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 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 8 | #include "SkBitmapProcState_opts_SSE2.h" |
Florin Malita | 9953737 | 2017-01-04 13:01:55 -0500 | [diff] [blame] | 9 | #include "SkBitmapProcState_utils.h" |
Cary Clark | a4083c9 | 2017-09-15 11:59:23 -0400 | [diff] [blame] | 10 | #include "SkColorData.h" |
reed@google.com | 9cfc83c | 2013-07-22 17:18:18 +0000 | [diff] [blame] | 11 | #include "SkPaint.h" |
Hal Canary | c640d0d | 2018-06-13 09:59:02 -0400 | [diff] [blame] | 12 | #include "SkTo.h" |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 13 | #include "SkUtils.h" |
| 14 | |
Hal Canary | c640d0d | 2018-06-13 09:59:02 -0400 | [diff] [blame] | 15 | #include <emmintrin.h> |
| 16 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 17 | void S32_opaque_D32_filter_DX_SSE2(const SkBitmapProcState& s, |
| 18 | const uint32_t* xy, |
| 19 | int count, uint32_t* colors) { |
halcanary | 96fcdcc | 2015-08-27 07:41:13 -0700 | [diff] [blame] | 20 | SkASSERT(count > 0 && colors != nullptr); |
reed | 05a5647 | 2016-03-02 09:49:02 -0800 | [diff] [blame] | 21 | SkASSERT(s.fFilterQuality != kNone_SkFilterQuality); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 22 | SkASSERT(kN32_SkColorType == s.fPixmap.colorType()); |
senorblanco@chromium.org | aa4f0c6 | 2009-12-01 13:36:19 +0000 | [diff] [blame] | 23 | SkASSERT(s.fAlphaScale == 256); |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 24 | |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 25 | const char* srcAddr = static_cast<const char*>(s.fPixmap.addr()); |
| 26 | size_t rb = s.fPixmap.rowBytes(); |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 27 | uint32_t XY = *xy++; |
| 28 | unsigned y0 = XY >> 14; |
| 29 | const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb); |
| 30 | const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb); |
| 31 | unsigned subY = y0 & 0xF; |
| 32 | |
| 33 | // ( 0, 0, 0, 0, 0, 0, 0, 16) |
| 34 | __m128i sixteen = _mm_cvtsi32_si128(16); |
| 35 | |
| 36 | // ( 0, 0, 0, 0, 16, 16, 16, 16) |
| 37 | sixteen = _mm_shufflelo_epi16(sixteen, 0); |
| 38 | |
| 39 | // ( 0, 0, 0, 0, 0, 0, 0, y) |
| 40 | __m128i allY = _mm_cvtsi32_si128(subY); |
| 41 | |
| 42 | // ( 0, 0, 0, 0, y, y, y, y) |
| 43 | allY = _mm_shufflelo_epi16(allY, 0); |
| 44 | |
| 45 | // ( 0, 0, 0, 0, 16-y, 16-y, 16-y, 16-y) |
| 46 | __m128i negY = _mm_sub_epi16(sixteen, allY); |
| 47 | |
| 48 | // (16-y, 16-y, 16-y, 16-y, y, y, y, y) |
| 49 | allY = _mm_unpacklo_epi64(allY, negY); |
| 50 | |
| 51 | // (16, 16, 16, 16, 16, 16, 16, 16 ) |
| 52 | sixteen = _mm_shuffle_epi32(sixteen, 0); |
| 53 | |
| 54 | // ( 0, 0, 0, 0, 0, 0, 0, 0) |
| 55 | __m128i zero = _mm_setzero_si128(); |
| 56 | do { |
| 57 | uint32_t XX = *xy++; // x0:14 | 4 | x1:14 |
| 58 | unsigned x0 = XX >> 18; |
| 59 | unsigned x1 = XX & 0x3FFF; |
| 60 | |
| 61 | // (0, 0, 0, 0, 0, 0, 0, x) |
| 62 | __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 63 | |
senorblanco@chromium.org | dc7de74 | 2009-11-30 20:00:29 +0000 | [diff] [blame] | 64 | // (0, 0, 0, 0, x, x, x, x) |
| 65 | allX = _mm_shufflelo_epi16(allX, 0); |
| 66 | |
| 67 | // (x, x, x, x, x, x, x, x) |
| 68 | allX = _mm_shuffle_epi32(allX, 0); |
| 69 | |
| 70 | // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x) |
| 71 | __m128i negX = _mm_sub_epi16(sixteen, allX); |
| 72 | |
| 73 | // Load 4 samples (pixels). |
| 74 | __m128i a00 = _mm_cvtsi32_si128(row0[x0]); |
| 75 | __m128i a01 = _mm_cvtsi32_si128(row0[x1]); |
| 76 | __m128i a10 = _mm_cvtsi32_si128(row1[x0]); |
| 77 | __m128i a11 = _mm_cvtsi32_si128(row1[x1]); |
| 78 | |
| 79 | // (0, 0, a00, a10) |
| 80 | __m128i a00a10 = _mm_unpacklo_epi32(a10, a00); |
| 81 | |
| 82 | // Expand to 16 bits per component. |
| 83 | a00a10 = _mm_unpacklo_epi8(a00a10, zero); |
| 84 | |
| 85 | // ((a00 * (16-y)), (a10 * y)). |
| 86 | a00a10 = _mm_mullo_epi16(a00a10, allY); |
| 87 | |
| 88 | // (a00 * (16-y) * (16-x), a10 * y * (16-x)). |
| 89 | a00a10 = _mm_mullo_epi16(a00a10, negX); |
| 90 | |
| 91 | // (0, 0, a01, a10) |
| 92 | __m128i a01a11 = _mm_unpacklo_epi32(a11, a01); |
| 93 | |
| 94 | // Expand to 16 bits per component. |
| 95 | a01a11 = _mm_unpacklo_epi8(a01a11, zero); |
| 96 | |
| 97 | // (a01 * (16-y)), (a11 * y) |
| 98 | a01a11 = _mm_mullo_epi16(a01a11, allY); |
| 99 | |
| 100 | // (a01 * (16-y) * x), (a11 * y * x) |
| 101 | a01a11 = _mm_mullo_epi16(a01a11, allX); |
| 102 | |
| 103 | // (a00*w00 + a01*w01, a10*w10 + a11*w11) |
| 104 | __m128i sum = _mm_add_epi16(a00a10, a01a11); |
| 105 | |
| 106 | // (DC, a00*w00 + a01*w01) |
| 107 | __m128i shifted = _mm_shuffle_epi32(sum, 0xEE); |
| 108 | |
| 109 | // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11) |
| 110 | sum = _mm_add_epi16(sum, shifted); |
| 111 | |
| 112 | // Divide each 16 bit component by 256. |
| 113 | sum = _mm_srli_epi16(sum, 8); |
| 114 | |
| 115 | // Pack lower 4 16 bit values of sum into lower 4 bytes. |
| 116 | sum = _mm_packus_epi16(sum, zero); |
| 117 | |
| 118 | // Extract low int and store. |
| 119 | *colors++ = _mm_cvtsi128_si32(sum); |
| 120 | } while (--count > 0); |
| 121 | } |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 122 | |
| 123 | void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s, |
| 124 | const uint32_t* xy, |
| 125 | int count, uint32_t* colors) { |
halcanary | 96fcdcc | 2015-08-27 07:41:13 -0700 | [diff] [blame] | 126 | SkASSERT(count > 0 && colors != nullptr); |
reed | 05a5647 | 2016-03-02 09:49:02 -0800 | [diff] [blame] | 127 | SkASSERT(s.fFilterQuality != kNone_SkFilterQuality); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 128 | SkASSERT(kN32_SkColorType == s.fPixmap.colorType()); |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 129 | SkASSERT(s.fAlphaScale < 256); |
| 130 | |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 131 | const char* srcAddr = static_cast<const char*>(s.fPixmap.addr()); |
| 132 | size_t rb = s.fPixmap.rowBytes(); |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 133 | uint32_t XY = *xy++; |
| 134 | unsigned y0 = XY >> 14; |
| 135 | const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb); |
| 136 | const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb); |
| 137 | unsigned subY = y0 & 0xF; |
| 138 | |
| 139 | // ( 0, 0, 0, 0, 0, 0, 0, 16) |
| 140 | __m128i sixteen = _mm_cvtsi32_si128(16); |
| 141 | |
| 142 | // ( 0, 0, 0, 0, 16, 16, 16, 16) |
| 143 | sixteen = _mm_shufflelo_epi16(sixteen, 0); |
| 144 | |
| 145 | // ( 0, 0, 0, 0, 0, 0, 0, y) |
| 146 | __m128i allY = _mm_cvtsi32_si128(subY); |
| 147 | |
| 148 | // ( 0, 0, 0, 0, y, y, y, y) |
| 149 | allY = _mm_shufflelo_epi16(allY, 0); |
| 150 | |
| 151 | // ( 0, 0, 0, 0, 16-y, 16-y, 16-y, 16-y) |
| 152 | __m128i negY = _mm_sub_epi16(sixteen, allY); |
| 153 | |
| 154 | // (16-y, 16-y, 16-y, 16-y, y, y, y, y) |
| 155 | allY = _mm_unpacklo_epi64(allY, negY); |
| 156 | |
| 157 | // (16, 16, 16, 16, 16, 16, 16, 16 ) |
| 158 | sixteen = _mm_shuffle_epi32(sixteen, 0); |
| 159 | |
| 160 | // ( 0, 0, 0, 0, 0, 0, 0, 0) |
| 161 | __m128i zero = _mm_setzero_si128(); |
| 162 | |
| 163 | // ( alpha, alpha, alpha, alpha, alpha, alpha, alpha, alpha ) |
| 164 | __m128i alpha = _mm_set1_epi16(s.fAlphaScale); |
| 165 | |
| 166 | do { |
| 167 | uint32_t XX = *xy++; // x0:14 | 4 | x1:14 |
| 168 | unsigned x0 = XX >> 18; |
| 169 | unsigned x1 = XX & 0x3FFF; |
| 170 | |
| 171 | // (0, 0, 0, 0, 0, 0, 0, x) |
| 172 | __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 173 | |
senorblanco@chromium.org | f3f0bd7 | 2009-12-10 22:46:31 +0000 | [diff] [blame] | 174 | // (0, 0, 0, 0, x, x, x, x) |
| 175 | allX = _mm_shufflelo_epi16(allX, 0); |
| 176 | |
| 177 | // (x, x, x, x, x, x, x, x) |
| 178 | allX = _mm_shuffle_epi32(allX, 0); |
| 179 | |
| 180 | // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x) |
| 181 | __m128i negX = _mm_sub_epi16(sixteen, allX); |
| 182 | |
| 183 | // Load 4 samples (pixels). |
| 184 | __m128i a00 = _mm_cvtsi32_si128(row0[x0]); |
| 185 | __m128i a01 = _mm_cvtsi32_si128(row0[x1]); |
| 186 | __m128i a10 = _mm_cvtsi32_si128(row1[x0]); |
| 187 | __m128i a11 = _mm_cvtsi32_si128(row1[x1]); |
| 188 | |
| 189 | // (0, 0, a00, a10) |
| 190 | __m128i a00a10 = _mm_unpacklo_epi32(a10, a00); |
| 191 | |
| 192 | // Expand to 16 bits per component. |
| 193 | a00a10 = _mm_unpacklo_epi8(a00a10, zero); |
| 194 | |
| 195 | // ((a00 * (16-y)), (a10 * y)). |
| 196 | a00a10 = _mm_mullo_epi16(a00a10, allY); |
| 197 | |
| 198 | // (a00 * (16-y) * (16-x), a10 * y * (16-x)). |
| 199 | a00a10 = _mm_mullo_epi16(a00a10, negX); |
| 200 | |
| 201 | // (0, 0, a01, a10) |
| 202 | __m128i a01a11 = _mm_unpacklo_epi32(a11, a01); |
| 203 | |
| 204 | // Expand to 16 bits per component. |
| 205 | a01a11 = _mm_unpacklo_epi8(a01a11, zero); |
| 206 | |
| 207 | // (a01 * (16-y)), (a11 * y) |
| 208 | a01a11 = _mm_mullo_epi16(a01a11, allY); |
| 209 | |
| 210 | // (a01 * (16-y) * x), (a11 * y * x) |
| 211 | a01a11 = _mm_mullo_epi16(a01a11, allX); |
| 212 | |
| 213 | // (a00*w00 + a01*w01, a10*w10 + a11*w11) |
| 214 | __m128i sum = _mm_add_epi16(a00a10, a01a11); |
| 215 | |
| 216 | // (DC, a00*w00 + a01*w01) |
| 217 | __m128i shifted = _mm_shuffle_epi32(sum, 0xEE); |
| 218 | |
| 219 | // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11) |
| 220 | sum = _mm_add_epi16(sum, shifted); |
| 221 | |
| 222 | // Divide each 16 bit component by 256. |
| 223 | sum = _mm_srli_epi16(sum, 8); |
| 224 | |
| 225 | // Multiply by alpha. |
| 226 | sum = _mm_mullo_epi16(sum, alpha); |
| 227 | |
| 228 | // Divide each 16 bit component by 256. |
| 229 | sum = _mm_srli_epi16(sum, 8); |
| 230 | |
| 231 | // Pack lower 4 16 bit values of sum into lower 4 bytes. |
| 232 | sum = _mm_packus_epi16(sum, zero); |
| 233 | |
| 234 | // Extract low int and store. |
| 235 | *colors++ = _mm_cvtsi128_si32(sum); |
| 236 | } while (--count > 0); |
| 237 | } |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 238 | |
Mike Reed | 2eab65b | 2018-04-17 12:01:10 -0400 | [diff] [blame] | 239 | // Temporarily go into 64bit so we don't overflow during the add. Since we shift down by 16 |
| 240 | // in the end, the result should always fit back in 32bits. |
| 241 | static inline int32_t safe_fixed_add_shift(SkFixed a, SkFixed b) { |
| 242 | int64_t tmp = a; |
| 243 | return SkToS32((tmp + b) >> 16); |
| 244 | } |
| 245 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 246 | static inline uint32_t ClampX_ClampY_pack_filter(SkFixed f, unsigned max, |
| 247 | SkFixed one) { |
| 248 | unsigned i = SkClampMax(f >> 16, max); |
| 249 | i = (i << 4) | ((f >> 12) & 0xF); |
Mike Reed | 2eab65b | 2018-04-17 12:01:10 -0400 | [diff] [blame] | 250 | return (i << 14) | SkClampMax(safe_fixed_add_shift(f, one), max); |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 251 | } |
| 252 | |
| 253 | /* SSE version of ClampX_ClampY_filter_scale() |
| 254 | * portable version is in core/SkBitmapProcState_matrix.h |
| 255 | */ |
| 256 | void ClampX_ClampY_filter_scale_SSE2(const SkBitmapProcState& s, uint32_t xy[], |
| 257 | int count, int x, int y) { |
| 258 | SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | |
| 259 | SkMatrix::kScale_Mask)) == 0); |
| 260 | SkASSERT(s.fInvKy == 0); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 261 | |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 262 | const unsigned maxX = s.fPixmap.width() - 1; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 263 | const SkFixed one = s.fFilterOneX; |
| 264 | const SkFixed dx = s.fInvSx; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 265 | |
fmalita | 2404f03 | 2016-02-03 05:44:21 -0800 | [diff] [blame] | 266 | const SkBitmapProcStateAutoMapper mapper(s, x, y); |
fmalita | be5cfa9 | 2016-02-03 10:21:33 -0800 | [diff] [blame] | 267 | const SkFixed fy = mapper.fixedY(); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 268 | const unsigned maxY = s.fPixmap.height() - 1; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 269 | // compute our two Y values up front |
| 270 | *xy++ = ClampX_ClampY_pack_filter(fy, maxY, s.fFilterOneY); |
| 271 | // now initialize fx |
fmalita | be5cfa9 | 2016-02-03 10:21:33 -0800 | [diff] [blame] | 272 | SkFixed fx = mapper.fixedX(); |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 273 | |
| 274 | // test if we don't need to apply the tile proc |
Florin Malita | 9953737 | 2017-01-04 13:01:55 -0500 | [diff] [blame] | 275 | if (can_truncate_to_fixed_for_decal(fx, dx, count, maxX)) { |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 276 | if (count >= 4) { |
| 277 | // SSE version of decal_filter_scale |
| 278 | while ((size_t(xy) & 0x0F) != 0) { |
| 279 | SkASSERT((fx >> (16 + 14)) == 0); |
| 280 | *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1); |
| 281 | fx += dx; |
| 282 | count--; |
| 283 | } |
| 284 | |
| 285 | __m128i wide_1 = _mm_set1_epi32(1); |
| 286 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 287 | __m128i wide_fx = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 288 | fx + dx, fx); |
| 289 | |
| 290 | while (count >= 4) { |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 291 | __m128i wide_out; |
| 292 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 293 | wide_out = _mm_slli_epi32(_mm_srai_epi32(wide_fx, 12), 14); |
| 294 | wide_out = _mm_or_si128(wide_out, _mm_add_epi32( |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 295 | _mm_srai_epi32(wide_fx, 16), wide_1)); |
| 296 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 297 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_out); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 298 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 299 | xy += 4; |
| 300 | fx += dx * 4; |
| 301 | wide_fx = _mm_add_epi32(wide_fx, wide_dx4); |
| 302 | count -= 4; |
| 303 | } // while count >= 4 |
| 304 | } // if count >= 4 |
| 305 | |
| 306 | while (count-- > 0) { |
| 307 | SkASSERT((fx >> (16 + 14)) == 0); |
| 308 | *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1); |
| 309 | fx += dx; |
| 310 | } |
| 311 | } else { |
| 312 | // SSE2 only support 16bit interger max & min, so only process the case |
| 313 | // 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] | 314 | // height, there should be rare bitmap whose height will be greater |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 315 | // than max 16bit interger in the real world. |
| 316 | if ((count >= 4) && (maxX <= 0xFFFF)) { |
| 317 | while (((size_t)xy & 0x0F) != 0) { |
| 318 | *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one); |
| 319 | fx += dx; |
| 320 | count--; |
| 321 | } |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 322 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 323 | __m128i wide_fx = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 324 | fx + dx, fx); |
| 325 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 326 | __m128i wide_one = _mm_set1_epi32(one); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 327 | __m128i wide_maxX = _mm_set1_epi32(maxX); |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 328 | __m128i wide_mask = _mm_set1_epi32(0xF); |
| 329 | |
| 330 | while (count >= 4) { |
| 331 | __m128i wide_i; |
| 332 | __m128i wide_lo; |
| 333 | __m128i wide_fx1; |
| 334 | |
| 335 | // i = SkClampMax(f>>16,maxX) |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 336 | wide_i = _mm_max_epi16(_mm_srli_epi32(wide_fx, 16), |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 337 | _mm_setzero_si128()); |
| 338 | wide_i = _mm_min_epi16(wide_i, wide_maxX); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 339 | |
Florin Malita | d1c550e | 2016-12-19 10:55:41 -0500 | [diff] [blame] | 340 | // i<<4 | EXTRACT_LOW_BITS(fx) |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 341 | wide_lo = _mm_srli_epi32(wide_fx, 12); |
| 342 | wide_lo = _mm_and_si128(wide_lo, wide_mask); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 343 | wide_i = _mm_slli_epi32(wide_i, 4); |
| 344 | wide_i = _mm_or_si128(wide_i, wide_lo); |
| 345 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 346 | // i<<14 |
| 347 | wide_i = _mm_slli_epi32(wide_i, 14); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 348 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 349 | // SkClampMax(((f+one))>>16,max) |
| 350 | wide_fx1 = _mm_add_epi32(wide_fx, wide_one); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 351 | wide_fx1 = _mm_max_epi16(_mm_srli_epi32(wide_fx1, 16), |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 352 | _mm_setzero_si128()); |
| 353 | wide_fx1 = _mm_min_epi16(wide_fx1, wide_maxX); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 354 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 355 | // final combination |
| 356 | wide_i = _mm_or_si128(wide_i, wide_fx1); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 357 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_i); |
| 358 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 359 | wide_fx = _mm_add_epi32(wide_fx, wide_dx4); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 360 | fx += dx * 4; |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 361 | xy += 4; |
| 362 | count -= 4; |
| 363 | } // while count >= 4 |
| 364 | } // if count >= 4 |
| 365 | |
Mike Reed | 010ce2b | 2018-05-09 13:53:59 -0400 | [diff] [blame] | 366 | /* |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 367 | while (count-- > 0) { |
| 368 | *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one); |
| 369 | fx += dx; |
| 370 | } |
Mike Reed | 010ce2b | 2018-05-09 13:53:59 -0400 | [diff] [blame] | 371 | We'd like to write this as above, but that form allows fx to get 1-iteration too big/small |
| 372 | when count is 0, and this can trigger a UBSAN error, even though we won't in fact use that |
| 373 | last (undefined) value for fx. |
| 374 | |
| 375 | Here is an alternative that should always be efficient, but seems much harder to read: |
| 376 | |
| 377 | if (count > 0) { |
| 378 | for (;;) { |
| 379 | *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one); |
| 380 | if (--count == 0) break; |
| 381 | fx += dx; |
| 382 | } |
| 383 | } |
| 384 | |
| 385 | For now, we'll try this variant: more compact than the if/for version, and we hope the |
| 386 | compiler will get rid of the integer multiply. |
| 387 | */ |
| 388 | for (int i = 0; i < count; ++i) { |
| 389 | *xy++ = ClampX_ClampY_pack_filter(fx + i*dx, maxX, one); |
| 390 | } |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 391 | } |
| 392 | } |
| 393 | |
| 394 | /* SSE version of ClampX_ClampY_nofilter_scale() |
| 395 | * portable version is in core/SkBitmapProcState_matrix.h |
| 396 | */ |
| 397 | void ClampX_ClampY_nofilter_scale_SSE2(const SkBitmapProcState& s, |
| 398 | uint32_t xy[], int count, int x, int y) { |
| 399 | SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | |
| 400 | SkMatrix::kScale_Mask)) == 0); |
| 401 | |
| 402 | // we store y, x, x, x, x, x |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 403 | const unsigned maxX = s.fPixmap.width() - 1; |
fmalita | eb54307 | 2016-02-02 10:17:24 -0800 | [diff] [blame] | 404 | const SkBitmapProcStateAutoMapper mapper(s, x, y); |
reed | ad7ae6c | 2015-06-04 14:12:25 -0700 | [diff] [blame] | 405 | const unsigned maxY = s.fPixmap.height() - 1; |
fmalita | be5cfa9 | 2016-02-03 10:21:33 -0800 | [diff] [blame] | 406 | *xy++ = SkClampMax(mapper.intY(), maxY); |
| 407 | SkFixed fx = mapper.fixedX(); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 408 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 409 | if (0 == maxX) { |
| 410 | // all of the following X values must be 0 |
| 411 | memset(xy, 0, count * sizeof(uint16_t)); |
| 412 | return; |
| 413 | } |
| 414 | |
| 415 | const SkFixed dx = s.fInvSx; |
| 416 | |
| 417 | // test if we don't need to apply the tile proc |
| 418 | if ((unsigned)(fx >> 16) <= maxX && |
| 419 | (unsigned)((fx + dx * (count - 1)) >> 16) <= maxX) { |
| 420 | // SSE version of decal_nofilter_scale |
| 421 | if (count >= 8) { |
| 422 | while (((size_t)xy & 0x0F) != 0) { |
| 423 | *xy++ = pack_two_shorts(fx >> 16, (fx + dx) >> 16); |
| 424 | fx += 2 * dx; |
| 425 | count -= 2; |
| 426 | } |
| 427 | |
| 428 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 429 | __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4); |
| 430 | |
| 431 | __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 432 | fx + dx, fx); |
| 433 | __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4); |
| 434 | |
| 435 | while (count >= 8) { |
| 436 | __m128i wide_out_low = _mm_srli_epi32(wide_low, 16); |
| 437 | __m128i wide_out_high = _mm_srli_epi32(wide_high, 16); |
| 438 | |
| 439 | __m128i wide_result = _mm_packs_epi32(wide_out_low, |
| 440 | wide_out_high); |
| 441 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result); |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 442 | |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 443 | wide_low = _mm_add_epi32(wide_low, wide_dx8); |
| 444 | wide_high = _mm_add_epi32(wide_high, wide_dx8); |
| 445 | |
| 446 | xy += 4; |
| 447 | fx += dx * 8; |
| 448 | count -= 8; |
| 449 | } |
| 450 | } // if count >= 8 |
| 451 | |
| 452 | uint16_t* xx = reinterpret_cast<uint16_t*>(xy); |
| 453 | while (count-- > 0) { |
| 454 | *xx++ = SkToU16(fx >> 16); |
| 455 | fx += dx; |
| 456 | } |
| 457 | } else { |
| 458 | // SSE2 only support 16bit interger max & min, so only process the case |
| 459 | // 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] | 460 | // height, there should be rare bitmap whose height will be greater |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 461 | // than max 16bit interger in the real world. |
| 462 | if ((count >= 8) && (maxX <= 0xFFFF)) { |
| 463 | while (((size_t)xy & 0x0F) != 0) { |
mike@reedtribe.org | 602f227 | 2012-03-14 02:04:40 +0000 | [diff] [blame] | 464 | *xy++ = pack_two_shorts(SkClampMax((fx + dx) >> 16, maxX), |
| 465 | SkClampMax(fx >> 16, maxX)); |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 466 | fx += 2 * dx; |
| 467 | count -= 2; |
| 468 | } |
| 469 | |
| 470 | __m128i wide_dx4 = _mm_set1_epi32(dx * 4); |
| 471 | __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4); |
| 472 | |
| 473 | __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2, |
| 474 | fx + dx, fx); |
| 475 | __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4); |
| 476 | __m128i wide_maxX = _mm_set1_epi32(maxX); |
| 477 | |
| 478 | while (count >= 8) { |
| 479 | __m128i wide_out_low = _mm_srli_epi32(wide_low, 16); |
| 480 | __m128i wide_out_high = _mm_srli_epi32(wide_high, 16); |
| 481 | |
rmistry@google.com | fbfcd56 | 2012-08-23 18:09:54 +0000 | [diff] [blame] | 482 | wide_out_low = _mm_max_epi16(wide_out_low, |
tomhudson@google.com | 06a7313 | 2012-02-22 18:30:43 +0000 | [diff] [blame] | 483 | _mm_setzero_si128()); |
| 484 | wide_out_low = _mm_min_epi16(wide_out_low, wide_maxX); |
| 485 | wide_out_high = _mm_max_epi16(wide_out_high, |
| 486 | _mm_setzero_si128()); |
| 487 | wide_out_high = _mm_min_epi16(wide_out_high, wide_maxX); |
| 488 | |
| 489 | __m128i wide_result = _mm_packs_epi32(wide_out_low, |
| 490 | wide_out_high); |
| 491 | _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result); |
| 492 | |
| 493 | wide_low = _mm_add_epi32(wide_low, wide_dx8); |
| 494 | wide_high = _mm_add_epi32(wide_high, wide_dx8); |
| 495 | |
| 496 | xy += 4; |
| 497 | fx += dx * 8; |
| 498 | count -= 8; |
| 499 | } |
| 500 | } // if count >= 8 |
| 501 | |
| 502 | uint16_t* xx = reinterpret_cast<uint16_t*>(xy); |
| 503 | while (count-- > 0) { |
| 504 | *xx++ = SkClampMax(fx >> 16, maxX); |
| 505 | fx += dx; |
| 506 | } |
| 507 | } |
| 508 | } |