cbuffer cb0 : register(b0) {
	float time;
	float dummy0, dummy1, dummy2;

	float3 viewPosition;
	float dummy3;

	float3 viewDirection;
	float dummy4;

	float3 viewUp;
	float dummy5;
}

#define FXAA_PC 1
#define FXAA_HLSL_5 1
#define FXAA_QUALITY__PRESET 39
//12

#include "fxaa311.hlsl"
#include "utils.hlsl"

RWTexture2D<float4> out0:register(u0);
SamplerState sampler0 : register(s0);
Texture2D tex0: register(t0);
Texture2D tex1: register(t1);
Texture2D tex2: register(t2);
Texture2D tex3: register(t3);
Texture2D tex4: register(t4);
Texture2D tex5: register(t5);
Texture2D tex6: register(t6);
Texture2D tex7: register(t7);

[numthreads(16, 16, 1)] void csFXAA(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);
	float2 tc = (id.xy + 0.5) / resolution;

	FxaaTex tex = {sampler0 , tex0};

	out0[id.xy] = FxaaPixelShader(tc, float4(0, 0, 0, 0), tex, tex, tex,
		1.0/resolution, float4(0, 0, 0, 0), float4(0, 0, 0, 0), float4(0, 0, 0, 0),
		0.75, 0.125, 0.0625, 0, 0, 0, 0);
}

//---
// Radial and circumferential blur
//---
static const float rcfMipmapFactor = 1.5;
static const float rcfStepFactor = 5;
static const float rcfStrengthRadial = 0*10;
static const float rcfStrengthCircumferential = 0*10;
static const float rcfRadius = 15;
static const float2 rcfCenter = {0.5, 0.5};

float rcfRadialFalloff(float radius, float amount) {
	return radius * (1 + amount*radius*0.01);
}

float4 rcfBlur(float2 tc, float2 resolution, float strength, float circumFerentialStrength) {
	// Direction from rcfCenter to the current tc, aspect ratio corrected
	float2 d = (tc - rcfCenter) * float2(resolution.x/resolution.y, 1);

	// Compute convolution step size
	float delta = length(d);
	float distortedDelta = rcfRadialFalloff(delta, strength);
	float stepSize = abs(delta - distortedDelta) * rcfStepFactor;

	float radius = rcfRadius;
	// Circumferential blur
	if (circumFerentialStrength > 0) {
		d = float2(d.y, -d.x);
		stepSize *= circumFerentialStrength;
		radius = floor(rcfRadius * circumFerentialStrength);
	}

	// Early exit
	if (radius <= 0.5) {// || stepSize <= 1/resolution.x) {
		return tex0.SampleLevel(sampler0, tc, 0);
	}

	// Adjust direction
	d = normalize(d) * stepSize / resolution;

	// Convolution
	float mipLevel = sqrt(stepSize) * rcfMipmapFactor;
	float sigma = radius/2;
	float twoSigmaSquared = 2 * sigma * sigma;
	float4 a = 0;
	for (float i = -radius; i <= radius; ++i) {
		float weight = exp(-i*i/twoSigmaSquared);
		a +=  tex0.SampleLevel(sampler0, i * d + tc, mipLevel) * weight;
	}

	return a / sqrt(twoSigmaSquared * PI);
}

[numthreads(16, 16, 1)] void csRadialBlur(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);

	float2 tc = (id.xy + 0.5) / resolution;

	out0[id.xy] = rcfBlur(tc, resolution, rcfStrengthRadial, 0);
}

[numthreads(16, 16, 1)] void csCircumferentialBlur(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);

	float2 tc = (id.xy + 0.5) / resolution;

	out0[id.xy] = rcfBlur(tc, resolution, rcfStrengthCircumferential, 1);
}

//---
// Sensor dirt
//---

[numthreads(16, 16, 1)] void csSensorDirt(int3 id:SV_DispatchThreadID) {
	float2 tc = id.xy + .5;
}

//---
// Lens dirt
//---
static const float ldMinRadius = 0.025;
static const float ldMaxRadius = 0.075;
static const float ldCellWidth = 3*ldMaxRadius;
static const float ldBorderWidthHalf = 0.00125;
static const float ldSeed = 3;
static const float ldNumLayers = 4;

[numthreads(16, 16, 1)] void csLensDirt(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);

	float2 tc = (id.xy + 0.5) / resolution;
	tc *= float2(resolution.x/resolution.y, 1);
	//float2 tc = id.xy + .5;

	rndSeed = ldSeed;
	
	float4 a = 0;
	for (int i = 0; i < ldNumLayers; ++i) {
		// Modulo repeat, find cell
		float2 p = mod(tc, ldCellWidth) - ldCellWidth / 2;
		float2 c = floor(tc / ldCellWidth);

		// Backup seed
		uint seedBackup = rndSeed;

		// Modifiy seed with cell coordinates
		rndSeed += (c.x + c.y*ldCellWidth + i*ldCellWidth*ldCellWidth) * resolution.x;

		// Determine radius and jitter
		float r = lerp(ldMinRadius, ldMaxRadius, rnd());
		p += float2(srnd(), srnd()) * (ldCellWidth / 2 - r);

		// Shapes
		// Hexagon
		//p = abs(p);
		//float d = max(p.y + p.x*0.57735, p.x*1.1547);

		// Circle
		//float d = length(p);

		// Pentagon
		//float phi = 0.145;
		//float2 n[] = {
		//	{ sin(0*PI/5 + phi), cos(0*PI/5 + phi) },
		//	{ sin(2*PI/5 + phi), cos(2*PI/5 + phi) },
		//	{ sin(4*PI/5 + phi), cos(4*PI/5 + phi) },
		//	{ sin(6*PI/5 + phi), cos(6*PI/5 + phi) },
		//	{ sin(8*PI/5 + phi), cos(8*PI/5 + phi) }
		//};
		//// Difference between circumcircle and incircle radii
		//float delta = r - (1 + sqrt(5)) / 4 * r;
		//float d = max(max(max(max(dot(p, n[0]), dot(p, n[1])), dot(p, n[2])), dot(p, n[3])), dot(p, n[4]));
		//// Mix between sphere and pentagon
		//d = lerp(length(p), d+ delta, 0.5);
		float d = length(p);

		// Random color
		a += (float4(rnd(), rnd(), rnd(), 1) + 0.5) * (smoothstep(r, r - ldBorderWidthHalf * 2, d) + 0.25 * smoothstep(ldBorderWidthHalf, 0, abs(d - r + ldBorderWidthHalf * 2)));
		//a += float4(c, 0, 1)*0.1;

		// Restore seed
		rndSeed = seedBackup;

		// Displace grid
		tc += ldCellWidth / ldNumLayers;
	}

	//out0[id.xy] = float4(a.xyz / ldNumLayers, 1);
	out0[id.xy] = float4(a.xyz / ldNumLayers, 1);
}

//---
// Blur
//---
static const float bRadius = 160;
static const float bSigma = bRadius / 3.5;
static const float bTwoSigmaSquared = 2 * bSigma * bSigma;

void bBlur(int3 id, float2 direction) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);

	float2 sourceResolution;
	tex0.GetDimensions(sourceResolution.x, sourceResolution.y);

	float mipLevel = log2(sourceResolution.x / resolution.x);

	float2 tc = (id.xy + 0.5) / resolution;
	float2 d = direction / resolution;

	float4 a = 0;
	for (float i = -bRadius; i <= bRadius; ++i) {
		float weight = exp(-i*i/bTwoSigmaSquared);
		a += tex0.SampleLevel(sampler0, i * d + tc, mipLevel) * weight;
	}

	out0[id.xy] = a / sqrt(bTwoSigmaSquared * PI);
}

[numthreads(16, 16, 1)] void csBlurH(int3 id:SV_DispatchThreadID) { bBlur(id, float2(1,0)); }
[numthreads(16, 16, 1)] void csBlurV(int3 id:SV_DispatchThreadID) { bBlur(id, float2(0,1)); }

//---
// Streaks
//---
static const float stRadius = 160;

[numthreads(16, 16, 1)] void csStreaks(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);
	
	float2 sourceResolution;
	tex0.GetDimensions(sourceResolution.x, sourceResolution.y);

	float mipLevel = log2(sourceResolution.x / resolution.x);

	float2 tc = (id.xy + 0.5) / resolution;
	float2 d = float2(1, 0) / resolution;

	// Convolution with parabolic tent filter
	float4 a = 0;
	for (float i = -stRadius+1; i < stRadius; ++i) {
		float weight = (stRadius - abs(i)) * (stRadius - abs(i));
		a += tex0.SampleLevel(sampler0, i * d + tc, mipLevel) * weight;
	}

	out0[id.xy] = a * 1.5 / (stRadius * stRadius * stRadius);
}

//---
// Distort Chroma
//---
static const float dcStrength = 0.0;

static const float2 dcCenter = {0.5, 0.5};
static const int dcSamples = 24;

float dcRadialFalloff(float radius, float amount) {
	return radius * (1 + amount*radius*0.01);
}

float4 dcSample(float2 tc, float amount, float2 resolution) {
	// Direction from cdCenter to the current tc, aspect ratio corrected
	float2 d = (tc - dcCenter) * float2(resolution.x/resolution.y, 1);
	float radius = length(d);
	d /= radius == 0 ? d : radius;
	return tex0.SampleLevel(sampler0, dcCenter + d * dcRadialFalloff(radius, amount) / float2(resolution.x/resolution.y, 1), 0);
}

[numthreads(16, 16, 1)] void csDistortChroma(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);

	float2 tc = (id.xy + 0.5) / resolution;

	float4 original = tex0.SampleLevel(sampler0, tc, 0);
	if (dcStrength == 0) {
		out0[id.xy] = original;
		return;
	}

	// Accumulate along spectrum
	float4 a = 0;
	for (int i = 1; i <= dcSamples; ++i) {
		float lambda = 0.4f * (i/float(dcSamples+1)) + 0.35f;
		float4 w = float4(pulse(0.65, 0.1, lambda), pulse(0.55, 0.1, lambda), pulse(0.45, 0.1, lambda), 0.25) * 4.0;
		a += w * dcSample(tc, -(i-1)/float(dcSamples) * dcStrength , resolution);
	}
	
	a /= dcSamples;

	out0[id.xy] = lerp(original, a, saturate(dcStrength));
}

//---
// Ghosts
//---
float4 ghRainbow(float angle) {
  return float4(sin(angle), sin(angle + 3.1415*2.0/3.0), sin(angle + 3.1415*4.0/3.0), 1) * 0.5 + 0.5;
}

[numthreads(16, 16, 1)] void csGhosts(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);

	float2 tc = (id.xy + 0.5) / resolution;
	tc = 1 - tc;
	float2 texelSize = 1 / resolution;
	float2 aspectCorrection = float2(resolution.x/resolution.y, 1);

	float2 direction = (0.5 - tc) * 2;
	float4 a = 0;
	for (int i = 0; i < 5; ++i) {
		float2 sc0 = 1 - frac(tc + direction * i * 0.2);
		float2 sc1 = 1 - frac(tc + direction * i * 0.1999);
		float r = length((sc0 - 0.5) * aspectCorrection);
		a += max(0, tex0.SampleLevel(sampler0, sc0, 0) - 0.125) * lerp(1, ghRainbow(r*20), 1.5*r);
		a += max(0, tex1.SampleLevel(sampler0, sc0, 0) - 0.125) * lerp(1, ghRainbow(r*15+i), 1.5*r);
	}

	out0[id.xy] = float4(a.xyz/5, 1);
}


//---
// Merge
//---
float3 saturation(float3 color, float saturation) {
	return lerp(dot(color, float3(0.2126, 0.7152, 0.0722)), color, saturation);
}

float3 liftGammaGain(float3 color, float3 lift, float3 gamma, float3 gain) {
	return pow(saturate(gain*color + lift*(-gain*color + 1)), 1/gamma);
}


[numthreads(16, 16, 1)] void csMerge(int3 id:SV_DispatchThreadID) {
	float2 resolution;
	out0.GetDimensions(resolution.x, resolution.y);

	float2 tc = (id.xy + 0.5) / resolution;

	rndSeed = (id.y * 720 + id.x) * int(time * 1000);
	float3 rndNoiseXYT = float3(rnd(), rnd(), rnd());

	rndSeed = id.y;
	float3 rndNoiseY = float3(rnd(), rnd(), rnd());


	float3 base = tex0.SampleLevel(sampler0, tc, 0).xyz;
	float3 blur1 = tex1.SampleLevel(sampler0, tc, 0).xyz;
	float3 blur8 = tex2.SampleLevel(sampler0, tc, 0).xyz;
	float3 blur32 = tex3.SampleLevel(sampler0, tc, 0).xyz;
	float3 streaks = tex4.SampleLevel(sampler0, tc, 0).xyz;
	//float3 lensDirt = tex5.SampleLevel(sampler0, tc, 0).xyz;
	//float3 ghosts = tex6.SampleLevel(sampler0, tc, 0).xyz;

	float3 color = base.xyz;

	color += blur1 *  0.025;
	color += blur8 *  0.025;
	color += blur32 * 0.025;

	// Streaks
	color += streaks * (lerp(1, rndNoiseY.x, 0.25)) * 0.25;// * rnd();
	
	// Lens dirt
	//color += lensDirt * 0.25 * max(0, dot(tex2.SampleLevel(sampler0, frac(1-tc), 0).xyz, float3(0.299, 0.587, 0.114)) - 1.0);

	// Ghosts
	//color += ghosts * 0.125;// * ghosts * 0.25;

	// Grain
	//color *= lerp(1, rndNoiseXYT, 0.35);

	//color = base;

	// Tonemapping
	color = max(0, uncharted2ToneMappingCompact(max(0, color*0.5)));
	//color *= color;

	color = liftGammaGain(color, float3(0, 0, 0), float3(1, 1, 1), float3(1, 1, 1));
	//color = saturation(color, 1.5);
	color = pow(color, 1/2.2);
	float4 result = float4(color, dot(color, float3(0.299, 0.587, 0.114)));

	out0[id.xy] = result;
}