1/* 2 * Copyright (c) 2017 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include <emmintrin.h> // SSE2 12 13#include "./vpx_config.h" 14#include "./vpx_dsp_rtcd.h" 15#include "vpx/vpx_integer.h" 16 17// ----------------------------------------------------------------------------- 18 19void vpx_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 20 const uint16_t *above, 21 const uint16_t *left, int bd) { 22 const __m128i left_u16 = _mm_loadl_epi64((const __m128i *)left); 23 const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); 24 const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); 25 const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); 26 const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); 27 (void)above; 28 (void)bd; 29 _mm_storel_epi64((__m128i *)dst, row0); 30 dst += stride; 31 _mm_storel_epi64((__m128i *)dst, row1); 32 dst += stride; 33 _mm_storel_epi64((__m128i *)dst, row2); 34 dst += stride; 35 _mm_storel_epi64((__m128i *)dst, row3); 36} 37 38void vpx_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, 39 const uint16_t *above, 40 const uint16_t *left, int bd) { 41 const __m128i left_u16 = _mm_load_si128((const __m128i *)left); 42 const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); 43 const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); 44 const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); 45 const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); 46 const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0); 47 const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55); 48 const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa); 49 const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff); 50 (void)above; 51 (void)bd; 52 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0)); 53 dst += stride; 54 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1)); 55 dst += stride; 56 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2)); 57 dst += stride; 58 _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3)); 59 dst += stride; 60 _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row4, row4)); 61 dst += stride; 62 _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row5, row5)); 63 dst += stride; 64 _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row6, row6)); 65 dst += stride; 66 _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row7, row7)); 67} 68 69static INLINE void h_store_16_unpacklo(uint16_t **dst, const ptrdiff_t stride, 70 const __m128i *row) { 71 const __m128i val = _mm_unpacklo_epi64(*row, *row); 72 _mm_store_si128((__m128i *)*dst, val); 73 _mm_store_si128((__m128i *)(*dst + 8), val); 74 *dst += stride; 75} 76 77static INLINE void h_store_16_unpackhi(uint16_t **dst, const ptrdiff_t stride, 78 const __m128i *row) { 79 const __m128i val = _mm_unpackhi_epi64(*row, *row); 80 _mm_store_si128((__m128i *)(*dst), val); 81 _mm_store_si128((__m128i *)(*dst + 8), val); 82 *dst += stride; 83} 84 85void vpx_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, 86 const uint16_t *above, 87 const uint16_t *left, int bd) { 88 int i; 89 (void)above; 90 (void)bd; 91 92 for (i = 0; i < 2; i++, left += 8) { 93 const __m128i left_u16 = _mm_load_si128((const __m128i *)left); 94 const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); 95 const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); 96 const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); 97 const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); 98 const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0); 99 const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55); 100 const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa); 101 const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff); 102 h_store_16_unpacklo(&dst, stride, &row0); 103 h_store_16_unpacklo(&dst, stride, &row1); 104 h_store_16_unpacklo(&dst, stride, &row2); 105 h_store_16_unpacklo(&dst, stride, &row3); 106 h_store_16_unpackhi(&dst, stride, &row4); 107 h_store_16_unpackhi(&dst, stride, &row5); 108 h_store_16_unpackhi(&dst, stride, &row6); 109 h_store_16_unpackhi(&dst, stride, &row7); 110 } 111} 112 113static INLINE void h_store_32_unpacklo(uint16_t **dst, const ptrdiff_t stride, 114 const __m128i *row) { 115 const __m128i val = _mm_unpacklo_epi64(*row, *row); 116 _mm_store_si128((__m128i *)(*dst), val); 117 _mm_store_si128((__m128i *)(*dst + 8), val); 118 _mm_store_si128((__m128i *)(*dst + 16), val); 119 _mm_store_si128((__m128i *)(*dst + 24), val); 120 *dst += stride; 121} 122 123static INLINE void h_store_32_unpackhi(uint16_t **dst, const ptrdiff_t stride, 124 const __m128i *row) { 125 const __m128i val = _mm_unpackhi_epi64(*row, *row); 126 _mm_store_si128((__m128i *)(*dst), val); 127 _mm_store_si128((__m128i *)(*dst + 8), val); 128 _mm_store_si128((__m128i *)(*dst + 16), val); 129 _mm_store_si128((__m128i *)(*dst + 24), val); 130 *dst += stride; 131} 132 133void vpx_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, 134 const uint16_t *above, 135 const uint16_t *left, int bd) { 136 int i; 137 (void)above; 138 (void)bd; 139 140 for (i = 0; i < 4; i++, left += 8) { 141 const __m128i left_u16 = _mm_load_si128((const __m128i *)left); 142 const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0); 143 const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55); 144 const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa); 145 const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff); 146 const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0); 147 const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55); 148 const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa); 149 const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff); 150 h_store_32_unpacklo(&dst, stride, &row0); 151 h_store_32_unpacklo(&dst, stride, &row1); 152 h_store_32_unpacklo(&dst, stride, &row2); 153 h_store_32_unpacklo(&dst, stride, &row3); 154 h_store_32_unpackhi(&dst, stride, &row4); 155 h_store_32_unpackhi(&dst, stride, &row5); 156 h_store_32_unpackhi(&dst, stride, &row6); 157 h_store_32_unpackhi(&dst, stride, &row7); 158 } 159} 160 161//------------------------------------------------------------------------------ 162// DC 4x4 163 164static INLINE __m128i dc_sum_4(const uint16_t *ref) { 165 const __m128i _dcba = _mm_loadl_epi64((const __m128i *)ref); 166 const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe); 167 const __m128i a = _mm_add_epi16(_dcba, _xxdc); 168 return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1)); 169} 170 171static INLINE void dc_store_4x4(uint16_t *dst, ptrdiff_t stride, 172 const __m128i *dc) { 173 const __m128i dc_dup = _mm_shufflelo_epi16(*dc, 0x0); 174 int i; 175 for (i = 0; i < 4; ++i, dst += stride) { 176 _mm_storel_epi64((__m128i *)dst, dc_dup); 177 } 178} 179 180void vpx_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 181 const uint16_t *above, 182 const uint16_t *left, int bd) { 183 const __m128i two = _mm_cvtsi32_si128(2); 184 const __m128i sum = dc_sum_4(left); 185 const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2); 186 (void)above; 187 (void)bd; 188 dc_store_4x4(dst, stride, &dc); 189} 190 191void vpx_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 192 const uint16_t *above, 193 const uint16_t *left, int bd) { 194 const __m128i two = _mm_cvtsi32_si128(2); 195 const __m128i sum = dc_sum_4(above); 196 const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2); 197 (void)left; 198 (void)bd; 199 dc_store_4x4(dst, stride, &dc); 200} 201 202void vpx_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 203 const uint16_t *above, 204 const uint16_t *left, int bd) { 205 const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); 206 const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); 207 (void)above; 208 (void)left; 209 dc_store_4x4(dst, stride, &dc_dup); 210} 211 212//------------------------------------------------------------------------------ 213// DC 8x8 214 215static INLINE __m128i dc_sum_8(const uint16_t *ref) { 216 const __m128i ref_u16 = _mm_load_si128((const __m128i *)ref); 217 const __m128i _dcba = _mm_add_epi16(ref_u16, _mm_srli_si128(ref_u16, 8)); 218 const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe); 219 const __m128i a = _mm_add_epi16(_dcba, _xxdc); 220 221 return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1)); 222} 223 224static INLINE void dc_store_8x8(uint16_t *dst, ptrdiff_t stride, 225 const __m128i *dc) { 226 const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0); 227 const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo); 228 int i; 229 for (i = 0; i < 8; ++i, dst += stride) { 230 _mm_store_si128((__m128i *)dst, dc_dup); 231 } 232} 233 234void vpx_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, 235 const uint16_t *above, 236 const uint16_t *left, int bd) { 237 const __m128i four = _mm_cvtsi32_si128(4); 238 const __m128i sum = dc_sum_8(left); 239 const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3); 240 (void)above; 241 (void)bd; 242 dc_store_8x8(dst, stride, &dc); 243} 244 245void vpx_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, 246 const uint16_t *above, 247 const uint16_t *left, int bd) { 248 const __m128i four = _mm_cvtsi32_si128(4); 249 const __m128i sum = dc_sum_8(above); 250 const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3); 251 (void)left; 252 (void)bd; 253 dc_store_8x8(dst, stride, &dc); 254} 255 256void vpx_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride, 257 const uint16_t *above, 258 const uint16_t *left, int bd) { 259 const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); 260 const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); 261 (void)above; 262 (void)left; 263 dc_store_8x8(dst, stride, &dc_dup); 264} 265 266//------------------------------------------------------------------------------ 267// DC 16x16 268 269static INLINE __m128i dc_sum_16(const uint16_t *ref) { 270 const __m128i sum_lo = dc_sum_8(ref); 271 const __m128i sum_hi = dc_sum_8(ref + 8); 272 return _mm_add_epi16(sum_lo, sum_hi); 273} 274 275static INLINE void dc_store_16x16(uint16_t *dst, ptrdiff_t stride, 276 const __m128i *dc) { 277 const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0); 278 const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo); 279 int i; 280 for (i = 0; i < 16; ++i, dst += stride) { 281 _mm_store_si128((__m128i *)dst, dc_dup); 282 _mm_store_si128((__m128i *)(dst + 8), dc_dup); 283 } 284} 285 286void vpx_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, 287 const uint16_t *above, 288 const uint16_t *left, int bd) { 289 const __m128i eight = _mm_cvtsi32_si128(8); 290 const __m128i sum = dc_sum_16(left); 291 const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); 292 (void)above; 293 (void)bd; 294 dc_store_16x16(dst, stride, &dc); 295} 296 297void vpx_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, 298 const uint16_t *above, 299 const uint16_t *left, int bd) { 300 const __m128i eight = _mm_cvtsi32_si128(8); 301 const __m128i sum = dc_sum_16(above); 302 const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4); 303 (void)left; 304 (void)bd; 305 dc_store_16x16(dst, stride, &dc); 306} 307 308void vpx_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride, 309 const uint16_t *above, 310 const uint16_t *left, int bd) { 311 const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); 312 const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); 313 (void)above; 314 (void)left; 315 dc_store_16x16(dst, stride, &dc_dup); 316} 317 318//------------------------------------------------------------------------------ 319// DC 32x32 320 321static INLINE __m128i dc_sum_32(const uint16_t *ref) { 322 const __m128i zero = _mm_setzero_si128(); 323 const __m128i sum_a = dc_sum_16(ref); 324 const __m128i sum_b = dc_sum_16(ref + 16); 325 // 12 bit bd will outrange, so expand to 32 bit before adding final total 326 return _mm_add_epi32(_mm_unpacklo_epi16(sum_a, zero), 327 _mm_unpacklo_epi16(sum_b, zero)); 328} 329 330static INLINE void dc_store_32x32(uint16_t *dst, ptrdiff_t stride, 331 const __m128i *dc) { 332 const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0); 333 const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo); 334 int i; 335 for (i = 0; i < 32; ++i, dst += stride) { 336 _mm_store_si128((__m128i *)dst, dc_dup); 337 _mm_store_si128((__m128i *)(dst + 8), dc_dup); 338 _mm_store_si128((__m128i *)(dst + 16), dc_dup); 339 _mm_store_si128((__m128i *)(dst + 24), dc_dup); 340 } 341} 342 343void vpx_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, 344 const uint16_t *above, 345 const uint16_t *left, int bd) { 346 const __m128i sixteen = _mm_cvtsi32_si128(16); 347 const __m128i sum = dc_sum_32(left); 348 const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5); 349 (void)above; 350 (void)bd; 351 dc_store_32x32(dst, stride, &dc); 352} 353 354void vpx_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, 355 const uint16_t *above, 356 const uint16_t *left, int bd) { 357 const __m128i sixteen = _mm_cvtsi32_si128(16); 358 const __m128i sum = dc_sum_32(above); 359 const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5); 360 (void)left; 361 (void)bd; 362 dc_store_32x32(dst, stride, &dc); 363} 364 365void vpx_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride, 366 const uint16_t *above, 367 const uint16_t *left, int bd) { 368 const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1)); 369 const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0); 370 (void)above; 371 (void)left; 372 dc_store_32x32(dst, stride, &dc_dup); 373} 374 375// ----------------------------------------------------------------------------- 376/* 377; ------------------------------------------ 378; input: x, y, z, result 379; 380; trick from pascal 381; (x+2y+z+2)>>2 can be calculated as: 382; result = avg(x,z) 383; result -= xor(x,z) & 1 384; result = avg(result,y) 385; ------------------------------------------ 386*/ 387static INLINE __m128i avg3_epu16(const __m128i *x, const __m128i *y, 388 const __m128i *z) { 389 const __m128i one = _mm_set1_epi16(1); 390 const __m128i a = _mm_avg_epu16(*x, *z); 391 const __m128i b = 392 _mm_subs_epu16(a, _mm_and_si128(_mm_xor_si128(*x, *z), one)); 393 return _mm_avg_epu16(b, *y); 394} 395 396void vpx_highbd_d117_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 397 const uint16_t *above, 398 const uint16_t *left, int bd) { 399 const int I = left[0]; 400 const int J = left[1]; 401 const int K = left[2]; 402 const __m128i XXXXABCD = _mm_loadu_si128((const __m128i *)(above - 4)); 403 const __m128i KXXXABCD = _mm_insert_epi16(XXXXABCD, K, 0); 404 const __m128i KJXXABCD = _mm_insert_epi16(KXXXABCD, J, 1); 405 const __m128i KJIXABCD = _mm_insert_epi16(KJXXABCD, I, 2); 406 const __m128i JIXABCD0 = _mm_srli_si128(KJIXABCD, 2); 407 const __m128i IXABCD00 = _mm_srli_si128(KJIXABCD, 4); 408 const __m128i avg2 = _mm_avg_epu16(KJIXABCD, JIXABCD0); 409 const __m128i avg3 = avg3_epu16(&KJIXABCD, &JIXABCD0, &IXABCD00); 410 const __m128i row0 = _mm_srli_si128(avg2, 6); 411 const __m128i row1 = _mm_srli_si128(avg3, 4); 412 const __m128i row2 = _mm_srli_si128(avg2, 4); 413 const __m128i row3 = _mm_srli_si128(avg3, 2); 414 (void)bd; 415 _mm_storel_epi64((__m128i *)dst, row0); 416 dst += stride; 417 _mm_storel_epi64((__m128i *)dst, row1); 418 dst += stride; 419 _mm_storel_epi64((__m128i *)dst, row2); 420 dst += stride; 421 _mm_storel_epi64((__m128i *)dst, row3); 422 423 dst -= stride; 424 dst[0] = _mm_extract_epi16(avg3, 1); 425 dst[stride] = _mm_extract_epi16(avg3, 0); 426} 427 428void vpx_highbd_d135_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 429 const uint16_t *above, 430 const uint16_t *left, int bd) { 431 const int I = left[0]; 432 const int J = left[1]; 433 const int K = left[2]; 434 const int L = left[3]; 435 const __m128i XXXXABCD = _mm_loadu_si128((const __m128i *)(above - 4)); 436 const __m128i KXXXABCD = _mm_insert_epi16(XXXXABCD, K, 0); 437 const __m128i KJXXABCD = _mm_insert_epi16(KXXXABCD, J, 1); 438 const __m128i KJIXABCD = _mm_insert_epi16(KJXXABCD, I, 2); 439 const __m128i JIXABCD0 = _mm_srli_si128(KJIXABCD, 2); 440 const __m128i LKJIXABC = _mm_insert_epi16(_mm_slli_si128(KJIXABCD, 2), L, 0); 441 const __m128i avg3 = avg3_epu16(&JIXABCD0, &KJIXABCD, &LKJIXABC); 442 const __m128i row0 = _mm_srli_si128(avg3, 6); 443 const __m128i row1 = _mm_srli_si128(avg3, 4); 444 const __m128i row2 = _mm_srli_si128(avg3, 2); 445 const __m128i row3 = avg3; 446 (void)bd; 447 _mm_storel_epi64((__m128i *)dst, row0); 448 dst += stride; 449 _mm_storel_epi64((__m128i *)dst, row1); 450 dst += stride; 451 _mm_storel_epi64((__m128i *)dst, row2); 452 dst += stride; 453 _mm_storel_epi64((__m128i *)dst, row3); 454} 455 456void vpx_highbd_d153_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 457 const uint16_t *above, 458 const uint16_t *left, int bd) { 459 const int I = left[0]; 460 const int J = left[1]; 461 const int K = left[2]; 462 const int L = left[3]; 463 const __m128i XXXXXABC = _mm_loadu_si128((const __m128i *)(above - 5)); 464 const __m128i LXXXXABC = _mm_insert_epi16(XXXXXABC, L, 0); 465 const __m128i LKXXXABC = _mm_insert_epi16(LXXXXABC, K, 1); 466 const __m128i LKJXXABC = _mm_insert_epi16(LKXXXABC, J, 2); 467 const __m128i LKJIXABC = _mm_insert_epi16(LKJXXABC, I, 3); 468 const __m128i KJIXABC0 = _mm_srli_si128(LKJIXABC, 2); 469 const __m128i JIXABC00 = _mm_srli_si128(LKJIXABC, 4); 470 const __m128i avg3 = avg3_epu16(&LKJIXABC, &KJIXABC0, &JIXABC00); 471 const __m128i avg2 = _mm_avg_epu16(LKJIXABC, KJIXABC0); 472 const __m128i row3 = _mm_unpacklo_epi16(avg2, avg3); 473 const __m128i row2 = _mm_srli_si128(row3, 4); 474 const __m128i row1 = _mm_srli_si128(row3, 8); 475 const __m128i row0 = _mm_srli_si128(avg3, 4); 476 (void)bd; 477 _mm_storel_epi64((__m128i *)dst, row0); 478 dst[0] = _mm_extract_epi16(avg2, 3); 479 dst += stride; 480 _mm_storel_epi64((__m128i *)dst, row1); 481 dst += stride; 482 _mm_storel_epi64((__m128i *)dst, row2); 483 dst += stride; 484 _mm_storel_epi64((__m128i *)dst, row3); 485} 486 487void vpx_highbd_d207_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 488 const uint16_t *above, 489 const uint16_t *left, int bd) { 490 const __m128i IJKL0000 = _mm_load_si128((const __m128i *)left); 491 const __m128i LLLL0000 = _mm_shufflelo_epi16(IJKL0000, 0xff); 492 const __m128i IJKLLLLL = _mm_unpacklo_epi64(IJKL0000, LLLL0000); 493 const __m128i JKLLLLL0 = _mm_srli_si128(IJKLLLLL, 2); 494 const __m128i KLLLLL00 = _mm_srli_si128(IJKLLLLL, 4); 495 const __m128i avg3 = avg3_epu16(&IJKLLLLL, &JKLLLLL0, &KLLLLL00); 496 const __m128i avg2 = _mm_avg_epu16(IJKLLLLL, JKLLLLL0); 497 const __m128i row0 = _mm_unpacklo_epi16(avg2, avg3); 498 const __m128i row1 = _mm_srli_si128(row0, 4); 499 const __m128i row2 = _mm_srli_si128(row0, 8); 500 const __m128i row3 = LLLL0000; 501 (void)above; 502 (void)bd; 503 _mm_storel_epi64((__m128i *)dst, row0); 504 dst += stride; 505 _mm_storel_epi64((__m128i *)dst, row1); 506 dst += stride; 507 _mm_storel_epi64((__m128i *)dst, row2); 508 dst += stride; 509 _mm_storel_epi64((__m128i *)dst, row3); 510} 511 512void vpx_highbd_d63_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride, 513 const uint16_t *above, 514 const uint16_t *left, int bd) { 515 const __m128i ABCDEFGH = _mm_loadu_si128((const __m128i *)above); 516 const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 2); 517 const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 4); 518 const __m128i avg3 = avg3_epu16(&ABCDEFGH, &BCDEFGH0, &CDEFGH00); 519 const __m128i avg2 = _mm_avg_epu16(ABCDEFGH, BCDEFGH0); 520 const __m128i row0 = avg2; 521 const __m128i row1 = avg3; 522 const __m128i row2 = _mm_srli_si128(avg2, 2); 523 const __m128i row3 = _mm_srli_si128(avg3, 2); 524 (void)left; 525 (void)bd; 526 _mm_storel_epi64((__m128i *)dst, row0); 527 dst += stride; 528 _mm_storel_epi64((__m128i *)dst, row1); 529 dst += stride; 530 _mm_storel_epi64((__m128i *)dst, row2); 531 dst += stride; 532 _mm_storel_epi64((__m128i *)dst, row3); 533} 534