1 2 /* 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. 7 */ 8 9 10 #include <emmintrin.h> 11 #include "SkBitmapProcState_opts_SSE2.h" 12 #include "SkPaint.h" 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) { 18 SkASSERT(count > 0 && colors != NULL); 19 SkASSERT(s.fFilterLevel != SkPaint::kNone_FilterLevel); 20 SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config); 21 SkASSERT(s.fAlphaScale == 256); 22 23 const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels()); 24 size_t rb = s.fBitmap->rowBytes(); 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); 61 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 } 120 121 void S32_alpha_D32_filter_DX_SSE2(const SkBitmapProcState& s, 122 const uint32_t* xy, 123 int count, uint32_t* colors) { 124 SkASSERT(count > 0 && colors != NULL); 125 SkASSERT(s.fFilterLevel != SkPaint::kNone_FilterLevel); 126 SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config); 127 SkASSERT(s.fAlphaScale < 256); 128 129 const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels()); 130 size_t rb = s.fBitmap->rowBytes(); 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); 171 172 // (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 } 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); 252 253 const unsigned maxX = s.fBitmap->width() - 1; 254 const SkFixed one = s.fFilterOneX; 255 const SkFixed dx = s.fInvSx; 256 SkFixed fx; 257 258 SkPoint pt; 259 s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, 260 SkIntToScalar(y) + SK_ScalarHalf, &pt); 261 const SkFixed fy = SkScalarToFixed(pt.fY) - (s.fFilterOneY >> 1); 262 const unsigned maxY = s.fBitmap->height() - 1; 263 // compute our two Y values up front 264 *xy++ = ClampX_ClampY_pack_filter(fy, maxY, s.fFilterOneY); 265 // now initialize fx 266 fx = SkScalarToFixed(pt.fX) - (one >> 1); 267 268 // test if we don't need to apply the tile proc 269 if (dx > 0 && (unsigned)(fx >> 16) <= maxX && 270 (unsigned)((fx + dx * (count - 1)) >> 16) < maxX) { 271 if (count >= 4) { 272 // SSE version of decal_filter_scale 273 while ((size_t(xy) & 0x0F) != 0) { 274 SkASSERT((fx >> (16 + 14)) == 0); 275 *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1); 276 fx += dx; 277 count--; 278 } 279 280 __m128i wide_1 = _mm_set1_epi32(1); 281 __m128i wide_dx4 = _mm_set1_epi32(dx * 4); 282 __m128i wide_fx = _mm_set_epi32(fx + dx * 3, fx + dx * 2, 283 fx + dx, fx); 284 285 while (count >= 4) { 286 __m128i wide_out; 287 288 wide_out = _mm_slli_epi32(_mm_srai_epi32(wide_fx, 12), 14); 289 wide_out = _mm_or_si128(wide_out, _mm_add_epi32( 290 _mm_srai_epi32(wide_fx, 16), wide_1)); 291 292 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_out); 293 294 xy += 4; 295 fx += dx * 4; 296 wide_fx = _mm_add_epi32(wide_fx, wide_dx4); 297 count -= 4; 298 } // while count >= 4 299 } // if count >= 4 300 301 while (count-- > 0) { 302 SkASSERT((fx >> (16 + 14)) == 0); 303 *xy++ = (fx >> 12 << 14) | ((fx >> 16) + 1); 304 fx += dx; 305 } 306 } else { 307 // SSE2 only support 16bit interger max & min, so only process the case 308 // maxX less than the max 16bit interger. Actually maxX is the bitmap's 309 // height, there should be rare bitmap whose height will be greater 310 // than max 16bit interger in the real world. 311 if ((count >= 4) && (maxX <= 0xFFFF)) { 312 while (((size_t)xy & 0x0F) != 0) { 313 *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one); 314 fx += dx; 315 count--; 316 } 317 318 __m128i wide_fx = _mm_set_epi32(fx + dx * 3, fx + dx * 2, 319 fx + dx, fx); 320 __m128i wide_dx4 = _mm_set1_epi32(dx * 4); 321 __m128i wide_one = _mm_set1_epi32(one); 322 __m128i wide_maxX = _mm_set1_epi32(maxX); 323 __m128i wide_mask = _mm_set1_epi32(0xF); 324 325 while (count >= 4) { 326 __m128i wide_i; 327 __m128i wide_lo; 328 __m128i wide_fx1; 329 330 // i = SkClampMax(f>>16,maxX) 331 wide_i = _mm_max_epi16(_mm_srli_epi32(wide_fx, 16), 332 _mm_setzero_si128()); 333 wide_i = _mm_min_epi16(wide_i, wide_maxX); 334 335 // i<<4 | TILEX_LOW_BITS(fx) 336 wide_lo = _mm_srli_epi32(wide_fx, 12); 337 wide_lo = _mm_and_si128(wide_lo, wide_mask); 338 wide_i = _mm_slli_epi32(wide_i, 4); 339 wide_i = _mm_or_si128(wide_i, wide_lo); 340 341 // i<<14 342 wide_i = _mm_slli_epi32(wide_i, 14); 343 344 // SkClampMax(((f+one))>>16,max) 345 wide_fx1 = _mm_add_epi32(wide_fx, wide_one); 346 wide_fx1 = _mm_max_epi16(_mm_srli_epi32(wide_fx1, 16), 347 _mm_setzero_si128()); 348 wide_fx1 = _mm_min_epi16(wide_fx1, wide_maxX); 349 350 // final combination 351 wide_i = _mm_or_si128(wide_i, wide_fx1); 352 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_i); 353 354 wide_fx = _mm_add_epi32(wide_fx, wide_dx4); 355 fx += dx * 4; 356 xy += 4; 357 count -= 4; 358 } // while count >= 4 359 } // if count >= 4 360 361 while (count-- > 0) { 362 *xy++ = ClampX_ClampY_pack_filter(fx, maxX, one); 363 fx += dx; 364 } 365 } 366 } 367 368 /* SSE version of ClampX_ClampY_nofilter_scale() 369 * portable version is in core/SkBitmapProcState_matrix.h 370 */ 371 void ClampX_ClampY_nofilter_scale_SSE2(const SkBitmapProcState& s, 372 uint32_t xy[], int count, int x, int y) { 373 SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | 374 SkMatrix::kScale_Mask)) == 0); 375 376 // we store y, x, x, x, x, x 377 const unsigned maxX = s.fBitmap->width() - 1; 378 SkFixed fx; 379 SkPoint pt; 380 s.fInvProc(s.fInvMatrix, SkIntToScalar(x) + SK_ScalarHalf, 381 SkIntToScalar(y) + SK_ScalarHalf, &pt); 382 fx = SkScalarToFixed(pt.fY); 383 const unsigned maxY = s.fBitmap->height() - 1; 384 *xy++ = SkClampMax(fx >> 16, maxY); 385 fx = SkScalarToFixed(pt.fX); 386 387 if (0 == maxX) { 388 // all of the following X values must be 0 389 memset(xy, 0, count * sizeof(uint16_t)); 390 return; 391 } 392 393 const SkFixed dx = s.fInvSx; 394 395 // test if we don't need to apply the tile proc 396 if ((unsigned)(fx >> 16) <= maxX && 397 (unsigned)((fx + dx * (count - 1)) >> 16) <= maxX) { 398 // SSE version of decal_nofilter_scale 399 if (count >= 8) { 400 while (((size_t)xy & 0x0F) != 0) { 401 *xy++ = pack_two_shorts(fx >> 16, (fx + dx) >> 16); 402 fx += 2 * dx; 403 count -= 2; 404 } 405 406 __m128i wide_dx4 = _mm_set1_epi32(dx * 4); 407 __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4); 408 409 __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2, 410 fx + dx, fx); 411 __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4); 412 413 while (count >= 8) { 414 __m128i wide_out_low = _mm_srli_epi32(wide_low, 16); 415 __m128i wide_out_high = _mm_srli_epi32(wide_high, 16); 416 417 __m128i wide_result = _mm_packs_epi32(wide_out_low, 418 wide_out_high); 419 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result); 420 421 wide_low = _mm_add_epi32(wide_low, wide_dx8); 422 wide_high = _mm_add_epi32(wide_high, wide_dx8); 423 424 xy += 4; 425 fx += dx * 8; 426 count -= 8; 427 } 428 } // if count >= 8 429 430 uint16_t* xx = reinterpret_cast<uint16_t*>(xy); 431 while (count-- > 0) { 432 *xx++ = SkToU16(fx >> 16); 433 fx += dx; 434 } 435 } else { 436 // SSE2 only support 16bit interger max & min, so only process the case 437 // maxX less than the max 16bit interger. Actually maxX is the bitmap's 438 // height, there should be rare bitmap whose height will be greater 439 // than max 16bit interger in the real world. 440 if ((count >= 8) && (maxX <= 0xFFFF)) { 441 while (((size_t)xy & 0x0F) != 0) { 442 *xy++ = pack_two_shorts(SkClampMax((fx + dx) >> 16, maxX), 443 SkClampMax(fx >> 16, maxX)); 444 fx += 2 * dx; 445 count -= 2; 446 } 447 448 __m128i wide_dx4 = _mm_set1_epi32(dx * 4); 449 __m128i wide_dx8 = _mm_add_epi32(wide_dx4, wide_dx4); 450 451 __m128i wide_low = _mm_set_epi32(fx + dx * 3, fx + dx * 2, 452 fx + dx, fx); 453 __m128i wide_high = _mm_add_epi32(wide_low, wide_dx4); 454 __m128i wide_maxX = _mm_set1_epi32(maxX); 455 456 while (count >= 8) { 457 __m128i wide_out_low = _mm_srli_epi32(wide_low, 16); 458 __m128i wide_out_high = _mm_srli_epi32(wide_high, 16); 459 460 wide_out_low = _mm_max_epi16(wide_out_low, 461 _mm_setzero_si128()); 462 wide_out_low = _mm_min_epi16(wide_out_low, wide_maxX); 463 wide_out_high = _mm_max_epi16(wide_out_high, 464 _mm_setzero_si128()); 465 wide_out_high = _mm_min_epi16(wide_out_high, wide_maxX); 466 467 __m128i wide_result = _mm_packs_epi32(wide_out_low, 468 wide_out_high); 469 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_result); 470 471 wide_low = _mm_add_epi32(wide_low, wide_dx8); 472 wide_high = _mm_add_epi32(wide_high, wide_dx8); 473 474 xy += 4; 475 fx += dx * 8; 476 count -= 8; 477 } 478 } // if count >= 8 479 480 uint16_t* xx = reinterpret_cast<uint16_t*>(xy); 481 while (count-- > 0) { 482 *xx++ = SkClampMax(fx >> 16, maxX); 483 fx += dx; 484 } 485 } 486 } 487 488 /* SSE version of ClampX_ClampY_filter_affine() 489 * portable version is in core/SkBitmapProcState_matrix.h 490 */ 491 void ClampX_ClampY_filter_affine_SSE2(const SkBitmapProcState& s, 492 uint32_t xy[], int count, int x, int y) { 493 SkPoint srcPt; 494 s.fInvProc(s.fInvMatrix, 495 SkIntToScalar(x) + SK_ScalarHalf, 496 SkIntToScalar(y) + SK_ScalarHalf, &srcPt); 497 498 SkFixed oneX = s.fFilterOneX; 499 SkFixed oneY = s.fFilterOneY; 500 SkFixed fx = SkScalarToFixed(srcPt.fX) - (oneX >> 1); 501 SkFixed fy = SkScalarToFixed(srcPt.fY) - (oneY >> 1); 502 SkFixed dx = s.fInvSx; 503 SkFixed dy = s.fInvKy; 504 unsigned maxX = s.fBitmap->width() - 1; 505 unsigned maxY = s.fBitmap->height() - 1; 506 507 if (count >= 2 && (maxX <= 0xFFFF)) { 508 SkFixed dx2 = dx + dx; 509 SkFixed dy2 = dy + dy; 510 511 __m128i wide_f = _mm_set_epi32(fx + dx, fy + dy, fx, fy); 512 __m128i wide_d2 = _mm_set_epi32(dx2, dy2, dx2, dy2); 513 __m128i wide_one = _mm_set_epi32(oneX, oneY, oneX, oneY); 514 __m128i wide_max = _mm_set_epi32(maxX, maxY, maxX, maxY); 515 __m128i wide_mask = _mm_set1_epi32(0xF); 516 517 while (count >= 2) { 518 // i = SkClampMax(f>>16,maxX) 519 __m128i wide_i = _mm_max_epi16(_mm_srli_epi32(wide_f, 16), 520 _mm_setzero_si128()); 521 wide_i = _mm_min_epi16(wide_i, wide_max); 522 523 // i<<4 | TILEX_LOW_BITS(f) 524 __m128i wide_lo = _mm_srli_epi32(wide_f, 12); 525 wide_lo = _mm_and_si128(wide_lo, wide_mask); 526 wide_i = _mm_slli_epi32(wide_i, 4); 527 wide_i = _mm_or_si128(wide_i, wide_lo); 528 529 // i<<14 530 wide_i = _mm_slli_epi32(wide_i, 14); 531 532 // SkClampMax(((f+one))>>16,max) 533 __m128i wide_f1 = _mm_add_epi32(wide_f, wide_one); 534 wide_f1 = _mm_max_epi16(_mm_srli_epi32(wide_f1, 16), 535 _mm_setzero_si128()); 536 wide_f1 = _mm_min_epi16(wide_f1, wide_max); 537 538 // final combination 539 wide_i = _mm_or_si128(wide_i, wide_f1); 540 _mm_storeu_si128(reinterpret_cast<__m128i*>(xy), wide_i); 541 542 wide_f = _mm_add_epi32(wide_f, wide_d2); 543 544 fx += dx2; 545 fy += dy2; 546 xy += 4; 547 count -= 2; 548 } // while count >= 2 549 } // if count >= 2 550 551 while (count-- > 0) { 552 *xy++ = ClampX_ClampY_pack_filter(fy, maxY, oneY); 553 fy += dy; 554 *xy++ = ClampX_ClampY_pack_filter(fx, maxX, oneX); 555 fx += dx; 556 } 557 } 558 559 /* SSE version of ClampX_ClampY_nofilter_affine() 560 * portable version is in core/SkBitmapProcState_matrix.h 561 */ 562 void ClampX_ClampY_nofilter_affine_SSE2(const SkBitmapProcState& s, 563 uint32_t xy[], int count, int x, int y) { 564 SkASSERT(s.fInvType & SkMatrix::kAffine_Mask); 565 SkASSERT((s.fInvType & ~(SkMatrix::kTranslate_Mask | 566 SkMatrix::kScale_Mask | 567 SkMatrix::kAffine_Mask)) == 0); 568 569 SkPoint srcPt; 570 s.fInvProc(s.fInvMatrix, 571 SkIntToScalar(x) + SK_ScalarHalf, 572 SkIntToScalar(y) + SK_ScalarHalf, &srcPt); 573 574 SkFixed fx = SkScalarToFixed(srcPt.fX); 575 SkFixed fy = SkScalarToFixed(srcPt.fY); 576 SkFixed dx = s.fInvSx; 577 SkFixed dy = s.fInvKy; 578 int maxX = s.fBitmap->width() - 1; 579 int maxY = s.fBitmap->height() - 1; 580 581 if (count >= 4 && (maxX <= 0xFFFF)) { 582 while (((size_t)xy & 0x0F) != 0) { 583 *xy++ = (SkClampMax(fy >> 16, maxY) << 16) | 584 SkClampMax(fx >> 16, maxX); 585 fx += dx; 586 fy += dy; 587 count--; 588 } 589 590 SkFixed dx4 = dx * 4; 591 SkFixed dy4 = dy * 4; 592 593 __m128i wide_fx = _mm_set_epi32(fx + dx * 3, fx + dx * 2, 594 fx + dx, fx); 595 __m128i wide_fy = _mm_set_epi32(fy + dy * 3, fy + dy * 2, 596 fy + dy, fy); 597 __m128i wide_dx4 = _mm_set1_epi32(dx4); 598 __m128i wide_dy4 = _mm_set1_epi32(dy4); 599 600 __m128i wide_maxX = _mm_set1_epi32(maxX); 601 __m128i wide_maxY = _mm_set1_epi32(maxY); 602 603 while (count >= 4) { 604 // SkClampMax(fx>>16,maxX) 605 __m128i wide_lo = _mm_max_epi16(_mm_srli_epi32(wide_fx, 16), 606 _mm_setzero_si128()); 607 wide_lo = _mm_min_epi16(wide_lo, wide_maxX); 608 609 // SkClampMax(fy>>16,maxY) 610 __m128i wide_hi = _mm_max_epi16(_mm_srli_epi32(wide_fy, 16), 611 _mm_setzero_si128()); 612 wide_hi = _mm_min_epi16(wide_hi, wide_maxY); 613 614 // final combination 615 __m128i wide_i = _mm_or_si128(_mm_slli_epi32(wide_hi, 16), 616 wide_lo); 617 _mm_store_si128(reinterpret_cast<__m128i*>(xy), wide_i); 618 619 wide_fx = _mm_add_epi32(wide_fx, wide_dx4); 620 wide_fy = _mm_add_epi32(wide_fy, wide_dy4); 621 622 fx += dx4; 623 fy += dy4; 624 xy += 4; 625 count -= 4; 626 } // while count >= 4 627 } // if count >= 4 628 629 while (count-- > 0) { 630 *xy++ = (SkClampMax(fy >> 16, maxY) << 16) | 631 SkClampMax(fx >> 16, maxX); 632 fx += dx; 633 fy += dy; 634 } 635 } 636 637 /* SSE version of S32_D16_filter_DX_SSE2 638 * Definition is in section of "D16 functions for SRC == 8888" in SkBitmapProcState.cpp 639 * It combines S32_opaque_D32_filter_DX_SSE2 and SkPixel32ToPixel16 640 */ 641 void S32_D16_filter_DX_SSE2(const SkBitmapProcState& s, 642 const uint32_t* xy, 643 int count, uint16_t* colors) { 644 SkASSERT(count > 0 && colors != NULL); 645 SkASSERT(s.fFilterLevel != SkPaint::kNone_FilterLevel); 646 SkASSERT(s.fBitmap->config() == SkBitmap::kARGB_8888_Config); 647 SkASSERT(s.fBitmap->isOpaque()); 648 649 SkPMColor dstColor; 650 const char* srcAddr = static_cast<const char*>(s.fBitmap->getPixels()); 651 size_t rb = s.fBitmap->rowBytes(); 652 uint32_t XY = *xy++; 653 unsigned y0 = XY >> 14; 654 const uint32_t* row0 = reinterpret_cast<const uint32_t*>(srcAddr + (y0 >> 4) * rb); 655 const uint32_t* row1 = reinterpret_cast<const uint32_t*>(srcAddr + (XY & 0x3FFF) * rb); 656 unsigned subY = y0 & 0xF; 657 658 // ( 0, 0, 0, 0, 0, 0, 0, 16) 659 __m128i sixteen = _mm_cvtsi32_si128(16); 660 661 // ( 0, 0, 0, 0, 16, 16, 16, 16) 662 sixteen = _mm_shufflelo_epi16(sixteen, 0); 663 664 // ( 0, 0, 0, 0, 0, 0, 0, y) 665 __m128i allY = _mm_cvtsi32_si128(subY); 666 667 // ( 0, 0, 0, 0, y, y, y, y) 668 allY = _mm_shufflelo_epi16(allY, 0); 669 670 // ( 0, 0, 0, 0, 16-y, 16-y, 16-y, 16-y) 671 __m128i negY = _mm_sub_epi16(sixteen, allY); 672 673 // (16-y, 16-y, 16-y, 16-y, y, y, y, y) 674 allY = _mm_unpacklo_epi64(allY, negY); 675 676 // (16, 16, 16, 16, 16, 16, 16, 16 ) 677 sixteen = _mm_shuffle_epi32(sixteen, 0); 678 679 // ( 0, 0, 0, 0, 0, 0, 0, 0) 680 __m128i zero = _mm_setzero_si128(); 681 682 do { 683 uint32_t XX = *xy++; // x0:14 | 4 | x1:14 684 unsigned x0 = XX >> 18; 685 unsigned x1 = XX & 0x3FFF; 686 687 // (0, 0, 0, 0, 0, 0, 0, x) 688 __m128i allX = _mm_cvtsi32_si128((XX >> 14) & 0x0F); 689 690 // (0, 0, 0, 0, x, x, x, x) 691 allX = _mm_shufflelo_epi16(allX, 0); 692 693 // (x, x, x, x, x, x, x, x) 694 allX = _mm_shuffle_epi32(allX, 0); 695 696 // (16-x, 16-x, 16-x, 16-x, 16-x, 16-x, 16-x) 697 __m128i negX = _mm_sub_epi16(sixteen, allX); 698 699 // Load 4 samples (pixels). 700 __m128i a00 = _mm_cvtsi32_si128(row0[x0]); 701 __m128i a01 = _mm_cvtsi32_si128(row0[x1]); 702 __m128i a10 = _mm_cvtsi32_si128(row1[x0]); 703 __m128i a11 = _mm_cvtsi32_si128(row1[x1]); 704 705 // (0, 0, a00, a10) 706 __m128i a00a10 = _mm_unpacklo_epi32(a10, a00); 707 708 // Expand to 16 bits per component. 709 a00a10 = _mm_unpacklo_epi8(a00a10, zero); 710 711 // ((a00 * (16-y)), (a10 * y)). 712 a00a10 = _mm_mullo_epi16(a00a10, allY); 713 714 // (a00 * (16-y) * (16-x), a10 * y * (16-x)). 715 a00a10 = _mm_mullo_epi16(a00a10, negX); 716 717 // (0, 0, a01, a10) 718 __m128i a01a11 = _mm_unpacklo_epi32(a11, a01); 719 720 // Expand to 16 bits per component. 721 a01a11 = _mm_unpacklo_epi8(a01a11, zero); 722 723 // (a01 * (16-y)), (a11 * y) 724 a01a11 = _mm_mullo_epi16(a01a11, allY); 725 726 // (a01 * (16-y) * x), (a11 * y * x) 727 a01a11 = _mm_mullo_epi16(a01a11, allX); 728 729 // (a00*w00 + a01*w01, a10*w10 + a11*w11) 730 __m128i sum = _mm_add_epi16(a00a10, a01a11); 731 732 // (DC, a00*w00 + a01*w01) 733 __m128i shifted = _mm_shuffle_epi32(sum, 0xEE); 734 735 // (DC, a00*w00 + a01*w01 + a10*w10 + a11*w11) 736 sum = _mm_add_epi16(sum, shifted); 737 738 // Divide each 16 bit component by 256. 739 sum = _mm_srli_epi16(sum, 8); 740 741 // Pack lower 4 16 bit values of sum into lower 4 bytes. 742 sum = _mm_packus_epi16(sum, zero); 743 744 // Extract low int and store. 745 dstColor = _mm_cvtsi128_si32(sum); 746 747 //*colors++ = SkPixel32ToPixel16(dstColor); 748 // below is much faster than the above. It's tested for Android benchmark--Softweg 749 __m128i _m_temp1 = _mm_set1_epi32(dstColor); 750 __m128i _m_temp2 = _mm_srli_epi32(_m_temp1, 3); 751 752 unsigned int r32 = _mm_cvtsi128_si32(_m_temp2); 753 unsigned r = (r32 & ((1<<5) -1)) << 11; 754 755 _m_temp2 = _mm_srli_epi32(_m_temp2, 7); 756 unsigned int g32 = _mm_cvtsi128_si32(_m_temp2); 757 unsigned g = (g32 & ((1<<6) -1)) << 5; 758 759 _m_temp2 = _mm_srli_epi32(_m_temp2, 9); 760 unsigned int b32 = _mm_cvtsi128_si32(_m_temp2); 761 unsigned b = (b32 & ((1<<5) -1)); 762 763 *colors++ = r | g | b; 764 765 } while (--count > 0); 766 } 767
