OLD | NEW |
1 /* | 1 /* |
2 * Copyright 2011 Google Inc. | 2 * Copyright 2011 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 #include "SkBlitRow.h" | 8 #include "SkBlitRow.h" |
9 #include "SkBlitMask.h" | 9 #include "SkBlitMask.h" |
10 #include "SkColorPriv.h" | 10 #include "SkColorPriv.h" |
(...skipping 113 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
124 flags &= kFlags32_Mask; | 124 flags &= kFlags32_Mask; |
125 | 125 |
126 SkBlitRow::Proc32 proc = PlatformProcs32(flags); | 126 SkBlitRow::Proc32 proc = PlatformProcs32(flags); |
127 if (NULL == proc) { | 127 if (NULL == proc) { |
128 proc = gDefault_Procs32[flags]; | 128 proc = gDefault_Procs32[flags]; |
129 } | 129 } |
130 SkASSERT(proc); | 130 SkASSERT(proc); |
131 return proc; | 131 return proc; |
132 } | 132 } |
133 | 133 |
134 // Color32 uses the blend_256_round_alt algorithm from tests/BlendTest.cpp. | 134 SkBlitRow::Proc32 SkBlitRow::ColorProcFactory() { |
135 // It's not quite perfect, but it's never wrong in the interesting edge cases, | 135 SkBlitRow::ColorProc proc = PlatformColorProc(); |
136 // and it's quite a bit faster than blend_perfect. | 136 if (NULL == proc) { |
| 137 proc = Color32; |
| 138 } |
| 139 SkASSERT(proc); |
| 140 return proc; |
| 141 } |
| 142 |
| 143 #define SK_SUPPORT_LEGACY_COLOR32_MATHx |
| 144 |
| 145 // Color32 and its SIMD specializations use the blend_256_round_alt algorithm |
| 146 // from tests/BlendTest.cpp. It's not quite perfect, but it's never wrong in th
e |
| 147 // interesting edge cases, and it's quite a bit faster than blend_perfect. |
137 // | 148 // |
138 // blend_256_round_alt is our currently blessed algorithm. Please use it or an
analogous one. | 149 // blend_256_round_alt is our currently blessed algorithm. Please use it or an
analogous one. |
139 void SkBlitRow::Color32(SkPMColor dst[], const SkPMColor src[], int count, SkPMC
olor color) { | 150 void SkBlitRow::Color32(SkPMColor* SK_RESTRICT dst, |
| 151 const SkPMColor* SK_RESTRICT src, |
| 152 int count, SkPMColor color) { |
140 switch (SkGetPackedA32(color)) { | 153 switch (SkGetPackedA32(color)) { |
141 case 0: memmove(dst, src, count * sizeof(SkPMColor)); return; | 154 case 0: memmove(dst, src, count * sizeof(SkPMColor)); return; |
142 case 255: sk_memset32(dst, color, count); return; | 155 case 255: sk_memset32(dst, color, count); return; |
143 } | 156 } |
144 | 157 |
145 unsigned invA = 255 - SkGetPackedA32(color); | 158 unsigned invA = 255 - SkGetPackedA32(color); |
| 159 #ifdef SK_SUPPORT_LEGACY_COLOR32_MATH // blend_256_plus1_trunc, busted |
| 160 unsigned round = 0; |
| 161 #else // blend_256_round_alt, good |
146 invA += invA >> 7; | 162 invA += invA >> 7; |
147 SkASSERT(invA < 256); // We've already handled alpha == 0 above. | |
148 | |
149 #if defined(SK_ARM_HAS_NEON) | |
150 uint16x8_t colorHigh = vshll_n_u8((uint8x8_t)vdup_n_u32(color), 8); | |
151 uint16x8_t colorAndRound = vaddq_u16(colorHigh, vdupq_n_u16(128)); | |
152 uint8x8_t invA8 = vdup_n_u8(invA); | |
153 | |
154 // Does the core work of blending color onto 4 pixels, returning the resulti
ng 4 pixels. | |
155 auto kernel = [&](const uint32x4_t& src4) -> uint32x4_t { | |
156 uint16x8_t lo = vmull_u8(vget_low_u8( (uint8x16_t)src4), invA8), | |
157 hi = vmull_u8(vget_high_u8((uint8x16_t)src4), invA8); | |
158 return (uint32x4_t) | |
159 vcombine_u8(vaddhn_u16(colorAndRound, lo), vaddhn_u16(colorAndRound,
hi)); | |
160 }; | |
161 | |
162 while (count >= 8) { | |
163 uint32x4_t dst0 = kernel(vld1q_u32(src+0)), | |
164 dst4 = kernel(vld1q_u32(src+4)); | |
165 vst1q_u32(dst+0, dst0); | |
166 vst1q_u32(dst+4, dst4); | |
167 src += 8; | |
168 dst += 8; | |
169 count -= 8; | |
170 } | |
171 if (count >= 4) { | |
172 vst1q_u32(dst, kernel(vld1q_u32(src))); | |
173 src += 4; | |
174 dst += 4; | |
175 count -= 4; | |
176 } | |
177 if (count >= 2) { | |
178 uint32x2_t src2 = vld1_u32(src); | |
179 vst1_u32(dst, vget_low_u32(kernel(vcombine_u32(src2, src2)))); | |
180 src += 2; | |
181 dst += 2; | |
182 count -= 2; | |
183 } | |
184 if (count >= 1) { | |
185 vst1q_lane_u32(dst, kernel(vdupq_n_u32(*src)), 0); | |
186 } | |
187 | |
188 #elif SK_CPU_SSE_LEVEL >= SK_CPU_LEVEL_SSE2 | |
189 __m128i colorHigh = _mm_unpacklo_epi8(_mm_setzero_si128(), _mm_set1_epi32(co
lor)); | |
190 __m128i colorAndRound = _mm_add_epi16(colorHigh, _mm_set1_epi16(128)); | |
191 __m128i invA16 = _mm_set1_epi16(invA); | |
192 | |
193 // Does the core work of blending color onto 4 pixels, returning the resulti
ng 4 pixels. | |
194 auto kernel = [&](const __m128i& src4) -> __m128i { | |
195 __m128i lo = _mm_mullo_epi16(invA16, _mm_unpacklo_epi8(src4, _mm_setzero
_si128())), | |
196 hi = _mm_mullo_epi16(invA16, _mm_unpackhi_epi8(src4, _mm_setzero
_si128())); | |
197 return _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(colorAndRound, lo),
8), | |
198 _mm_srli_epi16(_mm_add_epi16(colorAndRound, hi),
8)); | |
199 }; | |
200 | |
201 while (count >= 8) { | |
202 __m128i dst0 = kernel(_mm_loadu_si128((const __m128i*)(src+0))), | |
203 dst4 = kernel(_mm_loadu_si128((const __m128i*)(src+4))); | |
204 _mm_storeu_si128((__m128i*)(dst+0), dst0); | |
205 _mm_storeu_si128((__m128i*)(dst+4), dst4); | |
206 src += 8; | |
207 dst += 8; | |
208 count -= 8; | |
209 } | |
210 if (count >= 4) { | |
211 _mm_storeu_si128((__m128i*)dst, kernel(_mm_loadu_si128((const __m128i*)s
rc))); | |
212 src += 4; | |
213 dst += 4; | |
214 count -= 4; | |
215 } | |
216 if (count >= 2) { | |
217 _mm_storel_epi64((__m128i*)dst, kernel(_mm_loadl_epi64((const __m128i*)s
rc))); | |
218 src += 2; | |
219 dst += 2; | |
220 count -= 2; | |
221 } | |
222 if (count >= 1) { | |
223 *dst = _mm_cvtsi128_si32(kernel(_mm_cvtsi32_si128(*src))); | |
224 } | |
225 #else // Neither NEON nor SSE2. | |
226 unsigned round = (128 << 16) + (128 << 0); | 163 unsigned round = (128 << 16) + (128 << 0); |
| 164 #endif |
227 | 165 |
228 while (count --> 0) { | 166 while (count --> 0) { |
229 // Our math is 16-bit, so we can do a little bit of SIMD in 32-bit regis
ters. | 167 // Our math is 16-bit, so we can do a little bit of SIMD in 32-bit regis
ters. |
230 const uint32_t mask = 0x00FF00FF; | 168 const uint32_t mask = 0x00FF00FF; |
231 uint32_t rb = (((*src >> 0) & mask) * invA + round) >> 8, // _r_b | 169 uint32_t rb = (((*src >> 0) & mask) * invA + round) >> 8, // _r_b |
232 ag = (((*src >> 8) & mask) * invA + round) >> 0; // a_g_ | 170 ag = (((*src >> 8) & mask) * invA + round) >> 0; // a_g_ |
233 *dst = color + ((rb & mask) | (ag & ~mask)); | 171 *dst = color + ((rb & mask) | (ag & ~mask)); |
234 src++; | 172 src++; |
235 dst++; | 173 dst++; |
236 } | 174 } |
237 #endif | |
238 } | 175 } |
239 | 176 |
OLD | NEW |