@woosh/meep-engine 3.8.0 → 3.10.0

package/src/shade/renderer/shader/chunk/mboit/moments/geometric/chunk_generate_trigonometric_moments_4.js

@@ -1,35 +1,42 @@
-import { CodeChunk } from "../../../../compiler/CodeChunk.js";
-import { chunk_complex_multiply } from "../../../math/complex/chunk_complex_multiply.js";
-
-/**
- * This function handles the actual computation of the new vector of
- * trigonometric moments for NUM_MOMENTS = 4 with SINGLE_PRECISION = true.
- * @see https://github.com/AlicesoftHumika/UnityOIT/blob/612c3c04263acb22980f33fa3cb2d3653ed389ff/Assets/Shader/MBOIT/MomentOIT.hlsli#L123
- * @type {CodeChunk}
- */
-export const chunk_generate_trigonometric_moments_4 = CodeChunk.from(
-    //language=WGSL
-    `
-fn generate_trigonometric_moments_4(
-    b_0: ptr<function, f32>,
-    b_12: ptr<function, vec4<f32>>,
-    depth: f32,
-    transmittance: f32,
-    wrapping_zone_parameters: vec4<f32>
-) {
-    let absorbance = -log(transmittance);
-
-    let phase = depth * wrapping_zone_parameters.y + wrapping_zone_parameters.y;
-    let circle_point = vec2<f32>(cos(phase), sin(phase));
-    let circle_point_pow2 = complex_multiply(circle_point, circle_point);
-
-    let b_12_new = vec4<f32>(circle_point.x, circle_point.y, circle_point_pow2.x, circle_point_pow2.y) * absorbance;
-    
-    // SINGLE_PRECISION = true case
-    (*b_0) += absorbance;
-    (*b_12) += b_12_new;
-}
-    `, [
-        chunk_complex_multiply,
-    ]
-);
+import { CodeChunk } from "../../../../compiler/CodeChunk.js";
+import { chunk_complex_multiply } from "../../../math/complex/chunk_complex_multiply.js";
+
+/**
+ * This function computes one fragment's contribution to the vector of
+ * trigonometric moments, for NUM_MOMENTS = 4 with SINGLE_PRECISION = true.
+ * b_12 packs the two complex moments as (re_1, im_1, re_2, im_2). The
+ * contribution is returned raw; summing across fragments is the caller's
+ * business (additive blending into float render targets, or component-wise
+ * addition in a loop).
+ * @see https://github.com/AlicesoftHumika/UnityOIT/blob/612c3c04263acb22980f33fa3cb2d3653ed389ff/Assets/Shader/MBOIT/MomentOIT.hlsli#L123
+ * @type {CodeChunk}
+ */
+export const chunk_generate_trigonometric_moments_4 = CodeChunk.from(
+    //language=WGSL
+    `
+struct TrigonometricMoments4 {
+    b_0: f32,
+    b_12: vec4<f32>
+}
+
+fn generate_trigonometric_moments_4(
+    depth: f32,
+    transmittance: f32,
+    wrapping_zone_parameters: vec4<f32>
+) -> TrigonometricMoments4 {
+    let absorbance = -log(transmittance);
+
+    let phase = depth * wrapping_zone_parameters.y + wrapping_zone_parameters.y;
+    let circle_point = vec2<f32>(cos(phase), sin(phase));
+    let circle_point_pow2 = complex_multiply(circle_point, circle_point);
+
+    // SINGLE_PRECISION = true case
+    return TrigonometricMoments4(
+        absorbance,
+        vec4<f32>(circle_point.x, circle_point.y, circle_point_pow2.x, circle_point_pow2.y) * absorbance
+    );
+}
+    `, [
+        chunk_complex_multiply,
+    ]
+);

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_4_power_moments.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chunk_compute_transmittance_from_4_power_moments.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_4_power_moments.js"],"names":[],"mappings":"AAEA;;;;;GAKG;AACH,+DAFU,SAAS,CAsFjB;0BA5FwB,mCAAmC"}
+{"version":3,"file":"chunk_compute_transmittance_from_4_power_moments.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_4_power_moments.js"],"names":[],"mappings":"AAEA;;;;;GAKG;AACH,+DAFU,SAAS,CAmGjB;0BAzGwB,mCAAmC"}

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_4_power_moments.js

@@ -50,14 +50,27 @@ fn compute_transmittance_from_4_power_moments(
     c[1] -= l21 * c[2];
     c[0] -= dot(c.yz, b.xy);
     
-    // Solve the quadratic equation c[0]+c[1]*z+c[2]*z^2 to obtain solutions z[1] and z[2]
-    let inv_c2 = 1.0 / c[2];
+    // Solve the quadratic equation c[0]+c[1]*z+c[2]*z^2 to obtain solutions z[1] and z[2].
+    // c[2] crosses 0 where the third support point of the canonical distribution passes
+    // through infinity (its weight is 0 there), and the cancellation computing it snaps it
+    // to exactly 0.0 over a whole band of depths around the crossing — for a single layer
+    // the band sits at warp depth sqrt(0.375), from the bias vector. Unguarded, 1/0 turns
+    // the discriminant into inf - inf = NaN, knocking the pixel out of the composite as a
+    // camera-locked stripe. The magnitude clamp keeps the vanishing root huge-but-finite
+    // (its weight is negligible there); the discriminant floor keeps the roots distinct.
+    let c2_safe = select(min(c[2], -1e-10), max(c[2], 1e-10), c[2] >= 0.0);
+    let inv_c2 = 1.0 / c2_safe;
     let p = c[1] * inv_c2;
     let q = c[0] * inv_c2;
-    let d = (p * p * 0.25) - q;
+    let d = max((p * p * 0.25) - q, 1e-20);
     let r = sqrt(d);
-    z[1] = -p * 0.5 - r;
-    z[2] = -p * 0.5 + r;
+    // Citardauq form: the far root is computed additively (no cancellation) and the near
+    // root recovered via Vieta (z1*z2 = q) — the classic -p/2 ± r loses every significant
+    // digit of the near root when the clamp above makes |p| huge.
+    let z_far = -p * 0.5 - select(-r, r, p >= 0.0);
+    let z_near = q / z_far;
+    z[1] = min(z_far, z_near);
+    z[2] = max(z_far, z_near);
     
     // Compute the absorbance by summing the appropriate weights
     

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_6_power_moments.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chunk_compute_transmittance_from_6_power_moments.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_6_power_moments.js"],"names":[],"mappings":"AAGA;;;;;GAKG;AACH,+DAFU,SAAS,CAqGjB;0BA5GwB,mCAAmC"}
+{"version":3,"file":"chunk_compute_transmittance_from_6_power_moments.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_6_power_moments.js"],"names":[],"mappings":"AAIA;;;;;GAKG;AACH,+DAFU,SAAS,CAmHjB;0BA3HwB,mCAAmC"}

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_6_power_moments.js

@@ -1,4 +1,5 @@
 import { CodeChunk } from "../../../../compiler/CodeChunk.js";
+import { chunk_magnitude_floor } from "../../../math/chunk_magnitude_floor.js";
 import { chunk_solve_cubic } from "../../../math/polynomial/chunk_solve_cubic.js";
 
 /**
@@ -61,8 +62,17 @@ fn compute_transmittance_from_6_power_moments(
     c[1] -= dot(vec2<f32>(l21, l31), c.zw);
     c[0] -= dot(vec3<f32>(b[0], b[1], b[2]), c.yzw);
 
-    // Solve the cubic equation
-    let cubic_roots = solve_cubic(c);
+    // Solve the cubic equation.
+    // c[3] -> 0 where the fourth support point of the canonical distribution escapes to
+    // infinity (its weight vanishes there), and the cancellation computing it snaps it to
+    // exactly 0.0 over a whole band of depths around the crossing; the solver normalizes
+    // by it, which unguarded turns every root into inf/NaN and poisons the interpolation
+    // below. The floor keeps the escaped root huge-but-finite. It is relative to the
+    // largest lower coefficient — the pivot scaling above can amplify the whole vector —
+    // so the solver-normalized coefficients stay <= ~1e4: every solver intermediate then
+    // fits f32, while a root out at ~1e4 carries negligible weight in the reconstruction.
+    let c_scale = max(1.0, max(abs(c[0]), max(abs(c[1]), abs(c[2]))));
+    let cubic_roots = solve_cubic(vec4<f32>(c.xyz, magnitude_floor(c[3], 1e-4 * c_scale)));
     z[1] = cubic_roots.x;
     z[2] = cubic_roots.y;
     z[3] = cubic_roots.z;
@@ -80,12 +90,16 @@ fn compute_transmittance_from_6_power_moments(
     let f1 = weight_factor[1];
     let f2 = weight_factor[2];
     let f3 = weight_factor[3];
-    let f01 = (f1 - f0) / (z[1] - z[0]);
-    let f12 = (f2 - f1) / (z[2] - z[1]);
-    let f23 = (f3 - f2) / (z[3] - z[2]);
-    let f012 = (f12 - f01) / (z[2] - z[0]);
-    let f123 = (f23 - f12) / (z[3] - z[1]);
-    let f0123 = (f123 - f012) / (z[3] - z[0]);
+    // Coincident support points (or a root landing exactly on the query depth) zero
+    // these separations: equal weights then make the division 0/0 = NaN, unequal ones an
+    // inf that meets another inf two lines later. Floored, a merged pair contributes one
+    // bounded term instead.
+    let f01 = (f1 - f0) / magnitude_floor(z[1] - z[0], 1e-4);
+    let f12 = (f2 - f1) / magnitude_floor(z[2] - z[1], 1e-4);
+    let f23 = (f3 - f2) / magnitude_floor(z[3] - z[2], 1e-4);
+    let f012 = (f12 - f01) / magnitude_floor(z[2] - z[0], 1e-4);
+    let f123 = (f23 - f12) / magnitude_floor(z[3] - z[1], 1e-4);
+    let f0123 = (f123 - f012) / magnitude_floor(z[3] - z[0], 1e-4);
     var polynomial = vec4<f32>();
     // f012+f0123 *(z-z2)
     polynomial[0] = f0123 * (-z[2]) + f012;
@@ -104,6 +118,7 @@ fn compute_transmittance_from_6_power_moments(
     return saturate(exp(-b_0 * absorbance));
 }
     `, [
-        chunk_solve_cubic
+        chunk_solve_cubic,
+        chunk_magnitude_floor
     ]
 );

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_8_power_moments.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chunk_compute_transmittance_from_8_power_moments.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_8_power_moments.js"],"names":[],"mappings":"AAGA;;;;;GAKG;AACH,+DAFU,SAAS,CAiIjB;0BAxIwB,mCAAmC"}
+{"version":3,"file":"chunk_compute_transmittance_from_8_power_moments.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_8_power_moments.js"],"names":[],"mappings":"AAIA;;;;;GAKG;AACH,+DAFU,SAAS,CAgJjB;0BAxJwB,mCAAmC"}

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_compute_transmittance_from_8_power_moments.js

@@ -1,4 +1,5 @@
 import { CodeChunk } from "../../../../compiler/CodeChunk.js";
+import { chunk_magnitude_floor } from "../../../math/chunk_magnitude_floor.js";
 import { chunk_solve_quartic_neumark } from "../../../math/polynomial/chunk_solve_quartic_neumark.js";
 
 /**
@@ -81,8 +82,18 @@ fn compute_transmittance_from_8_power_moments(
     c[1] -= dot(vec3<f32>(l52, l42, l32), vec3<f32>(c[4], c[3], c[2]));
     c[0] -= dot(vec4<f32>(b[3], b[2], b[1], b[0]), vec4<f32>(c[4], c[3], c[2], c[1]));
 
-    // Solve the quartic equation
-    let quartic_roots = solve_quartic_neumark(c);
+    // Solve the quartic equation.
+    // c[4] -> 0 where the fifth support point of the canonical distribution escapes to
+    // infinity (its weight vanishes there), and the cancellation computing it snaps it to
+    // exactly 0.0 over a whole band of depths around the crossing; the solver normalizes
+    // by it, which unguarded turns every root into inf/NaN and poisons the interpolation
+    // below. The floor keeps the escaped root huge-but-finite. It is relative to the
+    // largest lower coefficient — the pivot scaling above can amplify the whole vector —
+    // so the solver-normalized coefficients stay <= ~1e4: every resolvent intermediate
+    // then fits f32, while a root out at ~1e4 carries negligible weight in the
+    // reconstruction.
+    let c_scale = max(1.0, max(max(abs(c[0]), abs(c[1])), max(abs(c[2]), abs(c[3]))));
+    let quartic_roots = solve_quartic_neumark(array<f32, 5>(c[0], c[1], c[2], c[3], magnitude_floor(c[4], 1e-4 * c_scale)));
     z[1] = quartic_roots[0];
     z[2] = quartic_roots[1];
     z[3] = quartic_roots[2];
@@ -96,16 +107,20 @@ fn compute_transmittance_from_8_power_moments(
     let f2 = weight_factor[1];
     let f3 = weight_factor[2];
     let f4 = weight_factor[3];
-    let f01 = (f1 - f0) / (z[1] - z[0]);
-    let f12 = (f2 - f1) / (z[2] - z[1]);
-    let f23 = (f3 - f2) / (z[3] - z[2]);
-    let f34 = (f4 - f3) / (z[4] - z[3]);
-    let f012 = (f12 - f01) / (z[2] - z[0]);
-    let f123 = (f23 - f12) / (z[3] - z[1]);
-    let f234 = (f34 - f23) / (z[4] - z[2]);
-    let f0123 = (f123 - f012) / (z[3] - z[0]);
-    let f1234 = (f234 - f123) / (z[4] - z[1]);
-    let f01234 = (f1234 - f0123) / (z[4] - z[0]);
+    // Coincident support points (or a root landing exactly on the query depth) zero
+    // these separations: equal weights then make the division 0/0 = NaN, unequal ones an
+    // inf that meets another inf further down the table. Floored, a merged pair
+    // contributes one bounded term instead.
+    let f01 = (f1 - f0) / magnitude_floor(z[1] - z[0], 1e-4);
+    let f12 = (f2 - f1) / magnitude_floor(z[2] - z[1], 1e-4);
+    let f23 = (f3 - f2) / magnitude_floor(z[3] - z[2], 1e-4);
+    let f34 = (f4 - f3) / magnitude_floor(z[4] - z[3], 1e-4);
+    let f012 = (f12 - f01) / magnitude_floor(z[2] - z[0], 1e-4);
+    let f123 = (f23 - f12) / magnitude_floor(z[3] - z[1], 1e-4);
+    let f234 = (f34 - f23) / magnitude_floor(z[4] - z[2], 1e-4);
+    let f0123 = (f123 - f012) / magnitude_floor(z[3] - z[0], 1e-4);
+    let f1234 = (f234 - f123) / magnitude_floor(z[4] - z[1], 1e-4);
+    let f01234 = (f1234 - f0123) / magnitude_floor(z[4] - z[0], 1e-4);
 
     var polynomial_0: f32;
     var polynomial = vec4<f32>();
@@ -132,6 +147,7 @@ fn compute_transmittance_from_8_power_moments(
     return saturate(exp(-b_0 * absorbance));
 }
     `, [
-        chunk_solve_quartic_neumark
+        chunk_solve_quartic_neumark,
+        chunk_magnitude_floor
     ]
 );

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_generate_power_moments_4.d.ts

@@ -1,3 +1,15 @@
+/**
+ * This function computes one fragment's contribution to the vector of power
+ * moments, for NUM_MOMENTS = 4 with SINGLE_PRECISION = true. The contribution
+ * is returned raw; summing across fragments is the caller's business (additive
+ * blending into float render targets, or component-wise addition in a loop).
+ * Quantization, normalization by b_0 and the [0, 1] offset belong to the
+ * quantized-storage path, which needs read-modify-write access to the already
+ * accumulated moments (an ROV in the reference) and cannot be expressed
+ * through additive blending.
+ * @see https://github.com/AlicesoftHumika/UnityOIT/blob/612c3c04263acb22980f33fa3cb2d3653ed389ff/Assets/Shader/MBOIT/MomentOIT.hlsli#L123
+ * @type {CodeChunk}
+ */
 export const chunk_generate_power_moments_4: CodeChunk;
 import { CodeChunk } from "../../../../compiler/CodeChunk.js";
 //# sourceMappingURL=chunk_generate_power_moments_4.d.ts.map

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_generate_power_moments_4.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"chunk_generate_power_moments_4.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_generate_power_moments_4.js"],"names":[],"mappings":"AAIA,uDA2CC;0BA/CyB,mCAAmC"}
+{"version":3,"file":"chunk_generate_power_moments_4.d.ts","sourceRoot":"","sources":["../../../../../../../../../src/shade/renderer/shader/chunk/mboit/moments/power/chunk_generate_power_moments_4.js"],"names":[],"mappings":"AAEA;;;;;;;;;;;GAWG;AACH,6CAFU,SAAS,CAyBjB;0BArCwB,mCAAmC"}

package/src/shade/renderer/shader/chunk/mboit/moments/power/chunk_generate_power_moments_4.js

@@ -1,48 +1,38 @@
-import { CodeChunk } from "../../../../compiler/CodeChunk.js";
-import { chunk_offset_moments_2 } from "../chunk_offset_moments_2.js";
-import { chunk_quantize_moments_2 } from "../chunk_quantize_moments_2.js";
-
-export const chunk_generate_power_moments_4 = CodeChunk.from(
-    //language=WGSL
-    `
-fn generate_power_moments_4(
-    b_0: ptr<function, f32>,
-    b_even: ptr<function, vec2<f32>>,
-    b_odd: ptr<function, vec2<f32>>,
-    depth: f32,
-    transmittance: f32,
-) {
-
-    let absorbance = -log(transmittance);
-	let depth_pow2 = depth * depth;
-	let depth_pow4 = depth_pow2 * depth_pow2;
-	
-	*b_0		+= absorbance;
-	*b_even	+= vec2(depth_pow2, depth_pow4)		* absorbance;
-	*b_odd	+= vec2(depth, depth_pow2 * depth)	* absorbance;
-
-	var b_even_new	= vec2(depth_pow2, depth_pow4);
-	var b_odd_new	= vec2(depth, depth_pow2 * depth);
-	
-	var b_even_new_q:vec2f;
-	var b_odd_new_q:vec2f;
-
-	quantize_moments_2(b_even_new, b_odd_new, &b_even_new_q, &b_odd_new_q);
-
-	// Combine Moments
-	*b_0 += absorbance;
-	*b_even += b_even_new_q * absorbance;
-	*b_odd += b_odd_new_q * absorbance;
-
-	// Go back to interval [0, 1]
-	*b_even /= *b_0;
-	*b_odd /= *b_0;
-	
-	offset_moments_2(b_even, b_odd, 1.0);
-
-}
-    `, [
-        chunk_quantize_moments_2,
-        chunk_offset_moments_2
-    ]
-)
+import { CodeChunk } from "../../../../compiler/CodeChunk.js";
+
+/**
+ * This function computes one fragment's contribution to the vector of power
+ * moments, for NUM_MOMENTS = 4 with SINGLE_PRECISION = true. The contribution
+ * is returned raw; summing across fragments is the caller's business (additive
+ * blending into float render targets, or component-wise addition in a loop).
+ * Quantization, normalization by b_0 and the [0, 1] offset belong to the
+ * quantized-storage path, which needs read-modify-write access to the already
+ * accumulated moments (an ROV in the reference) and cannot be expressed
+ * through additive blending.
+ * @see https://github.com/AlicesoftHumika/UnityOIT/blob/612c3c04263acb22980f33fa3cb2d3653ed389ff/Assets/Shader/MBOIT/MomentOIT.hlsli#L123
+ * @type {CodeChunk}
+ */
+export const chunk_generate_power_moments_4 = CodeChunk.from(
+    //language=WGSL
+    `
+struct PowerMoments4 {
+    b_0: f32,
+    b_even: vec2<f32>,
+    b_odd: vec2<f32>
+}
+
+fn generate_power_moments_4(depth: f32, transmittance: f32) -> PowerMoments4 {
+    let absorbance = -log(transmittance);
+
+    let depth_pow2 = depth * depth;
+    let depth_pow4 = depth_pow2 * depth_pow2;
+
+    // SINGLE_PRECISION = true case
+    return PowerMoments4(
+        absorbance,
+        vec2(depth_pow2, depth_pow4) * absorbance,
+        vec2(depth, depth_pow2 * depth) * absorbance
+    );
+}
+    `
+);