From 561d1dc446db7d1d3e02b92b43abedf1a5017850 Mon Sep 17 00:00:00 2001
From: skal <pascal.massimino@gmail.com>
Date: Fri, 13 Feb 2026 12:32:36 +0100
Subject: CNN v2: Refactor to uniform 12D→4D architecture
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

**Architecture changes:**
- Static features (8D): p0-p3 (parametric) + uv_x, uv_y, sin(10×uv_x), bias
- Input RGBD (4D): fed separately to all layers
- All layers: uniform 12D→4D (4 prev/input + 8 static → 4 output)
- Bias integrated in static features (bias=False in PyTorch)

**Weight calculations:**
- 3 layers × (12 × 3×3 × 4) = 1296 weights
- f16: 2.6 KB (vs old variable arch: ~6.4 KB)

**Updated files:**

*Training (Python):*
- train_cnn_v2.py: Uniform model, takes input_rgbd + static_features
- export_cnn_v2_weights.py: Binary export for storage buffers
- export_cnn_v2_shader.py: Per-layer shader export (debugging)

*Shaders (WGSL):*
- cnn_v2_static.wgsl: p0-p3 parametric features (mips/gradients)
- cnn_v2_compute.wgsl: 12D input, 4D output, vec4 packing

*Tools:*
- HTML tool (cnn_v2_test): Updated for 12D→4D, layer visualization

*Docs:*
- CNN_V2.md: Updated architecture, training, validation sections
- HOWTO.md: Reference HTML tool for validation

*Removed:*
- validate_cnn_v2.sh: Obsolete (used CNN v1 tool)

All code consistent with bias=False (bias in static features as 1.0).

handoff(Claude): CNN v2 architecture finalized and documented
---
 workspaces/main/shaders/cnn_v2/cnn_v2_static.wgsl | 23 +++++++++++++----------
 1 file changed, 13 insertions(+), 10 deletions(-)

(limited to 'workspaces/main/shaders/cnn_v2/cnn_v2_static.wgsl')

diff --git a/workspaces/main/shaders/cnn_v2/cnn_v2_static.wgsl b/workspaces/main/shaders/cnn_v2/cnn_v2_static.wgsl
index dd07f19..7a9e6de 100644
--- a/workspaces/main/shaders/cnn_v2/cnn_v2_static.wgsl
+++ b/workspaces/main/shaders/cnn_v2/cnn_v2_static.wgsl
@@ -1,5 +1,7 @@
 // CNN v2 Static Features Compute Shader
-// Generates 7D features + bias: [R, G, B, D, uv.x, uv.y, sin10_x, 1.0]
+// Generates 8D parametric features: [p0, p1, p2, p3, uv.x, uv.y, sin10_x, bias]
+// p0-p3: Parametric features (currently RGBD from mip0, could be mip1/2, gradients, etc.)
+// Note: Input image RGBD (mip0) fed separately to Layer 0
 
 @group(0) @binding(0) var input_tex: texture_2d<f32>;
 @group(0) @binding(1) var input_tex_mip1: texture_2d<f32>;
@@ -16,14 +18,14 @@ fn main(@builtin(global_invocation_id) id: vec3<u32>) {
     return;
   }
 
-  // Sample RGBA from mip 0
+  // Parametric features (p0-p3)
+  // TODO: Experiment with mip1 grayscale, Sobel gradients, etc.
+  // For now, use RGBD from mip 0 (same as input, but could differ)
   let rgba = textureLoad(input_tex, coord, 0);
-  let r = rgba.r;
-  let g = rgba.g;
-  let b = rgba.b;
-
-  // Sample depth
-  let d = textureLoad(depth_tex, coord, 0).r;
+  let p0 = rgba.r;
+  let p1 = rgba.g;
+  let p2 = rgba.b;
+  let p3 = textureLoad(depth_tex, coord, 0).r;
 
   // UV coordinates (normalized [0,1], bottom-left origin)
   let uv_x = f32(coord.x) / f32(dims.x);
@@ -36,9 +38,10 @@ fn main(@builtin(global_invocation_id) id: vec3<u32>) {
   let bias = 1.0;
 
   // Pack 8×f16 into 4×u32 (rgba32uint)
+  // [p0, p1, p2, p3, uv_x, uv_y, sin10_x, bias]
   let packed = vec4<u32>(
-    pack2x16float(vec2<f32>(r, g)),
-    pack2x16float(vec2<f32>(b, d)),
+    pack2x16float(vec2<f32>(p0, p1)),
+    pack2x16float(vec2<f32>(p2, p3)),
     pack2x16float(vec2<f32>(uv_x, uv_y)),
     pack2x16float(vec2<f32>(sin10_x, bias))
   );
-- 
cgit v1.2.3