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 #define ASSERT_EQ(a,b) SkASSERT(0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8((a), (b))))
14 
15 // Because no _mm_mul_epi32() in SSE2, we emulate it here.
16 // Multiplies 4 32-bit integers from a by 4 32-bit intergers from b.
17 // The 4 multiplication results should be represented within 32-bit
18 // integers, otherwise they would be overflow.
Multiply32_SSE2(const __m128i & a,const __m128i & b)19 static inline  __m128i Multiply32_SSE2(const __m128i& a, const __m128i& b) {
20     // Calculate results of a0 * b0 and a2 * b2.
21     __m128i r1 = _mm_mul_epu32(a, b);
22     // Calculate results of a1 * b1 and a3 * b3.
23     __m128i r2 = _mm_mul_epu32(_mm_srli_si128(a, 4), _mm_srli_si128(b, 4));
24     // Shuffle results to [63..0] and interleave the results.
25     __m128i r = _mm_unpacklo_epi32(_mm_shuffle_epi32(r1, _MM_SHUFFLE(0,0,2,0)),
26                                    _mm_shuffle_epi32(r2, _MM_SHUFFLE(0,0,2,0)));
27     return r;
28 }
29 
SkAlpha255To256_SSE2(const __m128i & alpha)30 static inline __m128i SkAlpha255To256_SSE2(const __m128i& alpha) {
31     return _mm_add_epi32(alpha, _mm_set1_epi32(1));
32 }
33 
34 // See #define SkAlphaMulAlpha(a, b)  SkMulDiv255Round(a, b) in SkXfermode.cpp.
SkAlphaMulAlpha_SSE2(const __m128i & a,const __m128i & b)35 static inline __m128i SkAlphaMulAlpha_SSE2(const __m128i& a,
36                                            const __m128i& b) {
37     __m128i prod = _mm_mullo_epi16(a, b);
38     prod = _mm_add_epi32(prod, _mm_set1_epi32(128));
39     prod = _mm_add_epi32(prod, _mm_srli_epi32(prod, 8));
40     prod = _mm_srli_epi32(prod, 8);
41 
42     return prod;
43 }
44 
45 // Portable version SkAlphaMulQ is in SkColorPriv.h.
SkAlphaMulQ_SSE2(const __m128i & c,const __m128i & scale)46 static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const __m128i& scale) {
47     const __m128i mask = _mm_set1_epi32(0xFF00FF);
48     __m128i s = _mm_or_si128(_mm_slli_epi32(scale, 16), scale);
49 
50     // uint32_t rb = ((c & mask) * scale) >> 8
51     __m128i rb = _mm_and_si128(mask, c);
52     rb = _mm_mullo_epi16(rb, s);
53     rb = _mm_srli_epi16(rb, 8);
54 
55     // uint32_t ag = ((c >> 8) & mask) * scale
56     __m128i ag = _mm_srli_epi16(c, 8);
57     ASSERT_EQ(ag, _mm_and_si128(mask, ag));  // ag = _mm_srli_epi16(c, 8) did this for us.
58     ag = _mm_mullo_epi16(ag, s);
59 
60     // (rb & mask) | (ag & ~mask)
61     ASSERT_EQ(rb, _mm_and_si128(mask, rb));  // rb = _mm_srli_epi16(rb, 8) did this for us.
62     ag = _mm_andnot_si128(mask, ag);
63     return _mm_or_si128(rb, ag);
64 }
65 
66 // Fast path for SkAlphaMulQ_SSE2 with a constant scale factor.
SkAlphaMulQ_SSE2(const __m128i & c,const unsigned scale)67 static inline __m128i SkAlphaMulQ_SSE2(const __m128i& c, const unsigned scale) {
68     const __m128i mask = _mm_set1_epi32(0xFF00FF);
69     __m128i s = _mm_set1_epi16(scale << 8); // Move scale factor to upper byte of word.
70 
71     // With mulhi, red and blue values are already in the right place and
72     // don't need to be divided by 256.
73     __m128i rb = _mm_and_si128(mask, c);
74     rb = _mm_mulhi_epu16(rb, s);
75 
76     __m128i ag = _mm_andnot_si128(mask, c);
77     ag = _mm_mulhi_epu16(ag, s);     // Alpha and green values are in the higher byte of each word.
78     ag = _mm_andnot_si128(mask, ag);
79 
80     return _mm_or_si128(rb, ag);
81 }
82 
SkGetPackedA32_SSE2(const __m128i & src)83 static inline __m128i SkGetPackedA32_SSE2(const __m128i& src) {
84 #if SK_A32_SHIFT == 24                // It's very common (universal?) that alpha is the top byte.
85     return _mm_srli_epi32(src, 24);   // You'd hope the compiler would remove the left shift then,
86 #else                                 // but I've seen Clang just do a dumb left shift of zero. :(
87     __m128i a = _mm_slli_epi32(src, (24 - SK_A32_SHIFT));
88     return _mm_srli_epi32(a, 24);
89 #endif
90 }
91 
SkGetPackedR32_SSE2(const __m128i & src)92 static inline __m128i SkGetPackedR32_SSE2(const __m128i& src) {
93     __m128i r = _mm_slli_epi32(src, (24 - SK_R32_SHIFT));
94     return _mm_srli_epi32(r, 24);
95 }
96 
SkGetPackedG32_SSE2(const __m128i & src)97 static inline __m128i SkGetPackedG32_SSE2(const __m128i& src) {
98     __m128i g = _mm_slli_epi32(src, (24 - SK_G32_SHIFT));
99     return _mm_srli_epi32(g, 24);
100 }
101 
SkGetPackedB32_SSE2(const __m128i & src)102 static inline __m128i SkGetPackedB32_SSE2(const __m128i& src) {
103     __m128i b = _mm_slli_epi32(src, (24 - SK_B32_SHIFT));
104     return _mm_srli_epi32(b, 24);
105 }
106 
SkMul16ShiftRound_SSE2(const __m128i & a,const __m128i & b,int shift)107 static inline __m128i SkMul16ShiftRound_SSE2(const __m128i& a,
108                                              const __m128i& b, int shift) {
109     __m128i prod = _mm_mullo_epi16(a, b);
110     prod = _mm_add_epi16(prod, _mm_set1_epi16(1 << (shift - 1)));
111     prod = _mm_add_epi16(prod, _mm_srli_epi16(prod, shift));
112     prod = _mm_srli_epi16(prod, shift);
113 
114     return prod;
115 }
116 
SkPackRGB16_SSE2(const __m128i & r,const __m128i & g,const __m128i & b)117 static inline __m128i SkPackRGB16_SSE2(const __m128i& r,
118                                        const __m128i& g, const __m128i& b) {
119     __m128i dr = _mm_slli_epi16(r, SK_R16_SHIFT);
120     __m128i dg = _mm_slli_epi16(g, SK_G16_SHIFT);
121     __m128i db = _mm_slli_epi16(b, SK_B16_SHIFT);
122 
123     __m128i c = _mm_or_si128(dr, dg);
124     return _mm_or_si128(c, db);
125 }
126 
SkPackARGB32_SSE2(const __m128i & a,const __m128i & r,const __m128i & g,const __m128i & b)127 static inline __m128i SkPackARGB32_SSE2(const __m128i& a, const __m128i& r,
128                                         const __m128i& g, const __m128i& b) {
129     __m128i da = _mm_slli_epi32(a, SK_A32_SHIFT);
130     __m128i dr = _mm_slli_epi32(r, SK_R32_SHIFT);
131     __m128i dg = _mm_slli_epi32(g, SK_G32_SHIFT);
132     __m128i db = _mm_slli_epi32(b, SK_B32_SHIFT);
133 
134     __m128i c = _mm_or_si128(da, dr);
135     c = _mm_or_si128(c, dg);
136     return _mm_or_si128(c, db);
137 }
138 
SkPacked16ToR32_SSE2(const __m128i & src)139 static inline __m128i SkPacked16ToR32_SSE2(const __m128i& src) {
140     __m128i r = _mm_srli_epi32(src, SK_R16_SHIFT);
141     r = _mm_and_si128(r, _mm_set1_epi32(SK_R16_MASK));
142     r = _mm_or_si128(_mm_slli_epi32(r, (8 - SK_R16_BITS)),
143                      _mm_srli_epi32(r, (2 * SK_R16_BITS - 8)));
144 
145     return r;
146 }
147 
SkPacked16ToG32_SSE2(const __m128i & src)148 static inline __m128i SkPacked16ToG32_SSE2(const __m128i& src) {
149     __m128i g = _mm_srli_epi32(src, SK_G16_SHIFT);
150     g = _mm_and_si128(g, _mm_set1_epi32(SK_G16_MASK));
151     g = _mm_or_si128(_mm_slli_epi32(g, (8 - SK_G16_BITS)),
152                      _mm_srli_epi32(g, (2 * SK_G16_BITS - 8)));
153 
154     return g;
155 }
156 
SkPacked16ToB32_SSE2(const __m128i & src)157 static inline __m128i SkPacked16ToB32_SSE2(const __m128i& src) {
158     __m128i b = _mm_srli_epi32(src, SK_B16_SHIFT);
159     b = _mm_and_si128(b, _mm_set1_epi32(SK_B16_MASK));
160     b = _mm_or_si128(_mm_slli_epi32(b, (8 - SK_B16_BITS)),
161                      _mm_srli_epi32(b, (2 * SK_B16_BITS - 8)));
162 
163     return b;
164 }
165 
SkPixel16ToPixel32_SSE2(const __m128i & src)166 static inline __m128i SkPixel16ToPixel32_SSE2(const __m128i& src) {
167     __m128i r = SkPacked16ToR32_SSE2(src);
168     __m128i g = SkPacked16ToG32_SSE2(src);
169     __m128i b = SkPacked16ToB32_SSE2(src);
170 
171     return SkPackARGB32_SSE2(_mm_set1_epi32(0xFF), r, g, b);
172 }
173 
SkPixel32ToPixel16_ToU16_SSE2(const __m128i & src_pixel1,const __m128i & src_pixel2)174 static inline __m128i SkPixel32ToPixel16_ToU16_SSE2(const __m128i& src_pixel1,
175                                                     const __m128i& src_pixel2) {
176     // Calculate result r.
177     __m128i r1 = _mm_srli_epi32(src_pixel1,
178                                 SK_R32_SHIFT + (8 - SK_R16_BITS));
179     r1 = _mm_and_si128(r1, _mm_set1_epi32(SK_R16_MASK));
180     __m128i r2 = _mm_srli_epi32(src_pixel2,
181                                 SK_R32_SHIFT + (8 - SK_R16_BITS));
182     r2 = _mm_and_si128(r2, _mm_set1_epi32(SK_R16_MASK));
183     __m128i r = _mm_packs_epi32(r1, r2);
184 
185     // Calculate result g.
186     __m128i g1 = _mm_srli_epi32(src_pixel1,
187                                 SK_G32_SHIFT + (8 - SK_G16_BITS));
188     g1 = _mm_and_si128(g1, _mm_set1_epi32(SK_G16_MASK));
189     __m128i g2 = _mm_srli_epi32(src_pixel2,
190                                 SK_G32_SHIFT + (8 - SK_G16_BITS));
191     g2 = _mm_and_si128(g2, _mm_set1_epi32(SK_G16_MASK));
192     __m128i g = _mm_packs_epi32(g1, g2);
193 
194     // Calculate result b.
195     __m128i b1 = _mm_srli_epi32(src_pixel1,
196                                 SK_B32_SHIFT + (8 - SK_B16_BITS));
197     b1 = _mm_and_si128(b1, _mm_set1_epi32(SK_B16_MASK));
198     __m128i b2 = _mm_srli_epi32(src_pixel2,
199                                 SK_B32_SHIFT + (8 - SK_B16_BITS));
200     b2 = _mm_and_si128(b2, _mm_set1_epi32(SK_B16_MASK));
201     __m128i b = _mm_packs_epi32(b1, b2);
202 
203     // Store 8 16-bit colors in dst.
204     __m128i d_pixel = SkPackRGB16_SSE2(r, g, b);
205 
206     return d_pixel;
207 }
208 
209 // Portable version is SkPMSrcOver in SkColorPriv.h.
SkPMSrcOver_SSE2(const __m128i & src,const __m128i & dst)210 static inline __m128i SkPMSrcOver_SSE2(const __m128i& src, const __m128i& dst) {
211     return _mm_add_epi32(src,
212                          SkAlphaMulQ_SSE2(dst, _mm_sub_epi32(_mm_set1_epi32(256),
213                                                              SkGetPackedA32_SSE2(src))));
214 }
215 
216 // Portable version is SkBlendARGB32 in SkColorPriv.h.
SkBlendARGB32_SSE2(const __m128i & src,const __m128i & dst,const __m128i & aa)217 static inline __m128i SkBlendARGB32_SSE2(const __m128i& src, const __m128i& dst,
218                                          const __m128i& aa) {
219     __m128i src_scale = SkAlpha255To256_SSE2(aa);
220     // SkAlpha255To256(255 - SkAlphaMul(SkGetPackedA32(src), src_scale))
221     __m128i dst_scale = SkGetPackedA32_SSE2(src);
222     dst_scale = _mm_mullo_epi16(dst_scale, src_scale);
223     dst_scale = _mm_srli_epi16(dst_scale, 8);
224     dst_scale = _mm_sub_epi32(_mm_set1_epi32(256), dst_scale);
225 
226     __m128i result = SkAlphaMulQ_SSE2(src, src_scale);
227     return _mm_add_epi8(result, SkAlphaMulQ_SSE2(dst, dst_scale));
228 }
229 
230 // Fast path for SkBlendARGB32_SSE2 with a constant alpha factor.
SkBlendARGB32_SSE2(const __m128i & src,const __m128i & dst,const unsigned aa)231 static inline __m128i SkBlendARGB32_SSE2(const __m128i& src, const __m128i& dst,
232                                          const unsigned aa) {
233     unsigned alpha = SkAlpha255To256(aa);
234     __m128i src_scale = _mm_set1_epi32(alpha);
235     // SkAlpha255To256(255 - SkAlphaMul(SkGetPackedA32(src), src_scale))
236     __m128i dst_scale = SkGetPackedA32_SSE2(src);
237     dst_scale = _mm_mullo_epi16(dst_scale, src_scale);
238     dst_scale = _mm_srli_epi16(dst_scale, 8);
239     dst_scale = _mm_sub_epi32(_mm_set1_epi32(256), dst_scale);
240 
241     __m128i result = SkAlphaMulQ_SSE2(src, alpha);
242     return _mm_add_epi8(result, SkAlphaMulQ_SSE2(dst, dst_scale));
243 }
244 
245 #undef ASSERT_EQ
246 #endif // SkColor_opts_SSE2_DEFINED
247