fix: Support variable kernel sizes in CNN layer generation

Training script was hardcoded to generate cnn_conv3x3_* calls regardless
of actual kernel size, causing shader validation errors when layer 1 used
5×5 kernel (100 weights) but called 3×3 function (expected 36).

Changes:
- train_cnn.py: Generate correct conv function based on kernel_sizes[i]
- cnn_conv5x5.wgsl: Add cnn_conv5x5_7to4 and cnn_conv5x5_7to1 variants
- Regenerate cnn_layer.wgsl with correct function calls for [3,5,3]
- Document kernel size→function mapping in HOWTO.md

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

1 files changed, 86 insertions, 0 deletions

diff --git a/workspaces/main/shaders/cnn/cnn_conv5x5.wgsl b/workspaces/main/shaders/cnn/cnn_conv5x5.wgsl
index bd9abfa..15eaf96 100644
--- a/workspaces/main/shaders/cnn/cnn_conv5x5.wgsl
+++ b/workspaces/main/shaders/cnn/cnn_conv5x5.wgsl
@@ -51,3 +51,89 @@ fn cnn_conv5x5_with_coord(
 
   return sum;
 }
+
+// 5×5 variant for 7→4 channels (RGBD output)
+// weights: array<array<f32, 8>, 100> (25 positions × 4 channels, each with 7 weights + bias)
+fn cnn_conv5x5_7to4(
+  tex: texture_2d<f32>,
+  samp: sampler,
+  uv: vec2<f32>,
+  resolution: vec2<f32>,
+  original: vec4<f32>,
+  weights: array<array<f32, 8>, 100>
+) -> vec4<f32> {
+  let step = 1.0 / resolution;
+
+  let gray_01 = 0.2126*original.r + 0.7152*original.g + 0.0722*original.b;
+  let gray = (gray_01 - 0.5) * 2.0;
+  let uv_norm = (uv - 0.5) * 2.0;
+
+  var sum = vec4<f32>(0.0);
+  var pos = 0;
+
+  for (var dy = -2; dy <= 2; dy++) {
+    for (var dx = -2; dx <= 2; dx++) {
+      let offset = vec2<f32>(f32(dx), f32(dy)) * step;
+      let rgbd_01 = textureSample(tex, samp, uv + offset);
+      let rgbd = (rgbd_01 - 0.5) * 2.0;
+
+      let inputs = array<f32, 7>(
+        rgbd.r, rgbd.g, rgbd.b, rgbd.a,
+        uv_norm.x, uv_norm.y, gray
+      );
+
+      for (var out_c = 0; out_c < 4; out_c++) {
+        let idx = pos * 4 + out_c;
+        var channel_sum = weights[idx][7];
+        for (var in_c = 0; in_c < 7; in_c++) {
+          channel_sum += weights[idx][in_c] * inputs[in_c];
+        }
+        sum[out_c] += channel_sum;
+      }
+      pos++;
+    }
+  }
+
+  return sum;
+}
+
+// 5×5 variant for 7→1 channel (scalar output)
+// weights: array<array<f32, 8>, 25> (25 positions, each with 7 weights + bias)
+fn cnn_conv5x5_7to1(
+  tex: texture_2d<f32>,
+  samp: sampler,
+  uv: vec2<f32>,
+  resolution: vec2<f32>,
+  original: vec4<f32>,
+  weights: array<array<f32, 8>, 25>
+) -> f32 {
+  let step = 1.0 / resolution;
+
+  let gray_01 = 0.2126*original.r + 0.7152*original.g + 0.0722*original.b;
+  let gray = (gray_01 - 0.5) * 2.0;
+  let uv_norm = (uv - 0.5) * 2.0;
+
+  var sum = 0.0;
+  var pos = 0;
+
+  for (var dy = -2; dy <= 2; dy++) {
+    for (var dx = -2; dx <= 2; dx++) {
+      let offset = vec2<f32>(f32(dx), f32(dy)) * step;
+      let rgbd_01 = textureSample(tex, samp, uv + offset);
+      let rgbd = (rgbd_01 - 0.5) * 2.0;
+
+      sum += weights[pos][0] * rgbd.r;
+      sum += weights[pos][1] * rgbd.g;
+      sum += weights[pos][2] * rgbd.b;
+      sum += weights[pos][3] * rgbd.a;
+      sum += weights[pos][4] * uv_norm.x;
+      sum += weights[pos][5] * uv_norm.y;
+      sum += weights[pos][6] * gray;
+      sum += weights[pos][7];  // Bias
+
+      pos++;
+    }
+  }
+
+  return sum;
+}