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 "include/core/SkColor.h" |
12 | #include "include/private/SkColorData.h" |
13 | #include "include/private/SkNx.h" |
14 | |
15 | // This file may be included multiple times by .cpp files with different flags, leading |
16 | // to different definitions. Usually that doesn't matter because it's all inlined, but |
17 | // in Debug modes the compilers may not inline everything. So wrap everything in an |
18 | // anonymous namespace to give each includer their own silo of this code (or the linker |
19 | // will probably pick one randomly for us, which is rarely correct). |
20 | namespace { // NOLINT(google-build-namespaces) |
21 | |
22 | // 1, 2 or 4 SkPMColors, generally vectorized. |
23 | class Sk4px : public Sk16b { |
24 | public: |
25 | Sk4px(const Sk16b& v) : INHERITED(v) {} |
26 | |
27 | static Sk4px DupPMColor(SkPMColor c) { |
28 | Sk4u splat(c); |
29 | |
30 | Sk4px v; |
31 | memcpy((void*)&v, &splat, 16); |
32 | return v; |
33 | } |
34 | |
35 | Sk4px alphas() const; // ARGB argb XYZW xyzw -> AAAA aaaa XXXX xxxx |
36 | Sk4px inv() const { return Sk16b(255) - *this; } |
37 | |
38 | // When loading or storing fewer than 4 SkPMColors, we use the low lanes. |
39 | static Sk4px Load4(const SkPMColor px[4]) { |
40 | Sk4px v; |
41 | memcpy((void*)&v, px, 16); |
42 | return v; |
43 | } |
44 | static Sk4px Load2(const SkPMColor px[2]) { |
45 | Sk4px v; |
46 | memcpy((void*)&v, px, 8); |
47 | return v; |
48 | } |
49 | static Sk4px Load1(const SkPMColor px[1]) { |
50 | Sk4px v; |
51 | memcpy((void*)&v, px, 4); |
52 | return v; |
53 | } |
54 | |
55 | // Ditto for Alphas... Load2Alphas fills the low two lanes of Sk4px. |
56 | static Sk4px Load4Alphas(const SkAlpha[4]); // AaXx -> AAAA aaaa XXXX xxxx |
57 | static Sk4px Load2Alphas(const SkAlpha[2]); // Aa -> AAAA aaaa ???? ???? |
58 | |
59 | void store4(SkPMColor px[4]) const { memcpy(px, this, 16); } |
60 | void store2(SkPMColor px[2]) const { memcpy(px, this, 8); } |
61 | void store1(SkPMColor px[1]) const { memcpy(px, this, 4); } |
62 | |
63 | // 1, 2, or 4 SkPMColors with 16-bit components. |
64 | // This is most useful as the result of a multiply, e.g. from mulWiden(). |
65 | class Wide : public Sk16h { |
66 | public: |
67 | Wide(const Sk16h& v) : Sk16h(v) {} |
68 | |
69 | // Add, then pack the top byte of each component back down into 4 SkPMColors. |
70 | Sk4px addNarrowHi(const Sk16h&) const; |
71 | |
72 | // Rounds, i.e. (x+127) / 255. |
73 | Sk4px div255() const; |
74 | |
75 | // These just keep the types as Wide so the user doesn't have to keep casting. |
76 | Wide operator * (const Wide& o) const { return INHERITED::operator*(o); } |
77 | Wide operator + (const Wide& o) const { return INHERITED::operator+(o); } |
78 | Wide operator - (const Wide& o) const { return INHERITED::operator-(o); } |
79 | Wide operator >> (int bits) const { return INHERITED::operator>>(bits); } |
80 | Wide operator << (int bits) const { return INHERITED::operator<<(bits); } |
81 | |
82 | private: |
83 | typedef Sk16h INHERITED; |
84 | }; |
85 | |
86 | Wide widen() const; // Widen 8-bit values to low 8-bits of 16-bit lanes. |
87 | Wide mulWiden(const Sk16b&) const; // 8-bit x 8-bit -> 16-bit components. |
88 | |
89 | // The only 8-bit multiply we use is 8-bit x 8-bit -> 16-bit. Might as well make it pithy. |
90 | Wide operator * (const Sk4px& o) const { return this->mulWiden(o); } |
91 | |
92 | // These just keep the types as Sk4px so the user doesn't have to keep casting. |
93 | Sk4px operator + (const Sk4px& o) const { return INHERITED::operator+(o); } |
94 | Sk4px operator - (const Sk4px& o) const { return INHERITED::operator-(o); } |
95 | Sk4px operator < (const Sk4px& o) const { return INHERITED::operator<(o); } |
96 | Sk4px thenElse(const Sk4px& t, const Sk4px& e) const { return INHERITED::thenElse(t,e); } |
97 | |
98 | // Generally faster than (*this * o).div255(). |
99 | // May be incorrect by +-1, but is always exactly correct when *this or o is 0 or 255. |
100 | Sk4px approxMulDiv255(const Sk16b& o) const { |
101 | // (x*y + x) / 256 meets these criteria. (As of course does (x*y + y) / 256 by symmetry.) |
102 | // FYI: (x*y + 255) / 256 also meets these criteria. In my brief testing, it was slower. |
103 | return this->widen().addNarrowHi(*this * o); |
104 | } |
105 | |
106 | // A generic driver that maps fn over a src array into a dst array. |
107 | // fn should take an Sk4px (4 src pixels) and return an Sk4px (4 dst pixels). |
108 | template <typename Fn> |
109 | static void MapSrc(int n, SkPMColor* dst, const SkPMColor* src, const Fn& fn) { |
110 | SkASSERT(dst); |
111 | SkASSERT(src); |
112 | // This looks a bit odd, but it helps loop-invariant hoisting across different calls to fn. |
113 | // Basically, we need to make sure we keep things inside a single loop. |
114 | while (n > 0) { |
115 | if (n >= 8) { |
116 | Sk4px dst0 = fn(Load4(src+0)), |
117 | dst4 = fn(Load4(src+4)); |
118 | dst0.store4(dst+0); |
119 | dst4.store4(dst+4); |
120 | dst += 8; src += 8; n -= 8; |
121 | continue; // Keep our stride at 8 pixels as long as possible. |
122 | } |
123 | SkASSERT(n <= 7); |
124 | if (n >= 4) { |
125 | fn(Load4(src)).store4(dst); |
126 | dst += 4; src += 4; n -= 4; |
127 | } |
128 | if (n >= 2) { |
129 | fn(Load2(src)).store2(dst); |
130 | dst += 2; src += 2; n -= 2; |
131 | } |
132 | if (n >= 1) { |
133 | fn(Load1(src)).store1(dst); |
134 | } |
135 | break; |
136 | } |
137 | } |
138 | |
139 | // As above, but with dst4' = fn(dst4, src4). |
140 | template <typename Fn> |
141 | static void MapDstSrc(int n, SkPMColor* dst, const SkPMColor* src, const Fn& fn) { |
142 | SkASSERT(dst); |
143 | SkASSERT(src); |
144 | while (n > 0) { |
145 | if (n >= 8) { |
146 | Sk4px dst0 = fn(Load4(dst+0), Load4(src+0)), |
147 | dst4 = fn(Load4(dst+4), Load4(src+4)); |
148 | dst0.store4(dst+0); |
149 | dst4.store4(dst+4); |
150 | dst += 8; src += 8; n -= 8; |
151 | continue; // Keep our stride at 8 pixels as long as possible. |
152 | } |
153 | SkASSERT(n <= 7); |
154 | if (n >= 4) { |
155 | fn(Load4(dst), Load4(src)).store4(dst); |
156 | dst += 4; src += 4; n -= 4; |
157 | } |
158 | if (n >= 2) { |
159 | fn(Load2(dst), Load2(src)).store2(dst); |
160 | dst += 2; src += 2; n -= 2; |
161 | } |
162 | if (n >= 1) { |
163 | fn(Load1(dst), Load1(src)).store1(dst); |
164 | } |
165 | break; |
166 | } |
167 | } |
168 | |
169 | // As above, but with dst4' = fn(dst4, alpha4). |
170 | template <typename Fn> |
171 | static void MapDstAlpha(int n, SkPMColor* dst, const SkAlpha* a, const Fn& fn) { |
172 | SkASSERT(dst); |
173 | SkASSERT(a); |
174 | while (n > 0) { |
175 | if (n >= 8) { |
176 | Sk4px dst0 = fn(Load4(dst+0), Load4Alphas(a+0)), |
177 | dst4 = fn(Load4(dst+4), Load4Alphas(a+4)); |
178 | dst0.store4(dst+0); |
179 | dst4.store4(dst+4); |
180 | dst += 8; a += 8; n -= 8; |
181 | continue; // Keep our stride at 8 pixels as long as possible. |
182 | } |
183 | SkASSERT(n <= 7); |
184 | if (n >= 4) { |
185 | fn(Load4(dst), Load4Alphas(a)).store4(dst); |
186 | dst += 4; a += 4; n -= 4; |
187 | } |
188 | if (n >= 2) { |
189 | fn(Load2(dst), Load2Alphas(a)).store2(dst); |
190 | dst += 2; a += 2; n -= 2; |
191 | } |
192 | if (n >= 1) { |
193 | fn(Load1(dst), Sk16b(*a)).store1(dst); |
194 | } |
195 | break; |
196 | } |
197 | } |
198 | |
199 | // As above, but with dst4' = fn(dst4, src4, alpha4). |
200 | template <typename Fn> |
201 | static void MapDstSrcAlpha(int n, SkPMColor* dst, const SkPMColor* src, const SkAlpha* a, |
202 | const Fn& fn) { |
203 | SkASSERT(dst); |
204 | SkASSERT(src); |
205 | SkASSERT(a); |
206 | while (n > 0) { |
207 | if (n >= 8) { |
208 | Sk4px dst0 = fn(Load4(dst+0), Load4(src+0), Load4Alphas(a+0)), |
209 | dst4 = fn(Load4(dst+4), Load4(src+4), Load4Alphas(a+4)); |
210 | dst0.store4(dst+0); |
211 | dst4.store4(dst+4); |
212 | dst += 8; src += 8; a += 8; n -= 8; |
213 | continue; // Keep our stride at 8 pixels as long as possible. |
214 | } |
215 | SkASSERT(n <= 7); |
216 | if (n >= 4) { |
217 | fn(Load4(dst), Load4(src), Load4Alphas(a)).store4(dst); |
218 | dst += 4; src += 4; a += 4; n -= 4; |
219 | } |
220 | if (n >= 2) { |
221 | fn(Load2(dst), Load2(src), Load2Alphas(a)).store2(dst); |
222 | dst += 2; src += 2; a += 2; n -= 2; |
223 | } |
224 | if (n >= 1) { |
225 | fn(Load1(dst), Load1(src), Sk16b(*a)).store1(dst); |
226 | } |
227 | break; |
228 | } |
229 | } |
230 | |
231 | private: |
232 | Sk4px() = default; |
233 | |
234 | typedef Sk16b INHERITED; |
235 | }; |
236 | |
237 | static_assert(sizeof(Sk4px) == sizeof(Sk16b)); |
238 | static_assert(sizeof(Sk4px) == 16); |
239 | |
240 | } // namespace |
241 | |
242 | #ifdef SKNX_NO_SIMD |
243 | #include "src/opts/Sk4px_none.h" |
244 | #else |
245 | #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
246 | #include "src/opts/Sk4px_SSE2.h" |
247 | #elif defined(SK_ARM_HAS_NEON) |
248 | #include "src/opts/Sk4px_NEON.h" |
249 | #else |
250 | #include "src/opts/Sk4px_none.h" |
251 | #endif |
252 | #endif |
253 | |
254 | #endif//Sk4px_DEFINED |
255 | |