OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2015 Google Inc. | 2 * Copyright 2015 Google Inc. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 namespace { // See Sk4px.h | 8 namespace { // See Sk4px.h |
9 | 9 |
10 inline Sk4px Sk4px::DupPMColor(SkPMColor px) { return Sk16b(_mm_set1_epi32(px));
} | 10 inline Sk4px Sk4px::DupPMColor(SkPMColor px) { return Sk16b(_mm_set1_epi32(px));
} |
(...skipping 83 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
94 | 94 |
95 inline Sk4px Sk4px::zeroColors() const { | 95 inline Sk4px Sk4px::zeroColors() const { |
96 return Sk16b(_mm_and_si128(_mm_set1_epi32(0xFF << SK_A32_SHIFT), this->fVec)
); | 96 return Sk16b(_mm_and_si128(_mm_set1_epi32(0xFF << SK_A32_SHIFT), this->fVec)
); |
97 } | 97 } |
98 | 98 |
99 inline Sk4px Sk4px::zeroAlphas() const { | 99 inline Sk4px Sk4px::zeroAlphas() const { |
100 // andnot(a,b) == ~a & b | 100 // andnot(a,b) == ~a & b |
101 return Sk16b(_mm_andnot_si128(_mm_set1_epi32(0xFF << SK_A32_SHIFT), this->fV
ec)); | 101 return Sk16b(_mm_andnot_si128(_mm_set1_epi32(0xFF << SK_A32_SHIFT), this->fV
ec)); |
102 } | 102 } |
103 | 103 |
104 static inline __m128i widen_low_half_to_8888(__m128i v) { | |
105 // RGB565 format: |R....|G.....|B....| | |
106 // Bit: 16 11 5 0 | |
107 | |
108 // First get each pixel into its own 32-bit lane. | |
109 // v == ____ ____ ____ ____ rgb3 rgb2 rgb1 rgb0 | |
110 // spread == 0000 rgb3 0000 rgb2 0000 rgb1 0000 rgb0 | |
111 auto spread = _mm_unpacklo_epi16(v, _mm_setzero_si128()); | |
112 | |
113 // Get each color independently, still in 565 precison but down at bit 0. | |
114 auto r5 = _mm_srli_epi32(spread, 11), | |
115 g6 = _mm_and_si128(_mm_set1_epi32(63), _mm_srli_epi32(spread, 5)), | |
116 b5 = _mm_and_si128(_mm_set1_epi32(31), spread); | |
117 | |
118 // Scale 565 precision up to 8-bit each, filling low 323 bits with high bits
of each component. | |
119 auto r8 = _mm_or_si128(_mm_slli_epi32(r5, 3), _mm_srli_epi32(r5, 2)), | |
120 g8 = _mm_or_si128(_mm_slli_epi32(g6, 2), _mm_srli_epi32(g6, 4)), | |
121 b8 = _mm_or_si128(_mm_slli_epi32(b5, 3), _mm_srli_epi32(b5, 2)); | |
122 | |
123 // Now put all the 8-bit components into SkPMColor order. | |
124 return _mm_or_si128(_mm_slli_epi32(r8, SK_R32_SHIFT), // TODO: one of thes
e shifts is zero... | |
125 _mm_or_si128(_mm_slli_epi32(g8, SK_G32_SHIFT), | |
126 _mm_or_si128(_mm_slli_epi32(b8, SK_B32_SHIFT), | |
127 _mm_set1_epi32(0xFF << SK_A32_SHIFT)))); | |
128 } | |
129 | |
130 static inline __m128i narrow_to_565(__m128i w) { | |
131 // Extract out top RGB 565 bits of each pixel, with no rounding. | |
132 auto r5 = _mm_and_si128(_mm_set1_epi32(31), _mm_srli_epi32(w, SK_R32_SHIFT +
3)), | |
133 g6 = _mm_and_si128(_mm_set1_epi32(63), _mm_srli_epi32(w, SK_G32_SHIFT +
2)), | |
134 b5 = _mm_and_si128(_mm_set1_epi32(31), _mm_srli_epi32(w, SK_B32_SHIFT +
3)); | |
135 | |
136 // Now put the bits in place in the low 16-bits of each 32-bit lane. | |
137 auto spread = _mm_or_si128(_mm_slli_epi32(r5, 11), | |
138 _mm_or_si128(_mm_slli_epi32(g6, 5), | |
139 b5)); | |
140 | |
141 // We want to pack the bottom 16-bits of spread down into the low half of th
e register, v. | |
142 // spread == 0000 rgb3 0000 rgb2 0000 rgb1 0000 rgb0 | |
143 // v == ____ ____ ____ ____ rgb3 rgb2 rgb1 rgb0 | |
144 | |
145 // Ideally now we'd use _mm_packus_epi32(spread, <anything>) to pack v. But
that's from SSE4. | |
146 // With only SSE2, we need to use _mm_packs_epi32. That does signed saturat
ion, and | |
147 // we need to preserve all 16 bits. So we pretend our data is signed by sig
n-extending first. | |
148 // TODO: is it faster to just _mm_shuffle_epi8 this when we have SSSE3? | |
149 auto signExtended = _mm_srai_epi32(_mm_slli_epi32(spread, 16), 16); | |
150 auto v = _mm_packs_epi32(signExtended, signExtended); | |
151 return v; | |
152 } | |
153 | |
154 inline Sk4px Sk4px::Load4(const SkPMColor16 src[4]) { | |
155 return Sk16b(widen_low_half_to_8888(_mm_loadl_epi64((const __m128i*)src))); | |
156 } | |
157 inline Sk4px Sk4px::Load2(const SkPMColor16 src[2]) { | |
158 auto src2 = ((uint32_t)src[0] ) | |
159 | ((uint32_t)src[1] << 16); | |
160 return Sk16b(widen_low_half_to_8888(_mm_cvtsi32_si128(src2))); | |
161 } | |
162 inline Sk4px Sk4px::Load1(const SkPMColor16 src[1]) { | |
163 return Sk16b(widen_low_half_to_8888(_mm_insert_epi16(_mm_setzero_si128(), sr
c[0], 0))); | |
164 } | |
165 | |
166 inline void Sk4px::store4(SkPMColor16 dst[4]) const { | |
167 _mm_storel_epi64((__m128i*)dst, narrow_to_565(this->fVec)); | |
168 } | |
169 inline void Sk4px::store2(SkPMColor16 dst[2]) const { | |
170 uint32_t dst2 = _mm_cvtsi128_si32(narrow_to_565(this->fVec)); | |
171 dst[0] = dst2; | |
172 dst[1] = dst2 >> 16; | |
173 } | |
174 inline void Sk4px::store1(SkPMColor16 dst[1]) const { | |
175 uint32_t dst2 = _mm_cvtsi128_si32(narrow_to_565(this->fVec)); | |
176 dst[0] = dst2; | |
177 } | |
178 | |
179 } // namespace | 104 } // namespace |
OLD | NEW |