1 /* 2 * Copyright (c) 2014 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_dsp_rtcd.h" 14 #include "vpx_ports/mem.h" 15 16 static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) { 17 __m128i ubounded; 18 __m128i lbounded; 19 __m128i retval; 20 21 const __m128i zero = _mm_set1_epi16(0); 22 const __m128i one = _mm_set1_epi16(1); 23 __m128i t80, max, min; 24 25 if (bd == 8) { 26 t80 = _mm_set1_epi16(0x80); 27 max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80); 28 } else if (bd == 10) { 29 t80 = _mm_set1_epi16(0x200); 30 max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80); 31 } else { // bd == 12 32 t80 = _mm_set1_epi16(0x800); 33 max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80); 34 } 35 36 min = _mm_subs_epi16(zero, t80); 37 38 ubounded = _mm_cmpgt_epi16(value, max); 39 lbounded = _mm_cmplt_epi16(value, min); 40 retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value); 41 ubounded = _mm_and_si128(ubounded, max); 42 lbounded = _mm_and_si128(lbounded, min); 43 retval = _mm_or_si128(retval, ubounded); 44 retval = _mm_or_si128(retval, lbounded); 45 return retval; 46 } 47 48 // TODO(debargha, peter): Break up large functions into smaller ones 49 // in this file. 50 void vpx_highbd_lpf_horizontal_16_sse2(uint16_t *s, int p, 51 const uint8_t *_blimit, 52 const uint8_t *_limit, 53 const uint8_t *_thresh, int bd) { 54 const __m128i zero = _mm_set1_epi16(0); 55 const __m128i one = _mm_set1_epi16(1); 56 __m128i blimit, limit, thresh; 57 __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0; 58 __m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0; 59 __m128i ps1, qs1, ps0, qs0; 60 __m128i abs_p0q0, abs_p1q1, ffff, work; 61 __m128i filt, work_a, filter1, filter2; 62 __m128i flat2_q6, flat2_p6, flat2_q5, flat2_p5, flat2_q4, flat2_p4; 63 __m128i flat2_q3, flat2_p3, flat2_q2, flat2_p2, flat2_q1, flat2_p1; 64 __m128i flat2_q0, flat2_p0; 65 __m128i flat_q2, flat_p2, flat_q1, flat_p1, flat_q0, flat_p0; 66 __m128i pixelFilter_p, pixelFilter_q; 67 __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0; 68 __m128i sum_p7, sum_q7, sum_p3, sum_q3; 69 __m128i t4, t3, t80, t1; 70 __m128i eight, four; 71 72 if (bd == 8) { 73 blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero); 74 limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero); 75 thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero); 76 } else if (bd == 10) { 77 blimit = _mm_slli_epi16( 78 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2); 79 limit = _mm_slli_epi16( 80 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2); 81 thresh = _mm_slli_epi16( 82 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2); 83 } else { // bd == 12 84 blimit = _mm_slli_epi16( 85 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4); 86 limit = _mm_slli_epi16( 87 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4); 88 thresh = _mm_slli_epi16( 89 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4); 90 } 91 92 q4 = _mm_load_si128((__m128i *)(s + 4 * p)); 93 p4 = _mm_load_si128((__m128i *)(s - 5 * p)); 94 q3 = _mm_load_si128((__m128i *)(s + 3 * p)); 95 p3 = _mm_load_si128((__m128i *)(s - 4 * p)); 96 q2 = _mm_load_si128((__m128i *)(s + 2 * p)); 97 p2 = _mm_load_si128((__m128i *)(s - 3 * p)); 98 q1 = _mm_load_si128((__m128i *)(s + 1 * p)); 99 p1 = _mm_load_si128((__m128i *)(s - 2 * p)); 100 q0 = _mm_load_si128((__m128i *)(s + 0 * p)); 101 p0 = _mm_load_si128((__m128i *)(s - 1 * p)); 102 103 // highbd_filter_mask 104 abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)); 105 abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1)); 106 107 ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0); 108 109 abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0)); 110 abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1)); 111 112 // highbd_hev_mask (in C code this is actually called from highbd_filter4) 113 flat = _mm_max_epi16(abs_p1p0, abs_q1q0); 114 hev = _mm_subs_epu16(flat, thresh); 115 hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff); 116 117 abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); // abs(p0 - q0) * 2 118 abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); // abs(p1 - q1) / 2 119 mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit); 120 mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); 121 mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one)); 122 work = _mm_max_epi16( 123 _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)), 124 _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1))); 125 mask = _mm_max_epi16(work, mask); 126 work = _mm_max_epi16( 127 _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)), 128 _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2))); 129 mask = _mm_max_epi16(work, mask); 130 work = _mm_max_epi16( 131 _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3)), 132 _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3))); 133 mask = _mm_max_epi16(work, mask); 134 135 mask = _mm_subs_epu16(mask, limit); 136 mask = _mm_cmpeq_epi16(mask, zero); // return ~mask 137 138 // lp filter 139 // highbd_filter4 140 t4 = _mm_set1_epi16(4); 141 t3 = _mm_set1_epi16(3); 142 if (bd == 8) 143 t80 = _mm_set1_epi16(0x80); 144 else if (bd == 10) 145 t80 = _mm_set1_epi16(0x200); 146 else // bd == 12 147 t80 = _mm_set1_epi16(0x800); 148 149 t1 = _mm_set1_epi16(0x1); 150 151 ps1 = _mm_subs_epi16(p1, t80); 152 qs1 = _mm_subs_epi16(q1, t80); 153 ps0 = _mm_subs_epi16(p0, t80); 154 qs0 = _mm_subs_epi16(q0, t80); 155 156 filt = _mm_and_si128(signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd), 157 hev); 158 work_a = _mm_subs_epi16(qs0, ps0); 159 filt = _mm_adds_epi16(filt, work_a); 160 filt = _mm_adds_epi16(filt, work_a); 161 filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd); 162 filt = _mm_and_si128(filt, mask); 163 filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd); 164 filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd); 165 166 // Filter1 >> 3 167 filter1 = _mm_srai_epi16(filter1, 0x3); 168 filter2 = _mm_srai_epi16(filter2, 0x3); 169 170 qs0 = _mm_adds_epi16( 171 signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80); 172 ps0 = _mm_adds_epi16( 173 signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80); 174 filt = _mm_adds_epi16(filter1, t1); 175 filt = _mm_srai_epi16(filt, 1); 176 filt = _mm_andnot_si128(hev, filt); 177 qs1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), 178 t80); 179 ps1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), 180 t80); 181 182 // end highbd_filter4 183 // loopfilter done 184 185 // highbd_flat_mask4 186 flat = _mm_max_epi16( 187 _mm_or_si128(_mm_subs_epu16(p2, p0), _mm_subs_epu16(p0, p2)), 188 _mm_or_si128(_mm_subs_epu16(p3, p0), _mm_subs_epu16(p0, p3))); 189 work = _mm_max_epi16( 190 _mm_or_si128(_mm_subs_epu16(q2, q0), _mm_subs_epu16(q0, q2)), 191 _mm_or_si128(_mm_subs_epu16(q3, q0), _mm_subs_epu16(q0, q3))); 192 flat = _mm_max_epi16(work, flat); 193 work = _mm_max_epi16(abs_p1p0, abs_q1q0); 194 flat = _mm_max_epi16(work, flat); 195 196 if (bd == 8) 197 flat = _mm_subs_epu16(flat, one); 198 else if (bd == 10) 199 flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2)); 200 else // bd == 12 201 flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4)); 202 203 flat = _mm_cmpeq_epi16(flat, zero); 204 // end flat_mask4 205 206 // flat & mask = flat && mask (as used in filter8) 207 // (because, in both vars, each block of 16 either all 1s or all 0s) 208 flat = _mm_and_si128(flat, mask); 209 210 p5 = _mm_load_si128((__m128i *)(s - 6 * p)); 211 q5 = _mm_load_si128((__m128i *)(s + 5 * p)); 212 p6 = _mm_load_si128((__m128i *)(s - 7 * p)); 213 q6 = _mm_load_si128((__m128i *)(s + 6 * p)); 214 p7 = _mm_load_si128((__m128i *)(s - 8 * p)); 215 q7 = _mm_load_si128((__m128i *)(s + 7 * p)); 216 217 // highbd_flat_mask5 (arguments passed in are p0, q0, p4-p7, q4-q7 218 // but referred to as p0-p4 & q0-q4 in fn) 219 flat2 = _mm_max_epi16( 220 _mm_or_si128(_mm_subs_epu16(p4, p0), _mm_subs_epu16(p0, p4)), 221 _mm_or_si128(_mm_subs_epu16(q4, q0), _mm_subs_epu16(q0, q4))); 222 223 work = _mm_max_epi16( 224 _mm_or_si128(_mm_subs_epu16(p5, p0), _mm_subs_epu16(p0, p5)), 225 _mm_or_si128(_mm_subs_epu16(q5, q0), _mm_subs_epu16(q0, q5))); 226 flat2 = _mm_max_epi16(work, flat2); 227 228 work = _mm_max_epi16( 229 _mm_or_si128(_mm_subs_epu16(p6, p0), _mm_subs_epu16(p0, p6)), 230 _mm_or_si128(_mm_subs_epu16(q6, q0), _mm_subs_epu16(q0, q6))); 231 flat2 = _mm_max_epi16(work, flat2); 232 233 work = _mm_max_epi16( 234 _mm_or_si128(_mm_subs_epu16(p7, p0), _mm_subs_epu16(p0, p7)), 235 _mm_or_si128(_mm_subs_epu16(q7, q0), _mm_subs_epu16(q0, q7))); 236 flat2 = _mm_max_epi16(work, flat2); 237 238 if (bd == 8) 239 flat2 = _mm_subs_epu16(flat2, one); 240 else if (bd == 10) 241 flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2)); 242 else // bd == 12 243 flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4)); 244 245 flat2 = _mm_cmpeq_epi16(flat2, zero); 246 flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask 247 // end highbd_flat_mask5 248 249 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 250 // flat and wide flat calculations 251 eight = _mm_set1_epi16(8); 252 four = _mm_set1_epi16(4); 253 254 pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5), _mm_add_epi16(p4, p3)); 255 pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5), _mm_add_epi16(q4, q3)); 256 257 pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1)); 258 pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); 259 260 pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1)); 261 pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); 262 pixelFilter_p = 263 _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); 264 pixetFilter_p2p1p0 = _mm_add_epi16( 265 four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); 266 flat2_p0 = 267 _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7, p0)), 4); 268 flat2_q0 = 269 _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7, q0)), 4); 270 flat_p0 = _mm_srli_epi16( 271 _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3, p0)), 3); 272 flat_q0 = _mm_srli_epi16( 273 _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3, q0)), 3); 274 275 sum_p7 = _mm_add_epi16(p7, p7); 276 sum_q7 = _mm_add_epi16(q7, q7); 277 sum_p3 = _mm_add_epi16(p3, p3); 278 sum_q3 = _mm_add_epi16(q3, q3); 279 280 pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6); 281 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6); 282 flat2_p1 = _mm_srli_epi16( 283 _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1)), 4); 284 flat2_q1 = _mm_srli_epi16( 285 _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1)), 4); 286 287 pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2); 288 pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2); 289 flat_p1 = _mm_srli_epi16( 290 _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1)), 3); 291 flat_q1 = _mm_srli_epi16( 292 _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1)), 3); 293 294 sum_p7 = _mm_add_epi16(sum_p7, p7); 295 sum_q7 = _mm_add_epi16(sum_q7, q7); 296 sum_p3 = _mm_add_epi16(sum_p3, p3); 297 sum_q3 = _mm_add_epi16(sum_q3, q3); 298 299 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5); 300 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5); 301 flat2_p2 = _mm_srli_epi16( 302 _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2)), 4); 303 flat2_q2 = _mm_srli_epi16( 304 _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2)), 4); 305 306 pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1); 307 pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1); 308 flat_p2 = _mm_srli_epi16( 309 _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2)), 3); 310 flat_q2 = _mm_srli_epi16( 311 _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2)), 3); 312 313 sum_p7 = _mm_add_epi16(sum_p7, p7); 314 sum_q7 = _mm_add_epi16(sum_q7, q7); 315 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4); 316 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4); 317 flat2_p3 = _mm_srli_epi16( 318 _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3)), 4); 319 flat2_q3 = _mm_srli_epi16( 320 _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3)), 4); 321 322 sum_p7 = _mm_add_epi16(sum_p7, p7); 323 sum_q7 = _mm_add_epi16(sum_q7, q7); 324 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3); 325 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3); 326 flat2_p4 = _mm_srli_epi16( 327 _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4)), 4); 328 flat2_q4 = _mm_srli_epi16( 329 _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4)), 4); 330 331 sum_p7 = _mm_add_epi16(sum_p7, p7); 332 sum_q7 = _mm_add_epi16(sum_q7, q7); 333 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2); 334 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2); 335 flat2_p5 = _mm_srli_epi16( 336 _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5)), 4); 337 flat2_q5 = _mm_srli_epi16( 338 _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5)), 4); 339 340 sum_p7 = _mm_add_epi16(sum_p7, p7); 341 sum_q7 = _mm_add_epi16(sum_q7, q7); 342 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1); 343 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1); 344 flat2_p6 = _mm_srli_epi16( 345 _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6)), 4); 346 flat2_q6 = _mm_srli_epi16( 347 _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6)), 4); 348 349 // wide flat 350 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 351 352 // highbd_filter8 353 p2 = _mm_andnot_si128(flat, p2); 354 // p2 remains unchanged if !(flat && mask) 355 flat_p2 = _mm_and_si128(flat, flat_p2); 356 // when (flat && mask) 357 p2 = _mm_or_si128(p2, flat_p2); // full list of p2 values 358 q2 = _mm_andnot_si128(flat, q2); 359 flat_q2 = _mm_and_si128(flat, flat_q2); 360 q2 = _mm_or_si128(q2, flat_q2); // full list of q2 values 361 362 ps1 = _mm_andnot_si128(flat, ps1); 363 // p1 takes the value assigned to in in filter4 if !(flat && mask) 364 flat_p1 = _mm_and_si128(flat, flat_p1); 365 // when (flat && mask) 366 p1 = _mm_or_si128(ps1, flat_p1); // full list of p1 values 367 qs1 = _mm_andnot_si128(flat, qs1); 368 flat_q1 = _mm_and_si128(flat, flat_q1); 369 q1 = _mm_or_si128(qs1, flat_q1); // full list of q1 values 370 371 ps0 = _mm_andnot_si128(flat, ps0); 372 // p0 takes the value assigned to in in filter4 if !(flat && mask) 373 flat_p0 = _mm_and_si128(flat, flat_p0); 374 // when (flat && mask) 375 p0 = _mm_or_si128(ps0, flat_p0); // full list of p0 values 376 qs0 = _mm_andnot_si128(flat, qs0); 377 flat_q0 = _mm_and_si128(flat, flat_q0); 378 q0 = _mm_or_si128(qs0, flat_q0); // full list of q0 values 379 // end highbd_filter8 380 381 // highbd_filter16 382 p6 = _mm_andnot_si128(flat2, p6); 383 // p6 remains unchanged if !(flat2 && flat && mask) 384 flat2_p6 = _mm_and_si128(flat2, flat2_p6); 385 // get values for when (flat2 && flat && mask) 386 p6 = _mm_or_si128(p6, flat2_p6); // full list of p6 values 387 q6 = _mm_andnot_si128(flat2, q6); 388 // q6 remains unchanged if !(flat2 && flat && mask) 389 flat2_q6 = _mm_and_si128(flat2, flat2_q6); 390 // get values for when (flat2 && flat && mask) 391 q6 = _mm_or_si128(q6, flat2_q6); // full list of q6 values 392 _mm_store_si128((__m128i *)(s - 7 * p), p6); 393 _mm_store_si128((__m128i *)(s + 6 * p), q6); 394 395 p5 = _mm_andnot_si128(flat2, p5); 396 // p5 remains unchanged if !(flat2 && flat && mask) 397 flat2_p5 = _mm_and_si128(flat2, flat2_p5); 398 // get values for when (flat2 && flat && mask) 399 p5 = _mm_or_si128(p5, flat2_p5); 400 // full list of p5 values 401 q5 = _mm_andnot_si128(flat2, q5); 402 // q5 remains unchanged if !(flat2 && flat && mask) 403 flat2_q5 = _mm_and_si128(flat2, flat2_q5); 404 // get values for when (flat2 && flat && mask) 405 q5 = _mm_or_si128(q5, flat2_q5); 406 // full list of q5 values 407 _mm_store_si128((__m128i *)(s - 6 * p), p5); 408 _mm_store_si128((__m128i *)(s + 5 * p), q5); 409 410 p4 = _mm_andnot_si128(flat2, p4); 411 // p4 remains unchanged if !(flat2 && flat && mask) 412 flat2_p4 = _mm_and_si128(flat2, flat2_p4); 413 // get values for when (flat2 && flat && mask) 414 p4 = _mm_or_si128(p4, flat2_p4); // full list of p4 values 415 q4 = _mm_andnot_si128(flat2, q4); 416 // q4 remains unchanged if !(flat2 && flat && mask) 417 flat2_q4 = _mm_and_si128(flat2, flat2_q4); 418 // get values for when (flat2 && flat && mask) 419 q4 = _mm_or_si128(q4, flat2_q4); // full list of q4 values 420 _mm_store_si128((__m128i *)(s - 5 * p), p4); 421 _mm_store_si128((__m128i *)(s + 4 * p), q4); 422 423 p3 = _mm_andnot_si128(flat2, p3); 424 // p3 takes value from highbd_filter8 if !(flat2 && flat && mask) 425 flat2_p3 = _mm_and_si128(flat2, flat2_p3); 426 // get values for when (flat2 && flat && mask) 427 p3 = _mm_or_si128(p3, flat2_p3); // full list of p3 values 428 q3 = _mm_andnot_si128(flat2, q3); 429 // q3 takes value from highbd_filter8 if !(flat2 && flat && mask) 430 flat2_q3 = _mm_and_si128(flat2, flat2_q3); 431 // get values for when (flat2 && flat && mask) 432 q3 = _mm_or_si128(q3, flat2_q3); // full list of q3 values 433 _mm_store_si128((__m128i *)(s - 4 * p), p3); 434 _mm_store_si128((__m128i *)(s + 3 * p), q3); 435 436 p2 = _mm_andnot_si128(flat2, p2); 437 // p2 takes value from highbd_filter8 if !(flat2 && flat && mask) 438 flat2_p2 = _mm_and_si128(flat2, flat2_p2); 439 // get values for when (flat2 && flat && mask) 440 p2 = _mm_or_si128(p2, flat2_p2); 441 // full list of p2 values 442 q2 = _mm_andnot_si128(flat2, q2); 443 // q2 takes value from highbd_filter8 if !(flat2 && flat && mask) 444 flat2_q2 = _mm_and_si128(flat2, flat2_q2); 445 // get values for when (flat2 && flat && mask) 446 q2 = _mm_or_si128(q2, flat2_q2); // full list of q2 values 447 _mm_store_si128((__m128i *)(s - 3 * p), p2); 448 _mm_store_si128((__m128i *)(s + 2 * p), q2); 449 450 p1 = _mm_andnot_si128(flat2, p1); 451 // p1 takes value from highbd_filter8 if !(flat2 && flat && mask) 452 flat2_p1 = _mm_and_si128(flat2, flat2_p1); 453 // get values for when (flat2 && flat && mask) 454 p1 = _mm_or_si128(p1, flat2_p1); // full list of p1 values 455 q1 = _mm_andnot_si128(flat2, q1); 456 // q1 takes value from highbd_filter8 if !(flat2 && flat && mask) 457 flat2_q1 = _mm_and_si128(flat2, flat2_q1); 458 // get values for when (flat2 && flat && mask) 459 q1 = _mm_or_si128(q1, flat2_q1); // full list of q1 values 460 _mm_store_si128((__m128i *)(s - 2 * p), p1); 461 _mm_store_si128((__m128i *)(s + 1 * p), q1); 462 463 p0 = _mm_andnot_si128(flat2, p0); 464 // p0 takes value from highbd_filter8 if !(flat2 && flat && mask) 465 flat2_p0 = _mm_and_si128(flat2, flat2_p0); 466 // get values for when (flat2 && flat && mask) 467 p0 = _mm_or_si128(p0, flat2_p0); // full list of p0 values 468 q0 = _mm_andnot_si128(flat2, q0); 469 // q0 takes value from highbd_filter8 if !(flat2 && flat && mask) 470 flat2_q0 = _mm_and_si128(flat2, flat2_q0); 471 // get values for when (flat2 && flat && mask) 472 q0 = _mm_or_si128(q0, flat2_q0); // full list of q0 values 473 _mm_store_si128((__m128i *)(s - 1 * p), p0); 474 _mm_store_si128((__m128i *)(s - 0 * p), q0); 475 } 476 477 void vpx_highbd_lpf_horizontal_16_dual_sse2(uint16_t *s, int p, 478 const uint8_t *_blimit, 479 const uint8_t *_limit, 480 const uint8_t *_thresh, int bd) { 481 vpx_highbd_lpf_horizontal_16_sse2(s, p, _blimit, _limit, _thresh, bd); 482 vpx_highbd_lpf_horizontal_16_sse2(s + 8, p, _blimit, _limit, _thresh, bd); 483 } 484 485 void vpx_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p, 486 const uint8_t *_blimit, 487 const uint8_t *_limit, 488 const uint8_t *_thresh, int bd) { 489 DECLARE_ALIGNED(16, uint16_t, flat_op2[16]); 490 DECLARE_ALIGNED(16, uint16_t, flat_op1[16]); 491 DECLARE_ALIGNED(16, uint16_t, flat_op0[16]); 492 DECLARE_ALIGNED(16, uint16_t, flat_oq2[16]); 493 DECLARE_ALIGNED(16, uint16_t, flat_oq1[16]); 494 DECLARE_ALIGNED(16, uint16_t, flat_oq0[16]); 495 const __m128i zero = _mm_set1_epi16(0); 496 __m128i blimit, limit, thresh; 497 __m128i mask, hev, flat; 498 __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p)); 499 __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p)); 500 __m128i p2 = _mm_load_si128((__m128i *)(s - 3 * p)); 501 __m128i q2 = _mm_load_si128((__m128i *)(s + 2 * p)); 502 __m128i p1 = _mm_load_si128((__m128i *)(s - 2 * p)); 503 __m128i q1 = _mm_load_si128((__m128i *)(s + 1 * p)); 504 __m128i p0 = _mm_load_si128((__m128i *)(s - 1 * p)); 505 __m128i q0 = _mm_load_si128((__m128i *)(s + 0 * p)); 506 const __m128i one = _mm_set1_epi16(1); 507 const __m128i ffff = _mm_cmpeq_epi16(one, one); 508 __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; 509 const __m128i four = _mm_set1_epi16(4); 510 __m128i workp_a, workp_b, workp_shft; 511 512 const __m128i t4 = _mm_set1_epi16(4); 513 const __m128i t3 = _mm_set1_epi16(3); 514 __m128i t80; 515 const __m128i t1 = _mm_set1_epi16(0x1); 516 __m128i ps1, ps0, qs0, qs1; 517 __m128i filt; 518 __m128i work_a; 519 __m128i filter1, filter2; 520 521 if (bd == 8) { 522 blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero); 523 limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero); 524 thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero); 525 t80 = _mm_set1_epi16(0x80); 526 } else if (bd == 10) { 527 blimit = _mm_slli_epi16( 528 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2); 529 limit = _mm_slli_epi16( 530 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2); 531 thresh = _mm_slli_epi16( 532 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2); 533 t80 = _mm_set1_epi16(0x200); 534 } else { // bd == 12 535 blimit = _mm_slli_epi16( 536 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4); 537 limit = _mm_slli_epi16( 538 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4); 539 thresh = _mm_slli_epi16( 540 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4); 541 t80 = _mm_set1_epi16(0x800); 542 } 543 544 ps1 = _mm_subs_epi16(p1, t80); 545 ps0 = _mm_subs_epi16(p0, t80); 546 qs0 = _mm_subs_epi16(q0, t80); 547 qs1 = _mm_subs_epi16(q1, t80); 548 549 // filter_mask and hev_mask 550 abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)); 551 abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1)); 552 553 abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0)); 554 abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1)); 555 flat = _mm_max_epi16(abs_p1p0, abs_q1q0); 556 hev = _mm_subs_epu16(flat, thresh); 557 hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff); 558 559 abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); 560 abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); 561 mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit); 562 mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); 563 // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 564 // So taking maximums continues to work: 565 mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one)); 566 mask = _mm_max_epi16(abs_p1p0, mask); 567 // mask |= (abs(p1 - p0) > limit) * -1; 568 mask = _mm_max_epi16(abs_q1q0, mask); 569 // mask |= (abs(q1 - q0) > limit) * -1; 570 571 work = _mm_max_epi16( 572 _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)), 573 _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2))); 574 mask = _mm_max_epi16(work, mask); 575 work = _mm_max_epi16( 576 _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3)), 577 _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3))); 578 mask = _mm_max_epi16(work, mask); 579 mask = _mm_subs_epu16(mask, limit); 580 mask = _mm_cmpeq_epi16(mask, zero); 581 582 // flat_mask4 583 flat = _mm_max_epi16( 584 _mm_or_si128(_mm_subs_epu16(p2, p0), _mm_subs_epu16(p0, p2)), 585 _mm_or_si128(_mm_subs_epu16(q2, q0), _mm_subs_epu16(q0, q2))); 586 work = _mm_max_epi16( 587 _mm_or_si128(_mm_subs_epu16(p3, p0), _mm_subs_epu16(p0, p3)), 588 _mm_or_si128(_mm_subs_epu16(q3, q0), _mm_subs_epu16(q0, q3))); 589 flat = _mm_max_epi16(work, flat); 590 flat = _mm_max_epi16(abs_p1p0, flat); 591 flat = _mm_max_epi16(abs_q1q0, flat); 592 593 if (bd == 8) 594 flat = _mm_subs_epu16(flat, one); 595 else if (bd == 10) 596 flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2)); 597 else // bd == 12 598 flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4)); 599 600 flat = _mm_cmpeq_epi16(flat, zero); 601 flat = _mm_and_si128(flat, mask); // flat & mask 602 603 // Added before shift for rounding part of ROUND_POWER_OF_TWO 604 605 workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1)); 606 workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0); 607 workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3); 608 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 609 _mm_store_si128((__m128i *)&flat_op2[0], workp_shft); 610 611 workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1); 612 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 613 _mm_store_si128((__m128i *)&flat_op1[0], workp_shft); 614 615 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2); 616 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0); 617 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 618 _mm_store_si128((__m128i *)&flat_op0[0], workp_shft); 619 620 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3); 621 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0); 622 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 623 _mm_store_si128((__m128i *)&flat_oq0[0], workp_shft); 624 625 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3); 626 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1); 627 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 628 _mm_store_si128((__m128i *)&flat_oq1[0], workp_shft); 629 630 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3); 631 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2); 632 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 633 _mm_store_si128((__m128i *)&flat_oq2[0], workp_shft); 634 635 // lp filter 636 filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd); 637 filt = _mm_and_si128(filt, hev); 638 work_a = _mm_subs_epi16(qs0, ps0); 639 filt = _mm_adds_epi16(filt, work_a); 640 filt = _mm_adds_epi16(filt, work_a); 641 filt = _mm_adds_epi16(filt, work_a); 642 // (vpx_filter + 3 * (qs0 - ps0)) & mask 643 filt = signed_char_clamp_bd_sse2(filt, bd); 644 filt = _mm_and_si128(filt, mask); 645 646 filter1 = _mm_adds_epi16(filt, t4); 647 filter2 = _mm_adds_epi16(filt, t3); 648 649 // Filter1 >> 3 650 filter1 = signed_char_clamp_bd_sse2(filter1, bd); 651 filter1 = _mm_srai_epi16(filter1, 3); 652 653 // Filter2 >> 3 654 filter2 = signed_char_clamp_bd_sse2(filter2, bd); 655 filter2 = _mm_srai_epi16(filter2, 3); 656 657 // filt >> 1 658 filt = _mm_adds_epi16(filter1, t1); 659 filt = _mm_srai_epi16(filt, 1); 660 // filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; 661 filt = _mm_andnot_si128(hev, filt); 662 663 work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd); 664 work_a = _mm_adds_epi16(work_a, t80); 665 q0 = _mm_load_si128((__m128i *)flat_oq0); 666 work_a = _mm_andnot_si128(flat, work_a); 667 q0 = _mm_and_si128(flat, q0); 668 q0 = _mm_or_si128(work_a, q0); 669 670 work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd); 671 work_a = _mm_adds_epi16(work_a, t80); 672 q1 = _mm_load_si128((__m128i *)flat_oq1); 673 work_a = _mm_andnot_si128(flat, work_a); 674 q1 = _mm_and_si128(flat, q1); 675 q1 = _mm_or_si128(work_a, q1); 676 677 work_a = _mm_loadu_si128((__m128i *)(s + 2 * p)); 678 q2 = _mm_load_si128((__m128i *)flat_oq2); 679 work_a = _mm_andnot_si128(flat, work_a); 680 q2 = _mm_and_si128(flat, q2); 681 q2 = _mm_or_si128(work_a, q2); 682 683 work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd); 684 work_a = _mm_adds_epi16(work_a, t80); 685 p0 = _mm_load_si128((__m128i *)flat_op0); 686 work_a = _mm_andnot_si128(flat, work_a); 687 p0 = _mm_and_si128(flat, p0); 688 p0 = _mm_or_si128(work_a, p0); 689 690 work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd); 691 work_a = _mm_adds_epi16(work_a, t80); 692 p1 = _mm_load_si128((__m128i *)flat_op1); 693 work_a = _mm_andnot_si128(flat, work_a); 694 p1 = _mm_and_si128(flat, p1); 695 p1 = _mm_or_si128(work_a, p1); 696 697 work_a = _mm_loadu_si128((__m128i *)(s - 3 * p)); 698 p2 = _mm_load_si128((__m128i *)flat_op2); 699 work_a = _mm_andnot_si128(flat, work_a); 700 p2 = _mm_and_si128(flat, p2); 701 p2 = _mm_or_si128(work_a, p2); 702 703 _mm_store_si128((__m128i *)(s - 3 * p), p2); 704 _mm_store_si128((__m128i *)(s - 2 * p), p1); 705 _mm_store_si128((__m128i *)(s - 1 * p), p0); 706 _mm_store_si128((__m128i *)(s + 0 * p), q0); 707 _mm_store_si128((__m128i *)(s + 1 * p), q1); 708 _mm_store_si128((__m128i *)(s + 2 * p), q2); 709 } 710 711 void vpx_highbd_lpf_horizontal_8_dual_sse2( 712 uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, 713 const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, 714 const uint8_t *_thresh1, int bd) { 715 vpx_highbd_lpf_horizontal_8_sse2(s, p, _blimit0, _limit0, _thresh0, bd); 716 vpx_highbd_lpf_horizontal_8_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd); 717 } 718 719 void vpx_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p, 720 const uint8_t *_blimit, 721 const uint8_t *_limit, 722 const uint8_t *_thresh, int bd) { 723 const __m128i zero = _mm_set1_epi16(0); 724 __m128i blimit, limit, thresh; 725 __m128i mask, hev, flat; 726 __m128i p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); 727 __m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); 728 __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); 729 __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); 730 __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); 731 __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); 732 __m128i q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); 733 __m128i q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); 734 const __m128i abs_p1p0 = 735 _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)); 736 const __m128i abs_q1q0 = 737 _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1)); 738 const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0); 739 const __m128i one = _mm_set1_epi16(1); 740 __m128i abs_p0q0 = 741 _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0)); 742 __m128i abs_p1q1 = 743 _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1)); 744 __m128i work; 745 const __m128i t4 = _mm_set1_epi16(4); 746 const __m128i t3 = _mm_set1_epi16(3); 747 __m128i t80; 748 __m128i tff80; 749 __m128i tffe0; 750 __m128i t1f; 751 // equivalent to shifting 0x1f left by bitdepth - 8 752 // and setting new bits to 1 753 const __m128i t1 = _mm_set1_epi16(0x1); 754 __m128i t7f; 755 // equivalent to shifting 0x7f left by bitdepth - 8 756 // and setting new bits to 1 757 __m128i ps1, ps0, qs0, qs1; 758 __m128i filt; 759 __m128i work_a; 760 __m128i filter1, filter2; 761 762 if (bd == 8) { 763 blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero); 764 limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero); 765 thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero); 766 t80 = _mm_set1_epi16(0x80); 767 tff80 = _mm_set1_epi16(0xff80); 768 tffe0 = _mm_set1_epi16(0xffe0); 769 t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8); 770 t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8); 771 } else if (bd == 10) { 772 blimit = _mm_slli_epi16( 773 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2); 774 limit = _mm_slli_epi16( 775 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2); 776 thresh = _mm_slli_epi16( 777 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2); 778 t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 2); 779 tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2); 780 tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2); 781 t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6); 782 t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6); 783 } else { // bd == 12 784 blimit = _mm_slli_epi16( 785 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4); 786 limit = _mm_slli_epi16( 787 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4); 788 thresh = _mm_slli_epi16( 789 _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4); 790 t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 4); 791 tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4); 792 tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4); 793 t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4); 794 t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4); 795 } 796 797 ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80); 798 ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80); 799 qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80); 800 qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80); 801 802 // filter_mask and hev_mask 803 flat = _mm_max_epi16(abs_p1p0, abs_q1q0); 804 hev = _mm_subs_epu16(flat, thresh); 805 hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff); 806 807 abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0); 808 abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); 809 mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit); 810 mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff); 811 // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 812 // So taking maximums continues to work: 813 mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one)); 814 mask = _mm_max_epi16(flat, mask); 815 // mask |= (abs(p1 - p0) > limit) * -1; 816 // mask |= (abs(q1 - q0) > limit) * -1; 817 work = _mm_max_epi16( 818 _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)), 819 _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3))); 820 mask = _mm_max_epi16(work, mask); 821 work = _mm_max_epi16( 822 _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2)), 823 _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3))); 824 mask = _mm_max_epi16(work, mask); 825 mask = _mm_subs_epu16(mask, limit); 826 mask = _mm_cmpeq_epi16(mask, zero); 827 828 // filter4 829 filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd); 830 filt = _mm_and_si128(filt, hev); 831 work_a = _mm_subs_epi16(qs0, ps0); 832 filt = _mm_adds_epi16(filt, work_a); 833 filt = _mm_adds_epi16(filt, work_a); 834 filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd); 835 836 // (vpx_filter + 3 * (qs0 - ps0)) & mask 837 filt = _mm_and_si128(filt, mask); 838 839 filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd); 840 filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd); 841 842 // Filter1 >> 3 843 work_a = _mm_cmpgt_epi16(zero, filter1); // get the values that are <0 844 filter1 = _mm_srli_epi16(filter1, 3); 845 work_a = _mm_and_si128(work_a, tffe0); // sign bits for the values < 0 846 filter1 = _mm_and_si128(filter1, t1f); // clamp the range 847 filter1 = _mm_or_si128(filter1, work_a); // reinsert the sign bits 848 849 // Filter2 >> 3 850 work_a = _mm_cmpgt_epi16(zero, filter2); 851 filter2 = _mm_srli_epi16(filter2, 3); 852 work_a = _mm_and_si128(work_a, tffe0); 853 filter2 = _mm_and_si128(filter2, t1f); 854 filter2 = _mm_or_si128(filter2, work_a); 855 856 // filt >> 1 857 filt = _mm_adds_epi16(filter1, t1); 858 work_a = _mm_cmpgt_epi16(zero, filt); 859 filt = _mm_srli_epi16(filt, 1); 860 work_a = _mm_and_si128(work_a, tff80); 861 filt = _mm_and_si128(filt, t7f); 862 filt = _mm_or_si128(filt, work_a); 863 864 filt = _mm_andnot_si128(hev, filt); 865 866 q0 = _mm_adds_epi16( 867 signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80); 868 q1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), 869 t80); 870 p0 = _mm_adds_epi16( 871 signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80); 872 p1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), 873 t80); 874 875 _mm_storeu_si128((__m128i *)(s - 2 * p), p1); 876 _mm_storeu_si128((__m128i *)(s - 1 * p), p0); 877 _mm_storeu_si128((__m128i *)(s + 0 * p), q0); 878 _mm_storeu_si128((__m128i *)(s + 1 * p), q1); 879 } 880 881 void vpx_highbd_lpf_horizontal_4_dual_sse2( 882 uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0, 883 const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1, 884 const uint8_t *_thresh1, int bd) { 885 vpx_highbd_lpf_horizontal_4_sse2(s, p, _blimit0, _limit0, _thresh0, bd); 886 vpx_highbd_lpf_horizontal_4_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd); 887 } 888 889 static INLINE void highbd_transpose(uint16_t *src[], int in_p, uint16_t *dst[], 890 int out_p, int num_8x8_to_transpose) { 891 int idx8x8 = 0; 892 __m128i p0, p1, p2, p3, p4, p5, p6, p7, x0, x1, x2, x3, x4, x5, x6, x7; 893 do { 894 uint16_t *in = src[idx8x8]; 895 uint16_t *out = dst[idx8x8]; 896 897 p0 = 898 _mm_loadu_si128((__m128i *)(in + 0 * in_p)); // 00 01 02 03 04 05 06 07 899 p1 = 900 _mm_loadu_si128((__m128i *)(in + 1 * in_p)); // 10 11 12 13 14 15 16 17 901 p2 = 902 _mm_loadu_si128((__m128i *)(in + 2 * in_p)); // 20 21 22 23 24 25 26 27 903 p3 = 904 _mm_loadu_si128((__m128i *)(in + 3 * in_p)); // 30 31 32 33 34 35 36 37 905 p4 = 906 _mm_loadu_si128((__m128i *)(in + 4 * in_p)); // 40 41 42 43 44 45 46 47 907 p5 = 908 _mm_loadu_si128((__m128i *)(in + 5 * in_p)); // 50 51 52 53 54 55 56 57 909 p6 = 910 _mm_loadu_si128((__m128i *)(in + 6 * in_p)); // 60 61 62 63 64 65 66 67 911 p7 = 912 _mm_loadu_si128((__m128i *)(in + 7 * in_p)); // 70 71 72 73 74 75 76 77 913 // 00 10 01 11 02 12 03 13 914 x0 = _mm_unpacklo_epi16(p0, p1); 915 // 20 30 21 31 22 32 23 33 916 x1 = _mm_unpacklo_epi16(p2, p3); 917 // 40 50 41 51 42 52 43 53 918 x2 = _mm_unpacklo_epi16(p4, p5); 919 // 60 70 61 71 62 72 63 73 920 x3 = _mm_unpacklo_epi16(p6, p7); 921 // 00 10 20 30 01 11 21 31 922 x4 = _mm_unpacklo_epi32(x0, x1); 923 // 40 50 60 70 41 51 61 71 924 x5 = _mm_unpacklo_epi32(x2, x3); 925 // 00 10 20 30 40 50 60 70 926 x6 = _mm_unpacklo_epi64(x4, x5); 927 // 01 11 21 31 41 51 61 71 928 x7 = _mm_unpackhi_epi64(x4, x5); 929 930 _mm_storeu_si128((__m128i *)(out + 0 * out_p), x6); 931 // 00 10 20 30 40 50 60 70 932 _mm_storeu_si128((__m128i *)(out + 1 * out_p), x7); 933 // 01 11 21 31 41 51 61 71 934 935 // 02 12 22 32 03 13 23 33 936 x4 = _mm_unpackhi_epi32(x0, x1); 937 // 42 52 62 72 43 53 63 73 938 x5 = _mm_unpackhi_epi32(x2, x3); 939 // 02 12 22 32 42 52 62 72 940 x6 = _mm_unpacklo_epi64(x4, x5); 941 // 03 13 23 33 43 53 63 73 942 x7 = _mm_unpackhi_epi64(x4, x5); 943 944 _mm_storeu_si128((__m128i *)(out + 2 * out_p), x6); 945 // 02 12 22 32 42 52 62 72 946 _mm_storeu_si128((__m128i *)(out + 3 * out_p), x7); 947 // 03 13 23 33 43 53 63 73 948 949 // 04 14 05 15 06 16 07 17 950 x0 = _mm_unpackhi_epi16(p0, p1); 951 // 24 34 25 35 26 36 27 37 952 x1 = _mm_unpackhi_epi16(p2, p3); 953 // 44 54 45 55 46 56 47 57 954 x2 = _mm_unpackhi_epi16(p4, p5); 955 // 64 74 65 75 66 76 67 77 956 x3 = _mm_unpackhi_epi16(p6, p7); 957 // 04 14 24 34 05 15 25 35 958 x4 = _mm_unpacklo_epi32(x0, x1); 959 // 44 54 64 74 45 55 65 75 960 x5 = _mm_unpacklo_epi32(x2, x3); 961 // 04 14 24 34 44 54 64 74 962 x6 = _mm_unpacklo_epi64(x4, x5); 963 // 05 15 25 35 45 55 65 75 964 x7 = _mm_unpackhi_epi64(x4, x5); 965 966 _mm_storeu_si128((__m128i *)(out + 4 * out_p), x6); 967 // 04 14 24 34 44 54 64 74 968 _mm_storeu_si128((__m128i *)(out + 5 * out_p), x7); 969 // 05 15 25 35 45 55 65 75 970 971 // 06 16 26 36 07 17 27 37 972 x4 = _mm_unpackhi_epi32(x0, x1); 973 // 46 56 66 76 47 57 67 77 974 x5 = _mm_unpackhi_epi32(x2, x3); 975 // 06 16 26 36 46 56 66 76 976 x6 = _mm_unpacklo_epi64(x4, x5); 977 // 07 17 27 37 47 57 67 77 978 x7 = _mm_unpackhi_epi64(x4, x5); 979 980 _mm_storeu_si128((__m128i *)(out + 6 * out_p), x6); 981 // 06 16 26 36 46 56 66 76 982 _mm_storeu_si128((__m128i *)(out + 7 * out_p), x7); 983 // 07 17 27 37 47 57 67 77 984 } while (++idx8x8 < num_8x8_to_transpose); 985 } 986 987 static INLINE void highbd_transpose8x16(uint16_t *in0, uint16_t *in1, int in_p, 988 uint16_t *out, int out_p) { 989 uint16_t *src0[1]; 990 uint16_t *src1[1]; 991 uint16_t *dest0[1]; 992 uint16_t *dest1[1]; 993 src0[0] = in0; 994 src1[0] = in1; 995 dest0[0] = out; 996 dest1[0] = out + 8; 997 highbd_transpose(src0, in_p, dest0, out_p, 1); 998 highbd_transpose(src1, in_p, dest1, out_p, 1); 999 } 1000 1001 void vpx_highbd_lpf_vertical_4_sse2(uint16_t *s, int p, const uint8_t *blimit, 1002 const uint8_t *limit, const uint8_t *thresh, 1003 int bd) { 1004 DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]); 1005 uint16_t *src[1]; 1006 uint16_t *dst[1]; 1007 1008 // Transpose 8x8 1009 src[0] = s - 4; 1010 dst[0] = t_dst; 1011 1012 highbd_transpose(src, p, dst, 8, 1); 1013 1014 // Loop filtering 1015 vpx_highbd_lpf_horizontal_4_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd); 1016 1017 src[0] = t_dst; 1018 dst[0] = s - 4; 1019 1020 // Transpose back 1021 highbd_transpose(src, 8, dst, p, 1); 1022 } 1023 1024 void vpx_highbd_lpf_vertical_4_dual_sse2( 1025 uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, 1026 const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, 1027 const uint8_t *thresh1, int bd) { 1028 DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]); 1029 uint16_t *src[2]; 1030 uint16_t *dst[2]; 1031 1032 // Transpose 8x16 1033 highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); 1034 1035 // Loop filtering 1036 vpx_highbd_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, 1037 thresh0, blimit1, limit1, thresh1, bd); 1038 src[0] = t_dst; 1039 src[1] = t_dst + 8; 1040 dst[0] = s - 4; 1041 dst[1] = s - 4 + p * 8; 1042 1043 // Transpose back 1044 highbd_transpose(src, 16, dst, p, 2); 1045 } 1046 1047 void vpx_highbd_lpf_vertical_8_sse2(uint16_t *s, int p, const uint8_t *blimit, 1048 const uint8_t *limit, const uint8_t *thresh, 1049 int bd) { 1050 DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]); 1051 uint16_t *src[1]; 1052 uint16_t *dst[1]; 1053 1054 // Transpose 8x8 1055 src[0] = s - 4; 1056 dst[0] = t_dst; 1057 1058 highbd_transpose(src, p, dst, 8, 1); 1059 1060 // Loop filtering 1061 vpx_highbd_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd); 1062 1063 src[0] = t_dst; 1064 dst[0] = s - 4; 1065 1066 // Transpose back 1067 highbd_transpose(src, 8, dst, p, 1); 1068 } 1069 1070 void vpx_highbd_lpf_vertical_8_dual_sse2( 1071 uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0, 1072 const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, 1073 const uint8_t *thresh1, int bd) { 1074 DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]); 1075 uint16_t *src[2]; 1076 uint16_t *dst[2]; 1077 1078 // Transpose 8x16 1079 highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); 1080 1081 // Loop filtering 1082 vpx_highbd_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, 1083 thresh0, blimit1, limit1, thresh1, bd); 1084 src[0] = t_dst; 1085 src[1] = t_dst + 8; 1086 1087 dst[0] = s - 4; 1088 dst[1] = s - 4 + p * 8; 1089 1090 // Transpose back 1091 highbd_transpose(src, 16, dst, p, 2); 1092 } 1093 1094 void vpx_highbd_lpf_vertical_16_sse2(uint16_t *s, int p, const uint8_t *blimit, 1095 const uint8_t *limit, 1096 const uint8_t *thresh, int bd) { 1097 DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 16]); 1098 uint16_t *src[2]; 1099 uint16_t *dst[2]; 1100 1101 src[0] = s - 8; 1102 src[1] = s; 1103 dst[0] = t_dst; 1104 dst[1] = t_dst + 8 * 8; 1105 1106 // Transpose 16x8 1107 highbd_transpose(src, p, dst, 8, 2); 1108 1109 // Loop filtering 1110 vpx_highbd_lpf_horizontal_16_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh, 1111 bd); 1112 src[0] = t_dst; 1113 src[1] = t_dst + 8 * 8; 1114 dst[0] = s - 8; 1115 dst[1] = s; 1116 1117 // Transpose back 1118 highbd_transpose(src, 8, dst, p, 2); 1119 } 1120 1121 void vpx_highbd_lpf_vertical_16_dual_sse2(uint16_t *s, int p, 1122 const uint8_t *blimit, 1123 const uint8_t *limit, 1124 const uint8_t *thresh, int bd) { 1125 DECLARE_ALIGNED(16, uint16_t, t_dst[256]); 1126 1127 // Transpose 16x16 1128 highbd_transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16); 1129 highbd_transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16); 1130 1131 // Loop filtering 1132 vpx_highbd_lpf_horizontal_16_dual_sse2(t_dst + 8 * 16, 16, blimit, limit, 1133 thresh, bd); 1134 1135 // Transpose back 1136 highbd_transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p); 1137 highbd_transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p); 1138 } 1139
