#include "../shaderenv.h"

////////////////////
// Screen Spaced Ambient Occlusion shader
// based on shader of Alexander Kusternig


struct fragment
{
	 // normalized screen position
	float4 pos: WPOS;
	float2 texCoord: TEXCOORD0; 
	float3 view: TEXCOORD1;
};


struct pixel
{
	float4 illum_col: COLOR0;
};


float2 myreflect(float2 pt, float2 n)
{
	// distance to plane
	float d = dot(n, pt);
	// reflect around plane
	float2 rpt = pt - d * 2.0f * n;

	return rpt;
}



/** The ssao shader returning the an intensity value between 0 and 1
*/
float2 ssao(fragment IN,
		   uniform sampler2D positions,
		   uniform sampler2D noiseTexture,
		   uniform float2 samples[NUM_SAMPLES],
		   uniform float3 currentNormal,
		   uniform float4 centerPosition,
		   uniform float scaleFactor
		   //,uniform float3 currentViewDir
		   )
{
	// Check in a circular area around the current position.
	// Shoot vectors to the positions there, and check the angle to these positions.
	// Summing up these angles gives an estimation of the occlusion at the current position.

	float total_ao = 0.0;
	float numSamples = 0;

	for (int i = 0; i < NUM_SAMPLES; ++ i) 
	{
		const float2 offset = samples[i];

#if 1
		////////////////////
		// add random noise: reflect around random normal vector (warning: slow!)
		float2 mynoise = tex2D(noiseTexture, IN.texCoord.xy).xy;
		const float2 offsetTransformed = myreflect(offset, mynoise);
#else
		const float2 offsetTransformed = offset;
#endif
		// weight with projected coordinate to reach similar kernel size for near and far
		float2 texcoord = IN.texCoord.xy + offsetTransformed * AREA_SIZE * scaleFactor;

		if ((texcoord.x <= 1.0f) && (texcoord.x >= 0.0f) && (texcoord.y <= 1.0f) && (texcoord.y >= 0.0f))
			++ numSamples;

		// sample downsampled texture in order to speed up texture accesses
		float3 sample_position = tex2Dlod(positions, float4(texcoord, 0, SSAO_MIPMAP_LEVEL)).xyz;
		//float3 sample_position = tex2D(positions, texcoord).xyz;

		float3 vector_to_sample = sample_position - centerPosition.xyz;
		const float length_to_sample = length(vector_to_sample);

		float3 direction_to_sample = vector_to_sample / length_to_sample;

		// angle between current normal and direction to sample controls AO intensity.
		const float cos_angle = max(dot(direction_to_sample, currentNormal), 0.0f);

		// the distance_scale offset is used to avoid singularity that occurs at global illumination when 
		// the distance to a sample approaches zero
		const float distance_intensity = 
			(SAMPLE_INTENSITY * DISTANCE_SCALE) / (DISTANCE_SCALE + length_to_sample * length_to_sample);

#if 0
		// if surface normal perpenticular to view dir, approx. half of the samples will not count
		// => compensate for this (on the other hand, projected sampling area could be larger!)
		const float view_correction = 1.0f + VIEW_CORRECTION_SCALE * (1.0f - dot(currentViewDir, currentNormal));
		total_ao += cos_angle * distance_intensity * view_correction;
#endif
		total_ao += cos_angle * distance_intensity;
	}

	return float2(max(0.0f, 1.0f - total_ao), numSamples);
	//return saturate(dot(currentViewDir, currentNormal));
}


/** The mrt shader for screen space ambient occlusion
*/
pixel main(fragment IN, 
		   uniform sampler2D colors,
		   uniform sampler2D positions,
		   uniform sampler2D normals,
		   uniform sampler2D noiseTexture,
		   uniform float2 samples[NUM_SAMPLES],
		   uniform sampler2D oldTex,
		   const uniform float4x4 oldModelViewProj,
		   uniform float maxDepth,
		   uniform float temporalCoherence
		   )
{
	pixel OUT;

	float4 norm = tex2D(normals, IN.texCoord.xy);
	float3 normal = normalize(norm.xyz);

	// a constant ambient term
	const float amb = norm.w;
	// the w coordinate from the persp. projection
	float w = norm.w;
	// the current world position
	const float4 centerPosition = tex2D(positions, IN.texCoord.xy);
	// the current color
	const float currentDepth = centerPosition.w;

	const float2 ao = ssao(IN, positions, noiseTexture, samples, normal, centerPosition, w);//, viewDir);
		

	/////////////////
	//-- compute temporally smoothing

	float4 realPos = centerPosition * maxDepth;
	realPos.w = 1.0f;


	///////////
	//-- reprojection from new frame into old one

	float4 oldPos = mul(oldModelViewProj, realPos);

	const float newDepth = oldPos.z / oldPos.w;

	float2 tex = (oldPos.xy / oldPos.w) * 0.5f + 0.5f;
	float4 oldCol = tex2D(oldTex, tex);

	const float oldDepth = oldCol.w;
	const float depthDif = 1.0f - newDepth / oldDepth;

	float oldWeight = clamp(oldCol.z, 0, temporalCoherence);
	float oldNumSamples = oldCol.y;

	float newWeight;

	if (//(temporalCoherence > 0) &&
		(tex.x >= 0.0f) && (tex.x < 1.0f) && 
		(tex.y >= 0.0f) && (tex.y < 1.0f) && 
		(abs(depthDif) < 1e-3f) 
		// check if visibility changed in the surrounding area:
		// then we have to recompute
		&& (oldNumSamples > ao.y - 1.5f) 
		)
	{
		// increase the weight for convergence
		newWeight = oldWeight + 1.0f;
		OUT.illum_col.xy = (ao.xy + oldCol.xy * oldWeight) / newWeight;

		//if (!(oldNumSamples > ao.y - 1.5f)) newWeight = 0;
	}
	else
	{
		OUT.illum_col.xy = ao.xy;
		newWeight = 0;
	}

	OUT.illum_col.z = newWeight;
	OUT.illum_col.w = currentDepth;

	return OUT;
}


pixel combine(fragment IN, 
			  uniform sampler2D colors,
			  uniform sampler2D ssaoTex,
			  uniform sampler2D positions)
{
	pixel OUT;

	float4 col = tex2Dlod(colors, float4(IN.texCoord.xy, 0, 0));
	float4 ao = tex2D(ssaoTex, IN.texCoord.xy);

	OUT.illum_col = col * ao.x;
	//OUT.illum_col.w = ao.w;
	OUT.illum_col.w = col.w;

	return OUT;
}