// 5x5 convolution with 25 samples
// Applies mat4 weights per sample

fn cnn_conv5x5(
  tex: texture_2d<f32>,
  samp: sampler,
  uv: vec2<f32>,
  resolution: vec2<f32>,
  weights: array<mat4x4<f32>, 25>,
  bias: vec4<f32>
) -> vec4<f32> {
  let step = 1.0 / resolution;
  var sum = bias;
  var idx = 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 sample = textureSample(tex, samp, uv + offset);
      sum += weights[idx] * sample;
      idx++;
    }
  }

  return sum;
}

fn cnn_conv5x5_with_coord(
  tex: texture_2d<f32>,
  samp: sampler,
  uv: vec2<f32>,
  resolution: vec2<f32>,
  rgba_weights: array<mat4x4<f32>, 25>,
  coord_weights: mat2x4<f32>,
  bias: vec4<f32>
) -> vec4<f32> {
  let step = 1.0 / resolution;
  var sum = bias;

  sum += coord_weights * uv;

  var idx = 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 rgba = textureSample(tex, samp, uv + offset);
      sum += rgba_weights[idx] * rgba;
      idx++;
    }
  }

  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;
}