senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 1 | /* |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 2 | * Copyright 2009 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. |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 6 | */ |
| 7 | |
| 8 | #include <emmintrin.h> |
| 9 | #include "SkBitmapProcState_opts_SSE2.h" |
Florin Malita | 9953737 | 2017-01-04 13:01:55 -0500 | [diff] [blame] | 10 | #include "SkBitmapProcState_utils.h" |
Cary Clark | a4083c9 | 2017-09-15 11:59:23 -0400 | [diff] [blame^] | 11 | #include "SkColorData.h" |
reed@google.com | 9cfc83c | 2013-07-22 17:18:18 +0000 | [diff] [blame] | 12 | #include "SkPaint.h" |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 13 | #include "SkUtils.h" |
| 14 | |
| 15 | void S32_opaque_D32_filter_DX_SSE2(const SkBitmapProcState& s, |
| 16 | const uint32_t* xy, |
| 17 | int count, uint32_t* colors) { |
halcanary | 96fcdcc | 2015-08-27 07:41:13 -0700 | [diff] [blame] | 18 | SkASSERT(count > 0 && colors != nullptr); |
reed | 05a5647 | 2016-03-02 09:49:02 -0800 | [diff] [blame] | 19 | SkASSERT(s.fFilterQuality != kNone_SkFilterQuality); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 20 | SkASSERT(kN32_SkColorType == s.fPixmap.colorType()); |
senorblanco@chromium.org | aa4f0c6 | 2009-12-01 13:36:19 +0000 | [diff] [blame] | 21 | SkASSERT(s.fAlphaScale == 256); |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 22 | |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 23 | const char* srcAddr = static_cast<const char*>(s.fPixmap.addr()); |
| 24 | size_t rb = s.fPixmap.rowBytes(); |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 25 | uint32_t XY = *xy++; |
| 26 | unsigned y0 = XY >> 14; |
| 27 | const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb); |
| 28 | const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb); |
| 29 | unsigned subY = y0 & 0xF; |
| 30 | |
| 31 | // ( 0, 0, 0, 0, 0, 0, 0, 16) |
| 32 | __m128i sixteen = _mm_cvtsi32_si128(16); |
| 33 | |
| 34 | // ( 0, 0, 0, 0, 16, 16, 16, 16) |
| 35 | sixteen = _mm_shufflelo_epi16(sixteen, 0); |
| 36 | |
| 37 | // ( 0, 0, 0, 0, 0, 0, 0, y) |
| 38 | __m128i allY = _mm_cvtsi32_si128(subY); |
| 39 | |
| 40 | // ( 0, 0, 0, 0, y, y, y, y) |
| 41 | allY = _mm_shufflelo_epi16(allY, 0); |
| 42 | |
| 43 | // ( 0, 0, 0, 0, 16-y, 16-y, 16-y, 16-y) |
| 44 | __m128i negY = _mm_sub_epi16(sixteen, allY); |
| 45 | |
| 46 | // (16-y, 16-y, 16-y, 16-y, y, y, y, y) |
| 47 | allY = _mm_unpacklo_epi64(allY, negY); |
| 48 | |
| 49 | // (16, 16, 16, 16, 16, 16, 16, 16 ) |
| 50 | sixteen = _mm_shuffle_epi32(sixteen, 0); |
| 51 | |
| 52 | // ( 0, 0, 0, 0, 0, 0, 0, 0) |
| 53 | __m128i zero = _mm_setzero_si128(); |
| 54 | do { |
| 55 | uint32_t XX = *xy++; // x0:14 | 4 | x1:14 |
| 56 | unsigned x0 = XX >> 18; |
| 57 | unsigned x1 = XX & 0x3FFF; |
| 58 | |
| 59 | // (0, 0, 0, 0, 0, 0, 0, x) |
| 60 | __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 61 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 62 | // (0, 0, 0, 0, x, x, x, x) |
| 63 | allX = _mm_shufflelo_epi16(allX, 0); |
| 64 | |
| 65 | // (x, x, x, x, x, x, x, x) |
| 66 | allX = _mm_shuffle_epi32(allX, 0); |
| 67 | |
| 68 | // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x) |
| 69 | __m128i negX = _mm_sub_epi16(sixteen, allX); |
| 70 | |
| 71 | // Load 4 samples (pixels). |
| 72 | __m128i a00 = _mm_cvtsi32_si128(row0[x0]); |
| 73 | __m128i a01 = _mm_cvtsi32_si128(row0[x1]); |
| 74 | __m128i a10 = _mm_cvtsi32_si128(row1[x0]); |
| 75 | __m128i a11 = _mm_cvtsi32_si128(row1[x1]); |
| 76 | |
| 77 | // (0, 0, a00, a10) |
| 78 | __m128i a00a10 = _mm_unpacklo_epi32(a10, a00); |
| 79 | |
| 80 | // Expand to 16 bits per component. |
| 81 | a00a10 = _mm_unpacklo_epi8(a00a10, zero); |
| 82 | |
| 83 | // ((a00 * (16-y)), (a10 * y)). |
| 84 | a00a10 = _mm_mullo_epi16(a00a10, allY); |
| 85 | |
| 86 | // (a00 * (16-y) * (16-x), a10 * y * (16-x)). |
| 87 | a00a10 = _mm_mullo_epi16(a00a10, negX); |
| 88 | |
| 89 | // (0, 0, a01, a10) |
| 90 | __m128i a01a11 = _mm_unpacklo_epi32(a11, a01); |
| 91 | |
| 92 | // Expand to 16 bits per component. |
| 93 | a01a11 = _mm_unpacklo_epi8(a01a11, zero); |
| 94 | |
| 95 | // (a01 * (16-y)), (a11 * y) |
| 96 | a01a11 = _mm_mullo_epi16(a01a11, allY); |
| 97 | |
| 98 | // (a01 * (16-y) * x), (a11 * y * x) |
| 99 | a01a11 = _mm_mullo_epi16(a01a11, allX); |
| 100 | |
| 101 | // (a00*w00 + a01*w01, a10*w10 + a11*w11) |
| 102 | __m128i sum = _mm_add_epi16(a00a10, a01a11); |
| 103 | |
| 104 | // (DC, a00*w00 + a01*w01) |
| 105 | __m128i shifted = _mm_shuffle_epi32(sum, 0xEE); |
| 106 | |
| 107 | // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11) |
| 108 | sum = _mm_add_epi16(sum, shifted); |
| 109 | |
| 110 | // Divide each 16 bit component by 256. |
| 111 | sum = _mm_srli_epi16(sum, 8); |
| 112 | |
| 113 | // Pack lower 4 16 bit values of sum into lower 4 bytes. |
| 114 | sum = _mm_packus_epi16(sum, zero); |
| 115 | |
| 116 | // Extract low int and store. |
| 117 | *colors++ = _mm_cvtsi128_si32(sum); |
| 118 | } while (--count > 0); |
| 119 | } |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 120 | |
| 121 | void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s, |
| 122 | const uint32_t* xy, |
| 123 | int count, uint32_t* colors) { |
halcanary | 96fcdcc | 2015-08-27 07:41:13 -0700 | [diff] [blame] | 124 | SkASSERT(count > 0 && colors != nullptr); |
reed | 05a5647 | 2016-03-02 09:49:02 -0800 | [diff] [blame] | 125 | SkASSERT(s.fFilterQuality != kNone_SkFilterQuality); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 126 | SkASSERT(kN32_SkColorType == s.fPixmap.colorType()); |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 127 | SkASSERT(s.fAlphaScale < 256); |
| 128 | |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 129 | const char* srcAddr = static_cast<const char*>(s.fPixmap.addr()); |
| 130 | size_t rb = s.fPixmap.rowBytes(); |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 131 | uint32_t XY = *xy++; |
| 132 | unsigned y0 = XY >> 14; |
| 133 | const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb); |
| 134 | const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb); |
| 135 | unsigned subY = y0 & 0xF; |
| 136 | |
| 137 | // ( 0, 0, 0, 0, 0, 0, 0, 16) |
| 138 | __m128i sixteen = _mm_cvtsi32_si128(16); |
| 139 | |
| 140 | // ( 0, 0, 0, 0, 16, 16, 16, 16) |
| 141 | sixteen = _mm_shufflelo_epi16(sixteen, 0); |
| 142 | |
| 143 | // ( 0, 0, 0, 0, 0, 0, 0, y) |
| 144 | __m128i allY = _mm_cvtsi32_si128(subY); |
| 145 | |
| 146 | // ( 0, 0, 0, 0, y, y, y, y) |
| 147 | allY = _mm_shufflelo_epi16(allY, 0); |
| 148 | |
| 149 | // ( 0, 0, 0, 0, 16-y, 16-y, 16-y, 16-y) |
| 150 | __m128i negY = _mm_sub_epi16(sixteen, allY); |
| 151 | |
| 152 | // (16-y, 16-y, 16-y, 16-y, y, y, y, y) |
| 153 | allY = _mm_unpacklo_epi64(allY, negY); |
| 154 | |
| 155 | // (16, 16, 16, 16, 16, 16, 16, 16 ) |
| 156 | sixteen = _mm_shuffle_epi32(sixteen, 0); |
| 157 | |
| 158 | // ( 0, 0, 0, 0, 0, 0, 0, 0) |
| 159 | __m128i zero = _mm_setzero_si128(); |
| 160 | |
| 161 | // ( alpha, alpha, alpha, alpha, alpha, alpha, alpha, alpha ) |
| 162 | __m128i alpha = _mm_set1_epi16(s.fAlphaScale); |
| 163 | |
| 164 | do { |
| 165 | uint32_t XX = *xy++; // x0:14 | 4 | x1:14 |
| 166 | unsigned x0 = XX >> 18; |
| 167 | unsigned x1 = XX & 0x3FFF; |
| 168 | |
| 169 | // (0, 0, 0, 0, 0, 0, 0, x) |
| 170 | __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 171 | |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 172 | // (0, 0, 0, 0, x, x, x, x) |
| 173 | allX = _mm_shufflelo_epi16(allX, 0); |
| 174 | |
| 175 | // (x, x, x, x, x, x, x, x) |
| 176 | allX = _mm_shuffle_epi32(allX, 0); |
| 177 | |
| 178 | // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x) |
| 179 | __m128i negX = _mm_sub_epi16(sixteen, allX); |
| 180 | |
| 181 | // Load 4 samples (pixels). |
| 182 | __m128i a00 = _mm_cvtsi32_si128(row0[x0]); |
| 183 | __m128i a01 = _mm_cvtsi32_si128(row0[x1]); |
| 184 | __m128i a10 = _mm_cvtsi32_si128(row1[x0]); |
| 185 | __m128i a11 = _mm_cvtsi32_si128(row1[x1]); |
| 186 | |
| 187 | // (0, 0, a00, a10) |
| 188 | __m128i a00a10 = _mm_unpacklo_epi32(a10, a00); |
| 189 | |
| 190 | // Expand to 16 bits per component. |
| 191 | a00a10 = _mm_unpacklo_epi8(a00a10, zero); |
| 192 | |
| 193 | // ((a00 * (16-y)), (a10 * y)). |
| 194 | a00a10 = _mm_mullo_epi16(a00a10, allY); |
| 195 | |
| 196 | // (a00 * (16-y) * (16-x), a10 * y * (16-x)). |
| 197 | a00a10 = _mm_mullo_epi16(a00a10, negX); |
| 198 | |
| 199 | // (0, 0, a01, a10) |
| 200 | __m128i a01a11 = _mm_unpacklo_epi32(a11, a01); |
| 201 | |
| 202 | // Expand to 16 bits per component. |
| 203 | a01a11 = _mm_unpacklo_epi8(a01a11, zero); |
| 204 | |
| 205 | // (a01 * (16-y)), (a11 * y) |
| 206 | a01a11 = _mm_mullo_epi16(a01a11, allY); |
| 207 | |
| 208 | // (a01 * (16-y) * x), (a11 * y * x) |
| 209 | a01a11 = _mm_mullo_epi16(a01a11, allX); |
| 210 | |
| 211 | // (a00*w00 + a01*w01, a10*w10 + a11*w11) |
| 212 | __m128i sum = _mm_add_epi16(a00a10, a01a11); |
| 213 | |
| 214 | // (DC, a00*w00 + a01*w01) |
| 215 | __m128i shifted = _mm_shuffle_epi32(sum, 0xEE); |
| 216 | |
| 217 | // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11) |
| 218 | sum = _mm_add_epi16(sum, shifted); |
| 219 | |
| 220 | // Divide each 16 bit component by 256. |
| 221 | sum = _mm_srli_epi16(sum, 8); |
| 222 | |
| 223 | // Multiply by alpha. |
| 224 | sum = _mm_mullo_epi16(sum, alpha); |
| 225 | |
| 226 | // Divide each 16 bit component by 256. |
| 227 | sum = _mm_srli_epi16(sum, 8); |
| 228 | |
| 229 | // Pack lower 4 16 bit values of sum into lower 4 bytes. |
| 230 | sum = _mm_packus_epi16(sum, zero); |
| 231 | |
| 232 | // Extract low int and store. |
| 233 | *colors++ = _mm_cvtsi128_si32(sum); |
| 234 | } while (--count > 0); |
| 235 | } |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 236 | |
| 237 | static inline uint32_t ClampX_ClampY_pack_filter(SkFixed f, unsigned max, |
| 238 | SkFixed one) { |
| 239 | unsigned i = SkClampMax(f >> 16, max); |
| 240 | i = (i << 4) | ((f >> 12) & 0xF); |
| 241 | return (i << 14) | SkClampMax((f + one) >> 16, max); |
| 242 | } |
| 243 | |
| 244 | /* SSE version of ClampX_ClampY_filter_scale() |
| 245 | * portable version is in core/SkBitmapProcState_matrix.h |
| 246 | */ |
| 247 | void ClampX_ClampY_filter_scale_SSE2(const SkBitmapProcState& s, uint32_t xy[], |
| 248 | int count, int x, int y) { |
| 249 | SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | |
| 250 | SkMatrix::kScale_Mask)) == 0); |
| 251 | SkASSERT(s.fInvKy == 0); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 252 | |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 253 | const unsigned maxX = s.fPixmap.width() - 1; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 254 | const SkFixed one = s.fFilterOneX; |
| 255 | const SkFixed dx = s.fInvSx; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 256 | |
fmalita | 2404f03 | 2016-02-03 05:44:21 -0800 | [diff] [blame] | 257 | const SkBitmapProcStateAutoMapper mapper(s, x, y); |
fmalita | be5cfa9 | 2016-02-03 10:21:33 -0800 | [diff] [blame] | 258 | const SkFixed fy = mapper.fixedY(); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 259 | const unsigned maxY = s.fPixmap.height() - 1; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 260 | // compute our two Y values up front |
| 261 | *xy++ = ClampX_ClampY_pack_filter(fy, maxY, s.fFilterOneY); |
| 262 | // now initialize fx |
fmalita | be5cfa9 | 2016-02-03 10:21:33 -0800 | [diff] [blame] | 263 | SkFixed fx = mapper.fixedX(); |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 264 | |
| 265 | // test if we don't need to apply the tile proc |
Florin Malita | 9953737 | 2017-01-04 13:01:55 -0500 | [diff] [blame] | 266 | if (can_truncate_to_fixed_for_decal(fx, dx, count, maxX)) { |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 267 | if (count >= 4) { |
| 268 | // SSE version of decal_filter_scale |
| 269 | while ((size_t(xy) & 0x0F) != 0) { |
| 270 | SkASSERT((fx >> (16 + 14)) == 0); |
| 271 | *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1); |
| 272 | fx += dx; |
| 273 | count--; |
| 274 | } |
| 275 | |
| 276 | __m128i wide_1 = _mm_set1_epi32(1); |
| 277 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 278 | __m128i wide_fx = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 279 | fx + dx, fx); |
| 280 | |
| 281 | while (count >= 4) { |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 282 | __m128i wide_out; |
| 283 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 284 | wide_out = _mm_slli_epi32(_mm_srai_epi32(wide_fx, 12), 14); |
| 285 | wide_out = _mm_or_si128(wide_out, _mm_add_epi32( |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 286 | _mm_srai_epi32(wide_fx, 16), wide_1)); |
| 287 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 288 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_out); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 289 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 290 | xy += 4; |
| 291 | fx += dx * 4; |
| 292 | wide_fx = _mm_add_epi32(wide_fx, wide_dx4); |
| 293 | count -= 4; |
| 294 | } // while count >= 4 |
| 295 | } // if count >= 4 |
| 296 | |
| 297 | while (count-- > 0) { |
| 298 | SkASSERT((fx >> (16 + 14)) == 0); |
| 299 | *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1); |
| 300 | fx += dx; |
| 301 | } |
| 302 | } else { |
| 303 | // SSE2 only support 16bit interger max & min, so only process the case |
| 304 | // maxX less than the max 16bit interger. Actually maxX is the bitmap's |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 305 | // height, there should be rare bitmap whose height will be greater |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 306 | // than max 16bit interger in the real world. |
| 307 | if ((count >= 4) && (maxX <= 0xFFFF)) { |
| 308 | while (((size_t)xy & 0x0F) != 0) { |
| 309 | *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one); |
| 310 | fx += dx; |
| 311 | count--; |
| 312 | } |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 313 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 314 | __m128i wide_fx = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 315 | fx + dx, fx); |
| 316 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 317 | __m128i wide_one = _mm_set1_epi32(one); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 318 | __m128i wide_maxX = _mm_set1_epi32(maxX); |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 319 | __m128i wide_mask = _mm_set1_epi32(0xF); |
| 320 | |
| 321 | while (count >= 4) { |
| 322 | __m128i wide_i; |
| 323 | __m128i wide_lo; |
| 324 | __m128i wide_fx1; |
| 325 | |
| 326 | // i = SkClampMax(f>>16,maxX) |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 327 | wide_i = _mm_max_epi16(_mm_srli_epi32(wide_fx, 16), |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 328 | _mm_setzero_si128()); |
| 329 | wide_i = _mm_min_epi16(wide_i, wide_maxX); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 330 | |
Florin Malita | d1c550e | 2016-12-19 10:55:41 -0500 | [diff] [blame] | 331 | // i<<4 | EXTRACT_LOW_BITS(fx) |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 332 | wide_lo = _mm_srli_epi32(wide_fx, 12); |
| 333 | wide_lo = _mm_and_si128(wide_lo, wide_mask); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 334 | wide_i = _mm_slli_epi32(wide_i, 4); |
| 335 | wide_i = _mm_or_si128(wide_i, wide_lo); |
| 336 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 337 | // i<<14 |
| 338 | wide_i = _mm_slli_epi32(wide_i, 14); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 339 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 340 | // SkClampMax(((f+one))>>16,max) |
| 341 | wide_fx1 = _mm_add_epi32(wide_fx, wide_one); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 342 | wide_fx1 = _mm_max_epi16(_mm_srli_epi32(wide_fx1, 16), |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 343 | _mm_setzero_si128()); |
| 344 | wide_fx1 = _mm_min_epi16(wide_fx1, wide_maxX); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 345 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 346 | // final combination |
| 347 | wide_i = _mm_or_si128(wide_i, wide_fx1); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 348 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_i); |
| 349 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 350 | wide_fx = _mm_add_epi32(wide_fx, wide_dx4); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 351 | fx += dx * 4; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 352 | xy += 4; |
| 353 | count -= 4; |
| 354 | } // while count >= 4 |
| 355 | } // if count >= 4 |
| 356 | |
| 357 | while (count-- > 0) { |
| 358 | *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one); |
| 359 | fx += dx; |
| 360 | } |
| 361 | } |
| 362 | } |
| 363 | |
| 364 | /* SSE version of ClampX_ClampY_nofilter_scale() |
| 365 | * portable version is in core/SkBitmapProcState_matrix.h |
| 366 | */ |
| 367 | void ClampX_ClampY_nofilter_scale_SSE2(const SkBitmapProcState& s, |
| 368 | uint32_t xy[], int count, int x, int y) { |
| 369 | SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | |
| 370 | SkMatrix::kScale_Mask)) == 0); |
| 371 | |
| 372 | // we store y, x, x, x, x, x |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 373 | const unsigned maxX = s.fPixmap.width() - 1; |
fmalita | eb54307 | 2016-02-02 10:17:24 -0800 | [diff] [blame] | 374 | const SkBitmapProcStateAutoMapper mapper(s, x, y); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 375 | const unsigned maxY = s.fPixmap.height() - 1; |
fmalita | be5cfa9 | 2016-02-03 10:21:33 -0800 | [diff] [blame] | 376 | *xy++ = SkClampMax(mapper.intY(), maxY); |
| 377 | SkFixed fx = mapper.fixedX(); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 378 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 379 | if (0 == maxX) { |
| 380 | // all of the following X values must be 0 |
| 381 | memset(xy, 0, count * sizeof(uint16_t)); |
| 382 | return; |
| 383 | } |
| 384 | |
| 385 | const SkFixed dx = s.fInvSx; |
| 386 | |
| 387 | // test if we don't need to apply the tile proc |
| 388 | if ((unsigned)(fx >> 16) <= maxX && |
| 389 | (unsigned)((fx + dx * (count - 1)) >> 16) <= maxX) { |
| 390 | // SSE version of decal_nofilter_scale |
| 391 | if (count >= 8) { |
| 392 | while (((size_t)xy & 0x0F) != 0) { |
| 393 | *xy++ = pack_two_shorts(fx >> 16, (fx + dx) >> 16); |
| 394 | fx += 2 * dx; |
| 395 | count -= 2; |
| 396 | } |
| 397 | |
| 398 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 399 | __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4); |
| 400 | |
| 401 | __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 402 | fx + dx, fx); |
| 403 | __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4); |
| 404 | |
| 405 | while (count >= 8) { |
| 406 | __m128i wide_out_low = _mm_srli_epi32(wide_low, 16); |
| 407 | __m128i wide_out_high = _mm_srli_epi32(wide_high, 16); |
| 408 | |
| 409 | __m128i wide_result = _mm_packs_epi32(wide_out_low, |
| 410 | wide_out_high); |
| 411 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 412 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 413 | wide_low = _mm_add_epi32(wide_low, wide_dx8); |
| 414 | wide_high = _mm_add_epi32(wide_high, wide_dx8); |
| 415 | |
| 416 | xy += 4; |
| 417 | fx += dx * 8; |
| 418 | count -= 8; |
| 419 | } |
| 420 | } // if count >= 8 |
| 421 | |
| 422 | uint16_t* xx = reinterpret_cast<uint16_t*>(xy); |
| 423 | while (count-- > 0) { |
| 424 | *xx++ = SkToU16(fx >> 16); |
| 425 | fx += dx; |
| 426 | } |
| 427 | } else { |
| 428 | // SSE2 only support 16bit interger max & min, so only process the case |
| 429 | // maxX less than the max 16bit interger. Actually maxX is the bitmap's |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 430 | // height, there should be rare bitmap whose height will be greater |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 431 | // than max 16bit interger in the real world. |
| 432 | if ((count >= 8) && (maxX <= 0xFFFF)) { |
| 433 | while (((size_t)xy & 0x0F) != 0) { |
mike@reedtribe.org | 602f227 | 2012-03-14 02:04:40 +0000 | [diff] [blame] | 434 | *xy++ = pack_two_shorts(SkClampMax((fx + dx) >> 16, maxX), |
| 435 | SkClampMax(fx >> 16, maxX)); |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 436 | fx += 2 * dx; |
| 437 | count -= 2; |
| 438 | } |
| 439 | |
| 440 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 441 | __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4); |
| 442 | |
| 443 | __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 444 | fx + dx, fx); |
| 445 | __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4); |
| 446 | __m128i wide_maxX = _mm_set1_epi32(maxX); |
| 447 | |
| 448 | while (count >= 8) { |
| 449 | __m128i wide_out_low = _mm_srli_epi32(wide_low, 16); |
| 450 | __m128i wide_out_high = _mm_srli_epi32(wide_high, 16); |
| 451 | |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 452 | wide_out_low = _mm_max_epi16(wide_out_low, |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 453 | _mm_setzero_si128()); |
| 454 | wide_out_low = _mm_min_epi16(wide_out_low, wide_maxX); |
| 455 | wide_out_high = _mm_max_epi16(wide_out_high, |
| 456 | _mm_setzero_si128()); |
| 457 | wide_out_high = _mm_min_epi16(wide_out_high, wide_maxX); |
| 458 | |
| 459 | __m128i wide_result = _mm_packs_epi32(wide_out_low, |
| 460 | wide_out_high); |
| 461 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result); |
| 462 | |
| 463 | wide_low = _mm_add_epi32(wide_low, wide_dx8); |
| 464 | wide_high = _mm_add_epi32(wide_high, wide_dx8); |
| 465 | |
| 466 | xy += 4; |
| 467 | fx += dx * 8; |
| 468 | count -= 8; |
| 469 | } |
| 470 | } // if count >= 8 |
| 471 | |
| 472 | uint16_t* xx = reinterpret_cast<uint16_t*>(xy); |
| 473 | while (count-- > 0) { |
| 474 | *xx++ = SkClampMax(fx >> 16, maxX); |
| 475 | fx += dx; |
| 476 | } |
| 477 | } |
| 478 | } |