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 Sk4px_DEFINED
9 #define Sk4px_DEFINED
10 
11 #include "SkNx.h"
12 #include "SkColor.h"
13 
14 // 1, 2 or 4 SkPMColors, generally vectorized.
15 class Sk4px : public Sk16b {
16 public:
Sk4px(SkAlpha a)17     Sk4px(SkAlpha a) : INHERITED(a) {} // Duplicate 16x: a    -> aaaa aaaa aaaa aaaa
18     Sk4px(SkPMColor);                  // Duplicate 4x:  argb -> argb argb argb argb
Sk4px(const Sk16b & v)19     Sk4px(const Sk16b& v) : INHERITED(v) {}
20 
21     Sk4px alphas() const;  // ARGB argb XYZW xyzw -> AAAA aaaa XXXX xxxx
22 
23     // Mask away color or alpha lanes.
24     Sk4px zeroColors() const;  // ARGB argb XYZW xyzw -> A000 a000 X000 x000
25     Sk4px zeroAlphas() const;  // ARGB argb XYZW xyzw -> 0RGB 0rgb 0YZW 0yzw
26 
inv()27     Sk4px inv() const { return Sk16b(255) - *this; }
28 
29     // When loading or storing fewer than 4 SkPMColors, we use the low lanes.
30     static Sk4px Load4(const SkPMColor[4]);  // PMColor[4] -> ARGB argb XYZW xyzw
31     static Sk4px Load2(const SkPMColor[2]);  // PMColor[2] -> ARGB argb ???? ????
32     static Sk4px Load1(const SkPMColor[1]);  // PMColor[1] -> ARGB ???? ???? ????
33 
34     // Ditto for Alphas... Load2Alphas fills the low two lanes of Sk4px.
35     static Sk4px Load4Alphas(const SkAlpha[4]);  // AaXx -> AAAA aaaa XXXX xxxx
36     static Sk4px Load2Alphas(const SkAlpha[2]);  // Aa   -> AAAA aaaa ???? ????
37 
38     void store4(SkPMColor[4]) const;
39     void store2(SkPMColor[2]) const;
40     void store1(SkPMColor[1]) const;
41 
42     // 1, 2, or 4 SkPMColors with 16-bit components.
43     // This is most useful as the result of a multiply, e.g. from mulWiden().
44     class Wide : public Sk16h {
45     public:
Wide(const Sk16h & v)46         Wide(const Sk16h& v) : Sk16h(v) {}
47 
48         // Pack the top byte of each component back down into 4 SkPMColors.
49         Sk4px addNarrowHi(const Sk16h&) const;
50 
div255TruncNarrow()51         Sk4px div255TruncNarrow() const { return this->addNarrowHi(*this >> 8); }
div255RoundNarrow()52         Sk4px div255RoundNarrow() const {
53             return Sk4px::Wide(*this + Sk16h(128)).div255TruncNarrow();
54         }
55 
56     private:
57         typedef Sk16h INHERITED;
58     };
59 
60     Wide widenLo() const;               // ARGB -> 0A 0R 0G 0B
61     Wide widenHi() const;               // ARGB -> A0 R0 G0 B0
62     Wide mulWiden(const Sk16b&) const;  // 8-bit x 8-bit -> 16-bit components.
mul255Widen()63     Wide mul255Widen() const {
64         // TODO: x*255 = x*256-x, so something like this->widenHi() - this->widenLo()?
65         return this->mulWiden(Sk16b(255));
66     }
67 
68     // A generic driver that maps fn over a src array into a dst array.
69     // fn should take an Sk4px (4 src pixels) and return an Sk4px (4 dst pixels).
70     template <typename Fn>
MapSrc(int count,SkPMColor * dst,const SkPMColor * src,Fn fn)71     static void MapSrc(int count, SkPMColor* dst, const SkPMColor* src, Fn fn) {
72         // This looks a bit odd, but it helps loop-invariant hoisting across different calls to fn.
73         // Basically, we need to make sure we keep things inside a single loop.
74         while (count > 0) {
75             if (count >= 8) {
76                 Sk4px dst0 = fn(Load4(src+0)),
77                       dst4 = fn(Load4(src+4));
78                 dst0.store4(dst+0);
79                 dst4.store4(dst+4);
80                 dst += 8; src += 8; count -= 8;
81                 continue;  // Keep our stride at 8 pixels as long as possible.
82             }
83             SkASSERT(count <= 7);
84             if (count >= 4) {
85                 fn(Load4(src)).store4(dst);
86                 dst += 4; src += 4; count -= 4;
87             }
88             if (count >= 2) {
89                 fn(Load2(src)).store2(dst);
90                 dst += 2; src += 2; count -= 2;
91             }
92             if (count >= 1) {
93                 fn(Load1(src)).store1(dst);
94             }
95             break;
96         }
97     }
98 
99     // As above, but with dst4' = fn(dst4, src4).
100     template <typename Fn>
MapDstSrc(int count,SkPMColor * dst,const SkPMColor * src,Fn fn)101     static void MapDstSrc(int count, SkPMColor* dst, const SkPMColor* src, Fn fn) {
102         while (count > 0) {
103             if (count >= 8) {
104                 Sk4px dst0 = fn(Load4(dst+0), Load4(src+0)),
105                       dst4 = fn(Load4(dst+4), Load4(src+4));
106                 dst0.store4(dst+0);
107                 dst4.store4(dst+4);
108                 dst += 8; src += 8; count -= 8;
109                 continue;  // Keep our stride at 8 pixels as long as possible.
110             }
111             SkASSERT(count <= 7);
112             if (count >= 4) {
113                 fn(Load4(dst), Load4(src)).store4(dst);
114                 dst += 4; src += 4; count -= 4;
115             }
116             if (count >= 2) {
117                 fn(Load2(dst), Load2(src)).store2(dst);
118                 dst += 2; src += 2; count -= 2;
119             }
120             if (count >= 1) {
121                 fn(Load1(dst), Load1(src)).store1(dst);
122             }
123             break;
124         }
125     }
126 
127     // As above, but with dst4' = fn(dst4, src4, alpha4).
128     template <typename Fn>
MapDstSrcAlpha(int count,SkPMColor * dst,const SkPMColor * src,const SkAlpha * a,Fn fn)129     static void MapDstSrcAlpha(
130             int count, SkPMColor* dst, const SkPMColor* src, const SkAlpha* a, Fn fn) {
131         while (count > 0) {
132             if (count >= 8) {
133                 Sk4px alpha0 = Load4Alphas(a+0),
134                       alpha4 = Load4Alphas(a+4);
135                 Sk4px dst0 = fn(Load4(dst+0), Load4(src+0), alpha0),
136                       dst4 = fn(Load4(dst+4), Load4(src+4), alpha4);
137                 dst0.store4(dst+0);
138                 dst4.store4(dst+4);
139                 dst += 8; src += 8; a += 8; count -= 8;
140                 continue;  // Keep our stride at 8 pixels as long as possible.
141             }
142             SkASSERT(count <= 7);
143             if (count >= 4) {
144                 Sk4px alpha = Load4Alphas(a);
145                 fn(Load4(dst), Load4(src), alpha).store4(dst);
146                 dst += 4; src += 4; a += 4; count -= 4;
147             }
148             if (count >= 2) {
149                 Sk4px alpha = Load2Alphas(a);
150                 fn(Load2(dst), Load2(src), alpha).store2(dst);
151                 dst += 2; src += 2; a += 2; count -= 2;
152             }
153             if (count >= 1) {
154                 Sk4px alpha(*a);
155                 fn(Load1(dst), Load1(src), alpha).store1(dst);
156             }
157             break;
158         }
159     }
160 
161 private:
162     typedef Sk16b INHERITED;
163 };
164 
165 #ifdef SKNX_NO_SIMD
166     #include "../opts/Sk4px_none.h"
167 #else
168     #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
169         #include "../opts/Sk4px_SSE2.h"
170     #elif defined(SK_ARM_HAS_NEON)
171         #include "../opts/Sk4px_NEON.h"
172     #else
173         #include "../opts/Sk4px_none.h"
174     #endif
175 #endif
176 
177 #endif//Sk4px_DEFINED
178