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 (at) gmail.com) 13 14 #include "./dsp.h" 15 16 #if defined(WEBP_USE_SSE2) 17 18 #include "./common_sse2.h" 19 #include "./lossless.h" 20 #include "./lossless_common.h" 21 #include <assert.h> 22 #include <emmintrin.h> 23 24 //------------------------------------------------------------------------------ 25 // Predictor Transform 26 27 static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, 28 uint32_t c2) { 29 const __m128i zero = _mm_setzero_si128(); 30 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); 31 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); 32 const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); 33 const __m128i V1 = _mm_add_epi16(C0, C1); 34 const __m128i V2 = _mm_sub_epi16(V1, C2); 35 const __m128i b = _mm_packus_epi16(V2, V2); 36 const uint32_t output = _mm_cvtsi128_si32(b); 37 return output; 38 } 39 40 static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, 41 uint32_t c2) { 42 const __m128i zero = _mm_setzero_si128(); 43 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); 44 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); 45 const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); 46 const __m128i avg = _mm_add_epi16(C1, C0); 47 const __m128i A0 = _mm_srli_epi16(avg, 1); 48 const __m128i A1 = _mm_sub_epi16(A0, B0); 49 const __m128i BgtA = _mm_cmpgt_epi16(B0, A0); 50 const __m128i A2 = _mm_sub_epi16(A1, BgtA); 51 const __m128i A3 = _mm_srai_epi16(A2, 1); 52 const __m128i A4 = _mm_add_epi16(A0, A3); 53 const __m128i A5 = _mm_packus_epi16(A4, A4); 54 const uint32_t output = _mm_cvtsi128_si32(A5); 55 return output; 56 } 57 58 static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { 59 int pa_minus_pb; 60 const __m128i zero = _mm_setzero_si128(); 61 const __m128i A0 = _mm_cvtsi32_si128(a); 62 const __m128i B0 = _mm_cvtsi32_si128(b); 63 const __m128i C0 = _mm_cvtsi32_si128(c); 64 const __m128i AC0 = _mm_subs_epu8(A0, C0); 65 const __m128i CA0 = _mm_subs_epu8(C0, A0); 66 const __m128i BC0 = _mm_subs_epu8(B0, C0); 67 const __m128i CB0 = _mm_subs_epu8(C0, B0); 68 const __m128i AC = _mm_or_si128(AC0, CA0); 69 const __m128i BC = _mm_or_si128(BC0, CB0); 70 const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c| 71 const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c| 72 const __m128i diff = _mm_sub_epi16(pb, pa); 73 { 74 int16_t out[8]; 75 _mm_storeu_si128((__m128i*)out, diff); 76 pa_minus_pb = out[0] + out[1] + out[2] + out[3]; 77 } 78 return (pa_minus_pb <= 0) ? a : b; 79 } 80 81 static WEBP_INLINE void Average2_m128i(const __m128i* const a0, 82 const __m128i* const a1, 83 __m128i* const avg) { 84 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1) 85 const __m128i ones = _mm_set1_epi8(1); 86 const __m128i avg1 = _mm_avg_epu8(*a0, *a1); 87 const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones); 88 *avg = _mm_sub_epi8(avg1, one); 89 } 90 91 static WEBP_INLINE void Average2_uint32(const uint32_t a0, const uint32_t a1, 92 __m128i* const avg) { 93 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1) 94 const __m128i ones = _mm_set1_epi8(1); 95 const __m128i A0 = _mm_cvtsi32_si128(a0); 96 const __m128i A1 = _mm_cvtsi32_si128(a1); 97 const __m128i avg1 = _mm_avg_epu8(A0, A1); 98 const __m128i one = _mm_and_si128(_mm_xor_si128(A0, A1), ones); 99 *avg = _mm_sub_epi8(avg1, one); 100 } 101 102 static WEBP_INLINE __m128i Average2_uint32_16(uint32_t a0, uint32_t a1) { 103 const __m128i zero = _mm_setzero_si128(); 104 const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero); 105 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); 106 const __m128i sum = _mm_add_epi16(A1, A0); 107 return _mm_srli_epi16(sum, 1); 108 } 109 110 static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { 111 __m128i output; 112 Average2_uint32(a0, a1, &output); 113 return _mm_cvtsi128_si32(output); 114 } 115 116 static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { 117 const __m128i zero = _mm_setzero_si128(); 118 const __m128i avg1 = Average2_uint32_16(a0, a2); 119 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); 120 const __m128i sum = _mm_add_epi16(avg1, A1); 121 const __m128i avg2 = _mm_srli_epi16(sum, 1); 122 const __m128i A2 = _mm_packus_epi16(avg2, avg2); 123 const uint32_t output = _mm_cvtsi128_si32(A2); 124 return output; 125 } 126 127 static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1, 128 uint32_t a2, uint32_t a3) { 129 const __m128i avg1 = Average2_uint32_16(a0, a1); 130 const __m128i avg2 = Average2_uint32_16(a2, a3); 131 const __m128i sum = _mm_add_epi16(avg2, avg1); 132 const __m128i avg3 = _mm_srli_epi16(sum, 1); 133 const __m128i A0 = _mm_packus_epi16(avg3, avg3); 134 const uint32_t output = _mm_cvtsi128_si32(A0); 135 return output; 136 } 137 138 static uint32_t Predictor5_SSE2(uint32_t left, const uint32_t* const top) { 139 const uint32_t pred = Average3(left, top[0], top[1]); 140 return pred; 141 } 142 static uint32_t Predictor6_SSE2(uint32_t left, const uint32_t* const top) { 143 const uint32_t pred = Average2(left, top[-1]); 144 return pred; 145 } 146 static uint32_t Predictor7_SSE2(uint32_t left, const uint32_t* const top) { 147 const uint32_t pred = Average2(left, top[0]); 148 return pred; 149 } 150 static uint32_t Predictor8_SSE2(uint32_t left, const uint32_t* const top) { 151 const uint32_t pred = Average2(top[-1], top[0]); 152 (void)left; 153 return pred; 154 } 155 static uint32_t Predictor9_SSE2(uint32_t left, const uint32_t* const top) { 156 const uint32_t pred = Average2(top[0], top[1]); 157 (void)left; 158 return pred; 159 } 160 static uint32_t Predictor10_SSE2(uint32_t left, const uint32_t* const top) { 161 const uint32_t pred = Average4(left, top[-1], top[0], top[1]); 162 return pred; 163 } 164 static uint32_t Predictor11_SSE2(uint32_t left, const uint32_t* const top) { 165 const uint32_t pred = Select(top[0], left, top[-1]); 166 return pred; 167 } 168 static uint32_t Predictor12_SSE2(uint32_t left, const uint32_t* const top) { 169 const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); 170 return pred; 171 } 172 static uint32_t Predictor13_SSE2(uint32_t left, const uint32_t* const top) { 173 const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); 174 return pred; 175 } 176 177 // Batch versions of those functions. 178 179 // Predictor0: ARGB_BLACK. 180 static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper, 181 int num_pixels, uint32_t* out) { 182 int i; 183 const __m128i black = _mm_set1_epi32(ARGB_BLACK); 184 for (i = 0; i + 4 <= num_pixels; i += 4) { 185 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 186 const __m128i res = _mm_add_epi8(src, black); 187 _mm_storeu_si128((__m128i*)&out[i], res); 188 } 189 if (i != num_pixels) { 190 VP8LPredictorsAdd_C[0](in + i, upper + i, num_pixels - i, out + i); 191 } 192 } 193 194 // Predictor1: left. 195 static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper, 196 int num_pixels, uint32_t* out) { 197 int i; 198 __m128i prev = _mm_set1_epi32(out[-1]); 199 for (i = 0; i + 4 <= num_pixels; i += 4) { 200 // a | b | c | d 201 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 202 // 0 | a | b | c 203 const __m128i shift0 = _mm_slli_si128(src, 4); 204 // a | a + b | b + c | c + d 205 const __m128i sum0 = _mm_add_epi8(src, shift0); 206 // 0 | 0 | a | a + b 207 const __m128i shift1 = _mm_slli_si128(sum0, 8); 208 // a | a + b | a + b + c | a + b + c + d 209 const __m128i sum1 = _mm_add_epi8(sum0, shift1); 210 const __m128i res = _mm_add_epi8(sum1, prev); 211 _mm_storeu_si128((__m128i*)&out[i], res); 212 // replicate prev output on the four lanes 213 prev = _mm_shuffle_epi32(res, (3 << 0) | (3 << 2) | (3 << 4) | (3 << 6)); 214 } 215 if (i != num_pixels) { 216 VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i); 217 } 218 } 219 220 // Macro that adds 32-bit integers from IN using mod 256 arithmetic 221 // per 8 bit channel. 222 #define GENERATE_PREDICTOR_1(X, IN) \ 223 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \ 224 int num_pixels, uint32_t* out) { \ 225 int i; \ 226 for (i = 0; i + 4 <= num_pixels; i += 4) { \ 227 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ 228 const __m128i other = _mm_loadu_si128((const __m128i*)&(IN)); \ 229 const __m128i res = _mm_add_epi8(src, other); \ 230 _mm_storeu_si128((__m128i*)&out[i], res); \ 231 } \ 232 if (i != num_pixels) { \ 233 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \ 234 } \ 235 } 236 237 // Predictor2: Top. 238 GENERATE_PREDICTOR_1(2, upper[i]) 239 // Predictor3: Top-right. 240 GENERATE_PREDICTOR_1(3, upper[i + 1]) 241 // Predictor4: Top-left. 242 GENERATE_PREDICTOR_1(4, upper[i - 1]) 243 #undef GENERATE_PREDICTOR_1 244 245 // Due to averages with integers, values cannot be accumulated in parallel for 246 // predictors 5 to 7. 247 GENERATE_PREDICTOR_ADD(Predictor5_SSE2, PredictorAdd5_SSE2) 248 GENERATE_PREDICTOR_ADD(Predictor6_SSE2, PredictorAdd6_SSE2) 249 GENERATE_PREDICTOR_ADD(Predictor7_SSE2, PredictorAdd7_SSE2) 250 251 #define GENERATE_PREDICTOR_2(X, IN) \ 252 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \ 253 int num_pixels, uint32_t* out) { \ 254 int i; \ 255 for (i = 0; i + 4 <= num_pixels; i += 4) { \ 256 const __m128i Tother = _mm_loadu_si128((const __m128i*)&(IN)); \ 257 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); \ 258 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \ 259 __m128i avg, res; \ 260 Average2_m128i(&T, &Tother, &avg); \ 261 res = _mm_add_epi8(avg, src); \ 262 _mm_storeu_si128((__m128i*)&out[i], res); \ 263 } \ 264 if (i != num_pixels) { \ 265 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \ 266 } \ 267 } 268 // Predictor8: average TL T. 269 GENERATE_PREDICTOR_2(8, upper[i - 1]) 270 // Predictor9: average T TR. 271 GENERATE_PREDICTOR_2(9, upper[i + 1]) 272 #undef GENERATE_PREDICTOR_2 273 274 // Predictor10: average of (average of (L,TL), average of (T, TR)). 275 static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper, 276 int num_pixels, uint32_t* out) { 277 int i, j; 278 __m128i L = _mm_cvtsi32_si128(out[-1]); 279 for (i = 0; i + 4 <= num_pixels; i += 4) { 280 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 281 __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); 282 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); 283 const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]); 284 __m128i avgTTR; 285 Average2_m128i(&T, &TR, &avgTTR); 286 for (j = 0; j < 4; ++j) { 287 __m128i avgLTL, avg; 288 Average2_m128i(&L, &TL, &avgLTL); 289 Average2_m128i(&avgTTR, &avgLTL, &avg); 290 L = _mm_add_epi8(avg, src); 291 out[i + j] = _mm_cvtsi128_si32(L); 292 // Rotate the pre-computed values for the next iteration. 293 avgTTR = _mm_srli_si128(avgTTR, 4); 294 TL = _mm_srli_si128(TL, 4); 295 src = _mm_srli_si128(src, 4); 296 } 297 } 298 if (i != num_pixels) { 299 VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i); 300 } 301 } 302 303 // Predictor11: select. 304 static void GetSumAbsDiff32(const __m128i* const A, const __m128i* const B, 305 __m128i* const out) { 306 // We can unpack with any value on the upper 32 bits, provided it's the same 307 // on both operands (to that their sum of abs diff is zero). Here we use *A. 308 const __m128i A_lo = _mm_unpacklo_epi32(*A, *A); 309 const __m128i B_lo = _mm_unpacklo_epi32(*B, *A); 310 const __m128i A_hi = _mm_unpackhi_epi32(*A, *A); 311 const __m128i B_hi = _mm_unpackhi_epi32(*B, *A); 312 const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo); 313 const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi); 314 *out = _mm_packs_epi32(s_lo, s_hi); 315 } 316 317 static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper, 318 int num_pixels, uint32_t* out) { 319 int i, j; 320 __m128i L = _mm_cvtsi32_si128(out[-1]); 321 for (i = 0; i + 4 <= num_pixels; i += 4) { 322 __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); 323 __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); 324 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 325 __m128i pa; 326 GetSumAbsDiff32(&T, &TL, &pa); // pa = sum |T-TL| 327 for (j = 0; j < 4; ++j) { 328 const __m128i L_lo = _mm_unpacklo_epi32(L, L); 329 const __m128i TL_lo = _mm_unpacklo_epi32(TL, L); 330 const __m128i pb = _mm_sad_epu8(L_lo, TL_lo); // pb = sum |L-TL| 331 const __m128i mask = _mm_cmpgt_epi32(pb, pa); 332 const __m128i A = _mm_and_si128(mask, L); 333 const __m128i B = _mm_andnot_si128(mask, T); 334 const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T 335 L = _mm_add_epi8(src, pred); 336 out[i + j] = _mm_cvtsi128_si32(L); 337 // Shift the pre-computed value for the next iteration. 338 T = _mm_srli_si128(T, 4); 339 TL = _mm_srli_si128(TL, 4); 340 src = _mm_srli_si128(src, 4); 341 pa = _mm_srli_si128(pa, 4); 342 } 343 } 344 if (i != num_pixels) { 345 VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i); 346 } 347 } 348 349 // Predictor12: ClampedAddSubtractFull. 350 #define DO_PRED12(DIFF, LANE, OUT) \ 351 do { \ 352 const __m128i all = _mm_add_epi16(L, (DIFF)); \ 353 const __m128i alls = _mm_packus_epi16(all, all); \ 354 const __m128i res = _mm_add_epi8(src, alls); \ 355 out[i + (OUT)] = _mm_cvtsi128_si32(res); \ 356 L = _mm_unpacklo_epi8(res, zero); \ 357 /* Shift the pre-computed value for the next iteration.*/ \ 358 if (LANE == 0) (DIFF) = _mm_srli_si128((DIFF), 8); \ 359 src = _mm_srli_si128(src, 4); \ 360 } while (0) 361 362 static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper, 363 int num_pixels, uint32_t* out) { 364 int i; 365 const __m128i zero = _mm_setzero_si128(); 366 const __m128i L8 = _mm_cvtsi32_si128(out[-1]); 367 __m128i L = _mm_unpacklo_epi8(L8, zero); 368 for (i = 0; i + 4 <= num_pixels; i += 4) { 369 // Load 4 pixels at a time. 370 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); 371 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); 372 const __m128i T_lo = _mm_unpacklo_epi8(T, zero); 373 const __m128i T_hi = _mm_unpackhi_epi8(T, zero); 374 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]); 375 const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero); 376 const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero); 377 __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo); 378 __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi); 379 DO_PRED12(diff_lo, 0, 0); 380 DO_PRED12(diff_lo, 1, 1); 381 DO_PRED12(diff_hi, 0, 2); 382 DO_PRED12(diff_hi, 1, 3); 383 } 384 if (i != num_pixels) { 385 VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i); 386 } 387 } 388 #undef DO_PRED12 389 390 // Due to averages with integers, values cannot be accumulated in parallel for 391 // predictors 13. 392 GENERATE_PREDICTOR_ADD(Predictor13_SSE2, PredictorAdd13_SSE2) 393 394 //------------------------------------------------------------------------------ 395 // Subtract-Green Transform 396 397 static void AddGreenToBlueAndRed(const uint32_t* const src, int num_pixels, 398 uint32_t* dst) { 399 int i; 400 for (i = 0; i + 4 <= num_pixels; i += 4) { 401 const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb 402 const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g 403 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); 404 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g 405 const __m128i out = _mm_add_epi8(in, C); 406 _mm_storeu_si128((__m128i*)&dst[i], out); 407 } 408 // fallthrough and finish off with plain-C 409 if (i != num_pixels) { 410 VP8LAddGreenToBlueAndRed_C(src + i, num_pixels - i, dst + i); 411 } 412 } 413 414 //------------------------------------------------------------------------------ 415 // Color Transform 416 417 static void TransformColorInverse(const VP8LMultipliers* const m, 418 const uint32_t* const src, int num_pixels, 419 uint32_t* dst) { 420 // sign-extended multiplying constants, pre-shifted by 5. 421 #define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend 422 const __m128i mults_rb = _mm_set_epi16( 423 CST(green_to_red_), CST(green_to_blue_), 424 CST(green_to_red_), CST(green_to_blue_), 425 CST(green_to_red_), CST(green_to_blue_), 426 CST(green_to_red_), CST(green_to_blue_)); 427 const __m128i mults_b2 = _mm_set_epi16( 428 CST(red_to_blue_), 0, CST(red_to_blue_), 0, 429 CST(red_to_blue_), 0, CST(red_to_blue_), 0); 430 #undef CST 431 const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks 432 int i; 433 for (i = 0; i + 4 <= num_pixels; i += 4) { 434 const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb 435 const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0 436 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); 437 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0 438 const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1 439 const __m128i E = _mm_add_epi8(in, D); // x r' x b' 440 const __m128i F = _mm_slli_epi16(E, 8); // r' 0 b' 0 441 const __m128i G = _mm_mulhi_epi16(F, mults_b2); // x db2 0 0 442 const __m128i H = _mm_srli_epi32(G, 8); // 0 x db2 0 443 const __m128i I = _mm_add_epi8(H, F); // r' x b'' 0 444 const __m128i J = _mm_srli_epi16(I, 8); // 0 r' 0 b'' 445 const __m128i out = _mm_or_si128(J, A); 446 _mm_storeu_si128((__m128i*)&dst[i], out); 447 } 448 // Fall-back to C-version for left-overs. 449 if (i != num_pixels) { 450 VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i); 451 } 452 } 453 454 //------------------------------------------------------------------------------ 455 // Color-space conversion functions 456 457 static void ConvertBGRAToRGB(const uint32_t* src, int num_pixels, 458 uint8_t* dst) { 459 const __m128i* in = (const __m128i*)src; 460 __m128i* out = (__m128i*)dst; 461 462 while (num_pixels >= 32) { 463 // Load the BGRA buffers. 464 __m128i in0 = _mm_loadu_si128(in + 0); 465 __m128i in1 = _mm_loadu_si128(in + 1); 466 __m128i in2 = _mm_loadu_si128(in + 2); 467 __m128i in3 = _mm_loadu_si128(in + 3); 468 __m128i in4 = _mm_loadu_si128(in + 4); 469 __m128i in5 = _mm_loadu_si128(in + 5); 470 __m128i in6 = _mm_loadu_si128(in + 6); 471 __m128i in7 = _mm_loadu_si128(in + 7); 472 VP8L32bToPlanar(&in0, &in1, &in2, &in3); 473 VP8L32bToPlanar(&in4, &in5, &in6, &in7); 474 // At this points, in1/in5 contains red only, in2/in6 green only ... 475 // Pack the colors in 24b RGB. 476 VP8PlanarTo24b(&in1, &in5, &in2, &in6, &in3, &in7); 477 _mm_storeu_si128(out + 0, in1); 478 _mm_storeu_si128(out + 1, in5); 479 _mm_storeu_si128(out + 2, in2); 480 _mm_storeu_si128(out + 3, in6); 481 _mm_storeu_si128(out + 4, in3); 482 _mm_storeu_si128(out + 5, in7); 483 in += 8; 484 out += 6; 485 num_pixels -= 32; 486 } 487 // left-overs 488 if (num_pixels > 0) { 489 VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 490 } 491 } 492 493 static void ConvertBGRAToRGBA(const uint32_t* src, 494 int num_pixels, uint8_t* dst) { 495 const __m128i* in = (const __m128i*)src; 496 __m128i* out = (__m128i*)dst; 497 while (num_pixels >= 8) { 498 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 499 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 500 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 501 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 502 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 503 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 504 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 505 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 506 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 507 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 508 const __m128i rg0 = _mm_unpacklo_epi8(rb0, ga0); // r0g0r1g1 ... r6g6r7g7 509 const __m128i ba0 = _mm_unpackhi_epi8(rb0, ga0); // b0a0b1a1 ... b6a6b7a7 510 const __m128i rgba0 = _mm_unpacklo_epi16(rg0, ba0); // rgba0|rgba1... 511 const __m128i rgba4 = _mm_unpackhi_epi16(rg0, ba0); // rgba4|rgba5... 512 _mm_storeu_si128(out++, rgba0); 513 _mm_storeu_si128(out++, rgba4); 514 num_pixels -= 8; 515 } 516 // left-overs 517 if (num_pixels > 0) { 518 VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 519 } 520 } 521 522 static void ConvertBGRAToRGBA4444(const uint32_t* src, 523 int num_pixels, uint8_t* dst) { 524 const __m128i mask_0x0f = _mm_set1_epi8(0x0f); 525 const __m128i mask_0xf0 = _mm_set1_epi8(0xf0); 526 const __m128i* in = (const __m128i*)src; 527 __m128i* out = (__m128i*)dst; 528 while (num_pixels >= 8) { 529 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 530 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 531 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 532 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 533 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 534 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 535 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 536 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 537 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 538 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 539 const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7- 540 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7 541 const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7- 542 const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7 543 const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0 544 #ifdef WEBP_SWAP_16BIT_CSP 545 const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7 546 #else 547 const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7 548 #endif 549 _mm_storeu_si128(out++, rgba); 550 num_pixels -= 8; 551 } 552 // left-overs 553 if (num_pixels > 0) { 554 VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 555 } 556 } 557 558 static void ConvertBGRAToRGB565(const uint32_t* src, 559 int num_pixels, uint8_t* dst) { 560 const __m128i mask_0xe0 = _mm_set1_epi8(0xe0); 561 const __m128i mask_0xf8 = _mm_set1_epi8(0xf8); 562 const __m128i mask_0x07 = _mm_set1_epi8(0x07); 563 const __m128i* in = (const __m128i*)src; 564 __m128i* out = (__m128i*)dst; 565 while (num_pixels >= 8) { 566 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 567 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 568 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 569 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 570 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 571 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 572 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 573 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 574 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 575 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 576 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7 577 const __m128i g_lo1 = _mm_srli_epi16(ga0, 5); 578 const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b) 579 const __m128i g_hi1 = _mm_slli_epi16(ga0, 3); 580 const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b) 581 const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0 582 const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx 583 const __m128i b1 = _mm_srli_epi16(b0, 3); 584 const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx 585 #ifdef WEBP_SWAP_16BIT_CSP 586 const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7 587 #else 588 const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7 589 #endif 590 _mm_storeu_si128(out++, rgba); 591 num_pixels -= 8; 592 } 593 // left-overs 594 if (num_pixels > 0) { 595 VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 596 } 597 } 598 599 static void ConvertBGRAToBGR(const uint32_t* src, 600 int num_pixels, uint8_t* dst) { 601 const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff); 602 const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0); 603 const __m128i* in = (const __m128i*)src; 604 const uint8_t* const end = dst + num_pixels * 3; 605 // the last storel_epi64 below writes 8 bytes starting at offset 18 606 while (dst + 26 <= end) { 607 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 608 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 609 const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0 610 const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0 611 const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0 612 const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0 613 const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00 614 const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00 615 const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00 616 const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00 617 const __m128i c2 = _mm_srli_si128(c0, 8); 618 const __m128i c6 = _mm_srli_si128(c4, 8); 619 _mm_storel_epi64((__m128i*)(dst + 0), c0); 620 _mm_storel_epi64((__m128i*)(dst + 6), c2); 621 _mm_storel_epi64((__m128i*)(dst + 12), c4); 622 _mm_storel_epi64((__m128i*)(dst + 18), c6); 623 dst += 24; 624 num_pixels -= 8; 625 } 626 // left-overs 627 if (num_pixels > 0) { 628 VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst); 629 } 630 } 631 632 //------------------------------------------------------------------------------ 633 // Entry point 634 635 extern void VP8LDspInitSSE2(void); 636 637 WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE2(void) { 638 VP8LPredictors[5] = Predictor5_SSE2; 639 VP8LPredictors[6] = Predictor6_SSE2; 640 VP8LPredictors[7] = Predictor7_SSE2; 641 VP8LPredictors[8] = Predictor8_SSE2; 642 VP8LPredictors[9] = Predictor9_SSE2; 643 VP8LPredictors[10] = Predictor10_SSE2; 644 VP8LPredictors[11] = Predictor11_SSE2; 645 VP8LPredictors[12] = Predictor12_SSE2; 646 VP8LPredictors[13] = Predictor13_SSE2; 647 648 VP8LPredictorsAdd[0] = PredictorAdd0_SSE2; 649 VP8LPredictorsAdd[1] = PredictorAdd1_SSE2; 650 VP8LPredictorsAdd[2] = PredictorAdd2_SSE2; 651 VP8LPredictorsAdd[3] = PredictorAdd3_SSE2; 652 VP8LPredictorsAdd[4] = PredictorAdd4_SSE2; 653 VP8LPredictorsAdd[5] = PredictorAdd5_SSE2; 654 VP8LPredictorsAdd[6] = PredictorAdd6_SSE2; 655 VP8LPredictorsAdd[7] = PredictorAdd7_SSE2; 656 VP8LPredictorsAdd[8] = PredictorAdd8_SSE2; 657 VP8LPredictorsAdd[9] = PredictorAdd9_SSE2; 658 VP8LPredictorsAdd[10] = PredictorAdd10_SSE2; 659 VP8LPredictorsAdd[11] = PredictorAdd11_SSE2; 660 VP8LPredictorsAdd[12] = PredictorAdd12_SSE2; 661 VP8LPredictorsAdd[13] = PredictorAdd13_SSE2; 662 663 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; 664 VP8LTransformColorInverse = TransformColorInverse; 665 666 VP8LConvertBGRAToRGB = ConvertBGRAToRGB; 667 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; 668 VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444; 669 VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565; 670 VP8LConvertBGRAToBGR = ConvertBGRAToBGR; 671 } 672 673 #else // !WEBP_USE_SSE2 674 675 WEBP_DSP_INIT_STUB(VP8LDspInitSSE2) 676 677 #endif // WEBP_USE_SSE2 678
