precision highp float;

#define MEAN_SAMPLE  8 // quality setting: 2, 4 or 8

varying vec2 v_TexCoords;
uniform sampler2D u_Texture;
uniform vec2 u_RenderSize;

// tweakable parameters
uniform float uSmoothing;

void main() 
{
  const int kernelSize = 15;
  int smallKernelSize = ((kernelSize + 1) / MEAN_SAMPLE) + 1;
  float smallKernelSizeRcp = 1.0 / (float(smallKernelSize) * float(smallKernelSize));
  
  vec2 uv = vec2(v_TexCoords.x, 1.0 - v_TexCoords.y);
  vec2 duv = 1.0 / u_RenderSize;
  vec2 uv0 = uv - 0.5 * duv * float(smallKernelSize - 1);
  vec4 meanIp = vec4(0.0);
  vec4 corrIp = vec4(0.0);

  for (int j = 0; j < smallKernelSize; j++) {
    for (int i = 0; i < smallKernelSize; i++) {
      vec2 texc = uv0 + duv * vec2(float(i), float(j));
      vec4 tex = texture2D(u_Texture, texc);
      meanIp += tex;
      corrIp += tex.w * tex;
    }
  }

  meanIp *= smallKernelSizeRcp;
  corrIp *= smallKernelSizeRcp;

  vec4 varIp = corrIp - meanIp.w * meanIp;
  float epsilon = uSmoothing;
  float epsilon2 = epsilon * epsilon;

  float a = varIp.w / (varIp.w + epsilon2);

  // clamp to prevent negative values
  a = max(a, 0.0);
  vec3 b = meanIp.xyz - a * meanIp.xyz;

  gl_FragColor = vec4(b, a);
}