SkColor_opts_SSE2.h revision 25f7455f3a7cf2c440509bead85486079f1e4b31 
1/*
2 * Copyright 2014 The Android Open Source Project
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 SkColor_opts_SSE2_DEFINED
9#define SkColor_opts_SSE2_DEFINED
10
11#include <emmintrin.h>
12
13// See #define SkAlphaMulAlpha(a, b)  SkMulDiv255Round(a, b) in SkXfermode.cpp.
14static inline __m128i SkAlphaMulAlpha_SSE2(__m128i a, __m128i b) {
15    __m128i prod = _mm_mullo_epi16(a, b);
16    prod = _mm_add_epi32(prod, _mm_set1_epi32(128));
17    prod = _mm_add_epi32(prod, _mm_srli_epi32(prod, 8));
18    prod = _mm_srli_epi32(prod, 8);
19
20    return prod;
21}
22
23static inline __m128i SkGetPackedA32_SSE2(__m128i src) {
24    __m128i a = _mm_slli_epi32(src, (24 - SK_A32_SHIFT));
25    return _mm_srli_epi32(a, 24);
26}
27
28static inline __m128i SkGetPackedR32_SSE2(__m128i src) {
29    __m128i r = _mm_slli_epi32(src, (24 - SK_R32_SHIFT));
30    return _mm_srli_epi32(r, 24);
31}
32
33static inline __m128i SkGetPackedG32_SSE2(__m128i src) {
34    __m128i g = _mm_slli_epi32(src, (24 - SK_G32_SHIFT));
35    return _mm_srli_epi32(g, 24);
36}
37
38static inline __m128i SkGetPackedB32_SSE2(__m128i src) {
39    __m128i b = _mm_slli_epi32(src, (24 - SK_B32_SHIFT));
40    return _mm_srli_epi32(b, 24);
41}
42
43static inline __m128i SkMul16ShiftRound_SSE2(__m128i a,
44                                             __m128i b, int shift) {
45    __m128i prod = _mm_mullo_epi16(a, b);
46    prod = _mm_add_epi16(prod, _mm_set1_epi16(1 << (shift - 1)));
47    prod = _mm_add_epi16(prod, _mm_srli_epi16(prod, shift));
48    prod = _mm_srli_epi16(prod, shift);
49
50    return prod;
51}
52
53static inline __m128i SkPackRGB16_SSE2(__m128i r, __m128i g, __m128i b) {
54    r = _mm_slli_epi16(r, SK_R16_SHIFT);
55    g = _mm_slli_epi16(g, SK_G16_SHIFT);
56    b = _mm_slli_epi16(b, SK_B16_SHIFT);
57
58    __m128i c = _mm_or_si128(r, g);
59    return _mm_or_si128(c, b);
60}
61
62static inline __m128i SkPackARGB32_SSE2(__m128i a, __m128i r,
63                                        __m128i g, __m128i b) {
64    a = _mm_slli_epi32(a, SK_A32_SHIFT);
65    r = _mm_slli_epi32(r, SK_R32_SHIFT);
66    g = _mm_slli_epi32(g, SK_G32_SHIFT);
67    b = _mm_slli_epi32(b, SK_B32_SHIFT);
68
69    __m128i c = _mm_or_si128(a, r);
70    c = _mm_or_si128(c, g);
71    return _mm_or_si128(c, b);
72}
73
74static inline __m128i SkPacked16ToR32_SSE2(__m128i src) {
75    __m128i r = _mm_srli_epi32(src, SK_R16_SHIFT);
76    r = _mm_and_si128(r, _mm_set1_epi32(SK_R16_MASK));
77    r = _mm_or_si128(_mm_slli_epi32(r, (8 - SK_R16_BITS)),
78                     _mm_srli_epi32(r, (2 * SK_R16_BITS - 8)));
79
80    return r;
81}
82
83static inline __m128i SkPacked16ToG32_SSE2(__m128i src) {
84    __m128i g = _mm_srli_epi32(src, SK_G16_SHIFT);
85    g = _mm_and_si128(g, _mm_set1_epi32(SK_G16_MASK));
86    g = _mm_or_si128(_mm_slli_epi32(g, (8 - SK_G16_BITS)),
87                     _mm_srli_epi32(g, (2 * SK_G16_BITS - 8)));
88
89    return g;
90}
91
92static inline __m128i SkPacked16ToB32_SSE2(__m128i src) {
93    __m128i b = _mm_srli_epi32(src, SK_B16_SHIFT);
94    b = _mm_and_si128(b, _mm_set1_epi32(SK_B16_MASK));
95    b = _mm_or_si128(_mm_slli_epi32(b, (8 - SK_B16_BITS)),
96                     _mm_srli_epi32(b, (2 * SK_B16_BITS - 8)));
97
98    return b;
99}
100
101static inline __m128i SkPixel16ToPixel32_SSE2(__m128i src) {
102    __m128i r = SkPacked16ToR32_SSE2(src);
103    __m128i g = SkPacked16ToG32_SSE2(src);
104    __m128i b = SkPacked16ToB32_SSE2(src);
105
106    return SkPackARGB32_SSE2(_mm_set1_epi32(0xFF), r, g, b);
107}
108
109static inline __m128i SkPixel32ToPixel16_ToU16_SSE2(__m128i src_pixel1,
110                                                    __m128i src_pixel2) {
111    // Calculate result r.
112    __m128i r1 = _mm_srli_epi32(src_pixel1,
113                                SK_R32_SHIFT + (8 - SK_R16_BITS));
114    r1 = _mm_and_si128(r1, _mm_set1_epi32(SK_R16_MASK));
115    __m128i r2 = _mm_srli_epi32(src_pixel2,
116                                SK_R32_SHIFT + (8 - SK_R16_BITS));
117    r2 = _mm_and_si128(r2, _mm_set1_epi32(SK_R16_MASK));
118    __m128i r = _mm_packs_epi32(r1, r2);
119
120    // Calculate result g.
121    __m128i g1 = _mm_srli_epi32(src_pixel1,
122                                SK_G32_SHIFT + (8 - SK_G16_BITS));
123    g1 = _mm_and_si128(g1, _mm_set1_epi32(SK_G16_MASK));
124    __m128i g2 = _mm_srli_epi32(src_pixel2,
125                                SK_G32_SHIFT + (8 - SK_G16_BITS));
126    g2 = _mm_and_si128(g2, _mm_set1_epi32(SK_G16_MASK));
127    __m128i g = _mm_packs_epi32(g1, g2);
128
129    // Calculate result b.
130    __m128i b1 = _mm_srli_epi32(src_pixel1,
131                                SK_B32_SHIFT + (8 - SK_B16_BITS));
132    b1 = _mm_and_si128(b1, _mm_set1_epi32(SK_B16_MASK));
133    __m128i b2 = _mm_srli_epi32(src_pixel2,
134                                SK_B32_SHIFT + (8 - SK_B16_BITS));
135    b2 = _mm_and_si128(b2, _mm_set1_epi32(SK_B16_MASK));
136    __m128i b = _mm_packs_epi32(b1, b2);
137
138    // Store 8 16-bit colors in dst.
139    __m128i d_pixel = SkPackRGB16_SSE2(r, g, b);
140
141    return d_pixel;
142}
143
144#endif // SkColor_opts_SSE2_DEFINED
145