// 3x3 convolution with weight indexing

// Source layers: 7→4 channels (RGBD output)
// Assumes 'tex' (the input) is *not* normalized to [-1,1], but is [0,1]
// UV coordinates remain in [0,1] and are normalized internally
// weights: array<array<f32, 8>, 36> (9 positions × 4 channels, each with 7 weights + bias)
fn cnn_conv3x3_7to4_src(
  tex: texture_2d<f32>,
  samp: sampler,
  uv: vec2<f32>,
  resolution: vec2<f32>,
  weights: array<array<f32, 8>, 36>
) -> vec4<f32> {
  let step = 1.0 / resolution;

  // Compute grayscale from original (converted in [-1,1])
  let original = (textureSample(tex, samp, uv) - 0.5) * 2.0;
  let gray = 0.2126*original.r + 0.7152*original.g + 0.0722*original.b;

  // Normalize UV to [-1,1]
  let uv_norm = (uv - 0.5) * 2.0;

  var sum = vec4<f32>(0.0);

  var pos = 0;
  for (var dy = -1; dy <= 1; dy++) {
    for (var dx = -1; dx <= 1; dx++) {
      let offset = vec2<f32>(f32(dx), f32(dy)) * step;
      let rgbd = (textureSample(tex, samp, uv + offset) - .5) * 2.0;  // convert to [-1,1]

      // 7-channel input: [R,G,B,D, uv.x, uv.y, gray] all in [-1,1]
      let inputs = array<f32, 7>(
        rgbd.r, rgbd.g, rgbd.b, rgbd.a,
        uv_norm.x, uv_norm.y, gray
      );

      // Accumulate for each output channel (RGBD)
      for (var out_c = 0; out_c < 4; out_c++) {
        let idx = pos * 4 + out_c;
        var channel_sum = weights[idx][7];  // Bias (8th element)
        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;  // Output in [-1,1] range
}

// Inner layers: 7→4 channels (RGBD output)
// Assumes 'tex' and 'original' are already normalized to [-1,1]
// UV coordinates remain in [0,1] and are normalized internally
// weights: array<array<f32, 8>, 36> (9 positions × 4 channels, each with 7 weights + bias)
fn cnn_conv3x3_7to4(
  tex: texture_2d<f32>,
  samp: sampler,
  uv: vec2<f32>,
  resolution: vec2<f32>,
  original: vec4<f32>,
  weights: array<array<f32, 8>, 36>
) -> vec4<f32> {
  let step = 1.0 / resolution;

  // Compute grayscale from original (already in [-1,1])
  let gray = 0.2126*original.r + 0.7152*original.g + 0.0722*original.b;

  // Normalize UV to [-1,1]
  let uv_norm = (uv - 0.5) * 2.0;

  var sum = vec4<f32>(0.0);

  var pos = 0;
  for (var dy = -1; dy <= 1; dy++) {
    for (var dx = -1; dx <= 1; dx++) {
      let offset = vec2<f32>(f32(dx), f32(dy)) * step;
      let rgbd = textureSample(tex, samp, uv + offset);  // Already in [-1,1]

      // 7-channel input: [R,G,B,D, uv.x, uv.y, gray] all in [-1,1]
      let inputs = array<f32, 7>(
        rgbd.r, rgbd.g, rgbd.b, rgbd.a,
        uv_norm.x, uv_norm.y, gray
      );

      // Accumulate for each output channel (RGBD)
      for (var out_c = 0; out_c < 4; out_c++) {
        let idx = pos * 4 + out_c;
        var channel_sum = weights[idx][7];  // Bias (8th element)
        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;  // Output in [-1,1] range
}

// Final layer: 7→1 channel (scalar output)
// Assumes 'tex' and 'original' are already normalized to [-1,1]
// UV coordinates remain in [0,1] and are normalized internally
// weights: array<array<f32, 8>, 9> (9 positions, each with 7 weights + bias)
fn cnn_conv3x3_7to1(
  tex: texture_2d<f32>,
  samp: sampler,
  uv: vec2<f32>,
  resolution: vec2<f32>,
  original: vec4<f32>,
  weights: array<array<f32, 8>, 9>
) -> f32 {
  let step = 1.0 / resolution;

  // Compute grayscale from original (already in [-1,1])
  let gray = 0.2126*original.r + 0.7152*original.g + 0.0722*original.b;

  // Normalize UV to [-1,1]
  let uv_norm = (uv - 0.5) * 2.0;

  var sum = 0.0;

  var pos = 0;
  for (var dy = -1; dy <= 1; dy++) {
    for (var dx = -1; dx <= 1; dx++) {
      let offset = vec2<f32>(f32(dx), f32(dy)) * step;
      let rgbd = textureSample(tex, samp, uv + offset);  // Already in [-1,1]

      // 7-channel input all in [-1,1]
      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;  // Output in [-1,1], needs denormalization
}