1 /* 2 * Copyright 2015 Google Inc. 3 * 4 * Use of this source code is governed by a BSD-style license that can be 5 * found in the LICENSE file. 6 */ 7 8 #ifndef Sk4pxXfermode_DEFINED 9 #define Sk4pxXfermode_DEFINED 10 11 #include "Sk4px.h" 12 #include "SkMSAN.h" 13 #include "SkNx.h" 14 #include "SkXfermodePriv.h" 15 16 namespace { 17 18 // Most xfermodes can be done most efficiently 4 pixels at a time in 8 or 16-bit fixed point. 19 #define XFERMODE(Xfermode) \ 20 struct Xfermode { Sk4px operator()(const Sk4px&, const Sk4px&) const; }; \ 21 inline Sk4px Xfermode::operator()(const Sk4px& d, const Sk4px& s) const 22 23 XFERMODE(Clear) { return Sk4px::DupPMColor(0); } 24 XFERMODE(Src) { return s; } 25 XFERMODE(Dst) { return d; } 26 XFERMODE(SrcIn) { return s.approxMulDiv255(d.alphas() ); } 27 XFERMODE(SrcOut) { return s.approxMulDiv255(d.alphas().inv()); } 28 XFERMODE(SrcOver) { return s + d.approxMulDiv255(s.alphas().inv()); } 29 XFERMODE(DstIn) { return SrcIn ()(s,d); } 30 XFERMODE(DstOut) { return SrcOut ()(s,d); } 31 XFERMODE(DstOver) { return SrcOver()(s,d); } 32 33 // [ S * Da + (1 - Sa) * D] 34 XFERMODE(SrcATop) { return (s * d.alphas() + d * s.alphas().inv()).div255(); } 35 XFERMODE(DstATop) { return SrcATop()(s,d); } 36 //[ S * (1 - Da) + (1 - Sa) * D ] 37 XFERMODE(Xor) { return (s * d.alphas().inv() + d * s.alphas().inv()).div255(); } 38 // [S + D ] 39 XFERMODE(Plus) { return s.saturatedAdd(d); } 40 // [S * D ] 41 XFERMODE(Modulate) { return s.approxMulDiv255(d); } 42 // [S + D - S * D] 43 XFERMODE(Screen) { 44 // Doing the math as S + (1-S)*D or S + (D - S*D) means the add and subtract can be done 45 // in 8-bit space without overflow. S + (1-S)*D is a touch faster because inv() is cheap. 46 return s + d.approxMulDiv255(s.inv()); 47 } 48 XFERMODE(Multiply) { return (s * d.alphas().inv() + d * s.alphas().inv() + s*d).div255(); } 49 // [ Sa + Da - Sa*Da, Sc + Dc - 2*min(Sc*Da, Dc*Sa) ] (And notice Sa*Da == min(Sa*Da, Da*Sa).) 50 XFERMODE(Difference) { 51 auto m = Sk4px::Wide::Min(s * d.alphas(), d * s.alphas()).div255(); 52 // There's no chance of underflow, and if we subtract m before adding s+d, no overflow. 53 return (s - m) + (d - m.zeroAlphas()); 54 } 55 // [ Sa + Da - Sa*Da, Sc + Dc - 2*Sc*Dc ] 56 XFERMODE(Exclusion) { 57 auto p = s.approxMulDiv255(d); 58 // There's no chance of underflow, and if we subtract p before adding src+dst, no overflow. 59 return (s - p) + (d - p.zeroAlphas()); 60 } 61 62 // We take care to use exact math for these next few modes where alphas 63 // and colors are calculated using significantly different math. We need 64 // to preserve premul invariants, and exact math makes this easier. 65 // 66 // TODO: Some of these implementations might be able to be sped up a bit 67 // while maintaining exact math, but let's follow up with that. 68 69 XFERMODE(HardLight) { 70 auto sa = s.alphas(), 71 da = d.alphas(); 72 73 auto srcover = s + (d * sa.inv()).div255(); 74 75 auto isLite = ((sa-s) < s).widenLoHi(); 76 77 auto lite = sa*da - ((da-d)*(sa-s) << 1), 78 dark = s*d << 1, 79 both = s*da.inv() + d*sa.inv(); 80 81 auto alphas = srcover; 82 auto colors = (both + isLite.thenElse(lite, dark)).div255(); 83 return alphas.zeroColors() + colors.zeroAlphas(); 84 } 85 XFERMODE(Overlay) { return HardLight()(s,d); } 86 87 XFERMODE(Darken) { 88 auto sa = s.alphas(), 89 da = d.alphas(); 90 91 auto sda = (s*da).div255(), 92 dsa = (d*sa).div255(); 93 94 auto srcover = s + (d * sa.inv()).div255(), 95 dstover = d + (s * da.inv()).div255(); 96 auto alphas = srcover, 97 colors = (sda < dsa).thenElse(srcover, dstover); 98 return alphas.zeroColors() + colors.zeroAlphas(); 99 } 100 XFERMODE(Lighten) { 101 auto sa = s.alphas(), 102 da = d.alphas(); 103 104 auto sda = (s*da).div255(), 105 dsa = (d*sa).div255(); 106 107 auto srcover = s + (d * sa.inv()).div255(), 108 dstover = d + (s * da.inv()).div255(); 109 auto alphas = srcover, 110 colors = (dsa < sda).thenElse(srcover, dstover); 111 return alphas.zeroColors() + colors.zeroAlphas(); 112 } 113 #undef XFERMODE 114 115 // Some xfermodes use math like divide or sqrt that's best done in floats 1 pixel at a time. 116 #define XFERMODE(Xfermode) \ 117 struct Xfermode { Sk4f operator()(const Sk4f&, const Sk4f&) const; }; \ 118 inline Sk4f Xfermode::operator()(const Sk4f& d, const Sk4f& s) const 119 120 static inline Sk4f a_rgb(const Sk4f& a, const Sk4f& rgb) { 121 static_assert(SK_A32_SHIFT == 24, ""); 122 return a * Sk4f(0,0,0,1) + rgb * Sk4f(1,1,1,0); 123 } 124 static inline Sk4f alphas(const Sk4f& f) { 125 return f[SK_A32_SHIFT/8]; 126 } 127 128 XFERMODE(ColorDodge) { 129 auto sa = alphas(s), 130 da = alphas(d), 131 isa = Sk4f(1)-sa, 132 ida = Sk4f(1)-da; 133 134 auto srcover = s + d*isa, 135 dstover = d + s*ida, 136 otherwise = sa * Sk4f::Min(da, (d*sa)*(sa-s).invert()) + s*ida + d*isa; 137 138 // Order matters here, preferring d==0 over s==sa. 139 auto colors = (d == Sk4f(0)).thenElse(dstover, 140 (s == sa).thenElse(srcover, 141 otherwise)); 142 return a_rgb(srcover, colors); 143 } 144 XFERMODE(ColorBurn) { 145 auto sa = alphas(s), 146 da = alphas(d), 147 isa = Sk4f(1)-sa, 148 ida = Sk4f(1)-da; 149 150 auto srcover = s + d*isa, 151 dstover = d + s*ida, 152 otherwise = sa*(da-Sk4f::Min(da, (da-d)*sa*s.invert())) + s*ida + d*isa; 153 154 // Order matters here, preferring d==da over s==0. 155 auto colors = (d == da).thenElse(dstover, 156 (s == Sk4f(0)).thenElse(srcover, 157 otherwise)); 158 return a_rgb(srcover, colors); 159 } 160 XFERMODE(SoftLight) { 161 auto sa = alphas(s), 162 da = alphas(d), 163 isa = Sk4f(1)-sa, 164 ida = Sk4f(1)-da; 165 166 // Some common terms. 167 auto m = (da > Sk4f(0)).thenElse(d / da, Sk4f(0)), 168 s2 = Sk4f(2)*s, 169 m4 = Sk4f(4)*m; 170 171 // The logic forks three ways: 172 // 1. dark src? 173 // 2. light src, dark dst? 174 // 3. light src, light dst? 175 auto darkSrc = d*(sa + (s2 - sa)*(Sk4f(1) - m)), // Used in case 1. 176 darkDst = (m4*m4 + m4)*(m - Sk4f(1)) + Sk4f(7)*m, // Used in case 2. 177 liteDst = m.sqrt() - m, // Used in case 3. 178 liteSrc = d*sa + da*(s2-sa)*(Sk4f(4)*d <= da).thenElse(darkDst, liteDst); // Case 2 or 3? 179 180 auto alpha = s + d*isa; 181 auto colors = s*ida + d*isa + (s2 <= sa).thenElse(darkSrc, liteSrc); // Case 1 or 2/3? 182 183 return a_rgb(alpha, colors); 184 } 185 #undef XFERMODE 186 187 // A reasonable fallback mode for doing AA is to simply apply the transfermode first, 188 // then linearly interpolate the AA. 189 template <typename Xfermode> 190 static Sk4px xfer_aa(const Sk4px& d, const Sk4px& s, const Sk4px& aa) { 191 Sk4px bw = Xfermode()(d, s); 192 return (bw * aa + d * aa.inv()).div255(); 193 } 194 195 // For some transfermodes we specialize AA, either for correctness or performance. 196 #define XFERMODE_AA(Xfermode) \ 197 template <> Sk4px xfer_aa<Xfermode>(const Sk4px& d, const Sk4px& s, const Sk4px& aa) 198 199 // Plus' clamp needs to happen after AA. skia:3852 200 XFERMODE_AA(Plus) { // [ clamp( (1-AA)D + (AA)(S+D) ) == clamp(D + AA*S) ] 201 return d.saturatedAdd(s.approxMulDiv255(aa)); 202 } 203 204 #undef XFERMODE_AA 205 206 // Src and Clear modes are safe to use with unitialized dst buffers, 207 // even if the implementation branches based on bytes from dst (e.g. asserts in Debug mode). 208 // For those modes, just lie to MSAN that dst is always intialized. 209 template <typename Xfermode> static void mark_dst_initialized_if_safe(void*, void*) {} 210 template <> void mark_dst_initialized_if_safe<Src>(void* dst, void* end) { 211 sk_msan_mark_initialized(dst, end, "Src doesn't read dst."); 212 } 213 template <> void mark_dst_initialized_if_safe<Clear>(void* dst, void* end) { 214 sk_msan_mark_initialized(dst, end, "Clear doesn't read dst."); 215 } 216 217 template <typename Xfermode> 218 class Sk4pxXfermode : public SkXfermode { 219 public: 220 Sk4pxXfermode() {} 221 222 void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override { 223 mark_dst_initialized_if_safe<Xfermode>(dst, dst+n); 224 if (nullptr == aa) { 225 Sk4px::MapDstSrc(n, dst, src, Xfermode()); 226 } else { 227 Sk4px::MapDstSrcAlpha(n, dst, src, aa, xfer_aa<Xfermode>); 228 } 229 } 230 }; 231 232 template <typename Xfermode> 233 class Sk4fXfermode : public SkXfermode { 234 public: 235 Sk4fXfermode() {} 236 237 void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override { 238 for (int i = 0; i < n; i++) { 239 dst[i] = Xfer32_1(dst[i], src[i], aa ? aa+i : nullptr); 240 } 241 } 242 243 private: 244 static SkPMColor Xfer32_1(SkPMColor dst, const SkPMColor src, const SkAlpha* aa) { 245 Sk4f d = Load(dst), 246 s = Load(src), 247 b = Xfermode()(d, s); 248 if (aa) { 249 Sk4f a = Sk4f(*aa) * Sk4f(1.0f/255); 250 b = b*a + d*(Sk4f(1)-a); 251 } 252 return Round(b); 253 } 254 255 static Sk4f Load(SkPMColor c) { 256 return SkNx_cast<float>(Sk4b::Load(&c)) * Sk4f(1.0f/255); 257 } 258 259 static SkPMColor Round(const Sk4f& f) { 260 SkPMColor c; 261 SkNx_cast<uint8_t>(f * Sk4f(255) + Sk4f(0.5f)).store(&c); 262 return c; 263 } 264 }; 265 266 } // namespace 267 268 namespace SK_OPTS_NS { 269 270 static SkXfermode* create_xfermode(SkBlendMode mode) { 271 switch (mode) { 272 #define CASE(Xfermode) \ 273 case SkBlendMode::k##Xfermode: return new Sk4pxXfermode<Xfermode>() 274 CASE(Clear); 275 CASE(Src); 276 CASE(Dst); 277 CASE(SrcOver); 278 CASE(DstOver); 279 CASE(SrcIn); 280 CASE(DstIn); 281 CASE(SrcOut); 282 CASE(DstOut); 283 CASE(SrcATop); 284 CASE(DstATop); 285 CASE(Xor); 286 CASE(Plus); 287 CASE(Modulate); 288 CASE(Screen); 289 CASE(Multiply); 290 CASE(Difference); 291 CASE(Exclusion); 292 CASE(HardLight); 293 CASE(Overlay); 294 CASE(Darken); 295 CASE(Lighten); 296 #undef CASE 297 298 #define CASE(Xfermode) \ 299 case SkBlendMode::k##Xfermode: return new Sk4fXfermode<Xfermode>() 300 CASE(ColorDodge); 301 CASE(ColorBurn); 302 CASE(SoftLight); 303 #undef CASE 304 305 default: break; 306 } 307 return nullptr; 308 } 309 310 } // namespace SK_OPTS_NS 311 312 #endif//Sk4pxXfermode_DEFINED 313