lossless_sse2.c (31910B)
1 // Copyright 2014 Google Inc. All Rights Reserved. 2 // 3 // Use of this source code is governed by a BSD-style license 4 // that can be found in the COPYING file in the root of the source 5 // tree. An additional intellectual property rights grant can be found 6 // in the file PATENTS. All contributing project authors may 7 // be found in the AUTHORS file in the root of the source tree. 8 // ----------------------------------------------------------------------------- 9 // 10 // SSE2 variant of methods for lossless decoder 11 // 12 // Author: Skal (pascal.massimino@gmail.com) 13 14 #include "src/dsp/dsp.h" 15 16 #if defined(WEBP_USE_SSE2) 17 18 #include <emmintrin.h> 19 #include <string.h> 20 21 #include "src/dsp/common_sse2.h" 22 #include "src/dsp/cpu.h" 23 #include "src/dsp/lossless.h" 24 #include "src/dsp/lossless_common.h" 25 #include "src/webp/format_constants.h" 26 #include "src/webp/types.h" 27 28 //------------------------------------------------------------------------------ 29 // Predictor Transform 30 31 static WEBP_INLINE uint32_t ClampedAddSubtractFull_SSE2(uint32_t c0, 32 uint32_t c1, 33 uint32_t c2) { 34 const __m128i zero = _mm_setzero_si128(); 35 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c0), zero); 36 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c1), zero); 37 const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c2), zero); 38 const __m128i V1 = _mm_add_epi16(C0, C1); 39 const __m128i V2 = _mm_sub_epi16(V1, C2); 40 const __m128i b = _mm_packus_epi16(V2, V2); 41 return (uint32_t)_mm_cvtsi128_si32(b); 42 } 43 44 static WEBP_INLINE uint32_t ClampedAddSubtractHalf_SSE2(uint32_t c0, 45 uint32_t c1, 46 uint32_t c2) { 47 const __m128i zero = _mm_setzero_si128(); 48 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c0), zero); 49 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c1), zero); 50 const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c2), zero); 51 const __m128i avg = _mm_add_epi16(C1, C0); 52 const __m128i A0 = _mm_srli_epi16(avg, 1); 53 const __m128i A1 = _mm_sub_epi16(A0, B0); 54 const __m128i BgtA = _mm_cmpgt_epi16(B0, A0); 55 const __m128i A2 = _mm_sub_epi16(A1, BgtA); 56 const __m128i A3 = _mm_srai_epi16(A2, 1); 57 const __m128i A4 = _mm_add_epi16(A0, A3); 58 const __m128i A5 = _mm_packus_epi16(A4, A4); 59 return (uint32_t)_mm_cvtsi128_si32(A5); 60 } 61 62 static WEBP_INLINE uint32_t Select_SSE2(uint32_t a, uint32_t b, uint32_t c) { 63 int pa_minus_pb; 64 const __m128i zero = _mm_setzero_si128(); 65 const __m128i A0 = _mm_cvtsi32_si128((int)a); 66 const __m128i B0 = _mm_cvtsi32_si128((int)b); 67 const __m128i C0 = _mm_cvtsi32_si128((int)c); 68 const __m128i AC0 = _mm_subs_epu8(A0, C0); 69 const __m128i CA0 = _mm_subs_epu8(C0, A0); 70 const __m128i BC0 = _mm_subs_epu8(B0, C0); 71 const __m128i CB0 = _mm_subs_epu8(C0, B0); 72 const __m128i AC = _mm_or_si128(AC0, CA0); 73 const __m128i BC = _mm_or_si128(BC0, CB0); 74 const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c| 75 const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c| 76 const __m128i diff = _mm_sub_epi16(pb, pa); 77 { 78 int16_t out[8]; 79 _mm_storeu_si128((__m128i*)out, diff); 80 pa_minus_pb = out[0] + out[1] + out[2] + out[3]; 81 } 82 return (pa_minus_pb <= 0) ? a : b; 83 } 84 85 static WEBP_INLINE void Average2_m128i(const __m128i* const a0, 86 const __m128i* const a1, 87 __m128i* const avg) { 88 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1) 89 const __m128i ones = _mm_set1_epi8(1); 90 const __m128i avg1 = _mm_avg_epu8(*a0, *a1); 91 const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones); 92 *avg = _mm_sub_epi8(avg1, one); 93 } 94 95 static WEBP_INLINE void Average2_uint32_SSE2(const uint32_t a0, 96 const uint32_t a1, 97 __m128i* const avg) { 98 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1) 99 const __m128i ones = _mm_set1_epi8(1); 100 const __m128i A0 = _mm_cvtsi32_si128((int)a0); 101 const __m128i A1 = _mm_cvtsi32_si128((int)a1); 102 const __m128i avg1 = _mm_avg_epu8(A0, A1); 103 const __m128i one = _mm_and_si128(_mm_xor_si128(A0, A1), ones); 104 *avg = _mm_sub_epi8(avg1, one); 105 } 106 107 static WEBP_INLINE __m128i Average2_uint32_16_SSE2(uint32_t a0, uint32_t a1) { 108 const __m128i zero = _mm_setzero_si128(); 109 const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a0), zero); 110 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a1), zero); 111 const __m128i sum = _mm_add_epi16(A1, A0); 112 return _mm_srli_epi16(sum, 1); 113 } 114 115 static WEBP_INLINE uint32_t Average2_SSE2(uint32_t a0, uint32_t a1) { 116 __m128i output; 117 Average2_uint32_SSE2(a0, a1, &output); 118 return (uint32_t)_mm_cvtsi128_si32(output); 119 } 120 121 static WEBP_INLINE uint32_t Average3_SSE2(uint32_t a0, uint32_t a1, 122 uint32_t a2) { 123 const __m128i zero = _mm_setzero_si128(); 124 const __m128i avg1 = Average2_uint32_16_SSE2(a0, a2); 125 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a1), zero); 126 const __m128i sum = _mm_add_epi16(avg1, A1); 127 const __m128i avg2 = _mm_srli_epi16(sum, 1); 128 const __m128i A2 = _mm_packus_epi16(avg2, avg2); 129 return (uint32_t)_mm_cvtsi128_si32(A2); 130 } 131 132 static WEBP_INLINE uint32_t Average4_SSE2(uint32_t a0, uint32_t a1, 133 uint32_t a2, uint32_t a3) { 134 const __m128i avg1 = Average2_uint32_16_SSE2(a0, a1); 135 const __m128i avg2 = Average2_uint32_16_SSE2(a2, a3); 136 const __m128i sum = _mm_add_epi16(avg2, avg1); 137 const __m128i avg3 = _mm_srli_epi16(sum, 1); 138 const __m128i A0 = _mm_packus_epi16(avg3, avg3); 139 return (uint32_t)_mm_cvtsi128_si32(A0); 140 } 141 142 static uint32_t Predictor5_SSE2(const uint32_t* const left, 143 const uint32_t* const top) { 144 const uint32_t pred = Average3_SSE2(*left, top[0], top[1]); 145 return pred; 146 } 147 static uint32_t Predictor6_SSE2(const uint32_t* const left, 148 const uint32_t* const top) { 149 const uint32_t pred = Average2_SSE2(*left, top[-1]); 150 return pred; 151 } 152 static uint32_t Predictor7_SSE2(const uint32_t* const left, 153 const uint32_t* const top) { 154 const uint32_t pred = Average2_SSE2(*left, top[0]); 155 return pred; 156 } 157 static uint32_t Predictor8_SSE2(const uint32_t* const left, 158 const uint32_t* const top) { 159 const uint32_t pred = Average2_SSE2(top[-1], top[0]); 160 (void)left; 161 return pred; 162 } 163 static uint32_t Predictor9_SSE2(const uint32_t* const left, 164 const uint32_t* const top) { 165 const uint32_t pred = Average2_SSE2(top[0], top[1]); 166 (void)left; 167 return pred; 168 } 169 static uint32_t Predictor10_SSE2(const uint32_t* const left, 170 const uint32_t* const top) { 171 const uint32_t pred = Average4_SSE2(*left, top[-1], top[0], top[1]); 172 return pred; 173 } 174 static uint32_t Predictor11_SSE2(const uint32_t* const left, 175 const uint32_t* const top) { 176 const uint32_t pred = Select_SSE2(top[0], *left, top[-1]); 177 return pred; 178 } 179 static uint32_t Predictor12_SSE2(const uint32_t* const left, 180 const uint32_t* const top) { 181 const uint32_t pred = ClampedAddSubtractFull_SSE2(*left, top[0], top[-1]); 182 return pred; 183 } 184 static uint32_t Predictor13_SSE2(const uint32_t* const left, 185 const uint32_t* const top) { 186 const uint32_t pred = ClampedAddSubtractHalf_SSE2(*left, top[0], top[-1]); 187 return pred; 188 } 189 190 // Batch versions of those functions. 191 192 // Predictor0: ARGB_BLACK. 193 static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper, 194 int num_pixels, uint32_t* WEBP_RESTRICT out) { 195 int i; 196 const __m128i black = _mm_set1_epi32((int)ARGB_BLACK); 197 for (i = 0; i + 4 <= num_pixels; i += 4) { 198 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 199 const __m128i res = _mm_add_epi8(src, black); 200 _mm_storeu_si128((__m128i*)&out[i], res); 201 } 202 if (i != num_pixels) { 203 VP8LPredictorsAdd_C[0](in + i, NULL, num_pixels - i, out + i); 204 } 205 (void)upper; 206 } 207 208 // Predictor1: left. 209 static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper, 210 int num_pixels, uint32_t* WEBP_RESTRICT out) { 211 int i; 212 __m128i prev = _mm_set1_epi32((int)out[-1]); 213 for (i = 0; i + 4 <= num_pixels; i += 4) { 214 // a | b | c | d 215 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 216 // 0 | a | b | c 217 const __m128i shift0 = _mm_slli_si128(src, 4); 218 // a | a + b | b + c | c + d 219 const __m128i sum0 = _mm_add_epi8(src, shift0); 220 // 0 | 0 | a | a + b 221 const __m128i shift1 = _mm_slli_si128(sum0, 8); 222 // a | a + b | a + b + c | a + b + c + d 223 const __m128i sum1 = _mm_add_epi8(sum0, shift1); 224 const __m128i res = _mm_add_epi8(sum1, prev); 225 _mm_storeu_si128((__m128i*)&out[i], res); 226 // replicate prev output on the four lanes 227 prev = _mm_shuffle_epi32(res, (3 << 0) | (3 << 2) | (3 << 4) | (3 << 6)); 228 } 229 if (i != num_pixels) { 230 VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i); 231 } 232 } 233 234 // Macro that adds 32-bit integers from IN using mod 256 arithmetic 235 // per 8 bit channel. 236 #define GENERATE_PREDICTOR_1(X, IN) \ 237 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \ 238 int num_pixels, \ 239 uint32_t* WEBP_RESTRICT out) { \ 240 int i; \ 241 for (i = 0; i + 4 <= num_pixels; i += 4) { \ 242 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ 243 const __m128i other = _mm_loadu_si128((const __m128i*)&(IN)); \ 244 const __m128i res = _mm_add_epi8(src, other); \ 245 _mm_storeu_si128((__m128i*)&out[i], res); \ 246 } \ 247 if (i != num_pixels) { \ 248 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \ 249 } \ 250 } 251 252 // Predictor2: Top. 253 GENERATE_PREDICTOR_1(2, upper[i]) 254 // Predictor3: Top-right. 255 GENERATE_PREDICTOR_1(3, upper[i + 1]) 256 // Predictor4: Top-left. 257 GENERATE_PREDICTOR_1(4, upper[i - 1]) 258 #undef GENERATE_PREDICTOR_1 259 260 // Due to averages with integers, values cannot be accumulated in parallel for 261 // predictors 5 to 7. 262 GENERATE_PREDICTOR_ADD(Predictor5_SSE2, PredictorAdd5_SSE2) 263 GENERATE_PREDICTOR_ADD(Predictor6_SSE2, PredictorAdd6_SSE2) 264 GENERATE_PREDICTOR_ADD(Predictor7_SSE2, PredictorAdd7_SSE2) 265 266 #define GENERATE_PREDICTOR_2(X, IN) \ 267 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \ 268 int num_pixels, \ 269 uint32_t* WEBP_RESTRICT out) { \ 270 int i; \ 271 for (i = 0; i + 4 <= num_pixels; i += 4) { \ 272 const __m128i Tother = _mm_loadu_si128((const __m128i*)&(IN)); \ 273 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); \ 274 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ 275 __m128i avg, res; \ 276 Average2_m128i(&T, &Tother, &avg); \ 277 res = _mm_add_epi8(avg, src); \ 278 _mm_storeu_si128((__m128i*)&out[i], res); \ 279 } \ 280 if (i != num_pixels) { \ 281 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \ 282 } \ 283 } 284 // Predictor8: average TL T. 285 GENERATE_PREDICTOR_2(8, upper[i - 1]) 286 // Predictor9: average T TR. 287 GENERATE_PREDICTOR_2(9, upper[i + 1]) 288 #undef GENERATE_PREDICTOR_2 289 290 // Predictor10: average of (average of (L,TL), average of (T, TR)). 291 #define DO_PRED10(OUT) do { \ 292 __m128i avgLTL, avg; \ 293 Average2_m128i(&L, &TL, &avgLTL); \ 294 Average2_m128i(&avgTTR, &avgLTL, &avg); \ 295 L = _mm_add_epi8(avg, src); \ 296 out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(L); \ 297 } while (0) 298 299 #define DO_PRED10_SHIFT do { \ 300 /* Rotate the pre-computed values for the next iteration.*/ \ 301 avgTTR = _mm_srli_si128(avgTTR, 4); \ 302 TL = _mm_srli_si128(TL, 4); \ 303 src = _mm_srli_si128(src, 4); \ 304 } while (0) 305 306 static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper, 307 int num_pixels, uint32_t* WEBP_RESTRICT out) { 308 int i; 309 __m128i L = _mm_cvtsi32_si128((int)out[-1]); 310 for (i = 0; i + 4 <= num_pixels; i += 4) { 311 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 312 __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); 313 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); 314 const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]); 315 __m128i avgTTR; 316 Average2_m128i(&T, &TR, &avgTTR); 317 DO_PRED10(0); 318 DO_PRED10_SHIFT; 319 DO_PRED10(1); 320 DO_PRED10_SHIFT; 321 DO_PRED10(2); 322 DO_PRED10_SHIFT; 323 DO_PRED10(3); 324 } 325 if (i != num_pixels) { 326 VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i); 327 } 328 } 329 #undef DO_PRED10 330 #undef DO_PRED10_SHIFT 331 332 // Predictor11: select. 333 #define DO_PRED11(OUT) do { \ 334 const __m128i L_lo = _mm_unpacklo_epi32(L, T); \ 335 const __m128i TL_lo = _mm_unpacklo_epi32(TL, T); \ 336 const __m128i pb = _mm_sad_epu8(L_lo, TL_lo); /* pb = sum |L-TL|*/ \ 337 const __m128i mask = _mm_cmpgt_epi32(pb, pa); \ 338 const __m128i A = _mm_and_si128(mask, L); \ 339 const __m128i B = _mm_andnot_si128(mask, T); \ 340 const __m128i pred = _mm_or_si128(A, B); /* pred = (pa > b)? L : T*/ \ 341 L = _mm_add_epi8(src, pred); \ 342 out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(L); \ 343 } while (0) 344 345 #define DO_PRED11_SHIFT do { \ 346 /* Shift the pre-computed value for the next iteration.*/ \ 347 T = _mm_srli_si128(T, 4); \ 348 TL = _mm_srli_si128(TL, 4); \ 349 src = _mm_srli_si128(src, 4); \ 350 pa = _mm_srli_si128(pa, 4); \ 351 } while (0) 352 353 static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper, 354 int num_pixels, uint32_t* WEBP_RESTRICT out) { 355 int i; 356 __m128i pa; 357 __m128i L = _mm_cvtsi32_si128((int)out[-1]); 358 for (i = 0; i + 4 <= num_pixels; i += 4) { 359 __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); 360 __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); 361 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 362 { 363 // We can unpack with any value on the upper 32 bits, provided it's the 364 // same on both operands (so that their sum of abs diff is zero). Here we 365 // use T. 366 const __m128i T_lo = _mm_unpacklo_epi32(T, T); 367 const __m128i TL_lo = _mm_unpacklo_epi32(TL, T); 368 const __m128i T_hi = _mm_unpackhi_epi32(T, T); 369 const __m128i TL_hi = _mm_unpackhi_epi32(TL, T); 370 const __m128i s_lo = _mm_sad_epu8(T_lo, TL_lo); 371 const __m128i s_hi = _mm_sad_epu8(T_hi, TL_hi); 372 pa = _mm_packs_epi32(s_lo, s_hi); // pa = sum |T-TL| 373 } 374 DO_PRED11(0); 375 DO_PRED11_SHIFT; 376 DO_PRED11(1); 377 DO_PRED11_SHIFT; 378 DO_PRED11(2); 379 DO_PRED11_SHIFT; 380 DO_PRED11(3); 381 } 382 if (i != num_pixels) { 383 VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i); 384 } 385 } 386 #undef DO_PRED11 387 #undef DO_PRED11_SHIFT 388 389 // Predictor12: ClampedAddSubtractFull. 390 #define DO_PRED12(DIFF, LANE, OUT) do { \ 391 const __m128i all = _mm_add_epi16(L, (DIFF)); \ 392 const __m128i alls = _mm_packus_epi16(all, all); \ 393 const __m128i res = _mm_add_epi8(src, alls); \ 394 out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(res); \ 395 L = _mm_unpacklo_epi8(res, zero); \ 396 } while (0) 397 398 #define DO_PRED12_SHIFT(DIFF, LANE) do { \ 399 /* Shift the pre-computed value for the next iteration.*/ \ 400 if ((LANE) == 0) (DIFF) = _mm_srli_si128((DIFF), 8); \ 401 src = _mm_srli_si128(src, 4); \ 402 } while (0) 403 404 static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper, 405 int num_pixels, uint32_t* WEBP_RESTRICT out) { 406 int i; 407 const __m128i zero = _mm_setzero_si128(); 408 const __m128i L8 = _mm_cvtsi32_si128((int)out[-1]); 409 __m128i L = _mm_unpacklo_epi8(L8, zero); 410 for (i = 0; i + 4 <= num_pixels; i += 4) { 411 // Load 4 pixels at a time. 412 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 413 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); 414 const __m128i T_lo = _mm_unpacklo_epi8(T, zero); 415 const __m128i T_hi = _mm_unpackhi_epi8(T, zero); 416 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); 417 const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero); 418 const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero); 419 __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo); 420 __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi); 421 DO_PRED12(diff_lo, 0, 0); 422 DO_PRED12_SHIFT(diff_lo, 0); 423 DO_PRED12(diff_lo, 1, 1); 424 DO_PRED12_SHIFT(diff_lo, 1); 425 DO_PRED12(diff_hi, 0, 2); 426 DO_PRED12_SHIFT(diff_hi, 0); 427 DO_PRED12(diff_hi, 1, 3); 428 } 429 if (i != num_pixels) { 430 VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i); 431 } 432 } 433 #undef DO_PRED12 434 #undef DO_PRED12_SHIFT 435 436 // Due to averages with integers, values cannot be accumulated in parallel for 437 // predictors 13. 438 GENERATE_PREDICTOR_ADD(Predictor13_SSE2, PredictorAdd13_SSE2) 439 440 //------------------------------------------------------------------------------ 441 // Subtract-Green Transform 442 443 static void AddGreenToBlueAndRed_SSE2(const uint32_t* const src, int num_pixels, 444 uint32_t* dst) { 445 int i; 446 for (i = 0; i + 4 <= num_pixels; i += 4) { 447 const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb 448 const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g 449 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); 450 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g 451 const __m128i out = _mm_add_epi8(in, C); 452 _mm_storeu_si128((__m128i*)&dst[i], out); 453 } 454 // fallthrough and finish off with plain-C 455 if (i != num_pixels) { 456 VP8LAddGreenToBlueAndRed_C(src + i, num_pixels - i, dst + i); 457 } 458 } 459 460 //------------------------------------------------------------------------------ 461 // Color Transform 462 463 static void TransformColorInverse_SSE2(const VP8LMultipliers* const m, 464 const uint32_t* const src, 465 int num_pixels, uint32_t* dst) { 466 // sign-extended multiplying constants, pre-shifted by 5. 467 #define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend 468 #define MK_CST_16(HI, LO) \ 469 _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff))) 470 const __m128i mults_rb = MK_CST_16(CST(green_to_red), CST(green_to_blue)); 471 const __m128i mults_b2 = MK_CST_16(CST(red_to_blue), 0); 472 #undef MK_CST_16 473 #undef CST 474 const __m128i mask_ag = _mm_set1_epi32((int)0xff00ff00); // alpha-green masks 475 int i; 476 for (i = 0; i + 4 <= num_pixels; i += 4) { 477 const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb 478 const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0 479 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); 480 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0 481 const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1 482 const __m128i E = _mm_add_epi8(in, D); // x r' x b' 483 const __m128i F = _mm_slli_epi16(E, 8); // r' 0 b' 0 484 const __m128i G = _mm_mulhi_epi16(F, mults_b2); // x db2 0 0 485 const __m128i H = _mm_srli_epi32(G, 8); // 0 x db2 0 486 const __m128i I = _mm_add_epi8(H, F); // r' x b'' 0 487 const __m128i J = _mm_srli_epi16(I, 8); // 0 r' 0 b'' 488 const __m128i out = _mm_or_si128(J, A); 489 _mm_storeu_si128((__m128i*)&dst[i], out); 490 } 491 // Fall-back to C-version for left-overs. 492 if (i != num_pixels) { 493 VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i); 494 } 495 } 496 497 //------------------------------------------------------------------------------ 498 // Color-space conversion functions 499 500 static void ConvertBGRAToRGB_SSE2(const uint32_t* WEBP_RESTRICT src, 501 int num_pixels, uint8_t* WEBP_RESTRICT dst) { 502 const __m128i* in = (const __m128i*)src; 503 __m128i* out = (__m128i*)dst; 504 505 while (num_pixels >= 32) { 506 // Load the BGRA buffers. 507 __m128i in0 = _mm_loadu_si128(in + 0); 508 __m128i in1 = _mm_loadu_si128(in + 1); 509 __m128i in2 = _mm_loadu_si128(in + 2); 510 __m128i in3 = _mm_loadu_si128(in + 3); 511 __m128i in4 = _mm_loadu_si128(in + 4); 512 __m128i in5 = _mm_loadu_si128(in + 5); 513 __m128i in6 = _mm_loadu_si128(in + 6); 514 __m128i in7 = _mm_loadu_si128(in + 7); 515 VP8L32bToPlanar_SSE2(&in0, &in1, &in2, &in3); 516 VP8L32bToPlanar_SSE2(&in4, &in5, &in6, &in7); 517 // At this points, in1/in5 contains red only, in2/in6 green only ... 518 // Pack the colors in 24b RGB. 519 VP8PlanarTo24b_SSE2(&in1, &in5, &in2, &in6, &in3, &in7); 520 _mm_storeu_si128(out + 0, in1); 521 _mm_storeu_si128(out + 1, in5); 522 _mm_storeu_si128(out + 2, in2); 523 _mm_storeu_si128(out + 3, in6); 524 _mm_storeu_si128(out + 4, in3); 525 _mm_storeu_si128(out + 5, in7); 526 in += 8; 527 out += 6; 528 num_pixels -= 32; 529 } 530 // left-overs 531 if (num_pixels > 0) { 532 VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 533 } 534 } 535 536 static void ConvertBGRAToRGBA_SSE2(const uint32_t* WEBP_RESTRICT src, 537 int num_pixels, uint8_t* WEBP_RESTRICT dst) { 538 const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ff); 539 const __m128i* in = (const __m128i*)src; 540 __m128i* out = (__m128i*)dst; 541 while (num_pixels >= 8) { 542 const __m128i A1 = _mm_loadu_si128(in++); 543 const __m128i A2 = _mm_loadu_si128(in++); 544 const __m128i B1 = _mm_and_si128(A1, red_blue_mask); // R 0 B 0 545 const __m128i B2 = _mm_and_si128(A2, red_blue_mask); // R 0 B 0 546 const __m128i C1 = _mm_andnot_si128(red_blue_mask, A1); // 0 G 0 A 547 const __m128i C2 = _mm_andnot_si128(red_blue_mask, A2); // 0 G 0 A 548 const __m128i D1 = _mm_shufflelo_epi16(B1, _MM_SHUFFLE(2, 3, 0, 1)); 549 const __m128i D2 = _mm_shufflelo_epi16(B2, _MM_SHUFFLE(2, 3, 0, 1)); 550 const __m128i E1 = _mm_shufflehi_epi16(D1, _MM_SHUFFLE(2, 3, 0, 1)); 551 const __m128i E2 = _mm_shufflehi_epi16(D2, _MM_SHUFFLE(2, 3, 0, 1)); 552 const __m128i F1 = _mm_or_si128(E1, C1); 553 const __m128i F2 = _mm_or_si128(E2, C2); 554 _mm_storeu_si128(out++, F1); 555 _mm_storeu_si128(out++, F2); 556 num_pixels -= 8; 557 } 558 // left-overs 559 if (num_pixels > 0) { 560 VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 561 } 562 } 563 564 static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* WEBP_RESTRICT src, 565 int num_pixels, 566 uint8_t* WEBP_RESTRICT dst) { 567 const __m128i mask_0x0f = _mm_set1_epi8(0x0f); 568 const __m128i mask_0xf0 = _mm_set1_epi8((char)0xf0); 569 const __m128i* in = (const __m128i*)src; 570 __m128i* out = (__m128i*)dst; 571 while (num_pixels >= 8) { 572 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 573 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 574 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 575 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 576 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 577 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 578 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 579 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 580 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 581 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 582 const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7- 583 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7 584 const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7- 585 const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7 586 const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0 587 #if (WEBP_SWAP_16BIT_CSP == 1) 588 const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7 589 #else 590 const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7 591 #endif 592 _mm_storeu_si128(out++, rgba); 593 num_pixels -= 8; 594 } 595 // left-overs 596 if (num_pixels > 0) { 597 VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 598 } 599 } 600 601 static void ConvertBGRAToRGB565_SSE2(const uint32_t* WEBP_RESTRICT src, 602 int num_pixels, 603 uint8_t* WEBP_RESTRICT dst) { 604 const __m128i mask_0xe0 = _mm_set1_epi8((char)0xe0); 605 const __m128i mask_0xf8 = _mm_set1_epi8((char)0xf8); 606 const __m128i mask_0x07 = _mm_set1_epi8(0x07); 607 const __m128i* in = (const __m128i*)src; 608 __m128i* out = (__m128i*)dst; 609 while (num_pixels >= 8) { 610 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 611 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 612 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 613 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 614 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 615 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 616 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 617 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 618 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 619 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 620 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7 621 const __m128i g_lo1 = _mm_srli_epi16(ga0, 5); 622 const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b) 623 const __m128i g_hi1 = _mm_slli_epi16(ga0, 3); 624 const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b) 625 const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0 626 const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx 627 const __m128i b1 = _mm_srli_epi16(b0, 3); 628 const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx 629 #if (WEBP_SWAP_16BIT_CSP == 1) 630 const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7 631 #else 632 const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7 633 #endif 634 _mm_storeu_si128(out++, rgba); 635 num_pixels -= 8; 636 } 637 // left-overs 638 if (num_pixels > 0) { 639 VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 640 } 641 } 642 643 static void ConvertBGRAToBGR_SSE2(const uint32_t* WEBP_RESTRICT src, 644 int num_pixels, uint8_t* WEBP_RESTRICT dst) { 645 const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff); 646 const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0); 647 const __m128i* in = (const __m128i*)src; 648 const uint8_t* const end = dst + num_pixels * 3; 649 // the last storel_epi64 below writes 8 bytes starting at offset 18 650 while (dst + 26 <= end) { 651 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 652 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 653 const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0 654 const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0 655 const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0 656 const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0 657 const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00 658 const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00 659 const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00 660 const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00 661 const __m128i c2 = _mm_srli_si128(c0, 8); 662 const __m128i c6 = _mm_srli_si128(c4, 8); 663 _mm_storel_epi64((__m128i*)(dst + 0), c0); 664 _mm_storel_epi64((__m128i*)(dst + 6), c2); 665 _mm_storel_epi64((__m128i*)(dst + 12), c4); 666 _mm_storel_epi64((__m128i*)(dst + 18), c6); 667 dst += 24; 668 num_pixels -= 8; 669 } 670 // left-overs 671 if (num_pixels > 0) { 672 VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst); 673 } 674 } 675 676 //------------------------------------------------------------------------------ 677 // Entry point 678 679 extern void VP8LDspInitSSE2(void); 680 681 WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE2(void) { 682 VP8LPredictors[5] = Predictor5_SSE2; 683 VP8LPredictors[6] = Predictor6_SSE2; 684 VP8LPredictors[7] = Predictor7_SSE2; 685 VP8LPredictors[8] = Predictor8_SSE2; 686 VP8LPredictors[9] = Predictor9_SSE2; 687 VP8LPredictors[10] = Predictor10_SSE2; 688 VP8LPredictors[11] = Predictor11_SSE2; 689 VP8LPredictors[12] = Predictor12_SSE2; 690 VP8LPredictors[13] = Predictor13_SSE2; 691 692 VP8LPredictorsAdd[0] = PredictorAdd0_SSE2; 693 VP8LPredictorsAdd[1] = PredictorAdd1_SSE2; 694 VP8LPredictorsAdd[2] = PredictorAdd2_SSE2; 695 VP8LPredictorsAdd[3] = PredictorAdd3_SSE2; 696 VP8LPredictorsAdd[4] = PredictorAdd4_SSE2; 697 VP8LPredictorsAdd[5] = PredictorAdd5_SSE2; 698 VP8LPredictorsAdd[6] = PredictorAdd6_SSE2; 699 VP8LPredictorsAdd[7] = PredictorAdd7_SSE2; 700 VP8LPredictorsAdd[8] = PredictorAdd8_SSE2; 701 VP8LPredictorsAdd[9] = PredictorAdd9_SSE2; 702 VP8LPredictorsAdd[10] = PredictorAdd10_SSE2; 703 VP8LPredictorsAdd[11] = PredictorAdd11_SSE2; 704 VP8LPredictorsAdd[12] = PredictorAdd12_SSE2; 705 VP8LPredictorsAdd[13] = PredictorAdd13_SSE2; 706 707 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_SSE2; 708 VP8LTransformColorInverse = TransformColorInverse_SSE2; 709 710 VP8LConvertBGRAToRGB = ConvertBGRAToRGB_SSE2; 711 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_SSE2; 712 VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444_SSE2; 713 VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565_SSE2; 714 VP8LConvertBGRAToBGR = ConvertBGRAToBGR_SSE2; 715 716 // SSE exports for AVX and above. 717 memcpy(VP8LPredictorsAdd_SSE, VP8LPredictorsAdd, sizeof(VP8LPredictorsAdd)); 718 719 VP8LAddGreenToBlueAndRed_SSE = AddGreenToBlueAndRed_SSE2; 720 VP8LTransformColorInverse_SSE = TransformColorInverse_SSE2; 721 722 VP8LConvertBGRAToRGB_SSE = ConvertBGRAToRGB_SSE2; 723 VP8LConvertBGRAToRGBA_SSE = ConvertBGRAToRGBA_SSE2; 724 } 725 726 #else // !WEBP_USE_SSE2 727 728 WEBP_DSP_INIT_STUB(VP8LDspInitSSE2) 729 730 #endif // WEBP_USE_SSE2