1#version 450 2 3/* Ported by Hyllian and hunterk - 2015 / 2017 */ 4 5// Copyright (c) 2015-2017, bacondither 6// All rights reserved. 7// 8// Redistribution and use in source and binary forms, with or without 9// modification, are permitted provided that the following conditions 10// are met: 11// 1. Redistributions of source code must retain the above copyright 12// notice, this list of conditions and the following disclaimer 13// in this position and unchanged. 14// 2. Redistributions in binary form must reproduce the above copyright 15// notice, this list of conditions and the following disclaimer in the 16// documentation and/or other materials provided with the distribution. 17// 18// THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR 19// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 20// OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21// IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 22// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 23// NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 24// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 25// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 26// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 27// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 28 29// Second pass, MUST BE PLACED IMMEDIATELY AFTER THE FIRST PASS IN THE CHAIN 30 31// Adaptive sharpen - version 2017-04-11 - (requires ps >= 3.0) 32// Tuned for use post-resize, EXPECTS FULL RANGE GAMMA LIGHT 33 34layout(push_constant) uniform Push 35{ 36 vec4 SourceSize; 37 vec4 OriginalSize; 38 vec4 OutputSize; 39 uint FrameCount; 40 float CURVE_HEIGHT; 41 float VIDEO_LEVEL_OUT; 42} params; 43 44#pragma parameter CURVE_HEIGHT "AS Curve Height" 1.0 0.3 2.0 0.1 45#pragma parameter VIDEO_LEVEL_OUT "AS Video Lvl Out" 0.0 0.0 1.0 1.0 46 47#define mul(a,b) (b*a) 48#define saturate(c) clamp(c, 0.0, 1.0) 49 50layout(std140, set = 0, binding = 0) uniform UBO 51{ 52 mat4 MVP; 53} global; 54 55//--------------------------------------- Settings ------------------------------------------------ 56 57#define curve_height params.CURVE_HEIGHT // Main sharpening strength, POSITIVE VALUES ONLY! 58 // 0.3 <-> 2.0 is a reasonable range of values 59 60#define video_level_out params.VIDEO_LEVEL_OUT // True to preserve BTB & WTW (minor summation error) 61 // Normally it should be set to false 62 63//------------------------------------------------------------------------------------------------- 64// Defined values under this row are "optimal" DO NOT CHANGE IF YOU DO NOT KNOW WHAT YOU ARE DOING! 65 66#define curveslope 0.4 // Sharpening curve slope, high edge values 67 68#define L_overshoot 0.003 // Max light overshoot before compression [>0.001] 69#define L_compr_low 0.169 // Light compression, default (0.169=~9x) 70#define L_compr_high 0.337 // Light compression, surrounded by edges (0.337=~4x) 71 72#define D_overshoot 0.009 // Max dark overshoot before compression [>0.001] 73#define D_compr_low 0.253 // Dark compression, default (0.253=~6x) 74#define D_compr_high 0.504 // Dark compression, surrounded by edges (0.504=~2.5x) 75 76#define scale_lim 0.1 // Abs max change before compression [>0.01] 77#define scale_cs 0.056 // Compression slope above scale_lim 78 79#define dW_lothr 0.3 // Start interpolating between W1 and W2 80#define dW_hithr 0.8 // When dW is equal to W2 81 82#define lowthr_mxw 0.11 // Edge value for max lowthr weight [>0.01] 83 84#define pm_p 0.75 // Power mean p-value [>0-1.0] 85 86#define alpha_out 1.0 // MPDN requires the alpha channel output to be 1.0 87 88//------------------------------------------------------------------------------------------------- 89#define w_offset 1.0 // Edge channel offset, must be the same in all passes 90#define bounds_check true // If edge data is outside bounds, make pixels green 91//------------------------------------------------------------------------------------------------- 92 93// Soft if, fast approx 94#define soft_if(a,b,c) ( saturate((a + b + c - 3*w_offset + 0.06)/(abs(maxedge) + 0.03) - 0.85) ) 95 96// Soft limit, modified tanh 97#define soft_lim(v,s) ( ((exp(2.*min(abs(v), s*24.)/s) - 1.)/(exp(2*min(abs(v), s*24.)/s) + 1.))*s ) 98 99// Weighted power mean 100#define wpmean(a,b,w) ( pow((w*pow(abs(a), pm_p) + abs(1-w)*pow(abs(b), pm_p)), (1.0/pm_p)) ) 101 102// Get destination pixel values 103#define get(x,y) ( texture(Source, coord + vec2(x*(px), y*(py))) ) 104#define sat(inp) ( vec4(saturate((inp).xyz), (inp).w) ) 105 106// Maximum of four values 107#define max4(a,b,c,d) ( max(max(a, b), max(c, d)) ) 108 109// Colour to luma, fast approx gamma, avg of rec. 709 & 601 luma coeffs 110#define CtL(RGB) ( sqrt(dot(vec3(0.2558, 0.6511, 0.0931), saturate((RGB)*abs(RGB)).rgb)) ) 111 112// Center pixel diff 113#define mdiff(a,b,c,d,e,f,g) ( abs(luma[g]-luma[a]) + abs(luma[g]-luma[b]) \ 114 + abs(luma[g]-luma[c]) + abs(luma[g]-luma[d]) \ 115 + 0.5*(abs(luma[g]-luma[e]) + abs(luma[g]-luma[f])) ) 116 117#pragma stage vertex 118layout(location = 0) in vec4 Position; 119layout(location = 1) in vec2 TexCoord; 120layout(location = 0) out vec2 vTexCoord; 121 122void main() 123{ 124 gl_Position = global.MVP * Position; 125 vTexCoord = TexCoord; 126} 127 128#pragma stage fragment 129layout(location = 0) in vec2 vTexCoord; 130layout(location = 0) out vec4 FragColor; 131layout(set = 0, binding = 2) uniform sampler2D Source; 132 133vec4 frag_op(vec4 orig, vec2 coord, float c_edge, float px, float py) 134{ 135 if (bounds_check == true) 136 { 137 if (c_edge > 24. || c_edge < -0.5) { return vec4( 0., 1.0, 0., alpha_out ); } 138 } 139 140 // Get points, clip out of range colour data in c[0] 141 // [ c22 ] 142 // [ c24, c9, c23 ] 143 // [ c21, c1, c2, c3, c18 ] 144 // [ c19, c10, c4, c0, c5, c11, c16 ] 145 // [ c20, c6, c7, c8, c17 ] 146 // [ c15, c12, c14 ] 147 // [ c13 ] 148 vec4 c[25] = { sat( orig), get(-1,-1), get( 0,-1), get( 1,-1), get(-1, 0), 149 get( 1, 0), get(-1, 1), get( 0, 1), get( 1, 1), get( 0,-2), 150 get(-2, 0), get( 2, 0), get( 0, 2), get( 0, 3), get( 1, 2), 151 get(-1, 2), get( 3, 0), get( 2, 1), get( 2,-1), get(-3, 0), 152 get(-2, 1), get(-2,-1), get( 0,-3), get( 1,-2), get(-1,-2) }; 153 154 // Allow for higher overshoot if the current edge pixel is surrounded by similar edge pixels 155 float maxedge = max4( max4(c[1].w,c[2].w,c[3].w,c[4].w), max4(c[5].w,c[6].w,c[7].w,c[8].w), 156 max4(c[9].w,c[10].w,c[11].w,c[12].w), c[0].w ) - w_offset; 157 158 // [ x ] 159 // [ z, x, w ] 160 // [ z, z, x, w, w ] 161 // [ y, y, y, 0, y, y, y ] 162 // [ w, w, x, z, z ] 163 // [ w, x, z ] 164 // [ x ] 165 float sbe = soft_if(c[2].w,c[9].w,c[22].w) *soft_if(c[7].w,c[12].w,c[13].w) // x dir 166 + soft_if(c[4].w,c[10].w,c[19].w)*soft_if(c[5].w,c[11].w,c[16].w) // y dir 167 + soft_if(c[1].w,c[24].w,c[21].w)*soft_if(c[8].w,c[14].w,c[17].w) // z dir 168 + soft_if(c[3].w,c[23].w,c[18].w)*soft_if(c[6].w,c[20].w,c[15].w); // w dir 169 170 vec2 cs = mix( vec2(L_compr_low, D_compr_low), 171 vec2(L_compr_high, D_compr_high), smoothstep(2, 3.1, sbe) ); 172 173 // RGB to luma 174 float c0_Y = CtL(c[0]); 175 176 float luma[25] = { c0_Y, CtL(c[1]), CtL(c[2]), CtL(c[3]), CtL(c[4]), CtL(c[5]), CtL(c[6]), 177 CtL(c[7]), CtL(c[8]), CtL(c[9]), CtL(c[10]), CtL(c[11]), CtL(c[12]), 178 CtL(c[13]), CtL(c[14]), CtL(c[15]), CtL(c[16]), CtL(c[17]), CtL(c[18]), 179 CtL(c[19]), CtL(c[20]), CtL(c[21]), CtL(c[22]), CtL(c[23]), CtL(c[24]) }; 180 181 // Pre-calculated default squared kernel weights 182 const vec3 W1 = vec3(0.5, 1.0, 1.41421356237); // 0.25, 1.0, 2.0 183 const vec3 W2 = vec3(0.86602540378, 1.0, 0.5477225575); // 0.75, 1.0, 0.3 184 185 // Transition to a concave kernel if the center edge val is above thr 186 vec3 dW = pow(mix( W1, W2, smoothstep(dW_lothr, dW_hithr, c_edge) ), vec3(2.0)); 187 188 float mdiff_c0 = 0.02 + 3*( abs(luma[0]-luma[2]) + abs(luma[0]-luma[4]) 189 + abs(luma[0]-luma[5]) + abs(luma[0]-luma[7]) 190 + 0.25*(abs(luma[0]-luma[1]) + abs(luma[0]-luma[3]) 191 +abs(luma[0]-luma[6]) + abs(luma[0]-luma[8])) ); 192 193 // Use lower weights for pixels in a more active area relative to center pixel area 194 // This results in narrower and less visible overshoots around sharp edges 195 float weights[12] = { ( min(mdiff_c0/mdiff(24, 21, 2, 4, 9, 10, 1), dW.y) ), // c1 196 ( dW.x ), // c2 197 ( min(mdiff_c0/mdiff(23, 18, 5, 2, 9, 11, 3), dW.y) ), // c3 198 ( dW.x ), // c4 199 ( dW.x ), // c5 200 ( min(mdiff_c0/mdiff(4, 20, 15, 7, 10, 12, 6), dW.y) ), // c6 201 ( dW.x ), // c7 202 ( min(mdiff_c0/mdiff(5, 7, 17, 14, 12, 11, 8), dW.y) ), // c8 203 ( min(mdiff_c0/mdiff(2, 24, 23, 22, 1, 3, 9), dW.z) ), // c9 204 ( min(mdiff_c0/mdiff(20, 19, 21, 4, 1, 6, 10), dW.z) ), // c10 205 ( min(mdiff_c0/mdiff(17, 5, 18, 16, 3, 8, 11), dW.z) ), // c11 206 ( min(mdiff_c0/mdiff(13, 15, 7, 14, 6, 8, 12), dW.z) ) }; // c12 207 208 weights[0] = (max(max((weights[8] + weights[9])/4, weights[0]), 0.25) + weights[0])/2; 209 weights[2] = (max(max((weights[8] + weights[10])/4, weights[2]), 0.25) + weights[2])/2; 210 weights[5] = (max(max((weights[9] + weights[11])/4, weights[5]), 0.25) + weights[5])/2; 211 weights[7] = (max(max((weights[10] + weights[11])/4, weights[7]), 0.25) + weights[7])/2; 212 213 // Calculate the negative part of the laplace kernel and the low threshold weight 214 float lowthrsum = 0.; 215 float weightsum = 0.; 216 float neg_laplace = 0.; 217 218 //[unroll] 219 for (int pix = 0; pix < 12; ++pix) 220 { 221 float x = saturate((c[pix + 1].w - w_offset - 0.01)/(lowthr_mxw - 0.01)); 222 float lowthr = x*x*(2.97 - 1.98*x) + 0.01; // x*x((3.0-c*3.) - (2.0-c*2.)*x) + c 223 224 neg_laplace += pow(luma[pix + 1] + 0.06, 2.4)*(weights[pix]*lowthr); 225 weightsum += weights[pix]*lowthr; 226 lowthrsum += lowthr/12.; 227 } 228 229 neg_laplace = pow(abs(neg_laplace/weightsum), (1.0/2.4)) - 0.06; 230 231 // Compute sharpening magnitude function 232 float sharpen_val = curve_height/(curve_height*curveslope*pow(abs(c_edge), 3.5) + 0.5); 233 234 // Calculate sharpening diff and scale 235 float sharpdiff = (c0_Y - neg_laplace)*(lowthrsum*sharpen_val*0.8 + 0.01); 236 237 // Calculate local near min & max, partial sort 238 //[unroll] 239 for (int i = 0; i < 3; ++i) 240 { 241 float temp; 242 243 for (int i1 = i; i1 < 24-i; i1 += 2) 244 { 245 temp = luma[i1]; 246 luma[i1] = min(luma[i1], luma[i1+1]); 247 luma[i1+1] = max(temp, luma[i1+1]); 248 } 249 250 for (int i2 = 24-i; i2 > i; i2 -= 2) 251 { 252 temp = luma[i]; 253 luma[i] = min(luma[i], luma[i2]); 254 luma[i2] = max(temp, luma[i2]); 255 256 temp = luma[24-i]; 257 luma[24-i] = max(luma[24-i], luma[i2-1]); 258 luma[i2-1] = min(temp, luma[i2-1]); 259 } 260 } 261 262 float nmax = (max(luma[22] + luma[23]*2., c0_Y*3.) + luma[24])/4.; 263 float nmin = (min(luma[2] + luma[1]*2., c0_Y*3.) + luma[0])/4.; 264 265 // Calculate tanh scale factor, pos/neg 266 float nmax_scale = nmax - c0_Y + min(L_overshoot, 1.0001 - nmax); 267 float nmin_scale = c0_Y - nmin + min(D_overshoot, 0.0001 + nmin); 268 269 nmax_scale = min(nmax_scale, scale_lim*(1. - scale_cs) + nmax_scale*scale_cs); 270 nmin_scale = min(nmin_scale, scale_lim*(1. - scale_cs) + nmin_scale*scale_cs); 271 272 // Soft limited anti-ringing with tanh, wpmean to control compression slope 273 sharpdiff = wpmean( max(sharpdiff, 0.), soft_lim( max(sharpdiff, 0.), nmax_scale ), cs.x ) 274 - wpmean( min(sharpdiff, 0.), soft_lim( min(sharpdiff, 0.), nmin_scale ), cs.y ); 275 276 // Compensate for saturation loss/gain while making pixels brighter/darker 277 float sharpdiff_lim = saturate(c0_Y + sharpdiff) - c0_Y; 278 float satmul = (c0_Y + sharpdiff_lim + 0.03)/(c0_Y + 0.03); 279 vec3 res = c0_Y + (sharpdiff_lim*3 + sharpdiff)/4 + (c[0].rgb - c0_Y)*satmul; 280 281 return vec4( (video_level_out == 1.0 ? orig.rgb + (res - c[0].rgb) : res), alpha_out ); 282} 283 284void main() 285{ 286 vec2 tex = vTexCoord; 287 288 float px = 1.0 / params.SourceSize.x; 289 float py = 1.0 / params.SourceSize.y; 290 291 vec4 orig = texture(Source, tex); 292 float c_edge = orig.w - w_offset; 293 294 FragColor = vec4(frag_op(orig, tex, c_edge, px, py)); 295}