// Copyright (c) 2017-2020 Xiamen Yaji Software Co., Ltd.

// Based on Kawase Dual Filter Blur
// https://community.arm.com/cfs-file/__key/communityserver-blogs-components-weblogfiles/00-00-00-20-66/siggraph2015_2D00_mmg_2D00_marius_2D00_slides.pdf

CCEffect %{
techniques:
  - passes:
      - vert: bloom-vs
        frag: prefilter-fs
        pass: cc-bloom-prefilter
        rasterizerState:
          cullMode: none
        depthStencilState:
          depthTest: false
          depthWrite: false
      - vert: bloom-vs
        frag: downsample-fs
        pass: cc-bloom-downsample
        rasterizerState:
          cullMode: none
        depthStencilState:
          depthTest: false
          depthWrite: false
      - vert: bloom-vs
        frag: upsample-fs
        pass: cc-bloom-upsample
        rasterizerState:
          cullMode: none
        depthStencilState:
          depthTest: false
          depthWrite: false
      - vert: bloom-vs
        frag: combine-fs
        pass: cc-bloom-combine
        rasterizerState:
          cullMode: none
        depthStencilState:
          depthTest: false
          depthWrite: false
        blendState:
          targets:
            - blend: true
              blendSrc: one
              blendDst: one
              blendSrcAlpha: zero
              blendDstAlpha: one

}%

CCProgram bloom-vs %{
precision highp float;
#include <legacy/decode-standard>
#include <post-process/pipeline>

out vec2 v_uv;

void main() {
  StandardVertInput In;
  CCDecode(In);
  FLIP_VULKAN_NDC(In.position);
  gl_Position = In.position;

  v_uv = a_texCoord;
  FLIP_SAMPLE_FROM_RT(v_uv);
}
}%

CCProgram prefilter-fs %{
precision highp float;
#include <common/color/gamma>

in vec2 v_uv;

#pragma rate BloomUBO pass
uniform BloomUBO {
  mediump vec4 bloomParams; // x: useHDRIlluminance z: threshold, w: enableAlphaMask
};

#pragma rate inputTexture pass
uniform sampler2D inputTexture; // hdr input

layout(location = 0)out vec4 fragColor;

float luminance(vec3 color) {
  return dot(color, vec3(0.3, 0.5, 0.2));
}

void main() {
  vec4 color = texture(inputTexture, v_uv);

  float contribute = step(bloomParams.z, luminance(color.rgb));
  contribute *= mix(1.0, step(253.0 / 255.0, color.a), bloomParams.w);

  fragColor = vec4(color.xyz * contribute, 1.0);
}
}%

CCProgram downsample-fs %{
precision highp float;

in vec2 v_uv;

#pragma rate BloomTexUBO pass
uniform BloomTexUBO {
  mediump vec4 bloomTexSize; // size
};
#pragma rate bloomTexture pass
uniform sampler2D bloomTexture;

layout(location = 0)out vec4 fragColor;

#pragma define SAM(uv) texture(bloomTexture, uv).xyz

vec3 downsample(vec2 uv, vec2 halfpixel) {
  // halfpixel is:
  // 1. half of the pixel size of the output texture
  // 2. pixel size of the input texture
  vec3 sum = SAM(uv) * 4.0;
  sum += SAM(uv - halfpixel.xy);
  sum += SAM(uv + halfpixel.xy);

  halfpixel.y = -halfpixel.y;
  sum += SAM(uv - halfpixel.xy);
  sum += SAM(uv + halfpixel.xy);
  return sum / 8.0;
}

vec3 downsample4taps(vec2 uv, vec2 halfpixel) {
  vec3 sum = SAM(uv + vec2(-halfpixel.x, halfpixel.y));
  sum += SAM(uv + vec2(halfpixel.x, halfpixel.y));
  sum += SAM(uv + vec2(halfpixel.x, - halfpixel.y));
  sum += SAM(uv + vec2(-halfpixel.x, - halfpixel.y));
  return sum / 4.0;
}

void main()
{
  vec3 result = downsample4taps(v_uv, 1.0 / bloomTexSize.xy).rgb;

  fragColor = vec4(result, 1.0);
}
}%

CCProgram upsample-fs %{
precision highp float;

in vec2 v_uv;

#pragma rate BloomTexUBO pass
uniform BloomTexUBO {
  mediump vec4 bloomTexSize;
};
#pragma rate bloomTexture pass
uniform sampler2D bloomTexture;

layout(location = 0)out vec4 fragColor;

#pragma define SAM(uv) texture(bloomTexture, uv).xyz

vec3 upsample(vec2 uv, vec2 halfpixel) {
  // halfpixel is:
  // 1. half of the pixel size of the input texture
  // 2. pixel size of the output texture
  vec3 sum = SAM(uv + vec2(-halfpixel.x * 2.0, 0.0));
  sum += SAM(uv + vec2(-halfpixel.x, halfpixel.y)) * 2.0;
  sum += SAM(uv + vec2(0.0, halfpixel.y * 2.0));
  sum += SAM(uv + vec2(halfpixel.x, halfpixel.y)) * 2.0;
  sum += SAM(uv + vec2(halfpixel.x * 2.0, 0.0));
  sum += SAM(uv + vec2(halfpixel.x, - halfpixel.y)) * 2.0;
  sum += SAM(uv + vec2(0.0, - halfpixel.y * 2.0));
  sum += SAM(uv + vec2(-halfpixel.x, - halfpixel.y)) * 2.0;
  return sum / 12.0;
}

vec3 upsample4taps(vec2 uv, vec2 halfpixel) {
  vec3 sum = SAM(uv + vec2(-halfpixel.x, halfpixel.y));
  sum += SAM(uv + vec2(halfpixel.x, halfpixel.y));
  sum += SAM(uv + vec2(halfpixel.x, - halfpixel.y));
  sum += SAM(uv + vec2(-halfpixel.x, - halfpixel.y));
  return sum / 4.0;
}

void main() {
  vec3 result = upsample(v_uv, 0.5 / bloomTexSize.xy).rgb;

  fragColor = vec4(result, 1.0);
}
}%

CCProgram combine-fs %{
precision highp float;

in vec2 v_uv;

#pragma rate BloomUBO pass
uniform BloomUBO {
  mediump vec4 bloomParams; // x: useHDRIlluminance z: threshold, w: enableAlphaMask
};
#pragma rate bloomTexture pass
uniform sampler2D bloomTexture;

layout(location = 0)out vec4 fragColor;

void main() {
  vec3 bloomColor = texture(bloomTexture, v_uv).rgb * bloomParams.w;
  fragColor = vec4(bloomColor, 0);
}
}%