fix: CNN bias accumulation and output format improvements

- Fix bias division bug: divide by num_positions to compensate for
  shader loop accumulation (affects all layers)
- train_cnn.py: Save RGBA output preserving alpha channel from input
- Add --debug-hex flag to both tools for pixel-level debugging
- Remove sRGB/linear_png debug code from cnn_test
- Regenerate weights with corrected bias export

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

1 files changed, 28 insertions, 0 deletions

diff --git a/tools/cnn_test.cc b/tools/cnn_test.cc
index fa4394f..59f1d22 100644
--- a/tools/cnn_test.cc
+++ b/tools/cnn_test.cc
@@ -43,6 +43,7 @@ struct Args {
   bool output_png = true; // Default to PNG
   const char* save_intermediates = nullptr;
   int num_layers = 3; // Default to 3 layers
+  bool debug_hex = false; // Print first 8 pixels as hex
 };
 
 // Parse command-line arguments
@@ -80,6 +81,8 @@ static bool parse_args(int argc, char** argv, Args* args) {
         fprintf(stderr, "Error: layers must be in range [1, 10]\n");
         return false;
       }
+    } else if (strcmp(argv[i], "--debug-hex") == 0) {
+      args->debug_hex = true;
     } else if (strcmp(argv[i], "--help") == 0) {
       return false;
     } else {
@@ -99,6 +102,7 @@ static void print_usage(const char* prog) {
   fprintf(stderr, "  --format ppm|png         Output format (default: png)\n");
   fprintf(stderr, "  --layers N               Number of CNN layers (1-10, default: 3)\n");
   fprintf(stderr, "  --save-intermediates DIR Save intermediate layers to directory\n");
+  fprintf(stderr, "  --debug-hex              Print first 8 pixels as hex (debug)\n");
   fprintf(stderr, "  --help                   Show this help\n");
 }
 
@@ -437,6 +441,18 @@ int main(int argc, char** argv) {
       printf("Reading pixels from GPU...\n");
       std::vector<uint8_t> pixels = rt.read_pixels();
 
+      // Debug: print first 8 pixels as hex
+      if (args.debug_hex && !pixels.empty()) {
+        printf("First 8 pixels (BGRA hex):\n");
+        for (int i = 0; i < 8 && i < width * height; ++i) {
+          const uint8_t b = pixels[i * 4 + 0];
+          const uint8_t g = pixels[i * 4 + 1];
+          const uint8_t r = pixels[i * 4 + 2];
+          const uint8_t a = pixels[i * 4 + 3];
+          printf("  [%d] 0x%02X%02X%02X%02X (RGBA)\n", i, r, g, b, a);
+        }
+      }
+
       if (pixels.empty()) {
         fprintf(stderr, "Error: GPU readback failed\n");
         wgpuTextureViewRelease(intermediate_views[0]);
@@ -513,6 +529,18 @@ int main(int argc, char** argv) {
         std::vector<uint8_t> pixels = texture_readback_fp16_to_u8(
             device, queue, intermediate_textures[dst_idx], width, height);
 
+        // Debug: print first 8 pixels as hex
+        if (args.debug_hex && !pixels.empty()) {
+          printf("Layer %d first 8 pixels (BGRA hex):\n", layer);
+          for (int i = 0; i < 8 && i < width * height; ++i) {
+            const uint8_t b = pixels[i * 4 + 0];
+            const uint8_t g = pixels[i * 4 + 1];
+            const uint8_t r = pixels[i * 4 + 2];
+            const uint8_t a = pixels[i * 4 + 3];
+            printf("  [%d] 0x%02X%02X%02X%02X (RGBA)\n", i, r, g, b, a);
+          }
+        }
+
         if (!pixels.empty()) {
           save_png(layer_path, pixels, width, height);
         } else {