// STATIC: "CONVERT_TO_SRGB" "0..1" [ps20b][= g_pHardwareConfig->NeedsShaderSRGBConversion()] [PC] // STATIC: "CONVERT_TO_SRGB" "0..0" [= 0] [XBOX] // STATIC: "LINEAR_INPUT" "0..1" [ps20b] // STATIC: "LINEAR_OUTPUT" "0..1" [ps20b] // DYNAMIC: "AA_ENABLE" "0..1" // rem DYNAMIC: "AA_DEBUG_MODE" "0..3" #define AA_DEBUG_MODE 0 // DYNAMIC: "AA_QUALITY_MODE" "0..0" [ps20] // DYNAMIC: "AA_QUALITY_MODE" "0..1" [ps20b] // DYNAMIC: "AA_QUALITY_MODE" "0..1" [ps30] // DYNAMIC: "AA_REDUCE_ONE_PIXEL_LINE_BLUR" "0..0" [ps20] // DYNAMIC: "AA_REDUCE_ONE_PIXEL_LINE_BLUR" "0..1" [ps20b] // DYNAMIC: "AA_REDUCE_ONE_PIXEL_LINE_BLUR" "0..1" [ps30] // DYNAMIC: "COL_CORRECT_NUM_LOOKUPS" "0..4" #define HDRTYPE HDR_TYPE_NONE #include "common_ps_fxc.h" #if !(defined(SHADER_MODEL_PS_2_B) || defined(SHADER_MODEL_PS_3_0)) // Only allow debug modes and high-quality mode if in ps2b or higher (not enough instruction slots in ps20) #undef AA_DEBUG_MODE #define AA_DEBUG_MODE 0 #endif /* * Engine_Post combines bloom (the final simple addition) with software anti-aliasing * and colour-correction. Combining them has these benefits: * (a) saves fillrate+bandwidth (big on PC) * (b) saves calls to UpdateScreenEffectTexture (big on 360) * (c) reduces quantization errors caused by multiple passes * (d) improves AA quality (it works better on sRGB values than linear) * * * Software AA Summary * ------------------- * * This AA process works by sampling neighbour pixels (4 or 8 of them): * * 5-tap filter: # 9-tap filter: ### * (AA_QUALITY_MODE 0) ### (AA_QUALITY_MODE 1) ### * # ### * * It then figures out which of these neighbours are 'unlike' the centre pixel. * This is based on RGB distance, weighted by the maximum luminance of the samples * (so the difference between 0.1 and 0.2 is the same as between 0.5 and 1.0). * This detects high-contrast edges in both dark and bright scenes. * * It then counts how many 'unlike' samples there are. Some example cases for 5-tap: * * O # # # # # * OOO OOO #OO OOO #O# #O# * O O O # O # * Zero One TwoA TwoB Three Four * * We then blend towards the average of the unlike neighbours, based on how many * unlike neighbours there are. The key case is 'TwoA' - this detects stairstep pixels * on non-axis-aligned edges. In that case, we blend the output colour towards the * average of the unlike samples by 33%. This yields a 3-pixel transition (0->33->66->100) * where before there was a 1-pixel transition (0->100). * * The 9-tap filter (which works the same as 5-tap, just with more samples and different * weights) has two advantages over the 5-tap filter: * - it can differentiate between stairsteps on 45-degree edges and near-horizontal edges * (so the 5-tap version smudges 45-degree edges more than you want, e.g. chain-link fences) * - it blurs less texture detail, by virtue of averaging out noise over more samples * * One problem case that both filters have to consider is one-pixel-thick lines (this is * case 'TwoB' above). Sometimes you do want to soften these lines (for slivers of brightly-lit * geometry in a dark area, e.g. a window frame), but sometimes you do NOT want to soften them * (for thin geometry which is alternating between 1-pixel and 2-pixel thickness, e.g. cables, * and also where 1-pixel lines appear in textures, e.g. roof tiles). So, blurring of 1-pixel * lines is tunable (it defaults to half-blurred as a compromise between the want/don't cases), * in the 'AA_REDUCE_ONE_PIXEL_LINE_BLUR' section below. Case TwoB is differentiated from TwoA by * computing the centroid of the unlike samples (the centroid will be at zero for case TwoB, * but not for TwoA). * */ sampler BaseTextureSampler : register( s0 ); sampler FBTextureSampler : register( s1 ); sampler ColorCorrectionVolumeTexture0 : register( s2 ); sampler ColorCorrectionVolumeTexture1 : register( s3 ); sampler ColorCorrectionVolumeTexture2 : register( s4 ); sampler ColorCorrectionVolumeTexture3 : register( s5 ); float4 psTapOffs_Packed : register( c0 ); // psTapOffs_packed contains 1-pixel offsets: ( +dX, 0, +dY, -dX ) float4 tweakables : register( c1 ); // (x - AA strength/unused) (y - reduction of 1-pixel-line blur) // (z - edge threshold multipler) (w - tap offset multiplier) float4 uvTransform : register( c2 ); // Transform BaseTexture UVs for use with the FBTexture float ColorCorrectionDefaultWeight : register( c3 ); float4 ColorCorrectionVolumeWeights : register( c4 ); float BloomFactor : register( c5 ); float4 GetBloomColor( float2 bloomUV ) { #if ( LINEAR_INPUT == 1 ) { // In this case, which is only used on OpenGL, we want sRGB data from this tex2D. // Hence, we have to undo the sRGB conversion that we are forced to apply by OpenGL return LinearToGamma( tex2D( BaseTextureSampler, bloomUV ) ); } #else { return tex2D( BaseTextureSampler, bloomUV ); } #endif } float4 PerformColorCorrection( float4 outColor, float2 fbTexCoord ) { #if ( COL_CORRECT_NUM_LOOKUPS > 0 ) { // NOTE: This code requires the color correction texture to be 32 units to be correct. // This code will cause (0,0,0) to be read from 0.5f/32 // and (1,1,1) to be read from 31.5f/32 float4 offsetOutColor = outColor*(31.0f/32.0f) + (0.5f/32.0f); outColor.rgb = outColor.rgb * ColorCorrectionDefaultWeight; outColor.rgb += tex3D( ColorCorrectionVolumeTexture0, offsetOutColor.rgb ) * ColorCorrectionVolumeWeights.x; #if ( COL_CORRECT_NUM_LOOKUPS > 1 ) { outColor.rgb += tex3D( ColorCorrectionVolumeTexture1, offsetOutColor.rgb ) * ColorCorrectionVolumeWeights.y; #if ( COL_CORRECT_NUM_LOOKUPS > 2 ) { outColor.rgb += tex3D( ColorCorrectionVolumeTexture2, offsetOutColor.rgb ) * ColorCorrectionVolumeWeights.z; #if ( COL_CORRECT_NUM_LOOKUPS > 3 ) { outColor.rgb += tex3D( ColorCorrectionVolumeTexture3, offsetOutColor.rgb ) * ColorCorrectionVolumeWeights.w; } #endif } #endif } #endif } #endif return outColor; } float3 PerformAA( float3 baseColor, float2 fbTexCoord, out float3 unlike, out float unlikeSum, out float lerpFactor ) { float3 a, b, c, d, e, f, g, h; float3 dA, dB, dC, dD, dE, dF, dG, dH; float4 deltas, deltas2; float4 weights, weights2; float4 lumS; float maxLumS; // Set FAST_DELTAS to '1' to use Manhattan distance (in colour-space) rather than Euclidean distance: const int FAST_DELTAS = 1; #if AA_QUALITY_MODE == 0 const float COLOUR_DELTA_BASE = (FAST_DELTAS == 0) ? 0.11f : 0.5f; const float COLOUR_DELTA_CONTRAST = 100; // Scaling down colour deltas (DELTA_SCALE) reduces the over-blurring of 45-degree edges // by the 5-tap filter. Conversely, increasing it smooths stairsteps more strongly. const float DELTA_SCALE = 0.75f; #else // AA_QUALITY_MODE == 0 const float COLOUR_DELTA_BASE = (FAST_DELTAS == 0) ? 0.24f : 0.65f; const float COLOUR_DELTA_CONTRAST = 100; const float DELTA_SCALE = 1.0f; #endif // AA_QUALITY_MODE == 0 const float MAX_LERP_FACTOR = 0.66f; const float SQRT3 = 1.73205080757f; float onePixelLineBlurReduction = tweakables.y; // psTapOffs_packed contains 1-pixel offsets: ( +dX, 0, +dY, -dX ) float4 texelDelta = psTapOffs_Packed*tweakables.w; // Allowed ps20 swizzles: // .xyzw on (+dX,0,+dY,-dX) gives: (+dX, 0) & (-dX, 0) (former with 'add', latter with 'sub') // .yzxw on (+dX,0,+dY,-dX) gives: ( 0,+dY) & ( 0,-dY) // .wzyx on (+dX,0,+dY,-dX) gives: (-dX,+dY) & (+dX,-dY) // .zxyw on (not used) // NOTE: These don't give us (+dX,+dY) and (-dX,-dY), we need to copy +dY: ( +dX, 0, +dY, -dX ) -> ( +dX, +dY, +dY, -dX ) // NOTE: tex2D() can't swizzle the source register in ps2x, so we have no // choice but to add each float2 offset to fbTexCoord one at a time :o/ a = tex2D( FBTextureSampler, fbTexCoord + texelDelta.yz ).rgb; // ( 0,+1) b = tex2D( FBTextureSampler, fbTexCoord + texelDelta.xy ).rgb; // (+1, 0) c = tex2D( FBTextureSampler, fbTexCoord - texelDelta.yz ).rgb; // ( 0,-1) d = tex2D( FBTextureSampler, fbTexCoord - texelDelta.xy ).rgb; // (-1, 0) #if AA_QUALITY_MODE == 1 // 9-tap method (do diagonal neighbours too) e = tex2D( FBTextureSampler, fbTexCoord + texelDelta.wz ).rgb; // (-1,+1) f = tex2D( FBTextureSampler, fbTexCoord - texelDelta.wz ).rgb; // (+1,-1) texelDelta.y = texelDelta.z; // Can't quite get all 8 sample offsets from a single float4 with the allowed swizzles! :o/ g = tex2D( FBTextureSampler, fbTexCoord + texelDelta.xy ).rgb; // (+1,+1) h = tex2D( FBTextureSampler, fbTexCoord - texelDelta.xy ).rgb; // (-1,-1) #endif // AA_QUALITY_MODE == 1 // Compute the like<-->unlike weights dA = a - baseColor; dB = b - baseColor; dC = c - baseColor; dD = d - baseColor; #if AA_QUALITY_MODE == 1 dE = e - baseColor; dF = f - baseColor; dG = g - baseColor; dH = h - baseColor; #endif // AA_QUALITY_MODE == 1 #if ( FAST_DELTAS == 0 ) { // Colour-space Euclidean distance deltas = float4( dot(dA, dA), dot(dB, dB), dot(dC, dC), dot(dD, dD) ); deltas = DELTA_SCALE*DELTA_SCALE*(deltas / 3); deltas = sqrt(deltas); } #else { // Colour-space Manhattan distance // OPT: to avoid the 'abs', try dividing colours by maxLumS then dotprodding w/ baseColor deltas.x = dot( abs( dA ), 1 ); deltas.y = dot( abs( dB ), 1 ); deltas.z = dot( abs( dC ), 1 ); deltas.w = dot( abs( dD ), 1 ); deltas *= DELTA_SCALE; } #endif weights = deltas; #if AA_QUALITY_MODE == 1 #if ( FAST_DELTAS == 0 ) { deltas2 = float4( dot(dE, dE), dot(dF, dF), dot(dG, dG), dot(dH, dH) ); deltas2 = DELTA_SCALE*DELTA_SCALE*(deltas2 / 3); deltas2 = sqrt(deltas2); } #else { deltas2.x = dot( abs( dE ), 1); deltas2.y = dot( abs( dF ), 1); deltas2.z = dot( abs( dG ), 1); deltas2.w = dot( abs( dH ), 1); deltas2 *= DELTA_SCALE; } #endif weights2 = deltas2; #endif // AA_QUALITY_MODE == 1 // Adjust weights relative to maximum sample luminance (local, relative contrast: 0.1 Vs 0.2 is the same as 0.5 Vs 1.0) lumS = float4( dot(a, a), dot(b, b), dot(c, c), dot(d, d) ); lumS.xy = max( lumS.xy, lumS.wz ); lumS.x = max( lumS.x, lumS.y ); maxLumS = max( lumS.x, dot( baseColor, baseColor ) ); #if AA_QUALITY_MODE == 1 lumS = float4( dot(e, e), dot(f, f), dot(g, g), dot(h, h) ); lumS.xy = max( lumS.xy, lumS.wz ); lumS.x = max( lumS.x, lumS.y ); maxLumS = max( lumS.x, maxLumS ); #endif // AA_QUALITY_MODE == 1 float lumScale = 1.0f / sqrt( maxLumS ); weights *= lumScale; #if AA_QUALITY_MODE == 1 weights2 *= lumScale; #endif // AA_QUALITY_MODE == 1 // Contrast-adjust weights such that only large contrast differences are taken into account // (pushes weights to 0.0 for 'like' neighbours and to 1.0 for 'unlike' neighbours) float colourDeltaBase = tweakables.z*COLOUR_DELTA_BASE; weights = saturate(colourDeltaBase + COLOUR_DELTA_CONTRAST*(weights - colourDeltaBase)); #if AA_QUALITY_MODE == 1 weights2 = saturate(colourDeltaBase + COLOUR_DELTA_CONTRAST*(weights2 - colourDeltaBase)); #endif // AA_QUALITY_MODE == 1 // Determine the average 'unlike' colour unlikeSum = dot(weights, 1); unlike = weights.x*a + weights.y*b + weights.z*c + weights.w*d; #if AA_QUALITY_MODE == 1 unlikeSum += dot(weights2, 1); unlike += weights2.x*e + weights2.y*f + weights2.z*g + weights2.w*h; #endif // AA_QUALITY_MODE == 1 // NOTE: this can cause div-by-zero, but lerpFactor ends up at zero in that case so it doesn't matter unlike = unlike / unlikeSum; #if AA_REDUCE_ONE_PIXEL_LINE_BLUR // Reduce lerpFactor for 1-pixel-thick lines - otherwise you lose texture detail, and it looks // really weird where geometry (e.g. cables) alternates between being 1 and 2 pixels thick. // [ The "*2" below is because the values here were tuned to reduce blurring one 1-pixel lines // by about half (which is a good compromise between the bad cases at either end). So you // want the controlling convar to default to 0.5 ] const float ONE_PIXEL_LINE_BIAS_BASE = 0.4f; const float ONE_PIXEL_LINE_BIAS_CONTRAST = 16.0f; float2 unlikeCentroid = 0; unlikeCentroid.x += dot( 1-weights, float4( 0, +1, 0, -1 ) ); // This 2x4 matrix is the transpose of unlikeCentroid.y += dot( 1-weights, float4( +1, 0, -1, 0 ) ); // the neighbour sample texel offsets #if AA_QUALITY_MODE == 0 unlikeCentroid /= 4 - unlikeSum; #else // AA_QUALITY_MODE == 0 unlikeCentroid.x += dot( 1-weights2, float4( -1, +1, +1, -1 ) ); unlikeCentroid.y += dot( 1-weights2, float4( +1, -1, +1, -1 ) ); unlikeCentroid /= 8 - unlikeSum; #endif // AA_QUALITY_MODE == 0 float onePixelLineBias = 1 - saturate( length(unlikeCentroid) ); // OPTIMIZE: try using distSquared, remove this sqrt onePixelLineBias = onePixelLineBlurReduction*saturate(ONE_PIXEL_LINE_BIAS_BASE + ONE_PIXEL_LINE_BIAS_CONTRAST*(onePixelLineBias - ONE_PIXEL_LINE_BIAS_BASE)); #if AA_QUALITY_MODE == 0 unlikeSum -= 2*onePixelLineBias*0.4f*saturate( 3 - unlikeSum ); // The 'min' thing avoids this affecting lone/pair pixels #else // AA_QUALITY_MODE == 0 unlikeSum -= 2*onePixelLineBias*1.9f*saturate( 7 - unlikeSum ); #endif // AA_QUALITY_MODE == 0 #endif // AA_REDUCE_ONE_PIXEL_LINE_BLUR // Compute the lerp factor we use to blend between 'baseColor' and 'unlike'. // We want to lerp 'stairstep' pixels (which have 2 unlike neighbours) // 33% towards the 'unlike' colour, such that these hard, 1-pixel transitions // (0% -> 100%) become soft, 3-pixel transitions (0% -> 33% -> 66% -> 100%). float strengthMultiplier = tweakables.x; #if ( AA_QUALITY_MODE == 0 ) { lerpFactor = saturate( strengthMultiplier*DELTA_SCALE*( (unlikeSum - 1) / 3 ) ); // Uncomment the following to blend slightly across vertical/horizontal edges (better for 45-degree edges, worse for 90-degree edges) //lerpFactor = saturate( strengthMultiplier*DELTA_SCALE*( unlikeSum / 6 ) ); } #else // AA_QUALITY_MODE != 0 { lerpFactor = saturate( strengthMultiplier*DELTA_SCALE*( (unlikeSum - 3) / 3 ) ); } #endif // Clamp the blend factor so that lone dot pixels aren't blurred into oblivion lerpFactor = min( lerpFactor, MAX_LERP_FACTOR ); baseColor = lerp( baseColor, unlike, lerpFactor ); return baseColor; } float4 GenerateAADebugColor( float4 outColor, float3 unlike, float unlikeSum, float lerpFactor ) { #if ( AA_DEBUG_MODE == 1 ) { // Debug: Visualize the number of 'unlike' samples outColor.rgb = 0; if ( AA_QUALITY_MODE == 0 ) { if (unlikeSum >= 0.95f) outColor.rgb = float3(1,0,0); if (unlikeSum >= 1.95f) outColor.rgb = float3(0,1,0); if (unlikeSum >= 2.95f) outColor.rgb = float3(0,0,1); } else { if (unlikeSum >= 2.95f) outColor.rgb = float3(1,0,0); if (unlikeSum >= 3.95f) outColor.rgb = float3(0,1,0); if (unlikeSum >= 4.95f) outColor.rgb = float3(0,0,1); } // Don't sRGB-write } #elif ( AA_DEBUG_MODE == 2 ) { // Debug: Visualize the strength of lerpFactor outColor.rgb = 0; outColor.g = lerpFactor; // Don't sRGB-write } #elif ( AA_DEBUG_MODE == 3 ) { // Debug: Visualize the 'unlike' colour that we blend towards outColor.rgb = lerp( 0, unlike, saturate(5*lerpFactor) ); // Do sRGB-write (if it's enabled) outColor = FinalOutput( outColor, 0, PIXEL_FOG_TYPE_NONE, TONEMAP_SCALE_NONE ); } #endif return outColor; } float2 PerformUVTransform( float2 bloomUVs ) { // NOTE: 'wz' is used since 'zw' is not a valid swizzle for ps20 shaders return bloomUVs*uvTransform.wz + uvTransform.xy; } struct PS_INPUT { float2 baseTexCoord : TEXCOORD0; #if defined( _X360 ) //avoid a shader patch on 360 due to pixel shader inputs being fewer than vertex shader outputs float2 ZeroTexCoord : TEXCOORD1; float2 bloomTexCoord : TEXCOORD2; #endif }; float4 main( PS_INPUT i ) : COLOR { float2 fbTexCoord = PerformUVTransform( i.baseTexCoord ); float3 baseColor = tex2D( FBTextureSampler, fbTexCoord ).rgb; #if ( LINEAR_INPUT == 1 ) { // In this case, which is only used on OpenGL, we want sRGB data from this tex2D. // Hence, we have to undo the sRGB conversion that we are forced to apply by OpenGL baseColor = LinearToGamma( baseColor ); } #endif float4 outColor = float4( baseColor, 1 ); #if ( AA_ENABLE == 1 ) { float unlikeSum, lerpFactor; float3 unlike; outColor.rgb = PerformAA( outColor.rgb, fbTexCoord, unlike, unlikeSum, lerpFactor ); #if ( AA_DEBUG_MODE > 0 ) { return GenerateAADebugColor( outColor, unlike, unlikeSum, lerpFactor ); } #endif } #endif float4 bloomColor = BloomFactor * GetBloomColor( i.baseTexCoord ); outColor.rgb += bloomColor.rgb; outColor = PerformColorCorrection( outColor, fbTexCoord ); outColor = FinalOutput( outColor, 0, PIXEL_FOG_TYPE_NONE, TONEMAP_SCALE_NONE ); // Go to linear since we're forced to do an sRGB write on OpenGL in ps2b #if ( LINEAR_OUTPUT == 1 ) { outColor = GammaToLinear( outColor ); } #endif return outColor; }