ffx_fsr1.h
1 //_____________________________________________________________/\_______________________________________________________________ 2 //============================================================================================================================== 3 // 4 // 5 // AMD FidelityFX SUPER RESOLUTION [FSR 1] ::: SPATIAL SCALING & EXTRAS - v1.20210629 6 // 7 // 8 //------------------------------------------------------------------------------------------------------------------------------ 9 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 10 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 11 //------------------------------------------------------------------------------------------------------------------------------ 12 // FidelityFX Super Resolution Sample 13 // 14 // Copyright (c) 2021 Advanced Micro Devices, Inc. All rights reserved. 15 // Permission is hereby granted, free of charge, to any person obtaining a copy 16 // of this software and associated documentation files(the "Software"), to deal 17 // in the Software without restriction, including without limitation the rights 18 // to use, copy, modify, merge, publish, distribute, sublicense, and / or sell 19 // copies of the Software, and to permit persons to whom the Software is 20 // furnished to do so, subject to the following conditions : 21 // The above copyright notice and this permission notice shall be included in 22 // all copies or substantial portions of the Software. 23 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 24 // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 25 // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE 26 // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 27 // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 28 // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 29 // THE SOFTWARE. 30 //------------------------------------------------------------------------------------------------------------------------------ 31 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 32 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 33 //------------------------------------------------------------------------------------------------------------------------------ 34 // ABOUT 35 // ===== 36 // FSR is a collection of algorithms relating to generating a higher resolution image. 37 // This specific header focuses on single-image non-temporal image scaling, and related tools. 38 // 39 // The core functions are EASU and RCAS: 40 // [EASU] Edge Adaptive Spatial Upsampling ....... 1x to 4x area range spatial scaling, clamped adaptive elliptical filter. 41 // [RCAS] Robust Contrast Adaptive Sharpening .... A non-scaling variation on CAS. 42 // RCAS needs to be applied after EASU as a separate pass. 43 // 44 // Optional utility functions are: 45 // [LFGA] Linear Film Grain Applicator ........... Tool to apply film grain after scaling. 46 // [SRTM] Simple Reversible Tone-Mapper .......... Linear HDR {0 to FP16_MAX} to {0 to 1} and back. 47 // [TEPD] Temporal Energy Preserving Dither ...... Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. 48 // See each individual sub-section for inline documentation. 49 //------------------------------------------------------------------------------------------------------------------------------ 50 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 51 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 52 //------------------------------------------------------------------------------------------------------------------------------ 53 // FUNCTION PERMUTATIONS 54 // ===================== 55 // *F() ..... Single item computation with 32-bit. 56 // *H() ..... Single item computation with 16-bit, with packing (aka two 16-bit ops in parallel) when possible. 57 // *Hx2() ... Processing two items in parallel with 16-bit, easier packing. 58 // Not all interfaces in this file have a *Hx2() form. 59 //============================================================================================================================== 60 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 61 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 62 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 63 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 64 //_____________________________________________________________/\_______________________________________________________________ 65 //============================================================================================================================== 66 // 67 // FSR - [EASU] EDGE ADAPTIVE SPATIAL UPSAMPLING 68 // 69 //------------------------------------------------------------------------------------------------------------------------------ 70 // EASU provides a high quality spatial-only scaling at relatively low cost. 71 // Meaning EASU is appropiate for laptops and other low-end GPUs. 72 // Quality from 1x to 4x area scaling is good. 73 //------------------------------------------------------------------------------------------------------------------------------ 74 // The scalar uses a modified fast approximation to the standard lanczos(size=2) kernel. 75 // EASU runs in a single pass, so it applies a directionally and anisotropically adaptive radial lanczos. 76 // This is also kept as simple as possible to have minimum runtime. 77 //------------------------------------------------------------------------------------------------------------------------------ 78 // The lanzcos filter has negative lobes, so by itself it will introduce ringing. 79 // To remove all ringing, the algorithm uses the nearest 2x2 input texels as a neighborhood, 80 // and limits output to the minimum and maximum of that neighborhood. 81 //------------------------------------------------------------------------------------------------------------------------------ 82 // Input image requirements: 83 // 84 // Color needs to be encoded as 3 channel[red, green, blue](e.g.XYZ not supported) 85 // Each channel needs to be in the range[0, 1] 86 // Any color primaries are supported 87 // Display / tonemapping curve needs to be as if presenting to sRGB display or similar(e.g.Gamma 2.0) 88 // There should be no banding in the input 89 // There should be no high amplitude noise in the input 90 // There should be no noise in the input that is not at input pixel granularity 91 // For performance purposes, use 32bpp formats 92 //------------------------------------------------------------------------------------------------------------------------------ 93 // Best to apply EASU at the end of the frame after tonemapping 94 // but before film grain or composite of the UI. 95 //------------------------------------------------------------------------------------------------------------------------------ 96 // Example of including this header for D3D HLSL : 97 // 98 // #define A_GPU 1 99 // #define A_HLSL 1 100 // #define A_HALF 1 101 // #include "ffx_a.h" 102 // #define FSR_EASU_H 1 103 // #define FSR_RCAS_H 1 104 // //declare input callbacks 105 // #include "ffx_fsr1.h" 106 // 107 // Example of including this header for Vulkan GLSL : 108 // 109 // #define A_GPU 1 110 // #define A_GLSL 1 111 // #define A_HALF 1 112 // #include "ffx_a.h" 113 // #define FSR_EASU_H 1 114 // #define FSR_RCAS_H 1 115 // //declare input callbacks 116 // #include "ffx_fsr1.h" 117 // 118 // Example of including this header for Vulkan HLSL : 119 // 120 // #define A_GPU 1 121 // #define A_HLSL 1 122 // #define A_HLSL_6_2 1 123 // #define A_NO_16_BIT_CAST 1 124 // #define A_HALF 1 125 // #include "ffx_a.h" 126 // #define FSR_EASU_H 1 127 // #define FSR_RCAS_H 1 128 // //declare input callbacks 129 // #include "ffx_fsr1.h" 130 // 131 // Example of declaring the required input callbacks for GLSL : 132 // The callbacks need to gather4 for each color channel using the specified texture coordinate 'p'. 133 // EASU uses gather4 to reduce position computation logic and for free Arrays of Structures to Structures of Arrays conversion. 134 // 135 // AH4 FsrEasuRH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,0));} 136 // AH4 FsrEasuGH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,1));} 137 // AH4 FsrEasuBH(AF2 p){return AH4(textureGather(sampler2D(tex,sam),p,2));} 138 // ... 139 // The FsrEasuCon function needs to be called from the CPU or GPU to set up constants. 140 // The difference in viewport and input image size is there to support Dynamic Resolution Scaling. 141 // To use FsrEasuCon() on the CPU, define A_CPU before including ffx_a and ffx_fsr1. 142 // Including a GPU example here, the 'con0' through 'con3' values would be stored out to a constant buffer. 143 // AU4 con0,con1,con2,con3; 144 // FsrEasuCon(con0,con1,con2,con3, 145 // 1920.0,1080.0, // Viewport size (top left aligned) in the input image which is to be scaled. 146 // 3840.0,2160.0, // The size of the input image. 147 // 2560.0,1440.0); // The output resolution. 148 //============================================================================================================================== 149 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 150 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 151 //_____________________________________________________________/\_______________________________________________________________ 152 //============================================================================================================================== 153 // CONSTANT SETUP 154 //============================================================================================================================== 155 // Call to setup required constant values (works on CPU or GPU). 156 A_STATIC void FsrEasuCon( 157 outAU4 con0, 158 outAU4 con1, 159 outAU4 con2, 160 outAU4 con3, 161 // This the rendered image resolution being upscaled 162 AF1 inputViewportInPixelsX, 163 AF1 inputViewportInPixelsY, 164 // This is the resolution of the resource containing the input image (useful for dynamic resolution) 165 AF1 inputSizeInPixelsX, 166 AF1 inputSizeInPixelsY, 167 // This is the display resolution which the input image gets upscaled to 168 AF1 outputSizeInPixelsX, 169 AF1 outputSizeInPixelsY){ 170 // Output integer position to a pixel position in viewport. 171 con0[0]=AU1_AF1(inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)); 172 con0[1]=AU1_AF1(inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)); 173 con0[2]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsX*ARcpF1(outputSizeInPixelsX)-AF1_(0.5)); 174 con0[3]=AU1_AF1(AF1_(0.5)*inputViewportInPixelsY*ARcpF1(outputSizeInPixelsY)-AF1_(0.5)); 175 // Viewport pixel position to normalized image space. 176 // This is used to get upper-left of 'F' tap. 177 con1[0]=AU1_AF1(ARcpF1(inputSizeInPixelsX)); 178 con1[1]=AU1_AF1(ARcpF1(inputSizeInPixelsY)); 179 // Centers of gather4, first offset from upper-left of 'F'. 180 // +---+---+ 181 // | | | 182 // +--(0)--+ 183 // | b | c | 184 // +---F---+---+---+ 185 // | e | f | g | h | 186 // +--(1)--+--(2)--+ 187 // | i | j | k | l | 188 // +---+---+---+---+ 189 // | n | o | 190 // +--(3)--+ 191 // | | | 192 // +---+---+ 193 con1[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); 194 con1[3]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsY)); 195 // These are from (0) instead of 'F'. 196 con2[0]=AU1_AF1(AF1_(-1.0)*ARcpF1(inputSizeInPixelsX)); 197 con2[1]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); 198 con2[2]=AU1_AF1(AF1_( 1.0)*ARcpF1(inputSizeInPixelsX)); 199 con2[3]=AU1_AF1(AF1_( 2.0)*ARcpF1(inputSizeInPixelsY)); 200 con3[0]=AU1_AF1(AF1_( 0.0)*ARcpF1(inputSizeInPixelsX)); 201 con3[1]=AU1_AF1(AF1_( 4.0)*ARcpF1(inputSizeInPixelsY)); 202 con3[2]=con3[3]=0;} 203 204 //If the an offset into the input image resource 205 A_STATIC void FsrEasuConOffset( 206 outAU4 con0, 207 outAU4 con1, 208 outAU4 con2, 209 outAU4 con3, 210 // This the rendered image resolution being upscaled 211 AF1 inputViewportInPixelsX, 212 AF1 inputViewportInPixelsY, 213 // This is the resolution of the resource containing the input image (useful for dynamic resolution) 214 AF1 inputSizeInPixelsX, 215 AF1 inputSizeInPixelsY, 216 // This is the display resolution which the input image gets upscaled to 217 AF1 outputSizeInPixelsX, 218 AF1 outputSizeInPixelsY, 219 // This is the input image offset into the resource containing it (useful for dynamic resolution) 220 AF1 inputOffsetInPixelsX, 221 AF1 inputOffsetInPixelsY) { 222 FsrEasuCon(con0, con1, con2, con3, inputViewportInPixelsX, inputViewportInPixelsY, inputSizeInPixelsX, inputSizeInPixelsY, outputSizeInPixelsX, outputSizeInPixelsY); 223 con0[2] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsX * ARcpF1(outputSizeInPixelsX) - AF1_(0.5) + inputOffsetInPixelsX); 224 con0[3] = AU1_AF1(AF1_(0.5) * inputViewportInPixelsY * ARcpF1(outputSizeInPixelsY) - AF1_(0.5) + inputOffsetInPixelsY); 225 } 226 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 227 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 228 //_____________________________________________________________/\_______________________________________________________________ 229 //============================================================================================================================== 230 // NON-PACKED 32-BIT VERSION 231 //============================================================================================================================== 232 #if defined(A_GPU)&&defined(FSR_EASU_F) 233 // Input callback prototypes, need to be implemented by calling shader 234 AF4 FsrEasuRF(AF2 p); 235 AF4 FsrEasuGF(AF2 p); 236 AF4 FsrEasuBF(AF2 p); 237 //------------------------------------------------------------------------------------------------------------------------------ 238 // Filtering for a given tap for the scalar. 239 void FsrEasuTapF( 240 inout AF3 aC, // Accumulated color, with negative lobe. 241 inout AF1 aW, // Accumulated weight. 242 AF2 off, // Pixel offset from resolve position to tap. 243 AF2 dir, // Gradient direction. 244 AF2 len, // Length. 245 AF1 lob, // Negative lobe strength. 246 AF1 clp, // Clipping point. 247 AF3 c){ // Tap color. 248 // Rotate offset by direction. 249 AF2 v; 250 v.x=(off.x*( dir.x))+(off.y*dir.y); 251 v.y=(off.x*(-dir.y))+(off.y*dir.x); 252 // Anisotropy. 253 v*=len; 254 // Compute distance^2. 255 AF1 d2=v.x*v.x+v.y*v.y; 256 // Limit to the window as at corner, 2 taps can easily be outside. 257 d2=min(d2,clp); 258 // Approximation of lancos2 without sin() or rcp(), or sqrt() to get x. 259 // (25/16 * (2/5 * x^2 - 1)^2 - (25/16 - 1)) * (1/4 * x^2 - 1)^2 260 // |_______________________________________| |_______________| 261 // base window 262 // The general form of the 'base' is, 263 // (a*(b*x^2-1)^2-(a-1)) 264 // Where 'a=1/(2*b-b^2)' and 'b' moves around the negative lobe. 265 AF1 wB=AF1_(2.0/5.0)*d2+AF1_(-1.0); 266 AF1 wA=lob*d2+AF1_(-1.0); 267 wB*=wB; 268 wA*=wA; 269 wB=AF1_(25.0/16.0)*wB+AF1_(-(25.0/16.0-1.0)); 270 AF1 w=wB*wA; 271 // Do weighted average. 272 aC+=c*w;aW+=w;} 273 //------------------------------------------------------------------------------------------------------------------------------ 274 // Accumulate direction and length. 275 void FsrEasuSetF( 276 inout AF2 dir, 277 inout AF1 len, 278 AF2 pp, 279 AP1 biS,AP1 biT,AP1 biU,AP1 biV, 280 AF1 lA,AF1 lB,AF1 lC,AF1 lD,AF1 lE){ 281 // Compute bilinear weight, branches factor out as predicates are compiler time immediates. 282 // s t 283 // u v 284 AF1 w = AF1_(0.0); 285 if(biS)w=(AF1_(1.0)-pp.x)*(AF1_(1.0)-pp.y); 286 if(biT)w= pp.x *(AF1_(1.0)-pp.y); 287 if(biU)w=(AF1_(1.0)-pp.x)* pp.y ; 288 if(biV)w= pp.x * pp.y ; 289 // Direction is the '+' diff. 290 // a 291 // b c d 292 // e 293 // Then takes magnitude from abs average of both sides of 'c'. 294 // Length converts gradient reversal to 0, smoothly to non-reversal at 1, shaped, then adding horz and vert terms. 295 AF1 dc=lD-lC; 296 AF1 cb=lC-lB; 297 AF1 lenX=max(abs(dc),abs(cb)); 298 lenX=APrxLoRcpF1(lenX); 299 AF1 dirX=lD-lB; 300 dir.x+=dirX*w; 301 lenX=ASatF1(abs(dirX)*lenX); 302 lenX*=lenX; 303 len+=lenX*w; 304 // Repeat for the y axis. 305 AF1 ec=lE-lC; 306 AF1 ca=lC-lA; 307 AF1 lenY=max(abs(ec),abs(ca)); 308 lenY=APrxLoRcpF1(lenY); 309 AF1 dirY=lE-lA; 310 dir.y+=dirY*w; 311 lenY=ASatF1(abs(dirY)*lenY); 312 lenY*=lenY; 313 len+=lenY*w;} 314 //------------------------------------------------------------------------------------------------------------------------------ 315 void FsrEasuF( 316 out AF3 pix, 317 AU2 ip, // Integer pixel position in output. 318 AU4 con0, // Constants generated by FsrEasuCon(). 319 AU4 con1, 320 AU4 con2, 321 AU4 con3){ 322 //------------------------------------------------------------------------------------------------------------------------------ 323 // Get position of 'f'. 324 AF2 pp=AF2(ip)*AF2_AU2(con0.xy)+AF2_AU2(con0.zw); 325 AF2 fp=floor(pp); 326 pp-=fp; 327 //------------------------------------------------------------------------------------------------------------------------------ 328 // 12-tap kernel. 329 // b c 330 // e f g h 331 // i j k l 332 // n o 333 // Gather 4 ordering. 334 // a b 335 // r g 336 // For packed FP16, need either {rg} or {ab} so using the following setup for gather in all versions, 337 // a b <- unused (z) 338 // r g 339 // a b a b 340 // r g r g 341 // a b 342 // r g <- unused (z) 343 // Allowing dead-code removal to remove the 'z's. 344 AF2 p0=fp*AF2_AU2(con1.xy)+AF2_AU2(con1.zw); 345 // These are from p0 to avoid pulling two constants on pre-Navi hardware. 346 AF2 p1=p0+AF2_AU2(con2.xy); 347 AF2 p2=p0+AF2_AU2(con2.zw); 348 AF2 p3=p0+AF2_AU2(con3.xy); 349 AF4 bczzR=FsrEasuRF(p0); 350 AF4 bczzG=FsrEasuGF(p0); 351 AF4 bczzB=FsrEasuBF(p0); 352 AF4 ijfeR=FsrEasuRF(p1); 353 AF4 ijfeG=FsrEasuGF(p1); 354 AF4 ijfeB=FsrEasuBF(p1); 355 AF4 klhgR=FsrEasuRF(p2); 356 AF4 klhgG=FsrEasuGF(p2); 357 AF4 klhgB=FsrEasuBF(p2); 358 AF4 zzonR=FsrEasuRF(p3); 359 AF4 zzonG=FsrEasuGF(p3); 360 AF4 zzonB=FsrEasuBF(p3); 361 //------------------------------------------------------------------------------------------------------------------------------ 362 // Simplest multi-channel approximate luma possible (luma times 2, in 2 FMA/MAD). 363 AF4 bczzL=bczzB*AF4_(0.5)+(bczzR*AF4_(0.5)+bczzG); 364 AF4 ijfeL=ijfeB*AF4_(0.5)+(ijfeR*AF4_(0.5)+ijfeG); 365 AF4 klhgL=klhgB*AF4_(0.5)+(klhgR*AF4_(0.5)+klhgG); 366 AF4 zzonL=zzonB*AF4_(0.5)+(zzonR*AF4_(0.5)+zzonG); 367 // Rename. 368 AF1 bL=bczzL.x; 369 AF1 cL=bczzL.y; 370 AF1 iL=ijfeL.x; 371 AF1 jL=ijfeL.y; 372 AF1 fL=ijfeL.z; 373 AF1 eL=ijfeL.w; 374 AF1 kL=klhgL.x; 375 AF1 lL=klhgL.y; 376 AF1 hL=klhgL.z; 377 AF1 gL=klhgL.w; 378 AF1 oL=zzonL.z; 379 AF1 nL=zzonL.w; 380 // Accumulate for bilinear interpolation. 381 AF2 dir=AF2_(0.0); 382 AF1 len=AF1_(0.0); 383 FsrEasuSetF(dir,len,pp,true, false,false,false,bL,eL,fL,gL,jL); 384 FsrEasuSetF(dir,len,pp,false,true ,false,false,cL,fL,gL,hL,kL); 385 FsrEasuSetF(dir,len,pp,false,false,true ,false,fL,iL,jL,kL,nL); 386 FsrEasuSetF(dir,len,pp,false,false,false,true ,gL,jL,kL,lL,oL); 387 //------------------------------------------------------------------------------------------------------------------------------ 388 // Normalize with approximation, and cleanup close to zero. 389 AF2 dir2=dir*dir; 390 AF1 dirR=dir2.x+dir2.y; 391 AP1 zro=dirR<AF1_(1.0/32768.0); 392 dirR=APrxLoRsqF1(dirR); 393 dirR=zro?AF1_(1.0):dirR; 394 dir.x=zro?AF1_(1.0):dir.x; 395 dir*=AF2_(dirR); 396 // Transform from {0 to 2} to {0 to 1} range, and shape with square. 397 len=len*AF1_(0.5); 398 len*=len; 399 // Stretch kernel {1.0 vert|horz, to sqrt(2.0) on diagonal}. 400 AF1 stretch=(dir.x*dir.x+dir.y*dir.y)*APrxLoRcpF1(max(abs(dir.x),abs(dir.y))); 401 // Anisotropic length after rotation, 402 // x := 1.0 lerp to 'stretch' on edges 403 // y := 1.0 lerp to 2x on edges 404 AF2 len2=AF2(AF1_(1.0)+(stretch-AF1_(1.0))*len,AF1_(1.0)+AF1_(-0.5)*len); 405 // Based on the amount of 'edge', 406 // the window shifts from +/-{sqrt(2.0) to slightly beyond 2.0}. 407 AF1 lob=AF1_(0.5)+AF1_((1.0/4.0-0.04)-0.5)*len; 408 // Set distance^2 clipping point to the end of the adjustable window. 409 AF1 clp=APrxLoRcpF1(lob); 410 //------------------------------------------------------------------------------------------------------------------------------ 411 // Accumulation mixed with min/max of 4 nearest. 412 // b c 413 // e f g h 414 // i j k l 415 // n o 416 AF3 min4=min(AMin3F3(AF3(ijfeR.z,ijfeG.z,ijfeB.z),AF3(klhgR.w,klhgG.w,klhgB.w),AF3(ijfeR.y,ijfeG.y,ijfeB.y)), 417 AF3(klhgR.x,klhgG.x,klhgB.x)); 418 AF3 max4=max(AMax3F3(AF3(ijfeR.z,ijfeG.z,ijfeB.z),AF3(klhgR.w,klhgG.w,klhgB.w),AF3(ijfeR.y,ijfeG.y,ijfeB.y)), 419 AF3(klhgR.x,klhgG.x,klhgB.x)); 420 // Accumulation. 421 AF3 aC=AF3_(0.0); 422 AF1 aW=AF1_(0.0); 423 FsrEasuTapF(aC,aW,AF2( 0.0,-1.0)-pp,dir,len2,lob,clp,AF3(bczzR.x,bczzG.x,bczzB.x)); // b 424 FsrEasuTapF(aC,aW,AF2( 1.0,-1.0)-pp,dir,len2,lob,clp,AF3(bczzR.y,bczzG.y,bczzB.y)); // c 425 FsrEasuTapF(aC,aW,AF2(-1.0, 1.0)-pp,dir,len2,lob,clp,AF3(ijfeR.x,ijfeG.x,ijfeB.x)); // i 426 FsrEasuTapF(aC,aW,AF2( 0.0, 1.0)-pp,dir,len2,lob,clp,AF3(ijfeR.y,ijfeG.y,ijfeB.y)); // j 427 FsrEasuTapF(aC,aW,AF2( 0.0, 0.0)-pp,dir,len2,lob,clp,AF3(ijfeR.z,ijfeG.z,ijfeB.z)); // f 428 FsrEasuTapF(aC,aW,AF2(-1.0, 0.0)-pp,dir,len2,lob,clp,AF3(ijfeR.w,ijfeG.w,ijfeB.w)); // e 429 FsrEasuTapF(aC,aW,AF2( 1.0, 1.0)-pp,dir,len2,lob,clp,AF3(klhgR.x,klhgG.x,klhgB.x)); // k 430 FsrEasuTapF(aC,aW,AF2( 2.0, 1.0)-pp,dir,len2,lob,clp,AF3(klhgR.y,klhgG.y,klhgB.y)); // l 431 FsrEasuTapF(aC,aW,AF2( 2.0, 0.0)-pp,dir,len2,lob,clp,AF3(klhgR.z,klhgG.z,klhgB.z)); // h 432 FsrEasuTapF(aC,aW,AF2( 1.0, 0.0)-pp,dir,len2,lob,clp,AF3(klhgR.w,klhgG.w,klhgB.w)); // g 433 FsrEasuTapF(aC,aW,AF2( 1.0, 2.0)-pp,dir,len2,lob,clp,AF3(zzonR.z,zzonG.z,zzonB.z)); // o 434 FsrEasuTapF(aC,aW,AF2( 0.0, 2.0)-pp,dir,len2,lob,clp,AF3(zzonR.w,zzonG.w,zzonB.w)); // n 435 //------------------------------------------------------------------------------------------------------------------------------ 436 // Normalize and dering. 437 pix=min(max4,max(min4,aC*AF3_(ARcpF1(aW))));} 438 #endif 439 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 440 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 441 //_____________________________________________________________/\_______________________________________________________________ 442 //============================================================================================================================== 443 // PACKED 16-BIT VERSION 444 //============================================================================================================================== 445 #if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_EASU_H) 446 // Input callback prototypes, need to be implemented by calling shader 447 AH4 FsrEasuRH(AF2 p); 448 AH4 FsrEasuGH(AF2 p); 449 AH4 FsrEasuBH(AF2 p); 450 //------------------------------------------------------------------------------------------------------------------------------ 451 // This runs 2 taps in parallel. 452 void FsrEasuTapH( 453 inout AH2 aCR,inout AH2 aCG,inout AH2 aCB, 454 inout AH2 aW, 455 AH2 offX,AH2 offY, 456 AH2 dir, 457 AH2 len, 458 AH1 lob, 459 AH1 clp, 460 AH2 cR,AH2 cG,AH2 cB){ 461 AH2 vX,vY; 462 vX=offX* dir.xx +offY*dir.yy; 463 vY=offX*(-dir.yy)+offY*dir.xx; 464 vX*=len.x;vY*=len.y; 465 AH2 d2=vX*vX+vY*vY; 466 d2=min(d2,AH2_(clp)); 467 AH2 wB=AH2_(2.0/5.0)*d2+AH2_(-1.0); 468 AH2 wA=AH2_(lob)*d2+AH2_(-1.0); 469 wB*=wB; 470 wA*=wA; 471 wB=AH2_(25.0/16.0)*wB+AH2_(-(25.0/16.0-1.0)); 472 AH2 w=wB*wA; 473 aCR+=cR*w;aCG+=cG*w;aCB+=cB*w;aW+=w;} 474 //------------------------------------------------------------------------------------------------------------------------------ 475 // This runs 2 taps in parallel. 476 void FsrEasuSetH( 477 inout AH2 dirPX,inout AH2 dirPY, 478 inout AH2 lenP, 479 AH2 pp, 480 AP1 biST,AP1 biUV, 481 AH2 lA,AH2 lB,AH2 lC,AH2 lD,AH2 lE){ 482 AH2 w = AH2_(0.0); 483 if(biST)w=(AH2(1.0,0.0)+AH2(-pp.x,pp.x))*AH2_(AH1_(1.0)-pp.y); 484 if(biUV)w=(AH2(1.0,0.0)+AH2(-pp.x,pp.x))*AH2_( pp.y); 485 // ABS is not free in the packed FP16 path. 486 AH2 dc=lD-lC; 487 AH2 cb=lC-lB; 488 AH2 lenX=max(abs(dc),abs(cb)); 489 lenX=ARcpH2(lenX); 490 AH2 dirX=lD-lB; 491 dirPX+=dirX*w; 492 lenX=ASatH2(abs(dirX)*lenX); 493 lenX*=lenX; 494 lenP+=lenX*w; 495 AH2 ec=lE-lC; 496 AH2 ca=lC-lA; 497 AH2 lenY=max(abs(ec),abs(ca)); 498 lenY=ARcpH2(lenY); 499 AH2 dirY=lE-lA; 500 dirPY+=dirY*w; 501 lenY=ASatH2(abs(dirY)*lenY); 502 lenY*=lenY; 503 lenP+=lenY*w;} 504 //------------------------------------------------------------------------------------------------------------------------------ 505 void FsrEasuH( 506 out AH3 pix, 507 AU2 ip, 508 AU4 con0, 509 AU4 con1, 510 AU4 con2, 511 AU4 con3){ 512 //------------------------------------------------------------------------------------------------------------------------------ 513 AF2 pp=AF2(ip)*AF2_AU2(con0.xy)+AF2_AU2(con0.zw); 514 AF2 fp=floor(pp); 515 pp-=fp; 516 AH2 ppp=AH2(pp); 517 //------------------------------------------------------------------------------------------------------------------------------ 518 AF2 p0=fp*AF2_AU2(con1.xy)+AF2_AU2(con1.zw); 519 AF2 p1=p0+AF2_AU2(con2.xy); 520 AF2 p2=p0+AF2_AU2(con2.zw); 521 AF2 p3=p0+AF2_AU2(con3.xy); 522 AH4 bczzR=FsrEasuRH(p0); 523 AH4 bczzG=FsrEasuGH(p0); 524 AH4 bczzB=FsrEasuBH(p0); 525 AH4 ijfeR=FsrEasuRH(p1); 526 AH4 ijfeG=FsrEasuGH(p1); 527 AH4 ijfeB=FsrEasuBH(p1); 528 AH4 klhgR=FsrEasuRH(p2); 529 AH4 klhgG=FsrEasuGH(p2); 530 AH4 klhgB=FsrEasuBH(p2); 531 AH4 zzonR=FsrEasuRH(p3); 532 AH4 zzonG=FsrEasuGH(p3); 533 AH4 zzonB=FsrEasuBH(p3); 534 //------------------------------------------------------------------------------------------------------------------------------ 535 AH4 bczzL=bczzB*AH4_(0.5)+(bczzR*AH4_(0.5)+bczzG); 536 AH4 ijfeL=ijfeB*AH4_(0.5)+(ijfeR*AH4_(0.5)+ijfeG); 537 AH4 klhgL=klhgB*AH4_(0.5)+(klhgR*AH4_(0.5)+klhgG); 538 AH4 zzonL=zzonB*AH4_(0.5)+(zzonR*AH4_(0.5)+zzonG); 539 AH1 bL=bczzL.x; 540 AH1 cL=bczzL.y; 541 AH1 iL=ijfeL.x; 542 AH1 jL=ijfeL.y; 543 AH1 fL=ijfeL.z; 544 AH1 eL=ijfeL.w; 545 AH1 kL=klhgL.x; 546 AH1 lL=klhgL.y; 547 AH1 hL=klhgL.z; 548 AH1 gL=klhgL.w; 549 AH1 oL=zzonL.z; 550 AH1 nL=zzonL.w; 551 // This part is different, accumulating 2 taps in parallel. 552 AH2 dirPX=AH2_(0.0); 553 AH2 dirPY=AH2_(0.0); 554 AH2 lenP=AH2_(0.0); 555 FsrEasuSetH(dirPX,dirPY,lenP,ppp,true, false,AH2(bL,cL),AH2(eL,fL),AH2(fL,gL),AH2(gL,hL),AH2(jL,kL)); 556 FsrEasuSetH(dirPX,dirPY,lenP,ppp,false,true ,AH2(fL,gL),AH2(iL,jL),AH2(jL,kL),AH2(kL,lL),AH2(nL,oL)); 557 AH2 dir=AH2(dirPX.r+dirPX.g,dirPY.r+dirPY.g); 558 AH1 len=lenP.r+lenP.g; 559 //------------------------------------------------------------------------------------------------------------------------------ 560 AH2 dir2=dir*dir; 561 AH1 dirR=dir2.x+dir2.y; 562 AP1 zro=dirR<AH1_(1.0/32768.0); 563 dirR=APrxLoRsqH1(dirR); 564 dirR=zro?AH1_(1.0):dirR; 565 dir.x=zro?AH1_(1.0):dir.x; 566 dir*=AH2_(dirR); 567 len=len*AH1_(0.5); 568 len*=len; 569 AH1 stretch=(dir.x*dir.x+dir.y*dir.y)*APrxLoRcpH1(max(abs(dir.x),abs(dir.y))); 570 AH2 len2=AH2(AH1_(1.0)+(stretch-AH1_(1.0))*len,AH1_(1.0)+AH1_(-0.5)*len); 571 AH1 lob=AH1_(0.5)+AH1_((1.0/4.0-0.04)-0.5)*len; 572 AH1 clp=APrxLoRcpH1(lob); 573 //------------------------------------------------------------------------------------------------------------------------------ 574 // FP16 is different, using packed trick to do min and max in same operation. 575 AH2 bothR=max(max(AH2(-ijfeR.z,ijfeR.z),AH2(-klhgR.w,klhgR.w)),max(AH2(-ijfeR.y,ijfeR.y),AH2(-klhgR.x,klhgR.x))); 576 AH2 bothG=max(max(AH2(-ijfeG.z,ijfeG.z),AH2(-klhgG.w,klhgG.w)),max(AH2(-ijfeG.y,ijfeG.y),AH2(-klhgG.x,klhgG.x))); 577 AH2 bothB=max(max(AH2(-ijfeB.z,ijfeB.z),AH2(-klhgB.w,klhgB.w)),max(AH2(-ijfeB.y,ijfeB.y),AH2(-klhgB.x,klhgB.x))); 578 // This part is different for FP16, working pairs of taps at a time. 579 AH2 pR=AH2_(0.0); 580 AH2 pG=AH2_(0.0); 581 AH2 pB=AH2_(0.0); 582 AH2 pW=AH2_(0.0); 583 FsrEasuTapH(pR,pG,pB,pW,AH2( 0.0, 1.0)-ppp.xx,AH2(-1.0,-1.0)-ppp.yy,dir,len2,lob,clp,bczzR.xy,bczzG.xy,bczzB.xy); 584 FsrEasuTapH(pR,pG,pB,pW,AH2(-1.0, 0.0)-ppp.xx,AH2( 1.0, 1.0)-ppp.yy,dir,len2,lob,clp,ijfeR.xy,ijfeG.xy,ijfeB.xy); 585 FsrEasuTapH(pR,pG,pB,pW,AH2( 0.0,-1.0)-ppp.xx,AH2( 0.0, 0.0)-ppp.yy,dir,len2,lob,clp,ijfeR.zw,ijfeG.zw,ijfeB.zw); 586 FsrEasuTapH(pR,pG,pB,pW,AH2( 1.0, 2.0)-ppp.xx,AH2( 1.0, 1.0)-ppp.yy,dir,len2,lob,clp,klhgR.xy,klhgG.xy,klhgB.xy); 587 FsrEasuTapH(pR,pG,pB,pW,AH2( 2.0, 1.0)-ppp.xx,AH2( 0.0, 0.0)-ppp.yy,dir,len2,lob,clp,klhgR.zw,klhgG.zw,klhgB.zw); 588 FsrEasuTapH(pR,pG,pB,pW,AH2( 1.0, 0.0)-ppp.xx,AH2( 2.0, 2.0)-ppp.yy,dir,len2,lob,clp,zzonR.zw,zzonG.zw,zzonB.zw); 589 AH3 aC=AH3(pR.x+pR.y,pG.x+pG.y,pB.x+pB.y); 590 AH1 aW=pW.x+pW.y; 591 //------------------------------------------------------------------------------------------------------------------------------ 592 // Slightly different for FP16 version due to combined min and max. 593 pix=min(AH3(bothR.y,bothG.y,bothB.y),max(-AH3(bothR.x,bothG.x,bothB.x),aC*AH3_(ARcpH1(aW))));} 594 #endif 595 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 596 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 597 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 598 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 599 //_____________________________________________________________/\_______________________________________________________________ 600 //============================================================================================================================== 601 // 602 // FSR - [RCAS] ROBUST CONTRAST ADAPTIVE SHARPENING 603 // 604 //------------------------------------------------------------------------------------------------------------------------------ 605 // CAS uses a simplified mechanism to convert local contrast into a variable amount of sharpness. 606 // RCAS uses a more exact mechanism, solving for the maximum local sharpness possible before clipping. 607 // RCAS also has a built in process to limit sharpening of what it detects as possible noise. 608 // RCAS sharper does not support scaling, as it should be applied after EASU scaling. 609 // Pass EASU output straight into RCAS, no color conversions necessary. 610 //------------------------------------------------------------------------------------------------------------------------------ 611 // RCAS is based on the following logic. 612 // RCAS uses a 5 tap filter in a cross pattern (same as CAS), 613 // w n 614 // w 1 w for taps w m e 615 // w s 616 // Where 'w' is the negative lobe weight. 617 // output = (w*(n+e+w+s)+m)/(4*w+1) 618 // RCAS solves for 'w' by seeing where the signal might clip out of the {0 to 1} input range, 619 // 0 == (w*(n+e+w+s)+m)/(4*w+1) -> w = -m/(n+e+w+s) 620 // 1 == (w*(n+e+w+s)+m)/(4*w+1) -> w = (1-m)/(n+e+w+s-4*1) 621 // Then chooses the 'w' which results in no clipping, limits 'w', and multiplies by the 'sharp' amount. 622 // This solution above has issues with MSAA input as the steps along the gradient cause edge detection issues. 623 // So RCAS uses 4x the maximum and 4x the minimum (depending on equation)in place of the individual taps. 624 // As well as switching from 'm' to either the minimum or maximum (depending on side), to help in energy conservation. 625 // This stabilizes RCAS. 626 // RCAS does a simple highpass which is normalized against the local contrast then shaped, 627 // 0.25 628 // 0.25 -1 0.25 629 // 0.25 630 // This is used as a noise detection filter, to reduce the effect of RCAS on grain, and focus on real edges. 631 // 632 // GLSL example for the required callbacks : 633 // 634 // AH4 FsrRcasLoadH(ASW2 p){return AH4(imageLoad(imgSrc,ASU2(p)));} 635 // void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b) 636 // { 637 // //do any simple input color conversions here or leave empty if none needed 638 // } 639 // 640 // FsrRcasCon need to be called from the CPU or GPU to set up constants. 641 // Including a GPU example here, the 'con' value would be stored out to a constant buffer. 642 // 643 // AU4 con; 644 // FsrRcasCon(con, 645 // 0.0); // The scale is {0.0 := maximum sharpness, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. 646 // --------------- 647 // RCAS sharpening supports a CAS-like pass-through alpha via, 648 // #define FSR_RCAS_PASSTHROUGH_ALPHA 1 649 // RCAS also supports a define to enable a more expensive path to avoid some sharpening of noise. 650 // Would suggest it is better to apply film grain after RCAS sharpening (and after scaling) instead of using this define, 651 // #define FSR_RCAS_DENOISE 1 652 //============================================================================================================================== 653 // This is set at the limit of providing unnatural results for sharpening. 654 #define FSR_RCAS_LIMIT (0.25-(1.0/16.0)) 655 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 656 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 657 //_____________________________________________________________/\_______________________________________________________________ 658 //============================================================================================================================== 659 // CONSTANT SETUP 660 //============================================================================================================================== 661 // Call to setup required constant values (works on CPU or GPU). 662 A_STATIC void FsrRcasCon( 663 outAU4 con, 664 // The scale is {0.0 := maximum, to N>0, where N is the number of stops (halving) of the reduction of sharpness}. 665 AF1 sharpness){ 666 // Transform from stops to linear value. 667 sharpness=AExp2F1(-sharpness); 668 varAF2(hSharp)=initAF2(sharpness,sharpness); 669 con[0]=AU1_AF1(sharpness); 670 con[1]=AU1_AH2_AF2(hSharp); 671 con[2]=0; 672 con[3]=0;} 673 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 674 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 675 //_____________________________________________________________/\_______________________________________________________________ 676 //============================================================================================================================== 677 // NON-PACKED 32-BIT VERSION 678 //============================================================================================================================== 679 #if defined(A_GPU)&&defined(FSR_RCAS_F) 680 // Input callback prototypes that need to be implemented by calling shader 681 AF4 FsrRcasLoadF(ASU2 p); 682 void FsrRcasInputF(inout AF1 r,inout AF1 g,inout AF1 b); 683 //------------------------------------------------------------------------------------------------------------------------------ 684 void FsrRcasF( 685 out AF1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. 686 out AF1 pixG, 687 out AF1 pixB, 688 #ifdef FSR_RCAS_PASSTHROUGH_ALPHA 689 out AF1 pixA, 690 #endif 691 AU2 ip, // Integer pixel position in output. 692 AU4 con){ // Constant generated by RcasSetup(). 693 // Algorithm uses minimal 3x3 pixel neighborhood. 694 // b 695 // d e f 696 // h 697 ASU2 sp=ASU2(ip); 698 AF3 b=FsrRcasLoadF(sp+ASU2( 0,-1)).rgb; 699 AF3 d=FsrRcasLoadF(sp+ASU2(-1, 0)).rgb; 700 #ifdef FSR_RCAS_PASSTHROUGH_ALPHA 701 AF4 ee=FsrRcasLoadF(sp); 702 AF3 e=ee.rgb;pixA=ee.a; 703 #else 704 AF3 e=FsrRcasLoadF(sp).rgb; 705 #endif 706 AF3 f=FsrRcasLoadF(sp+ASU2( 1, 0)).rgb; 707 AF3 h=FsrRcasLoadF(sp+ASU2( 0, 1)).rgb; 708 // Rename (32-bit) or regroup (16-bit). 709 AF1 bR=b.r; 710 AF1 bG=b.g; 711 AF1 bB=b.b; 712 AF1 dR=d.r; 713 AF1 dG=d.g; 714 AF1 dB=d.b; 715 AF1 eR=e.r; 716 AF1 eG=e.g; 717 AF1 eB=e.b; 718 AF1 fR=f.r; 719 AF1 fG=f.g; 720 AF1 fB=f.b; 721 AF1 hR=h.r; 722 AF1 hG=h.g; 723 AF1 hB=h.b; 724 // Run optional input transform. 725 FsrRcasInputF(bR,bG,bB); 726 FsrRcasInputF(dR,dG,dB); 727 FsrRcasInputF(eR,eG,eB); 728 FsrRcasInputF(fR,fG,fB); 729 FsrRcasInputF(hR,hG,hB); 730 // Luma times 2. 731 AF1 bL=bB*AF1_(0.5)+(bR*AF1_(0.5)+bG); 732 AF1 dL=dB*AF1_(0.5)+(dR*AF1_(0.5)+dG); 733 AF1 eL=eB*AF1_(0.5)+(eR*AF1_(0.5)+eG); 734 AF1 fL=fB*AF1_(0.5)+(fR*AF1_(0.5)+fG); 735 AF1 hL=hB*AF1_(0.5)+(hR*AF1_(0.5)+hG); 736 // Noise detection. 737 AF1 nz=AF1_(0.25)*bL+AF1_(0.25)*dL+AF1_(0.25)*fL+AF1_(0.25)*hL-eL; 738 nz=ASatF1(abs(nz)*APrxMedRcpF1(AMax3F1(AMax3F1(bL,dL,eL),fL,hL)-AMin3F1(AMin3F1(bL,dL,eL),fL,hL))); 739 nz=AF1_(-0.5)*nz+AF1_(1.0); 740 // Min and max of ring. 741 AF1 mn4R=min(AMin3F1(bR,dR,fR),hR); 742 AF1 mn4G=min(AMin3F1(bG,dG,fG),hG); 743 AF1 mn4B=min(AMin3F1(bB,dB,fB),hB); 744 AF1 mx4R=max(AMax3F1(bR,dR,fR),hR); 745 AF1 mx4G=max(AMax3F1(bG,dG,fG),hG); 746 AF1 mx4B=max(AMax3F1(bB,dB,fB),hB); 747 // Immediate constants for peak range. 748 AF2 peakC=AF2(1.0,-1.0*4.0); 749 // Limiters, these need to be high precision RCPs. 750 AF1 hitMinR=min(mn4R,eR)*ARcpF1(AF1_(4.0)*mx4R); 751 AF1 hitMinG=min(mn4G,eG)*ARcpF1(AF1_(4.0)*mx4G); 752 AF1 hitMinB=min(mn4B,eB)*ARcpF1(AF1_(4.0)*mx4B); 753 AF1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpF1(AF1_(4.0)*mn4R+peakC.y); 754 AF1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpF1(AF1_(4.0)*mn4G+peakC.y); 755 AF1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpF1(AF1_(4.0)*mn4B+peakC.y); 756 AF1 lobeR=max(-hitMinR,hitMaxR); 757 AF1 lobeG=max(-hitMinG,hitMaxG); 758 AF1 lobeB=max(-hitMinB,hitMaxB); 759 AF1 lobe=max(AF1_(-FSR_RCAS_LIMIT),min(AMax3F1(lobeR,lobeG,lobeB),AF1_(0.0)))*AF1_AU1(con.x); 760 // Apply noise removal. 761 #ifdef FSR_RCAS_DENOISE 762 lobe*=nz; 763 #endif 764 // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. 765 AF1 rcpL=APrxMedRcpF1(AF1_(4.0)*lobe+AF1_(1.0)); 766 pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; 767 pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; 768 pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL; 769 return;} 770 #endif 771 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 772 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 773 //_____________________________________________________________/\_______________________________________________________________ 774 //============================================================================================================================== 775 // NON-PACKED 16-BIT VERSION 776 //============================================================================================================================== 777 #if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_H) 778 // Input callback prototypes that need to be implemented by calling shader 779 AH4 FsrRcasLoadH(ASW2 p); 780 void FsrRcasInputH(inout AH1 r,inout AH1 g,inout AH1 b); 781 //------------------------------------------------------------------------------------------------------------------------------ 782 void FsrRcasH( 783 out AH1 pixR, // Output values, non-vector so port between RcasFilter() and RcasFilterH() is easy. 784 out AH1 pixG, 785 out AH1 pixB, 786 #ifdef FSR_RCAS_PASSTHROUGH_ALPHA 787 out AH1 pixA, 788 #endif 789 AU2 ip, // Integer pixel position in output. 790 AU4 con){ // Constant generated by RcasSetup(). 791 // Sharpening algorithm uses minimal 3x3 pixel neighborhood. 792 // b 793 // d e f 794 // h 795 ASW2 sp=ASW2(ip); 796 AH3 b=FsrRcasLoadH(sp+ASW2( 0,-1)).rgb; 797 AH3 d=FsrRcasLoadH(sp+ASW2(-1, 0)).rgb; 798 #ifdef FSR_RCAS_PASSTHROUGH_ALPHA 799 AH4 ee=FsrRcasLoadH(sp); 800 AH3 e=ee.rgb;pixA=ee.a; 801 #else 802 AH3 e=FsrRcasLoadH(sp).rgb; 803 #endif 804 AH3 f=FsrRcasLoadH(sp+ASW2( 1, 0)).rgb; 805 AH3 h=FsrRcasLoadH(sp+ASW2( 0, 1)).rgb; 806 // Rename (32-bit) or regroup (16-bit). 807 AH1 bR=b.r; 808 AH1 bG=b.g; 809 AH1 bB=b.b; 810 AH1 dR=d.r; 811 AH1 dG=d.g; 812 AH1 dB=d.b; 813 AH1 eR=e.r; 814 AH1 eG=e.g; 815 AH1 eB=e.b; 816 AH1 fR=f.r; 817 AH1 fG=f.g; 818 AH1 fB=f.b; 819 AH1 hR=h.r; 820 AH1 hG=h.g; 821 AH1 hB=h.b; 822 // Run optional input transform. 823 FsrRcasInputH(bR,bG,bB); 824 FsrRcasInputH(dR,dG,dB); 825 FsrRcasInputH(eR,eG,eB); 826 FsrRcasInputH(fR,fG,fB); 827 FsrRcasInputH(hR,hG,hB); 828 // Luma times 2. 829 AH1 bL=bB*AH1_(0.5)+(bR*AH1_(0.5)+bG); 830 AH1 dL=dB*AH1_(0.5)+(dR*AH1_(0.5)+dG); 831 AH1 eL=eB*AH1_(0.5)+(eR*AH1_(0.5)+eG); 832 AH1 fL=fB*AH1_(0.5)+(fR*AH1_(0.5)+fG); 833 AH1 hL=hB*AH1_(0.5)+(hR*AH1_(0.5)+hG); 834 // Noise detection. 835 AH1 nz=AH1_(0.25)*bL+AH1_(0.25)*dL+AH1_(0.25)*fL+AH1_(0.25)*hL-eL; 836 nz=ASatH1(abs(nz)*APrxMedRcpH1(AMax3H1(AMax3H1(bL,dL,eL),fL,hL)-AMin3H1(AMin3H1(bL,dL,eL),fL,hL))); 837 nz=AH1_(-0.5)*nz+AH1_(1.0); 838 // Min and max of ring. 839 AH1 mn4R=min(AMin3H1(bR,dR,fR),hR); 840 AH1 mn4G=min(AMin3H1(bG,dG,fG),hG); 841 AH1 mn4B=min(AMin3H1(bB,dB,fB),hB); 842 AH1 mx4R=max(AMax3H1(bR,dR,fR),hR); 843 AH1 mx4G=max(AMax3H1(bG,dG,fG),hG); 844 AH1 mx4B=max(AMax3H1(bB,dB,fB),hB); 845 // Immediate constants for peak range. 846 AH2 peakC=AH2(1.0,-1.0*4.0); 847 // Limiters, these need to be high precision RCPs. 848 AH1 hitMinR=min(mn4R,eR)*ARcpH1(AH1_(4.0)*mx4R); 849 AH1 hitMinG=min(mn4G,eG)*ARcpH1(AH1_(4.0)*mx4G); 850 AH1 hitMinB=min(mn4B,eB)*ARcpH1(AH1_(4.0)*mx4B); 851 AH1 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH1(AH1_(4.0)*mn4R+peakC.y); 852 AH1 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH1(AH1_(4.0)*mn4G+peakC.y); 853 AH1 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH1(AH1_(4.0)*mn4B+peakC.y); 854 AH1 lobeR=max(-hitMinR,hitMaxR); 855 AH1 lobeG=max(-hitMinG,hitMaxG); 856 AH1 lobeB=max(-hitMinB,hitMaxB); 857 AH1 lobe=max(AH1_(-FSR_RCAS_LIMIT),min(AMax3H1(lobeR,lobeG,lobeB),AH1_(0.0)))*AH2_AU1(con.y).x; 858 // Apply noise removal. 859 #ifdef FSR_RCAS_DENOISE 860 lobe*=nz; 861 #endif 862 // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. 863 AH1 rcpL=APrxMedRcpH1(AH1_(4.0)*lobe+AH1_(1.0)); 864 pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; 865 pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; 866 pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} 867 #endif 868 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 869 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 870 //_____________________________________________________________/\_______________________________________________________________ 871 //============================================================================================================================== 872 // PACKED 16-BIT VERSION 873 //============================================================================================================================== 874 #if defined(A_GPU)&&defined(A_HALF)&&defined(FSR_RCAS_HX2) 875 // Input callback prototypes that need to be implemented by the calling shader 876 AH4 FsrRcasLoadHx2(ASW2 p); 877 void FsrRcasInputHx2(inout AH2 r,inout AH2 g,inout AH2 b); 878 //------------------------------------------------------------------------------------------------------------------------------ 879 // Can be used to convert from packed Structures of Arrays to Arrays of Structures for store. 880 void FsrRcasDepackHx2(out AH4 pix0,out AH4 pix1,AH2 pixR,AH2 pixG,AH2 pixB){ 881 #ifdef A_HLSL 882 // Invoke a slower path for DX only, since it won't allow uninitialized values. 883 pix0.a=pix1.a=0.0; 884 #endif 885 pix0.rgb=AH3(pixR.x,pixG.x,pixB.x); 886 pix1.rgb=AH3(pixR.y,pixG.y,pixB.y);} 887 //------------------------------------------------------------------------------------------------------------------------------ 888 void FsrRcasHx2( 889 // Output values are for 2 8x8 tiles in a 16x8 region. 890 // pix<R,G,B>.x = left 8x8 tile 891 // pix<R,G,B>.y = right 8x8 tile 892 // This enables later processing to easily be packed as well. 893 out AH2 pixR, 894 out AH2 pixG, 895 out AH2 pixB, 896 #ifdef FSR_RCAS_PASSTHROUGH_ALPHA 897 out AH2 pixA, 898 #endif 899 AU2 ip, // Integer pixel position in output. 900 AU4 con){ // Constant generated by RcasSetup(). 901 // No scaling algorithm uses minimal 3x3 pixel neighborhood. 902 ASW2 sp0=ASW2(ip); 903 AH3 b0=FsrRcasLoadHx2(sp0+ASW2( 0,-1)).rgb; 904 AH3 d0=FsrRcasLoadHx2(sp0+ASW2(-1, 0)).rgb; 905 #ifdef FSR_RCAS_PASSTHROUGH_ALPHA 906 AH4 ee0=FsrRcasLoadHx2(sp0); 907 AH3 e0=ee0.rgb;pixA.r=ee0.a; 908 #else 909 AH3 e0=FsrRcasLoadHx2(sp0).rgb; 910 #endif 911 AH3 f0=FsrRcasLoadHx2(sp0+ASW2( 1, 0)).rgb; 912 AH3 h0=FsrRcasLoadHx2(sp0+ASW2( 0, 1)).rgb; 913 ASW2 sp1=sp0+ASW2(8,0); 914 AH3 b1=FsrRcasLoadHx2(sp1+ASW2( 0,-1)).rgb; 915 AH3 d1=FsrRcasLoadHx2(sp1+ASW2(-1, 0)).rgb; 916 #ifdef FSR_RCAS_PASSTHROUGH_ALPHA 917 AH4 ee1=FsrRcasLoadHx2(sp1); 918 AH3 e1=ee1.rgb;pixA.g=ee1.a; 919 #else 920 AH3 e1=FsrRcasLoadHx2(sp1).rgb; 921 #endif 922 AH3 f1=FsrRcasLoadHx2(sp1+ASW2( 1, 0)).rgb; 923 AH3 h1=FsrRcasLoadHx2(sp1+ASW2( 0, 1)).rgb; 924 // Arrays of Structures to Structures of Arrays conversion. 925 AH2 bR=AH2(b0.r,b1.r); 926 AH2 bG=AH2(b0.g,b1.g); 927 AH2 bB=AH2(b0.b,b1.b); 928 AH2 dR=AH2(d0.r,d1.r); 929 AH2 dG=AH2(d0.g,d1.g); 930 AH2 dB=AH2(d0.b,d1.b); 931 AH2 eR=AH2(e0.r,e1.r); 932 AH2 eG=AH2(e0.g,e1.g); 933 AH2 eB=AH2(e0.b,e1.b); 934 AH2 fR=AH2(f0.r,f1.r); 935 AH2 fG=AH2(f0.g,f1.g); 936 AH2 fB=AH2(f0.b,f1.b); 937 AH2 hR=AH2(h0.r,h1.r); 938 AH2 hG=AH2(h0.g,h1.g); 939 AH2 hB=AH2(h0.b,h1.b); 940 // Run optional input transform. 941 FsrRcasInputHx2(bR,bG,bB); 942 FsrRcasInputHx2(dR,dG,dB); 943 FsrRcasInputHx2(eR,eG,eB); 944 FsrRcasInputHx2(fR,fG,fB); 945 FsrRcasInputHx2(hR,hG,hB); 946 // Luma times 2. 947 AH2 bL=bB*AH2_(0.5)+(bR*AH2_(0.5)+bG); 948 AH2 dL=dB*AH2_(0.5)+(dR*AH2_(0.5)+dG); 949 AH2 eL=eB*AH2_(0.5)+(eR*AH2_(0.5)+eG); 950 AH2 fL=fB*AH2_(0.5)+(fR*AH2_(0.5)+fG); 951 AH2 hL=hB*AH2_(0.5)+(hR*AH2_(0.5)+hG); 952 // Noise detection. 953 AH2 nz=AH2_(0.25)*bL+AH2_(0.25)*dL+AH2_(0.25)*fL+AH2_(0.25)*hL-eL; 954 nz=ASatH2(abs(nz)*APrxMedRcpH2(AMax3H2(AMax3H2(bL,dL,eL),fL,hL)-AMin3H2(AMin3H2(bL,dL,eL),fL,hL))); 955 nz=AH2_(-0.5)*nz+AH2_(1.0); 956 // Min and max of ring. 957 AH2 mn4R=min(AMin3H2(bR,dR,fR),hR); 958 AH2 mn4G=min(AMin3H2(bG,dG,fG),hG); 959 AH2 mn4B=min(AMin3H2(bB,dB,fB),hB); 960 AH2 mx4R=max(AMax3H2(bR,dR,fR),hR); 961 AH2 mx4G=max(AMax3H2(bG,dG,fG),hG); 962 AH2 mx4B=max(AMax3H2(bB,dB,fB),hB); 963 // Immediate constants for peak range. 964 AH2 peakC=AH2(1.0,-1.0*4.0); 965 // Limiters, these need to be high precision RCPs. 966 AH2 hitMinR=min(mn4R,eR)*ARcpH2(AH2_(4.0)*mx4R); 967 AH2 hitMinG=min(mn4G,eG)*ARcpH2(AH2_(4.0)*mx4G); 968 AH2 hitMinB=min(mn4B,eB)*ARcpH2(AH2_(4.0)*mx4B); 969 AH2 hitMaxR=(peakC.x-max(mx4R,eR))*ARcpH2(AH2_(4.0)*mn4R+peakC.y); 970 AH2 hitMaxG=(peakC.x-max(mx4G,eG))*ARcpH2(AH2_(4.0)*mn4G+peakC.y); 971 AH2 hitMaxB=(peakC.x-max(mx4B,eB))*ARcpH2(AH2_(4.0)*mn4B+peakC.y); 972 AH2 lobeR=max(-hitMinR,hitMaxR); 973 AH2 lobeG=max(-hitMinG,hitMaxG); 974 AH2 lobeB=max(-hitMinB,hitMaxB); 975 AH2 lobe=max(AH2_(-FSR_RCAS_LIMIT),min(AMax3H2(lobeR,lobeG,lobeB),AH2_(0.0)))*AH2_(AH2_AU1(con.y).x); 976 // Apply noise removal. 977 #ifdef FSR_RCAS_DENOISE 978 lobe*=nz; 979 #endif 980 // Resolve, which needs the medium precision rcp approximation to avoid visible tonality changes. 981 AH2 rcpL=APrxMedRcpH2(AH2_(4.0)*lobe+AH2_(1.0)); 982 pixR=(lobe*bR+lobe*dR+lobe*hR+lobe*fR+eR)*rcpL; 983 pixG=(lobe*bG+lobe*dG+lobe*hG+lobe*fG+eG)*rcpL; 984 pixB=(lobe*bB+lobe*dB+lobe*hB+lobe*fB+eB)*rcpL;} 985 #endif 986 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 987 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 988 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 989 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 990 //_____________________________________________________________/\_______________________________________________________________ 991 //============================================================================================================================== 992 // 993 // FSR - [LFGA] LINEAR FILM GRAIN APPLICATOR 994 // 995 //------------------------------------------------------------------------------------------------------------------------------ 996 // Adding output-resolution film grain after scaling is a good way to mask both rendering and scaling artifacts. 997 // Suggest using tiled blue noise as film grain input, with peak noise frequency set for a specific look and feel. 998 // The 'Lfga*()' functions provide a convenient way to introduce grain. 999 // These functions limit grain based on distance to signal limits. 1000 // This is done so that the grain is temporally energy preserving, and thus won't modify image tonality. 1001 // Grain application should be done in a linear colorspace. 1002 // The grain should be temporally changing, but have a temporal sum per pixel that adds to zero (non-biased). 1003 //------------------------------------------------------------------------------------------------------------------------------ 1004 // Usage, 1005 // FsrLfga*( 1006 // color, // In/out linear colorspace color {0 to 1} ranged. 1007 // grain, // Per pixel grain texture value {-0.5 to 0.5} ranged, input is 3-channel to support colored grain. 1008 // amount); // Amount of grain (0 to 1} ranged. 1009 //------------------------------------------------------------------------------------------------------------------------------ 1010 // Example if grain texture is monochrome: 'FsrLfgaF(color,AF3_(grain),amount)' 1011 //============================================================================================================================== 1012 #if defined(A_GPU) 1013 // Maximum grain is the minimum distance to the signal limit. 1014 void FsrLfgaF(inout AF3 c,AF3 t,AF1 a){c+=(t*AF3_(a))*min(AF3_(1.0)-c,c);} 1015 #endif 1016 //============================================================================================================================== 1017 #if defined(A_GPU)&&defined(A_HALF) 1018 // Half precision version (slower). 1019 void FsrLfgaH(inout AH3 c,AH3 t,AH1 a){c+=(t*AH3_(a))*min(AH3_(1.0)-c,c);} 1020 //------------------------------------------------------------------------------------------------------------------------------ 1021 // Packed half precision version (faster). 1022 void FsrLfgaHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 tR,AH2 tG,AH2 tB,AH1 a){ 1023 cR+=(tR*AH2_(a))*min(AH2_(1.0)-cR,cR);cG+=(tG*AH2_(a))*min(AH2_(1.0)-cG,cG);cB+=(tB*AH2_(a))*min(AH2_(1.0)-cB,cB);} 1024 #endif 1025 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1026 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1027 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1028 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1029 //_____________________________________________________________/\_______________________________________________________________ 1030 //============================================================================================================================== 1031 // 1032 // FSR - [SRTM] SIMPLE REVERSIBLE TONE-MAPPER 1033 // 1034 //------------------------------------------------------------------------------------------------------------------------------ 1035 // This provides a way to take linear HDR color {0 to FP16_MAX} and convert it into a temporary {0 to 1} ranged post-tonemapped linear. 1036 // The tonemapper preserves RGB ratio, which helps maintain HDR color bleed during filtering. 1037 //------------------------------------------------------------------------------------------------------------------------------ 1038 // Reversible tonemapper usage, 1039 // FsrSrtm*(color); // {0 to FP16_MAX} converted to {0 to 1}. 1040 // FsrSrtmInv*(color); // {0 to 1} converted into {0 to 32768, output peak safe for FP16}. 1041 //============================================================================================================================== 1042 #if defined(A_GPU) 1043 void FsrSrtmF(inout AF3 c){c*=AF3_(ARcpF1(AMax3F1(c.r,c.g,c.b)+AF1_(1.0)));} 1044 // The extra max solves the c=1.0 case (which is a /0). 1045 void FsrSrtmInvF(inout AF3 c){c*=AF3_(ARcpF1(max(AF1_(1.0/32768.0),AF1_(1.0)-AMax3F1(c.r,c.g,c.b))));} 1046 #endif 1047 //============================================================================================================================== 1048 #if defined(A_GPU)&&defined(A_HALF) 1049 void FsrSrtmH(inout AH3 c){c*=AH3_(ARcpH1(AMax3H1(c.r,c.g,c.b)+AH1_(1.0)));} 1050 void FsrSrtmInvH(inout AH3 c){c*=AH3_(ARcpH1(max(AH1_(1.0/32768.0),AH1_(1.0)-AMax3H1(c.r,c.g,c.b))));} 1051 //------------------------------------------------------------------------------------------------------------------------------ 1052 void FsrSrtmHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ 1053 AH2 rcp=ARcpH2(AMax3H2(cR,cG,cB)+AH2_(1.0));cR*=rcp;cG*=rcp;cB*=rcp;} 1054 void FsrSrtmInvHx2(inout AH2 cR,inout AH2 cG,inout AH2 cB){ 1055 AH2 rcp=ARcpH2(max(AH2_(1.0/32768.0),AH2_(1.0)-AMax3H2(cR,cG,cB)));cR*=rcp;cG*=rcp;cB*=rcp;} 1056 #endif 1057 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1058 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1059 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1060 //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 1061 //_____________________________________________________________/\_______________________________________________________________ 1062 //============================================================================================================================== 1063 // 1064 // FSR - [TEPD] TEMPORAL ENERGY PRESERVING DITHER 1065 // 1066 //------------------------------------------------------------------------------------------------------------------------------ 1067 // Temporally energy preserving dithered {0 to 1} linear to gamma 2.0 conversion. 1068 // Gamma 2.0 is used so that the conversion back to linear is just to square the color. 1069 // The conversion comes in 8-bit and 10-bit modes, designed for output to 8-bit UNORM or 10:10:10:2 respectively. 1070 // Given good non-biased temporal blue noise as dither input, 1071 // the output dither will temporally conserve energy. 1072 // This is done by choosing the linear nearest step point instead of perceptual nearest. 1073 // See code below for details. 1074 //------------------------------------------------------------------------------------------------------------------------------ 1075 // DX SPEC RULES FOR FLOAT->UNORM 8-BIT CONVERSION 1076 // =============================================== 1077 // - Output is 'uint(floor(saturate(n)*255.0+0.5))'. 1078 // - Thus rounding is to nearest. 1079 // - NaN gets converted to zero. 1080 // - INF is clamped to {0.0 to 1.0}. 1081 //============================================================================================================================== 1082 #if defined(A_GPU) 1083 // Hand tuned integer position to dither value, with more values than simple checkerboard. 1084 // Only 32-bit has enough precision for this compddation. 1085 // Output is {0 to <1}. 1086 AF1 FsrTepdDitF(AU2 p,AU1 f){ 1087 AF1 x=AF1_(p.x+f); 1088 AF1 y=AF1_(p.y); 1089 // The 1.61803 golden ratio. 1090 AF1 a=AF1_((1.0+sqrt(5.0))/2.0); 1091 // Number designed to provide a good visual pattern. 1092 AF1 b=AF1_(1.0/3.69); 1093 x=x*a+(y*b); 1094 return AFractF1(x);} 1095 //------------------------------------------------------------------------------------------------------------------------------ 1096 // This version is 8-bit gamma 2.0. 1097 // The 'c' input is {0 to 1}. 1098 // Output is {0 to 1} ready for image store. 1099 void FsrTepdC8F(inout AF3 c,AF1 dit){ 1100 AF3 n=sqrt(c); 1101 n=floor(n*AF3_(255.0))*AF3_(1.0/255.0); 1102 AF3 a=n*n; 1103 AF3 b=n+AF3_(1.0/255.0);b=b*b; 1104 // Ratio of 'a' to 'b' required to produce 'c'. 1105 // APrxLoRcpF1() won't work here (at least for very high dynamic ranges). 1106 // APrxMedRcpF1() is an IADD,FMA,MUL. 1107 AF3 r=(c-b)*APrxMedRcpF3(a-b); 1108 // Use the ratio as a cutoff to choose 'a' or 'b'. 1109 // AGtZeroF1() is a MUL. 1110 c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/255.0));} 1111 //------------------------------------------------------------------------------------------------------------------------------ 1112 // This version is 10-bit gamma 2.0. 1113 // The 'c' input is {0 to 1}. 1114 // Output is {0 to 1} ready for image store. 1115 void FsrTepdC10F(inout AF3 c,AF1 dit){ 1116 AF3 n=sqrt(c); 1117 n=floor(n*AF3_(1023.0))*AF3_(1.0/1023.0); 1118 AF3 a=n*n; 1119 AF3 b=n+AF3_(1.0/1023.0);b=b*b; 1120 AF3 r=(c-b)*APrxMedRcpF3(a-b); 1121 c=ASatF3(n+AGtZeroF3(AF3_(dit)-r)*AF3_(1.0/1023.0));} 1122 #endif 1123 //============================================================================================================================== 1124 #if defined(A_GPU)&&defined(A_HALF) 1125 AH1 FsrTepdDitH(AU2 p,AU1 f){ 1126 AF1 x=AF1_(p.x+f); 1127 AF1 y=AF1_(p.y); 1128 AF1 a=AF1_((1.0+sqrt(5.0))/2.0); 1129 AF1 b=AF1_(1.0/3.69); 1130 x=x*a+(y*b); 1131 return AH1(AFractF1(x));} 1132 //------------------------------------------------------------------------------------------------------------------------------ 1133 void FsrTepdC8H(inout AH3 c,AH1 dit){ 1134 AH3 n=sqrt(c); 1135 n=floor(n*AH3_(255.0))*AH3_(1.0/255.0); 1136 AH3 a=n*n; 1137 AH3 b=n+AH3_(1.0/255.0);b=b*b; 1138 AH3 r=(c-b)*APrxMedRcpH3(a-b); 1139 c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/255.0));} 1140 //------------------------------------------------------------------------------------------------------------------------------ 1141 void FsrTepdC10H(inout AH3 c,AH1 dit){ 1142 AH3 n=sqrt(c); 1143 n=floor(n*AH3_(1023.0))*AH3_(1.0/1023.0); 1144 AH3 a=n*n; 1145 AH3 b=n+AH3_(1.0/1023.0);b=b*b; 1146 AH3 r=(c-b)*APrxMedRcpH3(a-b); 1147 c=ASatH3(n+AGtZeroH3(AH3_(dit)-r)*AH3_(1.0/1023.0));} 1148 //============================================================================================================================== 1149 // This computes dither for positions 'p' and 'p+{8,0}'. 1150 AH2 FsrTepdDitHx2(AU2 p,AU1 f){ 1151 AF2 x; 1152 x.x=AF1_(p.x+f); 1153 x.y=x.x+AF1_(8.0); 1154 AF1 y=AF1_(p.y); 1155 AF1 a=AF1_((1.0+sqrt(5.0))/2.0); 1156 AF1 b=AF1_(1.0/3.69); 1157 x=x*AF2_(a)+AF2_(y*b); 1158 return AH2(AFractF2(x));} 1159 //------------------------------------------------------------------------------------------------------------------------------ 1160 void FsrTepdC8Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ 1161 AH2 nR=sqrt(cR); 1162 AH2 nG=sqrt(cG); 1163 AH2 nB=sqrt(cB); 1164 nR=floor(nR*AH2_(255.0))*AH2_(1.0/255.0); 1165 nG=floor(nG*AH2_(255.0))*AH2_(1.0/255.0); 1166 nB=floor(nB*AH2_(255.0))*AH2_(1.0/255.0); 1167 AH2 aR=nR*nR; 1168 AH2 aG=nG*nG; 1169 AH2 aB=nB*nB; 1170 AH2 bR=nR+AH2_(1.0/255.0);bR=bR*bR; 1171 AH2 bG=nG+AH2_(1.0/255.0);bG=bG*bG; 1172 AH2 bB=nB+AH2_(1.0/255.0);bB=bB*bB; 1173 AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); 1174 AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); 1175 AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); 1176 cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/255.0)); 1177 cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/255.0)); 1178 cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/255.0));} 1179 //------------------------------------------------------------------------------------------------------------------------------ 1180 void FsrTepdC10Hx2(inout AH2 cR,inout AH2 cG,inout AH2 cB,AH2 dit){ 1181 AH2 nR=sqrt(cR); 1182 AH2 nG=sqrt(cG); 1183 AH2 nB=sqrt(cB); 1184 nR=floor(nR*AH2_(1023.0))*AH2_(1.0/1023.0); 1185 nG=floor(nG*AH2_(1023.0))*AH2_(1.0/1023.0); 1186 nB=floor(nB*AH2_(1023.0))*AH2_(1.0/1023.0); 1187 AH2 aR=nR*nR; 1188 AH2 aG=nG*nG; 1189 AH2 aB=nB*nB; 1190 AH2 bR=nR+AH2_(1.0/1023.0);bR=bR*bR; 1191 AH2 bG=nG+AH2_(1.0/1023.0);bG=bG*bG; 1192 AH2 bB=nB+AH2_(1.0/1023.0);bB=bB*bB; 1193 AH2 rR=(cR-bR)*APrxMedRcpH2(aR-bR); 1194 AH2 rG=(cG-bG)*APrxMedRcpH2(aG-bG); 1195 AH2 rB=(cB-bB)*APrxMedRcpH2(aB-bB); 1196 cR=ASatH2(nR+AGtZeroH2(dit-rR)*AH2_(1.0/1023.0)); 1197 cG=ASatH2(nG+AGtZeroH2(dit-rG)*AH2_(1.0/1023.0)); 1198 cB=ASatH2(nB+AGtZeroH2(dit-rB)*AH2_(1.0/1023.0));} 1199 #endif