source: GTP/trunk/App/Demos/Vis/FriendlyCulling/src/shaders/deferred.cg @ 3160

#include "../shaderenv.h"


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


struct pixel
{
        float4 color: 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;
}


/** function for standard deferred shading
*/
float4 shade(fragment IN, 
                         uniform float4 color,
                         uniform float3 normal,
                         float3 lightDir)
{
        // diffuse intensity
        const float angle = saturate(dot(normal, lightDir)); 
        
        float4 lightDiffuse = glstate.light[0].diffuse;
        float4 diffuse = angle * lightDiffuse;

        // global ambient
        const float4 ambient = glstate.light[0].ambient;
        
        float4 outColor;

        // hack: prevent shading the sky
        if (color.w > 1e19f) outColor = color;
        else outColor = (ambient + diffuse) * color;

        return outColor;
}



/** The mrt shader for standard rendering
*/
pixel main(fragment IN, 
                   uniform sampler2D colors,
                   uniform sampler2D normals,
                   uniform float3 lightDir
                   )
{
        pixel OUT;

        float4 norm = tex2D(normals, IN.texCoord);
        float4 color = tex2Dlod(colors, float4(IN.texCoord, 0, 0));
        
        float3 normal = normalize(norm.xyz);
        float4 col = shade(IN, color, normal, lightDir);
        
        OUT.color = col;
        // store scaled view vector so wie don't have to normalize for e.g., ssao
        OUT.color.w = color.w;// / length(IN.view);
        //OUT.color = color;
        return OUT;
}


float CalcShadowTerm(fragment IN,
                                         uniform sampler2D shadowMap,
                                         uniform float scale,
                                         uniform float2 lightSpacePos,
                                         uniform float depth,
                                         uniform float2 samples[NUM_PCF_TABS],
                                         uniform float weights[NUM_PCF_TABS],
                                         uniform sampler2D noiseTexture
                                         )
{
        //float shadowDepth = tex2D(shadowMap, lightSpacePos).x;
        //return step(depth, shadowDepth);

        float total_d = .0f;
        float total_w = .0f;

        for (int i = 0; i < NUM_PCF_TABS; ++ i) 
        {
                const float2 offset = samples[i];
                const float w = weights[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
                float2 texcoord = lightSpacePos + offsetTransformed * scale;

                float shadowDepth = tex2D(shadowMap, texcoord).x;

                total_d += w * step(depth, shadowDepth);
                total_w += w;
        }

        total_d /= (float)total_w;

        return total_d;
}


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;
}


pixel main_shadow(fragment IN, 
                                  uniform sampler2D colors,
                                  uniform sampler2D positions,
                                  uniform sampler2D normals,               
                                  uniform sampler2D shadowMap,
                                  uniform float4x4 shadowMatrix,
                                  uniform float sampleWidth,
                                  uniform sampler2D noiseTex,
                                  uniform float2 samples[NUM_PCF_TABS],
                                  uniform float weights[NUM_PCF_TABS],
                                  uniform float3 lightDir,
                                  uniform float3 eyePos,
                                  uniform float3 bl,
                                  uniform float3 br,
                                  uniform float3 tl,
                                  uniform float3 tr
                                  )
{
        pixel OUT;

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

        float4 color = tex2Dlod(colors, float4(IN.texCoord, 0, 0));

        /// reconstruct position from the eye space depth
        float3 viewDir = IN.view;
        const float lenView = length(viewDir);
        viewDir /= lenView;

        const float eyeDepth = tex2Dlod(colors, float4(IN.texCoord, 0, 0)).w;

        const float4 worldPos = float4(eyePos - viewDir * eyeDepth, 1);
        
        // diffuse intensity
        const float angle = saturate(dot(normal, lightDir)); 
        const float4 lightDiffuse = glstate.light[0].diffuse;
        
        float4 diffuse = lightDiffuse * angle;

        // hack: prevent shadowing the sky      
        const bool useShading = (color.w < 1e19f);

        // calc diffuse illumination + shadow term
        if (useShading && 
                (angle > 1e-3f) // shadow only if diffuse color has some minimum intensity
                )
        {
                float4 lightSpacePos = mul(shadowMatrix, worldPos);
                lightSpacePos /= lightSpacePos.w;

                float shadowTerm = CalcShadowTerm(IN, shadowMap, sampleWidth, lightSpacePos.xy, lightSpacePos.z, samples, weights, noiseTex);
        
                diffuse *= shadowTerm;
        }

        // light ambient term
        const float4 ambient = glstate.light[0].ambient;
        // compute shading
        OUT.color = useShading ? (ambient + diffuse) * color : color;
        // store scaled view vector from now on so wie don't have to normalize later (e.g., for ssao)
        //OUT.color.w = color.w / lenView;

        return OUT;
}


float4 Output(fragment IN, uniform sampler2D colors): COLOR
{    
        return tex2Dlod(colors, float4(IN.texCoord, 0, 0));
}


float4 ScaleDepth(fragment IN,
                                  uniform sampler2D colors): COLOR
{    
        float4 color = tex2Dlod(colors, float4(IN.texCoord, 0, 0));
        // store scaled view vector so wie don't have to normalize for e.g., ssao
        color.w /= length(IN.view);
        
        return color;
}



inline float SqrLen(float3 v)
{
        return v.x * v.x + v.y * v.y + v.z * v.z;
}



/** This shader computes the reprojection and checks
        if the reprojected pixel from last frame is still
        valid in the current frame
        */
inline float PixelValid(sampler2D diffVals,
                                                sampler2D oldTex,
                                                float4 color,
                                                float2 texCoord,
                                                float3 viewDir,
                                                float3 oldEyePos,
                                                float4x4 modelViewProj,
                                                float4x4 oldModelViewProj,
                                                float3 oldbl,
                                                float3 oldbr,
                                                float3 oldtl,
                                                float3 oldtr
                                                )
{
        // reconstruct position from the eye space depth
        const float eyeSpaceDepth = color.w;
        const float4 worldPos = float4(-viewDir * eyeSpaceDepth, 1.0f);

        float3 diffVec = tex2Dlod(diffVals, float4(texCoord, 0, 0)).xyz;
        

        ////////////////
        //-- calculcate the current projected posiion (also used for next frame)
        
        float4 projPos = mul(modelViewProj, worldPos);
        const float invw = 1.0f / projPos.w;
        projPos *= invw;

        // compute position from old frame for dynamic objects + translational portion
        const float3 translatedPos = diffVec - oldEyePos + worldPos.xyz;


        /////////////////
        //-- reproject into old frame and calculate texture position of sample in old frame

        // note: the old model view matrix only holds the view orientation part
        float4 backProjPos = mul(oldModelViewProj, float4(translatedPos, 1.0f));
        backProjPos /= backProjPos.w;
        
        // fit from unit cube into 0 .. 1
        const float2 oldTexCoords = backProjPos.xy * 0.5f + 0.5f;
        //const float2 oldTexCoords = texCoord;
        // retrieve the sample from the last frame
        const float4 oldPixel = tex2Dlod(oldTex, float4(oldTexCoords, .0f, .0f));

        // calculate eye space position of sample in old frame
        const float oldEyeSpaceDepth = oldPixel.w;

        // vector from eye pos to old sample 
        const float3 oldViewDir = Interpol(oldTexCoords, oldbl, oldbr, oldtl, oldtr);
        const float invLen = 1.0f / length(oldViewDir);
        const float projectedEyeSpaceDepth = invLen * length(translatedPos);
        
        const float depthDif = abs(1.0f - oldEyeSpaceDepth / projectedEyeSpaceDepth);

        const float squaredLen = SqrLen(diffVec);
        
        // test if this pixel was not valid in the old frame
        float validPixel;

        if ((((squaredLen <= DYNAMIC_OBJECTS_THRESHOLD) && (oldPixel.z <= DYNAMIC_OBJECTS_THRESHOLD)) || 
                (depthDif <= 1e-2))
                && (oldTexCoords.x >= 0.0f) && (oldTexCoords.x < 1.0f)
                && (oldTexCoords.y >= 0.0f) && (oldTexCoords.y < 1.0f)
                )
        {
                validPixel = 0.0f;
        }
        else
        {       
                validPixel = 10.5f;
        }

        //return depthDif;
        return validPixel;
}


float4 PrepareSsao(fragment IN,
                                   uniform sampler2D colorsTex,
                                   uniform sampler2D oldTex,
                                   uniform float4x4 modelViewProj,
                                   uniform float4x4 oldModelViewProj,
                                   uniform float3 oldbl,
                                   uniform float3 oldbr,
                                   uniform float3 oldtl,
                                   uniform float3 oldtr,
                                   uniform float3 oldEyePos,
                                   uniform sampler2D diffVals
                                   ): COLOR
{   
        //const float3 normal = normalize(tex2Dlod(normals, float4(IN.texCoord, 0 ,0)).xyz);
        const float3 viewDir = IN.view;
        float4 color = tex2Dlod(colorsTex, float4(IN.texCoord, 0, 0));

        // store scaled view vector so wie don't have to normalize for e.g., ssao
        color.w /= length(IN.view);
        //color.w = 1;

        // do reprojection and filter out the pixels that are not save
        float pValid = PixelValid(diffVals,
                oldTex,
                color, 
                IN.texCoord,
                viewDir,
                oldEyePos,
                modelViewProj,
                oldModelViewProj,
                oldbl, oldbr, oldtl, oldtr
                );

        color.x = pValid;
        return color;
}


float4 DownSample(fragment IN,
                                  uniform sampler2D colors,
                                  uniform float2 downSampleOffs[NUM_DOWNSAMPLES]): COLOR
{    
        // let bilinear filtering do its work
        float4 color = tex2Dlod(colors, float4(IN.texCoord, 0, 0));
        return color;
}