@woosh/meep-engine 3.60.3 → 3.60.5

package/src/shade/renderer/postprocess/nss/shader_nss_extract_feedback.js

@@ -1,49 +1,48 @@
-import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
-import { ComputeShader } from "../../shader/ComputeShader.js";
-import { ShaderResourceSetDescriptor } from "../../shader/resource/ShaderResourceSetDescriptor.js";
-import { NSS_WORKGROUP_SIZE } from "./NSS_CONSTANTS.js";
-
-/**
- * Copies the feedback channel-group (z = 0 — head slot 0 in ARM's
- * quantized model output ordering, see NSS_NETWORK_SCHEMA) of the
- * network's 3D output tensor into a standalone 2D rgba16f texture, so
- * the next frame's preprocess pass can sample it as a regular 2D
- * texture (with bilinear filtering for the warp).
- *
- * Trivial pass — included as a separate shader rather than folded into
- * postprocess because postprocess runs at output-res while feedback is
- * needed at render-res.
- */
-const resources = new ShaderResourceSetDescriptor();
-resources.createGroup()
-    .addTexture("tNetworkOut", "float", "3d")
-    .addStorageTexture("tFeedback", "rgba16float", "2d", "write-only")
-;
-
-const body = CodeChunk.from(
-    //language=WGSL
-    `
-@compute @workgroup_size(${NSS_WORKGROUP_SIZE}, ${NSS_WORKGROUP_SIZE}, 1)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    let dim = textureDimensions(tFeedback);
-
-    if (any(gid.xy >= dim)) {
-        return;
-    }
-
-    let st = vec2<i32>(gid.xy);
-
-    // Slot 0 of the network output is the feedback tensor (ARM's
-    // quantized-model OutputIndices.Feedback = 0).
-    let fb = textureLoad(tNetworkOut, vec3<i32>(st, 0), 0);
-
-    textureStore(tFeedback, st, fb);
-}
-    `
-);
-
-export const shader_nss_extract_feedback = ComputeShader.from({
-    label: "NSS / extract feedback",
-    body,
-    resources,
-});
+import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
+import { ComputeShader } from "../../shader/ComputeShader.js";
+import { ShaderResourceSetDescriptor } from "../../shader/resource/ShaderResourceSetDescriptor.js";
+import { NSS_WORKGROUP_SIZE } from "./NSS_CONSTANTS.js";
+
+/**
+ * Copies the feedback channel-group (z = 0 — head slot 0 in ARM's
+ * quantized model output ordering, see NSS_NETWORK_SCHEMA) of the
+ * network's 3D output tensor into a standalone 2D rgba16f texture, so
+ * the next frame's preprocess pass can sample it as a regular 2D
+ * texture (with bilinear filtering for the warp).
+ *
+ * Trivial pass — included as a separate shader rather than folded into
+ * postprocess because postprocess runs at output-res while feedback is
+ * needed at render-res.
+ */
+const resources = new ShaderResourceSetDescriptor();
+resources.createGroup()
+    .addTexture("tNetworkOut", "float", "3d")
+    .addStorageTexture("tFeedback", "rgba16float", "2d", "write-only")
+;
+
+const body = CodeChunk.from(
+    //language=WGSL
+    `
+@compute @workgroup_size(${NSS_WORKGROUP_SIZE}, ${NSS_WORKGROUP_SIZE}, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let dim = textureDimensions(tFeedback);
+
+    if (any(gid.xy >= dim)) {
+        return;
+    }
+
+    let st = vec2<i32>(gid.xy);
+
+    // Slot 0 of the network output is the feedback tensor.
+    let fb = textureLoad(tNetworkOut, vec3<i32>(st, 0), 0);
+
+    textureStore(tFeedback, st, fb);
+}
+    `
+);
+
+export const shader_nss_extract_feedback = ComputeShader.from({
+    label: "NSS / extract feedback",
+    body,
+    resources,
+});

package/src/shade/renderer/postprocess/nss/shader_nss_postprocess.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"shader_nss_postprocess.d.ts","sourceRoot":"","sources":["../../../../../../src/shade/renderer/postprocess/nss/shader_nss_postprocess.js"],"names":[],"mappings":"AAqSA,mDAIG;8BAvS2B,+BAA+B"}
+{"version":3,"file":"shader_nss_postprocess.d.ts","sourceRoot":"","sources":["../../../../../../src/shade/renderer/postprocess/nss/shader_nss_postprocess.js"],"names":[],"mappings":"AAiSA,mDAIG;8BAnS2B,+BAA+B"}

package/src/shade/renderer/postprocess/nss/shader_nss_postprocess.js

@@ -1,4 +1,4 @@
-import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
+import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
 import { F16_MAX_DEFINITION } from "../../shader/chunk/math/F16_MAX_DEFINITION.js";
 import { ComputeShader } from "../../shader/ComputeShader.js";
 import { ShaderResourceSetDescriptor } from "../../shader/resource/ShaderResourceSetDescriptor.js";
@@ -19,7 +19,7 @@ import { NSS_SETTINGS_STRUCT } from "./NSS_SETTINGS_STRUCT.js";
  *   - the warped history at output-res
  *   - the nearest-depth offset from preprocess (used to dilate velocity)
  *
- * Per-output-pixel logic mirrors ARM's reference (UE5.5 NSSPostprocess.usf):
+ * Per-output-pixel logic mirrors ARM's reference:
  *   1. Solve which render-res pixel feeds each of the 4x4 KPN taps.
  *   2. Apply the KPN to those taps, normalising by the weight sum.
  *   3. Warp the history (bilinear) using velocity dilated by the nearest
@@ -27,9 +27,6 @@ import { NSS_SETTINGS_STRUCT } from "./NSS_SETTINGS_STRUCT.js";
  *   4. Blend the KPN colour with warped history using theta.
  *   5. In tonemapped (Karis) space, accumulate the center tap with alpha.
  *   6. Inverse tonemap, divide by EXPOSURE, write.
- *
- * @see ARM reference: UE5.5/Shaders/Private/NSSPostprocess.usf and
- *      NSSCommon.ush in github.com/arm/neural-graphics-for-unreal
  */
 
 const resources = new ShaderResourceSetDescriptor();
@@ -53,9 +50,9 @@ resources.createGroup()
 const body = CodeChunk.from(
     //language=WGSL
     `
-const NSS_EXPOSURE: f32 = 7.38905609893; // e^2, matches ARM's NSSCommon.ush
+const NSS_EXPOSURE: f32 = 7.38905609893; // e^2
 
-// Karis tonemap (max-of-channels) — matches ARM's NSSCommon.ush.
+// Karis tonemap (max-of-channels).
 // The guard is a FINITE clamp, not just a non-negative one: this target is rgba16float, so shading
 // above its range stores as +Inf and rgb / (1.0 + m) is Inf/Inf = NaN, which goes straight into the
 // next frame's history and stays there. NaN fails every comparison including with itself, which is
@@ -106,7 +103,7 @@ fn nss_sample_layer_bilinear(render_pos: vec2<f32>, render_res_max: vec2<i32>, l
 
 // SolveInputLocation: figure out which render-res pixel maps to a given
 // output-res pixel position, given upscale ratio and jitter. Direct port
-// of ARM's helper (NSSPostprocess.usf:80). Returns -1 when no render-res
+// of ARM's reference helper. Returns -1 when no render-res
 // sample lands at this output position.
 fn nss_solve_input_location(out_idx: i32, upscale: f32, jitter: f32) -> i32 {
     let lower = f32(out_idx) / upscale - (-jitter + 0.5);
@@ -144,8 +141,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let coef_st_clamped = clamp(coef_st, vec2<i32>(0, 0), render_res_max);
 
     // ── KPN columns ──
-    // ARM's quantized model maps OutputIndices.KPNFilterCol0 = 5,
-    // KPNFilterCol1 = 4, KPNFilterCol2 = 3, KPNFilterCol3 = 2 (NSS.cpp:243).
+    // ARM's quantized model maps KPN columns 0..3 to head slots 5, 4, 3, 2.
     // Each column has 4 channels = the 4 row weights for that column.
     // Bilinear at render_centre (matches ARM's ManualBilinear4) — without
     // this, every 2x2 output block in 2x mode shares one set of weights and
@@ -155,7 +151,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let col2 = nss_sample_layer_bilinear(render_centre, render_res_max, 3);
     let col3 = nss_sample_layer_bilinear(render_centre, render_res_max, 2);
     // Force tiny positive — ARM does this to dodge 0/0 in the
-    // weight-normalisation divide (NSSPostprocess.usf:118).
+    // weight-normalisation divide.
     let kpn_cols = array<vec4<f32>, 4>(
         max(col0, vec4<f32>(1e-7)),
         max(col1, vec4<f32>(1e-7)),
@@ -208,7 +204,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let nearest_offset = vec2<i32>(round(nearest_packed * 2.0 - 1.0));
 
     var velocity = textureLoad(tVelocity, coef_st_clamped + nearest_offset, 0).rg;
-    // Motion zero-clamp (ARM NSSPostprocess.usf:215). Sub-pixel velocity on
+    // Motion zero-clamp. Sub-pixel velocity on
     // near-static frames drifts prev_uv by fractions of a pixel each frame
     // and bilinear-resamples the history at slightly different sub-pixel
     // positions — visible as flicker even when nothing in the scene is
@@ -225,10 +221,10 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let on_screen_f = select(0.0, 1.0, all(prev_uv >= vec2<f32>(0.0)) && all(prev_uv <= vec2<f32>(1.0)));
 
     // ── ThetaAlpha (head slot 1) ──
-    // ARM's quantized model maps OutputIndices.ThetaAlpha = 1
-    // (NSS.cpp:242). Channel x = theta (history blend weight, [0, 1] from
+    // ARM's quantized model maps ThetaAlpha to head slot 1.
+    // Channel x = theta (history blend weight, [0, 1] from
     // sigmoid LUT), channel y = alpha (raw [0, 1], remapped to
-    // [0.05, 0.4] per NSSPostprocess.usf:248).
+    // [0.05, 0.4]).
     // Bilinear at render_centre — same reasoning as the KPN cols above.
     let ta = nss_sample_layer_bilinear(render_centre, render_res_max, 1);
     let theta_network = saturate(ta.x);
@@ -286,7 +282,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     textureStore(tOutput, out_st, vec4<f32>(final_rgb, theta));
 }
     `, [
-        F16_MAX_DEFINITION,
+        F16_MAX_DEFINITION,
         NSS_SETTINGS_STRUCT.declaration_chunk,
     ]
 );