<?php
include 'Source/functions.php';
$fullWidth = $argv[1];
$fullHeight = $argv[2];
$scaleFactorX = always_decimal_format($fullWidth / 1280.0);
$scaleFactorY = always_decimal_format($fullHeight / 720.0);
?>

#version 420
#extension GL_ARB_texture_gather : enable
// shader d8e69e8df8c227f5 
// Bloom/blur 1st step, 1280->640->1280
// To do, go back and remove all test blurs.. 
uniform ivec4 uf_remappedPS[3];
layout(binding = 0) uniform sampler2D textureUnitPS0;// Tex0 addr 0xf470a000 res 1280x720x1 dim 1 tm: 4 format 0816 compSel: 0 1 2 5 mipView: 0x0 (num 0x1) sliceView: 0x0 (num 0x1) Sampler0 ClampX/Y/Z: 2 2 2 border: 0
layout(binding = 1) uniform sampler2D textureUnitPS1;// Tex1 addr 0xf4386000 res 1280x720x1 dim 1 tm: 4 format 001a compSel: 0 1 2 3 mipView: 0x0 (num 0x1) sliceView: 0x0 (num 0x1) Sampler1 ClampX/Y/Z: 2 2 2 border: 0
layout(location = 0) in vec4 passParameterSem0;
layout(location = 0) out vec4 passPixelColor0;
uniform vec2 uf_fragCoordScale;
const float resScale = <?=$scaleFactorX?>;
//const float resScale = 3.5; //3.5 looks nicer

highp float lineRand(vec2 co)
{
	highp float a = 12.9898;
	highp float b = 78.233;
	highp float c = 43758.5453;
	highp float dt = dot(co.xy, vec2(a, b));
	highp float sn = mod(dt, 3.14);
	return fract(sin(sn) * c);
}

vec4 blur21Vert(sampler2D image, vec2 vuv, vec2 res) {
	// Note, must use gl_FragCord not emulated coordinates to sample between textels
	// Two pass implementation
	
	//iq CC  ,     http://rastergrid.com/blog/2010/09/efficient-gaussian-blur-with-linear-sampling/
	vec3 blr = vec3(0.0);
	
	blr += 0.026109*texture(image, (vuv + vec2(0.0, -15.5)) / res.xy).xyz;
	blr += 0.034202*texture(image, (vuv + vec2(0.0, -13.5)) / res.xy).xyz;
	blr += 0.043219*texture(image, (vuv + vec2(0.0, -11.5)) / res.xy).xyz;
	blr += 0.052683*texture(image, (vuv + vec2(0.0, -9.5)) / res.xy).xyz;
	blr += 0.061948*texture(image, (vuv + vec2(0.0, -7.5)) / res.xy).xyz;
	blr += 0.070266*texture(image, (vuv + vec2(0.0, -5.5)) / res.xy).xyz;
	blr += 0.076883*texture(image, (vuv + vec2(0.0, -3.5)) / res.xy).xyz;
	blr += 0.081149*texture(image, (vuv + vec2(0.0, -1.5)) / res.xy).xyz;
	blr += 0.041312*texture(image, (vuv + vec2(0.0, 0.0)) / res.xy).xyz;
	blr += 0.081149*texture(image, (vuv + vec2(0.0, 1.5)) / res.xy).xyz;
	blr += 0.076883*texture(image, (vuv + vec2(0.0, 3.5)) / res.xy).xyz;
	blr += 0.070266*texture(image, (vuv + vec2(0.0, 5.5)) / res.xy).xyz;
	blr += 0.061948*texture(image, (vuv + vec2(0.0, 7.5)) / res.xy).xyz;
	blr += 0.052683*texture(image, (vuv + vec2(0.0, 9.5)) / res.xy).xyz;
	blr += 0.043219*texture(image, (vuv + vec2(0.0, 11.5)) / res.xy).xyz;
	blr += 0.034202*texture(image, (vuv + vec2(0.0, 13.5)) / res.xy).xyz;
	blr += 0.026109*texture(image, (vuv + vec2(0.0, 15.5)) / res.xy).xyz;

	blr /= 0.93423; // renormalize to compensate for the 4 taps I skipped
	return  vec4(blr, 1.0);
}


vec4 blur21Horiz(sampler2D image, vec2 vuv, vec2 res) {
	//iq CC  
	vec3 blr = vec3(0.0);

	blr += 0.026109*texture(image, (vuv + vec2(-15.5, 0.0)) / res.xy).xyz;
	blr += 0.034202*texture(image, (vuv + vec2(-13.5, 0.0)) / res.xy).xyz;
	blr += 0.043219*texture(image, (vuv + vec2(-11.5, 0.0)) / res.xy).xyz;
	blr += 0.052683*texture(image, (vuv + vec2(-9.5, 0.0)) / res.xy).xyz;
	blr += 0.061948*texture(image, (vuv + vec2(-7.5, 0.0)) / res.xy).xyz;
	blr += 0.070266*texture(image, (vuv + vec2(-5.5, 0.0)) / res.xy).xyz;
	blr += 0.076883*texture(image, (vuv + vec2(-3.5, 0.0)) / res.xy).xyz;
	blr += 0.081149*texture(image, (vuv + vec2(-1.5, 0.0)) / res.xy).xyz;
	blr += 0.041312*texture(image, (vuv + vec2(0.0, 0.0)) / res.xy).xyz;
	blr += 0.081149*texture(image, (vuv + vec2(1.5, 0.0)) / res.xy).xyz;
	blr += 0.076883*texture(image, (vuv + vec2(3.5, 0.0)) / res.xy).xyz;
	blr += 0.070266*texture(image, (vuv + vec2(5.5, 0.0)) / res.xy).xyz;
	blr += 0.061948*texture(image, (vuv + vec2(7.5, 0.0)) / res.xy).xyz;
	blr += 0.052683*texture(image, (vuv + vec2(9.5, 0.0)) / res.xy).xyz;
	blr += 0.043219*texture(image, (vuv + vec2(11.5, 0.0)) / res.xy).xyz;
	blr += 0.034202*texture(image, (vuv + vec2(13.5, 0.0)) / res.xy).xyz;
	blr += 0.026109*texture(image, (vuv + vec2(15.5, 0.0)) / res.xy).xyz;

	blr /= 0.93423; // renormalize to compensate for the 4 taps I skipped
	return  vec4(blr, 1.0);
}

vec2 Circle(float Start, float Points, float Point)
{
	float Rad = (3.141592 * 2.0 * (1.0 / Points)) * (Point + Start);
	return vec2(sin(Rad), cos(Rad));
}

vec4 circleBlur(sampler2D image, vec2 vuv, vec2 res){
	//original by jcant0n ?
	//vec2 uv = gl_FragCoord.xy / res.xy;
	vec2 uv = vuv;
	vec2 PixelOffset = 1.9  / res.xy; //1.0

	float Start = 2.0 / 14.0;
	vec2 Scale = 0.66 * 4.0 * 2.0 * PixelOffset.xy;

	vec3 N0 = texture(image, uv + Circle(Start, 14.0, 0.0) * Scale).rgb;
	vec3 N1 = texture(image, uv + Circle(Start, 14.0, 1.0) * Scale).rgb;
	vec3 N2 = texture(image, uv + Circle(Start, 14.0, 2.0) * Scale).rgb;
	vec3 N3 = texture(image, uv + Circle(Start, 14.0, 3.0) * Scale).rgb;
	vec3 N4 = texture(image, uv + Circle(Start, 14.0, 4.0) * Scale).rgb;
	vec3 N5 = texture(image, uv + Circle(Start, 14.0, 5.0) * Scale).rgb;
	vec3 N6 = texture(image, uv + Circle(Start, 14.0, 6.0) * Scale).rgb;
	vec3 N7 = texture(image, uv + Circle(Start, 14.0, 7.0) * Scale).rgb;
	vec3 N8 = texture(image, uv + Circle(Start, 14.0, 8.0) * Scale).rgb;
	vec3 N9 = texture(image, uv + Circle(Start, 14.0, 9.0) * Scale).rgb;
	vec3 N10 = texture(image, uv + Circle(Start, 14.0, 10.0) * Scale).rgb;
	vec3 N11 = texture(image, uv + Circle(Start, 14.0, 11.0) * Scale).rgb;
	vec3 N12 = texture(image, uv + Circle(Start, 14.0, 12.0) * Scale).rgb;
	vec3 N13 = texture(image, uv + Circle(Start, 14.0, 13.0) * Scale).rgb;
	vec3 N14 = texture(image, uv).rgb;

	vec3 color = vec3(0, 0, 0);
	color.rgb =
		(N0 * 0.01478636) +
		(N1 * 0.03696590) +
		(N2 * 0.07393179) +
		(N3 * 0.07393179) +
		(N4 * 0.12013917) +
		(N5 * 0.16018555) +
		(N6 * 0.17620411) +
		(N7 * 0.17620411) +
		(N8 * 0.16018555) +
		(N9 * 0.12013917) +
		(N10 * 0.07393179) +
		(N11 * 0.07393179) +
		(N12 * 0.03696590) +
		(N13 * 0.03696590) +
		(N14 * 0.01478636);

	return vec4(color.rgb, 1.0);
}


vec4 blur25(sampler2D image, vec2 uv, vec2 resolution, vec2 direction) {
	//curent implentation doesn't work in cemu.. 
	//mattdesl MIT  - https://github.com/Jam3/glsl-fast-gaussian-blur
	vec4 color = vec4(0.0);
	vec2 off1 = vec2(1.44827586) * direction;
	vec2 off2 = vec2(3.37931034) * direction;
	vec2 off3 = vec2(5.31034483) * direction;
	vec2 off4 = vec2(7.24137931) * direction;
	vec2 off5 = vec2(9.17241379) * direction;
	vec2 off6 = vec2(11.10344828) * direction;

	//color += texture2D(image, uv - (off6 / resolution)) * 0.00001361;
	color += texture2D(image, uv - (off5 / resolution)) * 0.00044240;
	color += texture2D(image, uv - (off4 / resolution)) * 0.00581436;
	color += texture2D(image, uv - (off3 / resolution)) * 0.03730881;
	color += texture2D(image, uv - (off2 / resolution)) * 0.12888498;
	color += texture2D(image, uv - (off1 / resolution)) * 0.25281284;
	color += texture2D(image, uv)                       * 0.14944601;
	color += texture2D(image, uv + (off1 / resolution)) * 0.25281284;
	color += texture2D(image, uv + (off2 / resolution)) * 0.12888498;
	color += texture2D(image, uv + (off3 / resolution)) * 0.03730881;
	color += texture2D(image, uv + (off4 / resolution)) * 0.00581436;
	color += texture2D(image, uv + (off5 / resolution)) * 0.00044240;
	//color += texture2D(image, uv + (off6 / resolution)) * 0.00001361;
	///= 0.93423
	return color /= 0.93423;
}

int clampFI32(int v)
{
if( v == 0x7FFFFFFF )
	return floatBitsToInt(1.0);
else if( v == 0xFFFFFFFF )
	return floatBitsToInt(0.0);
return floatBitsToInt(clamp(intBitsToFloat(v), 0.0, 1.0));
}
float mul_nonIEEE(float a, float b){ return min(a*b,min(abs(a)*3.40282347E+38F,abs(b)*3.40282347E+38F)); }
void main()
{
vec4 R0f = vec4(0.0);
vec4 R1f = vec4(0.0);
vec4 R2f = vec4(0.0);
vec4 R3f = vec4(0.0);
vec4 R4f = vec4(0.0);
vec4 R5f = vec4(0.0);
vec4 R6f = vec4(0.0);
vec4 R7f = vec4(0.0);
vec4 R123f = vec4(0.0);
vec4 R127f = vec4(0.0);
float backupReg0f, backupReg1f, backupReg2f, backupReg3f, backupReg4f;
vec4 PV0f = vec4(0.0), PV1f = vec4(0.0);
float PS0f = 0.0, PS1f = 0.0;
vec4 tempf = vec4(0.0);
float tempResultf;
int tempResulti;
ivec4 ARi = ivec4(0);
bool predResult = true;
vec3 cubeMapSTM;
int cubeMapFaceId;
R0f = passParameterSem0;
R0f.xy = vec2((passParameterSem0.x + passParameterSem0.z) , (passParameterSem0.y + passParameterSem0.w) ); //center point

// 0
R1f.x = R0f.x + intBitsToFloat(uf_remappedPS[0].x)/resScale;
R1f.y = R0f.y + intBitsToFloat(uf_remappedPS[0].y)/resScale;
R0f.z = R0f.x + intBitsToFloat(uf_remappedPS[0].z)/resScale;
R0f.w = R0f.y + intBitsToFloat(uf_remappedPS[0].w)/resScale;
// 1
backupReg0f = R0f.x;
backupReg1f = R0f.y;
backupReg0f = R0f.x;
backupReg1f = R0f.y;
R0f.x = backupReg0f + intBitsToFloat(uf_remappedPS[1].x)/resScale;
R0f.y = backupReg1f + intBitsToFloat(uf_remappedPS[1].y)/resScale;
R1f.z = backupReg0f + intBitsToFloat(uf_remappedPS[1].z)/resScale;
R1f.w = backupReg1f + intBitsToFloat(uf_remappedPS[1].w)/resScale;


R2f.xyz = (texture(textureUnitPS1, R1f.xy).xyz);
R3f.xyz = (texture(textureUnitPS1, R0f.zw).xyz);
R4f.xyz = (texture(textureUnitPS1, R0f.xy).xyz);
R5f.xyz = (texture(textureUnitPS1, R1f.zw).xyz);

//vec2 iResolution = textureSize(textureUnitPS0, 0);

R6f = circleBlur(textureUnitPS0, R1f.xy, textureSize(textureUnitPS0, 0));
R7f = circleBlur(textureUnitPS0, R0f.zw, textureSize(textureUnitPS0, 0));
R0f = circleBlur(textureUnitPS0, R0f.xy, textureSize(textureUnitPS0, 0));
R1f = circleBlur(textureUnitPS0, R1f.zw, textureSize(textureUnitPS0, 0));


// Sample distance fails :( 
//R6f = mix(blur25(textureUnitPS0, R1f.xy, vec2(2560, 1440), vec2(0.0, 1.0)), blur13(textureUnitPS0, R1f.xy, textureSize(textureUnitPS0, 0), vec2(1.0,0.0)), 0.5);
//R7f = mix(blur25(textureUnitPS0, R0f.zw, textureSize(textureUnitPS0, 0), vec2(2.0, -3.0)), blur13(textureUnitPS0, R0f.zw, textureSize(textureUnitPS0, 0), vec2(4.0, 6.0)), 0.5);
//R0f = mix(blur25(textureUnitPS0, R0f.xy, textureSize(textureUnitPS0, 0), vec2(4.0, -6.0)), blur13(textureUnitPS0, R0f.xy, textureSize(textureUnitPS0, 0), vec2(2.0, 3.0)), 0.5);
//R1f = mix(blur25(textureUnitPS0, R1f.zw, textureSize(textureUnitPS0, 0), vec2(8.0, -9.0)), blur13(textureUnitPS0, R1f.zw, textureSize(textureUnitPS0, 0), vec2(1.1, 1.5)), 0.5);


tempf.x = dot(vec4(R2f.x,R2f.y,R2f.z,-0.0),vec4(intBitsToFloat(0x3e000000),intBitsToFloat(0x41ff0000),intBitsToFloat(0x45fe0100),0.0));
PV0f.x = tempf.x;
PV0f.y = tempf.x;
PV0f.z = tempf.x;
PV0f.w = tempf.x;
R127f.w = tempf.x;
R127f.z = R6f.x + R7f.x;
PS0f = R127f.z;
// 1
tempf.x = dot(vec4(R3f.x,R3f.y,R3f.z,-0.0),vec4(intBitsToFloat(0x3e000000),intBitsToFloat(0x41ff0000),intBitsToFloat(0x45fe0100),0.0));
PV1f.x = tempf.x;
PV1f.y = tempf.x;
PV1f.z = tempf.x;
PV1f.w = tempf.x;
R127f.y = R6f.y + R7f.y;
PS1f = R127f.y;
// 2
tempf.x = dot(vec4(R4f.x,R4f.y,R4f.z,-0.0),vec4(intBitsToFloat(0x3e000000),intBitsToFloat(0x41ff0000),intBitsToFloat(0x45fe0100),0.0));
PV0f.x = tempf.x;
PV0f.y = tempf.x;
PV0f.z = tempf.x;
PV0f.w = tempf.x;
PS0f = R127f.w + PV1f.x;
// 3
tempf.x = dot(vec4(R5f.x,R5f.y,R5f.z,-0.0),vec4(intBitsToFloat(0x3e000000),intBitsToFloat(0x41ff0000),intBitsToFloat(0x45fe0100),0.0));
PV1f.x = tempf.x;
PV1f.y = tempf.x;
PV1f.z = tempf.x;
PV1f.w = tempf.x;
PS1f = PS0f + PV0f.x;
// 4
backupReg0f = R0f.y;
PV0f.y = PS1f + PV1f.x;
PV0f.z = R127f.y + backupReg0f;
PV0f.w = R6f.z + R7f.z;
// 5
backupReg0f = R0f.z;
R123f.x = (PV0f.y * 0.25 + -(intBitsToFloat(uf_remappedPS[2].x)));
PV1f.x = R123f.x;
PV1f.y = R127f.z + R0f.x;
PV1f.z = PV0f.z + R1f.y;
PV1f.w = PV0f.w + backupReg0f;
// 6
PV0f.x = PV1f.y + R1f.x;
PV0f.y = PV1f.w + R1f.z;
PV0f.z = PV1f.z * 0.25;
PV0f.w = max(PV1f.x, -(PV1f.x));
// 7
PV1f.x = PV0f.x * 0.25;
PV1f.y = PV0f.y * 0.25;
PV1f.z = PV0f.w * intBitsToFloat(uf_remappedPS[2].y);
PV1f.z = clamp(PV1f.z, 0.0, 1.0);
PV1f.w = max(PV0f.z, 0.0);
// 8
PV0f.x = max(PV1f.y, 0.0);
R1f.y = min(PV1f.w, 4.0);
PV0f.z = max(PV1f.x, 0.0);
tempResultf = log2(PV1f.z);
if( isinf(tempResultf) == true ) tempResultf = -3.40282347E+38F;
PS0f = tempResultf;
// 9
PV1f.x = PS0f * intBitsToFloat(uf_remappedPS[2].w);
R1f.z = min(PV0f.x, 4.0);
R1f.x = min(PV0f.z, 4.0);
PS1f = R1f.x;
// 10
PS0f = exp2(PV1f.x);
// 11
R1f.w = PS0f * intBitsToFloat(uf_remappedPS[2].z);
// export
passPixelColor0 = vec4(R1f.x, R1f.y, R1f.z, R1f.w);
}