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 "SkXfermode_proccoeff.h"
15
16namespace {
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
23XFERMODE(Clear) { return Sk4px::DupPMColor(0); }
24XFERMODE(Src)   { return s; }
25XFERMODE(Dst)   { return d; }
26XFERMODE(SrcIn)   { return     s.approxMulDiv255(d.alphas()      ); }
27XFERMODE(SrcOut)  { return     s.approxMulDiv255(d.alphas().inv()); }
28XFERMODE(SrcOver) { return s + d.approxMulDiv255(s.alphas().inv()); }
29XFERMODE(DstIn)   { return SrcIn  ()(s,d); }
30XFERMODE(DstOut)  { return SrcOut ()(s,d); }
31XFERMODE(DstOver) { return SrcOver()(s,d); }
32
33// [ S * Da + (1 - Sa) * D]
34XFERMODE(SrcATop) { return (s * d.alphas() + d * s.alphas().inv()).div255(); }
35XFERMODE(DstATop) { return SrcATop()(s,d); }
36//[ S * (1 - Da) + (1 - Sa) * D ]
37XFERMODE(Xor) { return (s * d.alphas().inv() + d * s.alphas().inv()).div255(); }
38// [S + D ]
39XFERMODE(Plus) { return s.saturatedAdd(d); }
40// [S * D ]
41XFERMODE(Modulate) { return s.approxMulDiv255(d); }
42// [S + D - S * D]
43XFERMODE(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}
48XFERMODE(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).)
50XFERMODE(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 ]
56XFERMODE(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
69XFERMODE(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}
85XFERMODE(Overlay) { return HardLight()(s,d); }
86
87XFERMODE(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}
100XFERMODE(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
120static 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}
124static inline Sk4f alphas(const Sk4f& f) {
125    return f[SK_A32_SHIFT/8];
126}
127
128XFERMODE(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).approxInvert()) + 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}
144XFERMODE(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.approxInvert())) + 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}
160XFERMODE(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.
189template <typename Xfermode>
190static 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
200XFERMODE_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.
209template <typename Xfermode> static void mark_dst_initialized_if_safe(void*, void*) {}
210template <> void mark_dst_initialized_if_safe<Src>(void* dst, void* end) {
211    sk_msan_mark_initialized(dst, end, "Src doesn't read dst.");
212}
213template <> 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
217template <typename Xfermode>
218class Sk4pxXfermode : public SkProcCoeffXfermode {
219public:
220    Sk4pxXfermode(const ProcCoeff& rec, SkXfermode::Mode mode)
221        : INHERITED(rec, mode) {}
222
223    void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
224        mark_dst_initialized_if_safe<Xfermode>(dst, dst+n);
225        if (nullptr == aa) {
226            Sk4px::MapDstSrc(n, dst, src, Xfermode());
227        } else {
228            Sk4px::MapDstSrcAlpha(n, dst, src, aa, xfer_aa<Xfermode>);
229        }
230    }
231
232    void xfer16(uint16_t dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
233        mark_dst_initialized_if_safe<Xfermode>(dst, dst+n);
234        SkPMColor dst32[4];
235        while (n >= 4) {
236            dst32[0] = SkPixel16ToPixel32(dst[0]);
237            dst32[1] = SkPixel16ToPixel32(dst[1]);
238            dst32[2] = SkPixel16ToPixel32(dst[2]);
239            dst32[3] = SkPixel16ToPixel32(dst[3]);
240
241            this->xfer32(dst32, src, 4, aa);
242
243            dst[0] = SkPixel32ToPixel16(dst32[0]);
244            dst[1] = SkPixel32ToPixel16(dst32[1]);
245            dst[2] = SkPixel32ToPixel16(dst32[2]);
246            dst[3] = SkPixel32ToPixel16(dst32[3]);
247
248            dst += 4;
249            src += 4;
250            aa  += aa ? 4 : 0;
251            n -= 4;
252        }
253        while (n) {
254            SkPMColor dst32 = SkPixel16ToPixel32(*dst);
255            this->xfer32(&dst32, src, 1, aa);
256            *dst = SkPixel32ToPixel16(dst32);
257
258            dst += 1;
259            src += 1;
260            aa  += aa ? 1 : 0;
261            n   -= 1;
262        }
263    }
264
265private:
266    typedef SkProcCoeffXfermode INHERITED;
267};
268
269template <typename Xfermode>
270class Sk4fXfermode : public SkProcCoeffXfermode {
271public:
272    Sk4fXfermode(const ProcCoeff& rec, SkXfermode::Mode mode)
273        : INHERITED(rec, mode) {}
274
275    void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
276        for (int i = 0; i < n; i++) {
277            dst[i] = Xfer32_1(dst[i], src[i], aa ? aa+i : nullptr);
278        }
279    }
280
281    void xfer16(uint16_t dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
282        for (int i = 0; i < n; i++) {
283            SkPMColor dst32 = SkPixel16ToPixel32(dst[i]);
284            dst32 = Xfer32_1(dst32, src[i], aa ? aa+i : nullptr);
285            dst[i] = SkPixel32ToPixel16(dst32);
286        }
287    }
288
289private:
290    static SkPMColor Xfer32_1(SkPMColor dst, const SkPMColor src, const SkAlpha* aa) {
291        Sk4f d = Load(dst),
292             s = Load(src),
293             b = Xfermode()(d, s);
294        if (aa) {
295            Sk4f a = Sk4f(*aa) * Sk4f(1.0f/255);
296            b = b*a + d*(Sk4f(1)-a);
297        }
298        return Round(b);
299    }
300
301    static Sk4f Load(SkPMColor c) {
302        return SkNx_cast<float>(Sk4b::Load(&c)) * Sk4f(1.0f/255);
303    }
304
305    static SkPMColor Round(const Sk4f& f) {
306        SkPMColor c;
307        SkNx_cast<uint8_t>(f * Sk4f(255) + Sk4f(0.5f)).store(&c);
308        return c;
309    }
310
311    typedef SkProcCoeffXfermode INHERITED;
312};
313
314} // namespace
315
316namespace SK_OPTS_NS {
317
318static SkXfermode* create_xfermode(const ProcCoeff& rec, SkXfermode::Mode mode) {
319    switch (mode) {
320#define CASE(Xfermode) \
321    case SkXfermode::k##Xfermode##_Mode: return new Sk4pxXfermode<Xfermode>(rec, mode)
322        CASE(Clear);
323        CASE(Src);
324        CASE(Dst);
325        CASE(SrcOver);
326        CASE(DstOver);
327        CASE(SrcIn);
328        CASE(DstIn);
329        CASE(SrcOut);
330        CASE(DstOut);
331        CASE(SrcATop);
332        CASE(DstATop);
333        CASE(Xor);
334        CASE(Plus);
335        CASE(Modulate);
336        CASE(Screen);
337        CASE(Multiply);
338        CASE(Difference);
339        CASE(Exclusion);
340        CASE(HardLight);
341        CASE(Overlay);
342        CASE(Darken);
343        CASE(Lighten);
344    #undef CASE
345
346#define CASE(Xfermode) \
347    case SkXfermode::k##Xfermode##_Mode: return new Sk4fXfermode<Xfermode>(rec, mode)
348        CASE(ColorDodge);
349        CASE(ColorBurn);
350        CASE(SoftLight);
351    #undef CASE
352
353        default: break;
354    }
355    return nullptr;
356}
357
358} // namespace SK_OPTS_NS
359
360#endif//Sk4pxXfermode_DEFINED
361