samples; both sides pick the
+    // same count.
+    for depth in [BitDepth::Bpc8, BitDepth::Bpc32] {
+        let mut s = crossing(depth, BlendMode::Normal);
+        for l in &mut s.p.comp_mut(s.cid).unwrap().layers {
+            if l.switches.three_d {
+                set(l, "materialOptions/castsShadows", Value::Enum(1));
+            }
+        }
+        light(&mut s, LightKind::Point, [150.0, 20.0, -80.0], &[("castsShadows", Value::Bool(true)), ("shadowDiffusion", n(60.0))]);
+        check(&format!("wide shadow diffusion {depth:?}"), compare_at(&s, opts(), Tick::ZERO), 0.005);
+    }
+}
+
 #[test]
 fn depth_of_field_and_motion_blur() {
     for depth in [BitDepth::Bpc8, BitDepth::Bpc32] {

crates/render/src/three_d/compose.rs

@@ -645,6 +645,13 @@ fn shade_texel(it: &Item, g: &Geo, u: f64, v: f64, t: Px, lights: &[LightState],
     [lit[0] * a, lit[1] * a, lit[2] * a, a]
 }
 
+/// Samples per side of the Shadow Diffusion kernel for a penumbra radius of `r_px` caster pixels:
+/// odd, at least 3, and enough that neighbouring samples are at most 1.5 pixels apart, up to 15.
+pub(crate) fn shadow_taps(r_px: f64) -> u32 {
+    let n = ((2.0 * r_px / 1.5).ceil() as u32).clamp(3, 15);
+    n | 1
+}
+
 /// RGB transmittance of light `l` reaching `p` past caster `c` (1 = unoccluded).
 fn occlusion(c: &Item, p: Vec3, l: &LightState) -> [f32; 3] {
     const FAR: f64 = 1.0e6;
@@ -672,11 +679,17 @@ fn occlusion(c: &Item, p: Vec3, l: &LightState) -> [f32; 3] {
     if r * c.buf.scale < 0.5 {
         px = c.texel(hit.x, hit.y);
     } else {
-        for j in -1..=1 {
-            for i in -1..=1 {
-                let q = c.texel(hit.x + i as f64 * r * 0.66, hit.y + j as f64 * r * 0.66);
+        // Box filter over the penumbra square: n × n cell centres, dense enough that a wide
+        // diffusion blurs the caster instead of stamping copies of it (#2444).
+        let n = shadow_taps(r * c.buf.scale);
+        let step = 2.0 * r / n as f64;
+        let h = (n / 2) as i32;
+        let w = 1.0 / (n * n) as f32;
+        for j in -h..=h {
+            for i in -h..=h {
+                let q = c.texel(hit.x + f64::from(i) * step, hit.y + f64::from(j) * step);
                 for k in 0..4 {
-                    px[k] += q[k] / 9.0;
+                    px[k] += q[k] * w;
                 }
             }
         }

crates/render/src/three_d/tests.rs

@@ -344,6 +344,33 @@ fn shadow_darkness_and_accepts_off() {
     assert!(close(v, 0.5, 1e-3), "no shadow accepted: {v}");
 }
 
+#[test]
+fn shadow_diffusion_is_a_soft_ramp_not_copies() {
+    // #2444: a wide Shadow Diffusion stamped 3 × 3 copies of the caster, so the penumbra stepped
+    // through three flat levels instead of ramping smoothly from lit to shadowed.
+    let (mut p, cid, comp) = setup();
+    let mut recv = solid3(&mut p, &comp, [1.0, 1.0, 1.0], 400, 200, [100.0, 50.0, 100.0]);
+    no_specular(&mut recv);
+    let mut caster = solid3(&mut p, &comp, [1.0, 1.0, 1.0], 100, 200, [40.0, 50.0, 0.0]);
+    set(&mut caster, "materialOptions/castsShadows", Value::Enum(2)); // Only: invisible caster
+    let mut l = light(&mut p, &comp, LightKind::Parallel, [100.0, 50.0, -500.0]);
+    set(&mut l, "transform/poi", Value::Vec3([100.0, 50.0, 0.0]));
+    set(&mut l, "lightOptions/castsShadows", Value::Bool(true));
+    // 100 px to the caster over the parallel light's 1000: a 15 px penumbra radius.
+    set(&mut l, "lightOptions/shadowDiffusion", Value::Scalar(150.0));
+    push(&mut p, cid, l);
+    push(&mut p, cid, caster);
+    push(&mut p, cid, recv);
+    let img = render(&p, cid);
+    let row: Vec<f32> = (0..i64::from(W)).map(|x| img.get(x, 50)[0]).collect();
+    // The row crosses both vertical edges of the caster's shadow: lit, ramp, shadowed, ramp, lit.
+    assert!(row.iter().any(|v| *v < 0.01) && close(row[W as usize - 1], 0.5, 1e-3), "shadow and light: {row:?}");
+    let ramp = row.iter().filter(|v| **v > 0.02 && **v < 0.48).count();
+    assert!(ramp >= 16, "two penumbras several pixels wide: {row:?}");
+    let step = row.windows(2).map(|w| (w[1] - w[0]).abs()).fold(0.0f32, f32::max);
+    assert!(step < 0.08, "no jumps between copies of the caster (largest step {step}): {row:?}");
+}
+
 // ---------------------------------------------------------------- cameras
 
 fn camera(p: &mut Project, comp: &Comp) -> Layer {