epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 1 | |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 2 | /* |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 3 | * Copyright 2009 The Android Open Source Project |
| 4 | * |
| 5 | * Use of this source code is governed by a BSD-style license that can be |
| 6 | * found in the LICENSE file. |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 7 | */ |
| 8 | |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 9 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 10 | #include "SkBlitRow_opts_SSE2.h" |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 11 | #include "SkColorPriv.h" |
senorblanco@chromium.org | c385638 | 2010-12-13 15:27:20 +0000 | [diff] [blame] | 12 | #include "SkUtils.h" |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 13 | |
| 14 | #include <emmintrin.h> |
| 15 | |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 16 | /* SSE2 version of S32_Blend_BlitRow32() |
| 17 | * portable version is in core/SkBlitRow_D32.cpp |
| 18 | */ |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 19 | void S32_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, |
| 20 | const SkPMColor* SK_RESTRICT src, |
| 21 | int count, U8CPU alpha) { |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 22 | SkASSERT(alpha <= 255); |
| 23 | if (count <= 0) { |
| 24 | return; |
| 25 | } |
| 26 | |
| 27 | uint32_t src_scale = SkAlpha255To256(alpha); |
| 28 | uint32_t dst_scale = 256 - src_scale; |
| 29 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 30 | if (count >= 4) { |
| 31 | SkASSERT(((size_t)dst & 0x03) == 0); |
| 32 | while (((size_t)dst & 0x0F) != 0) { |
| 33 | *dst = SkAlphaMulQ(*src, src_scale) + SkAlphaMulQ(*dst, dst_scale); |
| 34 | src++; |
| 35 | dst++; |
| 36 | count--; |
| 37 | } |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 38 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 39 | const __m128i *s = reinterpret_cast<const __m128i*>(src); |
| 40 | __m128i *d = reinterpret_cast<__m128i*>(dst); |
| 41 | __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
| 42 | __m128i src_scale_wide = _mm_set1_epi16(src_scale); |
| 43 | __m128i dst_scale_wide = _mm_set1_epi16(dst_scale); |
| 44 | while (count >= 4) { |
| 45 | // Load 4 pixels each of src and dest. |
| 46 | __m128i src_pixel = _mm_loadu_si128(s); |
| 47 | __m128i dst_pixel = _mm_load_si128(d); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 48 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 49 | // Get red and blue pixels into lower byte of each word. |
| 50 | __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); |
| 51 | __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 52 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 53 | // Get alpha and green into lower byte of each word. |
| 54 | __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); |
| 55 | __m128i src_ag = _mm_srli_epi16(src_pixel, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 56 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 57 | // Multiply by scale. |
| 58 | src_rb = _mm_mullo_epi16(src_rb, src_scale_wide); |
| 59 | src_ag = _mm_mullo_epi16(src_ag, src_scale_wide); |
| 60 | dst_rb = _mm_mullo_epi16(dst_rb, dst_scale_wide); |
| 61 | dst_ag = _mm_mullo_epi16(dst_ag, dst_scale_wide); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 62 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 63 | // Divide by 256. |
| 64 | src_rb = _mm_srli_epi16(src_rb, 8); |
| 65 | dst_rb = _mm_srli_epi16(dst_rb, 8); |
| 66 | src_ag = _mm_andnot_si128(rb_mask, src_ag); |
| 67 | dst_ag = _mm_andnot_si128(rb_mask, dst_ag); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 68 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 69 | // Combine back into RGBA. |
| 70 | src_pixel = _mm_or_si128(src_rb, src_ag); |
| 71 | dst_pixel = _mm_or_si128(dst_rb, dst_ag); |
| 72 | |
| 73 | // Add result |
| 74 | __m128i result = _mm_add_epi8(src_pixel, dst_pixel); |
| 75 | _mm_store_si128(d, result); |
| 76 | s++; |
| 77 | d++; |
| 78 | count -= 4; |
| 79 | } |
| 80 | src = reinterpret_cast<const SkPMColor*>(s); |
| 81 | dst = reinterpret_cast<SkPMColor*>(d); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 82 | } |
| 83 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 84 | while (count > 0) { |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 85 | *dst = SkAlphaMulQ(*src, src_scale) + SkAlphaMulQ(*dst, dst_scale); |
| 86 | src++; |
| 87 | dst++; |
| 88 | count--; |
| 89 | } |
| 90 | } |
| 91 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 92 | void S32A_Opaque_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, |
| 93 | const SkPMColor* SK_RESTRICT src, |
| 94 | int count, U8CPU alpha) { |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 95 | SkASSERT(alpha == 255); |
| 96 | if (count <= 0) { |
| 97 | return; |
| 98 | } |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 99 | |
| 100 | if (count >= 4) { |
| 101 | SkASSERT(((size_t)dst & 0x03) == 0); |
| 102 | while (((size_t)dst & 0x0F) != 0) { |
| 103 | *dst = SkPMSrcOver(*src, *dst); |
| 104 | src++; |
| 105 | dst++; |
| 106 | count--; |
| 107 | } |
| 108 | |
| 109 | const __m128i *s = reinterpret_cast<const __m128i*>(src); |
| 110 | __m128i *d = reinterpret_cast<__m128i*>(dst); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 111 | #ifdef SK_USE_ACCURATE_BLENDING |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 112 | __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
| 113 | __m128i c_128 = _mm_set1_epi16(128); // 8 copies of 128 (16-bit) |
| 114 | __m128i c_255 = _mm_set1_epi16(255); // 8 copies of 255 (16-bit) |
| 115 | while (count >= 4) { |
| 116 | // Load 4 pixels |
| 117 | __m128i src_pixel = _mm_loadu_si128(s); |
| 118 | __m128i dst_pixel = _mm_load_si128(d); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 119 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 120 | __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 121 | __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 122 | // Shift alphas down to lower 8 bits of each quad. |
| 123 | __m128i alpha = _mm_srli_epi32(src_pixel, 24); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 124 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 125 | // Copy alpha to upper 3rd byte of each quad |
| 126 | alpha = _mm_or_si128(alpha, _mm_slli_epi32(alpha, 16)); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 127 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 128 | // Subtract alphas from 255, to get 0..255 |
| 129 | alpha = _mm_sub_epi16(c_255, alpha); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 130 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 131 | // Multiply by red and blue by src alpha. |
| 132 | dst_rb = _mm_mullo_epi16(dst_rb, alpha); |
| 133 | // Multiply by alpha and green by src alpha. |
| 134 | dst_ag = _mm_mullo_epi16(dst_ag, alpha); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 135 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 136 | // dst_rb_low = (dst_rb >> 8) |
| 137 | __m128i dst_rb_low = _mm_srli_epi16(dst_rb, 8); |
| 138 | __m128i dst_ag_low = _mm_srli_epi16(dst_ag, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 139 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 140 | // dst_rb = (dst_rb + dst_rb_low + 128) >> 8 |
| 141 | dst_rb = _mm_add_epi16(dst_rb, dst_rb_low); |
| 142 | dst_rb = _mm_add_epi16(dst_rb, c_128); |
| 143 | dst_rb = _mm_srli_epi16(dst_rb, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 144 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 145 | // dst_ag = (dst_ag + dst_ag_low + 128) & ag_mask |
| 146 | dst_ag = _mm_add_epi16(dst_ag, dst_ag_low); |
| 147 | dst_ag = _mm_add_epi16(dst_ag, c_128); |
| 148 | dst_ag = _mm_andnot_si128(rb_mask, dst_ag); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 149 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 150 | // Combine back into RGBA. |
| 151 | dst_pixel = _mm_or_si128(dst_rb, dst_ag); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 152 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 153 | // Add result |
| 154 | __m128i result = _mm_add_epi8(src_pixel, dst_pixel); |
| 155 | _mm_store_si128(d, result); |
| 156 | s++; |
| 157 | d++; |
| 158 | count -= 4; |
| 159 | } |
| 160 | #else |
| 161 | __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
| 162 | __m128i c_256 = _mm_set1_epi16(0x0100); // 8 copies of 256 (16-bit) |
| 163 | while (count >= 4) { |
| 164 | // Load 4 pixels |
| 165 | __m128i src_pixel = _mm_loadu_si128(s); |
| 166 | __m128i dst_pixel = _mm_load_si128(d); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 167 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 168 | __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 169 | __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 170 | |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 171 | // (a0, g0, a1, g1, a2, g2, a3, g3) (low byte of each word) |
| 172 | __m128i alpha = _mm_srli_epi16(src_pixel, 8); |
| 173 | |
| 174 | // (a0, a0, a1, a1, a2, g2, a3, g3) |
| 175 | alpha = _mm_shufflehi_epi16(alpha, 0xF5); |
| 176 | |
| 177 | // (a0, a0, a1, a1, a2, a2, a3, a3) |
| 178 | alpha = _mm_shufflelo_epi16(alpha, 0xF5); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 179 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 180 | // Subtract alphas from 256, to get 1..256 |
| 181 | alpha = _mm_sub_epi16(c_256, alpha); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 182 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 183 | // Multiply by red and blue by src alpha. |
| 184 | dst_rb = _mm_mullo_epi16(dst_rb, alpha); |
| 185 | // Multiply by alpha and green by src alpha. |
| 186 | dst_ag = _mm_mullo_epi16(dst_ag, alpha); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 187 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 188 | // Divide by 256. |
| 189 | dst_rb = _mm_srli_epi16(dst_rb, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 190 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 191 | // Mask out high bits (already in the right place) |
| 192 | dst_ag = _mm_andnot_si128(rb_mask, dst_ag); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 193 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 194 | // Combine back into RGBA. |
| 195 | dst_pixel = _mm_or_si128(dst_rb, dst_ag); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 196 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 197 | // Add result |
| 198 | __m128i result = _mm_add_epi8(src_pixel, dst_pixel); |
| 199 | _mm_store_si128(d, result); |
| 200 | s++; |
| 201 | d++; |
| 202 | count -= 4; |
| 203 | } |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 204 | #endif |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 205 | src = reinterpret_cast<const SkPMColor*>(s); |
| 206 | dst = reinterpret_cast<SkPMColor*>(d); |
| 207 | } |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 208 | |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 209 | while (count > 0) { |
| 210 | *dst = SkPMSrcOver(*src, *dst); |
| 211 | src++; |
| 212 | dst++; |
| 213 | count--; |
| 214 | } |
| 215 | } |
| 216 | |
senorblanco@chromium.org | 4e75355 | 2009-11-16 21:09:00 +0000 | [diff] [blame] | 217 | void S32A_Blend_BlitRow32_SSE2(SkPMColor* SK_RESTRICT dst, |
| 218 | const SkPMColor* SK_RESTRICT src, |
| 219 | int count, U8CPU alpha) { |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 220 | SkASSERT(alpha <= 255); |
| 221 | if (count <= 0) { |
| 222 | return; |
| 223 | } |
| 224 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 225 | if (count >= 4) { |
| 226 | while (((size_t)dst & 0x0F) != 0) { |
| 227 | *dst = SkBlendARGB32(*src, *dst, alpha); |
| 228 | src++; |
| 229 | dst++; |
| 230 | count--; |
| 231 | } |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 232 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 233 | uint32_t src_scale = SkAlpha255To256(alpha); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 234 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 235 | const __m128i *s = reinterpret_cast<const __m128i*>(src); |
| 236 | __m128i *d = reinterpret_cast<__m128i*>(dst); |
| 237 | __m128i src_scale_wide = _mm_set1_epi16(src_scale); |
| 238 | __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
| 239 | __m128i c_256 = _mm_set1_epi16(256); // 8 copies of 256 (16-bit) |
| 240 | while (count >= 4) { |
| 241 | // Load 4 pixels each of src and dest. |
| 242 | __m128i src_pixel = _mm_loadu_si128(s); |
| 243 | __m128i dst_pixel = _mm_load_si128(d); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 244 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 245 | // Get red and blue pixels into lower byte of each word. |
| 246 | __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); |
| 247 | __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 248 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 249 | // Get alpha and green into lower byte of each word. |
| 250 | __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); |
| 251 | __m128i src_ag = _mm_srli_epi16(src_pixel, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 252 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 253 | // Put per-pixel alpha in low byte of each word. |
| 254 | __m128i dst_alpha = _mm_shufflehi_epi16(src_ag, 0xF5); |
| 255 | dst_alpha = _mm_shufflelo_epi16(dst_alpha, 0xF5); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 256 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 257 | // dst_alpha = dst_alpha * src_scale |
| 258 | dst_alpha = _mm_mullo_epi16(dst_alpha, src_scale_wide); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 259 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 260 | // Divide by 256. |
| 261 | dst_alpha = _mm_srli_epi16(dst_alpha, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 262 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 263 | // Subtract alphas from 256, to get 1..256 |
| 264 | dst_alpha = _mm_sub_epi16(c_256, dst_alpha); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 265 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 266 | // Multiply red and blue by dst pixel alpha. |
| 267 | dst_rb = _mm_mullo_epi16(dst_rb, dst_alpha); |
| 268 | // Multiply alpha and green by dst pixel alpha. |
| 269 | dst_ag = _mm_mullo_epi16(dst_ag, dst_alpha); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 270 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 271 | // Multiply red and blue by global alpha. |
| 272 | src_rb = _mm_mullo_epi16(src_rb, src_scale_wide); |
| 273 | // Multiply alpha and green by global alpha. |
| 274 | src_ag = _mm_mullo_epi16(src_ag, src_scale_wide); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 275 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 276 | // Divide by 256. |
| 277 | dst_rb = _mm_srli_epi16(dst_rb, 8); |
| 278 | src_rb = _mm_srli_epi16(src_rb, 8); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 279 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 280 | // Mask out low bits (goodies already in the right place; no need to divide) |
| 281 | dst_ag = _mm_andnot_si128(rb_mask, dst_ag); |
| 282 | src_ag = _mm_andnot_si128(rb_mask, src_ag); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 283 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 284 | // Combine back into RGBA. |
| 285 | dst_pixel = _mm_or_si128(dst_rb, dst_ag); |
| 286 | src_pixel = _mm_or_si128(src_rb, src_ag); |
| 287 | |
| 288 | // Add two pixels into result. |
| 289 | __m128i result = _mm_add_epi8(src_pixel, dst_pixel); |
| 290 | _mm_store_si128(d, result); |
| 291 | s++; |
| 292 | d++; |
| 293 | count -= 4; |
| 294 | } |
| 295 | src = reinterpret_cast<const SkPMColor*>(s); |
| 296 | dst = reinterpret_cast<SkPMColor*>(d); |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 297 | } |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 298 | |
senorblanco@chromium.org | 9272761 | 2009-11-04 20:51:06 +0000 | [diff] [blame] | 299 | while (count > 0) { |
| 300 | *dst = SkBlendARGB32(*src, *dst, alpha); |
| 301 | src++; |
| 302 | dst++; |
| 303 | count--; |
| 304 | } |
| 305 | } |
senorblanco@chromium.org | c385638 | 2010-12-13 15:27:20 +0000 | [diff] [blame] | 306 | |
| 307 | /* SSE2 version of Color32() |
| 308 | * portable version is in core/SkBlitRow_D32.cpp |
| 309 | */ |
| 310 | void Color32_SSE2(SkPMColor dst[], const SkPMColor src[], int count, |
| 311 | SkPMColor color) { |
| 312 | |
| 313 | if (count <= 0) { |
| 314 | return; |
| 315 | } |
| 316 | |
| 317 | if (0 == color) { |
| 318 | if (src != dst) { |
| 319 | memcpy(dst, src, count * sizeof(SkPMColor)); |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | unsigned colorA = SkGetPackedA32(color); |
| 324 | if (255 == colorA) { |
| 325 | sk_memset32(dst, color, count); |
| 326 | } else { |
| 327 | unsigned scale = 256 - SkAlpha255To256(colorA); |
| 328 | |
| 329 | if (count >= 4) { |
| 330 | SkASSERT(((size_t)dst & 0x03) == 0); |
| 331 | while (((size_t)dst & 0x0F) != 0) { |
| 332 | *dst = color + SkAlphaMulQ(*src, scale); |
| 333 | src++; |
| 334 | dst++; |
| 335 | count--; |
| 336 | } |
| 337 | |
| 338 | const __m128i *s = reinterpret_cast<const __m128i*>(src); |
| 339 | __m128i *d = reinterpret_cast<__m128i*>(dst); |
| 340 | __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
| 341 | __m128i src_scale_wide = _mm_set1_epi16(scale); |
| 342 | __m128i color_wide = _mm_set1_epi32(color); |
| 343 | while (count >= 4) { |
| 344 | // Load 4 pixels each of src and dest. |
| 345 | __m128i src_pixel = _mm_loadu_si128(s); |
| 346 | |
| 347 | // Get red and blue pixels into lower byte of each word. |
| 348 | __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); |
reed@google.com | 981d479 | 2011-03-09 12:55:47 +0000 | [diff] [blame] | 349 | |
senorblanco@chromium.org | c385638 | 2010-12-13 15:27:20 +0000 | [diff] [blame] | 350 | // Get alpha and green into lower byte of each word. |
| 351 | __m128i src_ag = _mm_srli_epi16(src_pixel, 8); |
| 352 | |
| 353 | // Multiply by scale. |
| 354 | src_rb = _mm_mullo_epi16(src_rb, src_scale_wide); |
| 355 | src_ag = _mm_mullo_epi16(src_ag, src_scale_wide); |
| 356 | |
| 357 | // Divide by 256. |
| 358 | src_rb = _mm_srli_epi16(src_rb, 8); |
| 359 | src_ag = _mm_andnot_si128(rb_mask, src_ag); |
| 360 | |
| 361 | // Combine back into RGBA. |
| 362 | src_pixel = _mm_or_si128(src_rb, src_ag); |
| 363 | |
| 364 | // Add color to result. |
| 365 | __m128i result = _mm_add_epi8(color_wide, src_pixel); |
| 366 | |
| 367 | // Store result. |
| 368 | _mm_store_si128(d, result); |
| 369 | s++; |
| 370 | d++; |
| 371 | count -= 4; |
| 372 | } |
| 373 | src = reinterpret_cast<const SkPMColor*>(s); |
| 374 | dst = reinterpret_cast<SkPMColor*>(d); |
| 375 | } |
| 376 | |
| 377 | while (count > 0) { |
| 378 | *dst = color + SkAlphaMulQ(*src, scale); |
| 379 | src += 1; |
| 380 | dst += 1; |
| 381 | count--; |
reed@google.com | 981d479 | 2011-03-09 12:55:47 +0000 | [diff] [blame] | 382 | } |
senorblanco@chromium.org | c385638 | 2010-12-13 15:27:20 +0000 | [diff] [blame] | 383 | } |
| 384 | } |
reed@google.com | 981d479 | 2011-03-09 12:55:47 +0000 | [diff] [blame] | 385 | |
| 386 | void SkARGB32_BlitMask_SSE2(void* device, size_t dstRB, |
| 387 | SkBitmap::Config dstConfig, const uint8_t* mask, |
reed@google.com | ee467ee | 2011-03-09 13:23:57 +0000 | [diff] [blame] | 388 | size_t maskRB, SkColor origColor, |
reed@google.com | 981d479 | 2011-03-09 12:55:47 +0000 | [diff] [blame] | 389 | int width, int height) |
| 390 | { |
reed@google.com | ee467ee | 2011-03-09 13:23:57 +0000 | [diff] [blame] | 391 | SkPMColor color = SkPreMultiplyColor(origColor); |
reed@google.com | 981d479 | 2011-03-09 12:55:47 +0000 | [diff] [blame] | 392 | size_t dstOffset = dstRB - (width << 2); |
| 393 | size_t maskOffset = maskRB - width; |
| 394 | SkPMColor* dst = (SkPMColor *)device; |
| 395 | do { |
| 396 | int count = width; |
| 397 | if (count >= 4) { |
| 398 | while (((size_t)dst & 0x0F) != 0 && (count > 0)) { |
| 399 | *dst = SkBlendARGB32(color, *dst, *mask); |
| 400 | mask++; |
| 401 | dst++; |
| 402 | count--; |
| 403 | } |
| 404 | __m128i *d = reinterpret_cast<__m128i*>(dst); |
| 405 | __m128i rb_mask = _mm_set1_epi32(0x00FF00FF); |
| 406 | __m128i c_256 = _mm_set1_epi16(256); |
| 407 | __m128i c_1 = _mm_set1_epi16(1); |
| 408 | __m128i src_pixel = _mm_set1_epi32(color); |
| 409 | while (count >= 4) { |
| 410 | // Load 4 pixels each of src and dest. |
| 411 | __m128i dst_pixel = _mm_load_si128(d); |
| 412 | |
| 413 | //set the aphla value |
| 414 | __m128i src_scale_wide = _mm_set_epi8(0, *(mask+3),\ |
| 415 | 0, *(mask+3),0, \ |
| 416 | *(mask+2),0, *(mask+2),\ |
| 417 | 0,*(mask+1), 0,*(mask+1),\ |
| 418 | 0, *mask,0,*mask); |
| 419 | |
| 420 | //call SkAlpha255To256() |
| 421 | src_scale_wide = _mm_add_epi16(src_scale_wide, c_1); |
| 422 | |
| 423 | // Get red and blue pixels into lower byte of each word. |
| 424 | __m128i dst_rb = _mm_and_si128(rb_mask, dst_pixel); |
| 425 | __m128i src_rb = _mm_and_si128(rb_mask, src_pixel); |
| 426 | |
| 427 | // Get alpha and green into lower byte of each word. |
| 428 | __m128i dst_ag = _mm_srli_epi16(dst_pixel, 8); |
| 429 | __m128i src_ag = _mm_srli_epi16(src_pixel, 8); |
| 430 | |
| 431 | // Put per-pixel alpha in low byte of each word. |
| 432 | __m128i dst_alpha = _mm_shufflehi_epi16(src_ag, 0xF5); |
| 433 | dst_alpha = _mm_shufflelo_epi16(dst_alpha, 0xF5); |
| 434 | |
| 435 | // dst_alpha = dst_alpha * src_scale |
| 436 | dst_alpha = _mm_mullo_epi16(dst_alpha, src_scale_wide); |
| 437 | |
| 438 | // Divide by 256. |
| 439 | dst_alpha = _mm_srli_epi16(dst_alpha, 8); |
| 440 | |
| 441 | // Subtract alphas from 256, to get 1..256 |
| 442 | dst_alpha = _mm_sub_epi16(c_256, dst_alpha); |
| 443 | // Multiply red and blue by dst pixel alpha. |
| 444 | dst_rb = _mm_mullo_epi16(dst_rb, dst_alpha); |
| 445 | // Multiply alpha and green by dst pixel alpha. |
| 446 | dst_ag = _mm_mullo_epi16(dst_ag, dst_alpha); |
| 447 | |
| 448 | // Multiply red and blue by global alpha. |
| 449 | src_rb = _mm_mullo_epi16(src_rb, src_scale_wide); |
| 450 | // Multiply alpha and green by global alpha. |
| 451 | src_ag = _mm_mullo_epi16(src_ag, src_scale_wide); |
| 452 | // Divide by 256. |
| 453 | dst_rb = _mm_srli_epi16(dst_rb, 8); |
| 454 | src_rb = _mm_srli_epi16(src_rb, 8); |
| 455 | |
| 456 | // Mask out low bits (goodies already in the right place; no need to divide) |
| 457 | dst_ag = _mm_andnot_si128(rb_mask, dst_ag); |
| 458 | src_ag = _mm_andnot_si128(rb_mask, src_ag); |
| 459 | |
| 460 | // Combine back into RGBA. |
| 461 | dst_pixel = _mm_or_si128(dst_rb, dst_ag); |
| 462 | __m128i tmp_src_pixel = _mm_or_si128(src_rb, src_ag); |
| 463 | |
| 464 | // Add two pixels into result. |
| 465 | __m128i result = _mm_add_epi8(tmp_src_pixel, dst_pixel); |
| 466 | _mm_store_si128(d, result); |
| 467 | // load the next 4 pixel |
| 468 | mask = mask + 4; |
| 469 | d++; |
| 470 | count -= 4; |
| 471 | } |
| 472 | dst = reinterpret_cast<SkPMColor *>(d); |
| 473 | } |
| 474 | while(count > 0) { |
| 475 | *dst= SkBlendARGB32(color, *dst, *mask); |
| 476 | dst += 1; |
| 477 | mask++; |
| 478 | count --; |
| 479 | } |
| 480 | dst = (SkPMColor *)((char*)dst + dstOffset); |
| 481 | mask += maskOffset; |
| 482 | } while (--height != 0); |
| 483 | } |