source: GTP/trunk/App/Demos/Vis/FriendlyCulling/src/shaders/ssao.cg @ 3034

#include "../shaderenv.h"

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


#define USE_EYE_SPACE_DEPTH 1


struct fragment
{
        float2 texCoord: TEXCOORD0; 
        float3 view: TEXCOORD1;
};


struct pixel
{
        float4 illum_col: COLOR0;
};


inline 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;
}


inline float3 Interpol(float2 w, float3 bl, float3 br, float3 tl, float3 tr)
{
        float3 x1 = lerp(bl, tl, w.y);
        float3 x2 = lerp(br, tr, w.y); 
        float3 v = lerp(x1, x2, w.x); 

        return v;
}


// reconstruct world space position
inline float3 ReconstructSamplePos(uniform sampler2D colors,
                                                                   float2 texcoord, 
                                                                   float3 bl, float3 br, float3 tl, float3 tr)
{
        //const float eyeSpaceDepth = tex2Dlod(colors, float4(texcoord, 0, 0)).w;
        const float eyeSpaceDepth = tex2D(colors, texcoord).w;
        float3 viewVec = Interpol(texcoord, bl, br, tl, tr);
        float3 samplePos = -viewVec * eyeSpaceDepth;

        return samplePos;
}


/** The ssao shader returning the an intensity value between 0 and 1
*/
float2 ssao(fragment IN,
                   uniform sampler2D colors,
                   uniform sampler2D noiseTexture,
                   uniform float2 samples[NUM_SAMPLES],
                   uniform float3 currentNormal,
                   uniform float3 centerPosition,
                   uniform float scaleFactor,
                   uniform float3 bl,
                   uniform float3 br,
                   uniform float3 tl,
                   uniform float3 tr, 
                   uniform float3 viewDir
                   )
{
        // 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;
                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
                const float2 texcoord = IN.texCoord.xy + offsetTransformed * scaleFactor;

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

                const float3 samplePos = ReconstructSamplePos(colors, texcoord, bl, br, tl, tr);


                ////////////////
                //-- compute contribution of sample using the direction and angle

                float3 dirSample = samplePos - centerPosition;
                const float magSample = length(dirSample);
                // normalize
                dirSample /= magSample;

                // angle between current normal and direction to sample controls AO intensity.
                const float cosAngle = max(dot(dirSample, 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 intensity = SAMPLE_INTENSITY / (DISTANCE_SCALE + magSample * magSample);

#if 1
                // 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 viewCorrection = 1.0f + VIEW_CORRECTION_SCALE * dot(viewDir, currentNormal);
                total_ao += cosAngle * intensity * viewCorrection;
#else
                total_ao += cosAngle * intensity;
#endif
        }

        return float2(max(0.0f, 1.0f - total_ao), numSamples);
}

#pragma position_invariant main

/** The mrt shader for screen space ambient occlusion
*/
pixel main(fragment IN, 
                   uniform sampler2D colors,
                   uniform sampler2D normals,
                   uniform sampler2D noise,
                   uniform float2 samples[NUM_SAMPLES],
                   uniform sampler2D oldTex,
                   const uniform float4x4 oldModelViewProj,
                   const uniform float4x4 modelViewProj,
                   uniform float temporalCoherence,
                   uniform float3 eyePos,
                   uniform float3 bl,
                   uniform float3 br,
                   uniform float3 tl,
                   uniform float3 tr
                   )
{
        pixel OUT;

        float4 norm = tex2Dlod(normals, float4(IN.texCoord, 0 ,0));
        const float3 normal = normalize(norm.xyz);

        /// reconstruct position from the eye space depth
        float3 viewDir = IN.view;
        const float eyeDepth = tex2Dlod(colors, float4(IN.texCoord, 0, 0)).w;
        const float3 eyeSpacePos = -viewDir * eyeDepth;
        const float4 worldPos = float4(eyePos + eyeSpacePos, 1.0f);


        ////////////////
        //-- calculcate the current projected depth for next frame
        
        float4 currentPos = mul(modelViewProj, worldPos);
        
        const float w = SAMPLE_RADIUS / currentPos.w;
        currentPos /= currentPos.w;
        
        const float precisionScale = 1e-3f;
        const float currentDepth = currentPos.z * precisionScale;

        const float2 ao = ssao(IN, colors, noise, samples, normal, eyeSpacePos, w, bl, br, tl, tr, normalize(viewDir));


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


        // reproject new frame into old one 
        
        // calculate projected depth
        float4 projPos = mul(oldModelViewProj, worldPos);
        projPos /= projPos.w;
        
        // the current depth projected into the old frame
        const float projDepth = projPos.z * precisionScale;

        // fit from unit cube into 0 .. 1
        const float2 tex = projPos.xy * 0.5f + 0.5f;

        // retrieve the sample from the last frame
        float4 oldCol = tex2D(oldTex, tex);

        const float oldDepth = oldCol.z;
        //const float depthDif = 1.0f - projDepth / oldDepth;
        const float depthDif = projDepth - oldDepth;


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

        float newWeight;

        // the number of valid samples in this frame
        //const float newNumSamples = ao.y;

        if ((temporalCoherence > 0) &&
                (tex.x >= 0.0f) && (tex.x < 1.0f) && 
                (tex.y >= 0.0f) && (tex.y < 1.0f) && 
                (abs(depthDif) < MIN_DEPTH_DIFF) 
                // if visibility changed in the surrounding area we have to recompute
                //&& (oldNumSamples > 0.8f * newNumSamples)
                )
        {
                // increase the weight for convergence
                newWeight = oldWeight + 1.0f;
                OUT.illum_col.x = (ao.x + oldCol.x * oldWeight) / newWeight;
                //if (!(oldNumSamples > ao.y - 1.5f)) newWeight = 0;
        }
        else
        {       
                OUT.illum_col.x = ao.x;
                newWeight = 0;
        }

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

        return OUT;
}


float Filter(float2 texCoord, 
                         uniform sampler2D ssaoTex,
                         uniform float2 filterOffs[NUM_DOWNSAMPLES],
                         uniform float filterWeights[NUM_DOWNSAMPLES]
)
{
        float average = .0f;
        float w = .0f;

        for (int i = 0; i < NUM_DOWNSAMPLES; ++ i)
        {
                average += filterWeights[i] * tex2Dlod(ssaoTex, float4(texCoord + filterOffs[i], 0, 0)).x;
                w += filterWeights[i];
        }

        average *= 1.0f / (float)w;

        return average;
}


pixel combine(fragment IN, 
                          uniform sampler2D colors,
                          uniform sampler2D ssaoTex,
                          uniform float2 filterOffs[NUM_DOWNSAMPLES],
                          uniform float filterWeights[NUM_DOWNSAMPLES]
                          )
{
        pixel OUT;

        float4 col = tex2Dlod(colors, float4(IN.texCoord, 0, 0));
        float3 ao = tex2Dlod(ssaoTex, float4(IN.texCoord, 0, 0));

        //if (ao.y < 2000.0f) 
        //      ao.x = Filter(IN.texCoord, ssaoTex, filterOffs, filterWeights);

        OUT.illum_col = col * ao.x;
        //OUT.illum_col.xyz = float3(ao.x,1-ao.y*1e-2f, 0);
        OUT.illum_col.w = col.w;

        return OUT;
}