1/* 2 * Copyright (c) 2010 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#include "vpx_ports/emmintrin_compat.h" 16 17static INLINE __m128i abs_diff(__m128i a, __m128i b) { 18 return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a)); 19} 20 21static void mb_lpf_horizontal_edge_w_sse2_8(unsigned char *s, 22 int p, 23 const unsigned char *_blimit, 24 const unsigned char *_limit, 25 const unsigned char *_thresh) { 26 const __m128i zero = _mm_set1_epi16(0); 27 const __m128i one = _mm_set1_epi8(1); 28 const __m128i blimit = _mm_load_si128((const __m128i *)_blimit); 29 const __m128i limit = _mm_load_si128((const __m128i *)_limit); 30 const __m128i thresh = _mm_load_si128((const __m128i *)_thresh); 31 __m128i mask, hev, flat, flat2; 32 __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1; 33 __m128i abs_p1p0; 34 35 q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p)); 36 q4p4 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q4p4), 37 (__m64 *)(s + 4 * p))); 38 q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p)); 39 q3p3 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q3p3), 40 (__m64 *)(s + 3 * p))); 41 q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p)); 42 q2p2 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q2p2), 43 (__m64 *)(s + 2 * p))); 44 q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p)); 45 q1p1 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q1p1), 46 (__m64 *)(s + 1 * p))); 47 p1q1 = _mm_shuffle_epi32(q1p1, 78); 48 q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p)); 49 q0p0 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q0p0), 50 (__m64 *)(s - 0 * p))); 51 p0q0 = _mm_shuffle_epi32(q0p0, 78); 52 53 { 54 __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work; 55 abs_p1p0 = abs_diff(q1p1, q0p0); 56 abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); 57 fe = _mm_set1_epi8(0xfe); 58 ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); 59 abs_p0q0 = abs_diff(q0p0, p0q0); 60 abs_p1q1 = abs_diff(q1p1, p1q1); 61 flat = _mm_max_epu8(abs_p1p0, abs_q1q0); 62 hev = _mm_subs_epu8(flat, thresh); 63 hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); 64 65 abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0); 66 abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); 67 mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); 68 mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); 69 // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 70 mask = _mm_max_epu8(abs_p1p0, mask); 71 // mask |= (abs(p1 - p0) > limit) * -1; 72 // mask |= (abs(q1 - q0) > limit) * -1; 73 74 work = _mm_max_epu8(abs_diff(q2p2, q1p1), 75 abs_diff(q3p3, q2p2)); 76 mask = _mm_max_epu8(work, mask); 77 mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); 78 mask = _mm_subs_epu8(mask, limit); 79 mask = _mm_cmpeq_epi8(mask, zero); 80 } 81 82 // lp filter 83 { 84 const __m128i t4 = _mm_set1_epi8(4); 85 const __m128i t3 = _mm_set1_epi8(3); 86 const __m128i t80 = _mm_set1_epi8(0x80); 87 const __m128i t1 = _mm_set1_epi16(0x1); 88 __m128i qs1ps1 = _mm_xor_si128(q1p1, t80); 89 __m128i qs0ps0 = _mm_xor_si128(q0p0, t80); 90 __m128i qs0 = _mm_xor_si128(p0q0, t80); 91 __m128i qs1 = _mm_xor_si128(p1q1, t80); 92 __m128i filt; 93 __m128i work_a; 94 __m128i filter1, filter2; 95 __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2; 96 __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0; 97 98 filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev); 99 work_a = _mm_subs_epi8(qs0, qs0ps0); 100 filt = _mm_adds_epi8(filt, work_a); 101 filt = _mm_adds_epi8(filt, work_a); 102 filt = _mm_adds_epi8(filt, work_a); 103 // (vpx_filter + 3 * (qs0 - ps0)) & mask 104 filt = _mm_and_si128(filt, mask); 105 106 filter1 = _mm_adds_epi8(filt, t4); 107 filter2 = _mm_adds_epi8(filt, t3); 108 109 filter1 = _mm_unpacklo_epi8(zero, filter1); 110 filter1 = _mm_srai_epi16(filter1, 0xB); 111 filter2 = _mm_unpacklo_epi8(zero, filter2); 112 filter2 = _mm_srai_epi16(filter2, 0xB); 113 114 // Filter1 >> 3 115 filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1)); 116 qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80); 117 118 // filt >> 1 119 filt = _mm_adds_epi16(filter1, t1); 120 filt = _mm_srai_epi16(filt, 1); 121 filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8), 122 filt); 123 filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt)); 124 qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80); 125 // loopfilter done 126 127 { 128 __m128i work; 129 flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0)); 130 flat = _mm_max_epu8(abs_p1p0, flat); 131 flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); 132 flat = _mm_subs_epu8(flat, one); 133 flat = _mm_cmpeq_epi8(flat, zero); 134 flat = _mm_and_si128(flat, mask); 135 136 q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p)); 137 q5p5 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q5p5), 138 (__m64 *)(s + 5 * p))); 139 140 q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p)); 141 q6p6 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q6p6), 142 (__m64 *)(s + 6 * p))); 143 flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0)); 144 145 q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p)); 146 q7p7 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q7p7), 147 (__m64 *)(s + 7 * p))); 148 work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0)); 149 flat2 = _mm_max_epu8(work, flat2); 150 flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8)); 151 flat2 = _mm_subs_epu8(flat2, one); 152 flat2 = _mm_cmpeq_epi8(flat2, zero); 153 flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask 154 } 155 156 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 157 // flat and wide flat calculations 158 { 159 const __m128i eight = _mm_set1_epi16(8); 160 const __m128i four = _mm_set1_epi16(4); 161 __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16; 162 __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16; 163 __m128i pixelFilter_p, pixelFilter_q; 164 __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0; 165 __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q; 166 167 p7_16 = _mm_unpacklo_epi8(q7p7, zero);; 168 p6_16 = _mm_unpacklo_epi8(q6p6, zero); 169 p5_16 = _mm_unpacklo_epi8(q5p5, zero); 170 p4_16 = _mm_unpacklo_epi8(q4p4, zero); 171 p3_16 = _mm_unpacklo_epi8(q3p3, zero); 172 p2_16 = _mm_unpacklo_epi8(q2p2, zero); 173 p1_16 = _mm_unpacklo_epi8(q1p1, zero); 174 p0_16 = _mm_unpacklo_epi8(q0p0, zero); 175 q0_16 = _mm_unpackhi_epi8(q0p0, zero); 176 q1_16 = _mm_unpackhi_epi8(q1p1, zero); 177 q2_16 = _mm_unpackhi_epi8(q2p2, zero); 178 q3_16 = _mm_unpackhi_epi8(q3p3, zero); 179 q4_16 = _mm_unpackhi_epi8(q4p4, zero); 180 q5_16 = _mm_unpackhi_epi8(q5p5, zero); 181 q6_16 = _mm_unpackhi_epi8(q6p6, zero); 182 q7_16 = _mm_unpackhi_epi8(q7p7, zero); 183 184 pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16), 185 _mm_add_epi16(p4_16, p3_16)); 186 pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16), 187 _mm_add_epi16(q4_16, q3_16)); 188 189 pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16)); 190 pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); 191 192 pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16)); 193 pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); 194 pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, 195 pixelFilter_q)); 196 pixetFilter_p2p1p0 = _mm_add_epi16(four, 197 _mm_add_epi16(pixetFilter_p2p1p0, 198 pixetFilter_q2q1q0)); 199 res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 200 _mm_add_epi16(p7_16, p0_16)), 4); 201 res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 202 _mm_add_epi16(q7_16, q0_16)), 4); 203 flat2_q0p0 = _mm_packus_epi16(res_p, res_q); 204 res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0, 205 _mm_add_epi16(p3_16, p0_16)), 3); 206 res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0, 207 _mm_add_epi16(q3_16, q0_16)), 3); 208 209 flat_q0p0 = _mm_packus_epi16(res_p, res_q); 210 211 sum_p7 = _mm_add_epi16(p7_16, p7_16); 212 sum_q7 = _mm_add_epi16(q7_16, q7_16); 213 sum_p3 = _mm_add_epi16(p3_16, p3_16); 214 sum_q3 = _mm_add_epi16(q3_16, q3_16); 215 216 pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16); 217 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16); 218 res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 219 _mm_add_epi16(sum_p7, p1_16)), 4); 220 res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q, 221 _mm_add_epi16(sum_q7, q1_16)), 4); 222 flat2_q1p1 = _mm_packus_epi16(res_p, res_q); 223 224 pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16); 225 pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16); 226 res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0, 227 _mm_add_epi16(sum_p3, p1_16)), 3); 228 res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0, 229 _mm_add_epi16(sum_q3, q1_16)), 3); 230 flat_q1p1 = _mm_packus_epi16(res_p, res_q); 231 232 sum_p7 = _mm_add_epi16(sum_p7, p7_16); 233 sum_q7 = _mm_add_epi16(sum_q7, q7_16); 234 sum_p3 = _mm_add_epi16(sum_p3, p3_16); 235 sum_q3 = _mm_add_epi16(sum_q3, q3_16); 236 237 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16); 238 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16); 239 res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 240 _mm_add_epi16(sum_p7, p2_16)), 4); 241 res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q, 242 _mm_add_epi16(sum_q7, q2_16)), 4); 243 flat2_q2p2 = _mm_packus_epi16(res_p, res_q); 244 245 pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16); 246 pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16); 247 248 res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0, 249 _mm_add_epi16(sum_p3, p2_16)), 3); 250 res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0, 251 _mm_add_epi16(sum_q3, q2_16)), 3); 252 flat_q2p2 = _mm_packus_epi16(res_p, res_q); 253 254 sum_p7 = _mm_add_epi16(sum_p7, p7_16); 255 sum_q7 = _mm_add_epi16(sum_q7, q7_16); 256 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16); 257 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16); 258 res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 259 _mm_add_epi16(sum_p7, p3_16)), 4); 260 res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q, 261 _mm_add_epi16(sum_q7, q3_16)), 4); 262 flat2_q3p3 = _mm_packus_epi16(res_p, res_q); 263 264 sum_p7 = _mm_add_epi16(sum_p7, p7_16); 265 sum_q7 = _mm_add_epi16(sum_q7, q7_16); 266 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16); 267 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16); 268 res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 269 _mm_add_epi16(sum_p7, p4_16)), 4); 270 res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q, 271 _mm_add_epi16(sum_q7, q4_16)), 4); 272 flat2_q4p4 = _mm_packus_epi16(res_p, res_q); 273 274 sum_p7 = _mm_add_epi16(sum_p7, p7_16); 275 sum_q7 = _mm_add_epi16(sum_q7, q7_16); 276 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16); 277 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16); 278 res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 279 _mm_add_epi16(sum_p7, p5_16)), 4); 280 res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q, 281 _mm_add_epi16(sum_q7, q5_16)), 4); 282 flat2_q5p5 = _mm_packus_epi16(res_p, res_q); 283 284 sum_p7 = _mm_add_epi16(sum_p7, p7_16); 285 sum_q7 = _mm_add_epi16(sum_q7, q7_16); 286 pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16); 287 pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16); 288 res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, 289 _mm_add_epi16(sum_p7, p6_16)), 4); 290 res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q, 291 _mm_add_epi16(sum_q7, q6_16)), 4); 292 flat2_q6p6 = _mm_packus_epi16(res_p, res_q); 293 } 294 // wide flat 295 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 296 297 flat = _mm_shuffle_epi32(flat, 68); 298 flat2 = _mm_shuffle_epi32(flat2, 68); 299 300 q2p2 = _mm_andnot_si128(flat, q2p2); 301 flat_q2p2 = _mm_and_si128(flat, flat_q2p2); 302 q2p2 = _mm_or_si128(q2p2, flat_q2p2); 303 304 qs1ps1 = _mm_andnot_si128(flat, qs1ps1); 305 flat_q1p1 = _mm_and_si128(flat, flat_q1p1); 306 q1p1 = _mm_or_si128(qs1ps1, flat_q1p1); 307 308 qs0ps0 = _mm_andnot_si128(flat, qs0ps0); 309 flat_q0p0 = _mm_and_si128(flat, flat_q0p0); 310 q0p0 = _mm_or_si128(qs0ps0, flat_q0p0); 311 312 q6p6 = _mm_andnot_si128(flat2, q6p6); 313 flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6); 314 q6p6 = _mm_or_si128(q6p6, flat2_q6p6); 315 _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6); 316 _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6)); 317 318 q5p5 = _mm_andnot_si128(flat2, q5p5); 319 flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5); 320 q5p5 = _mm_or_si128(q5p5, flat2_q5p5); 321 _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5); 322 _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5)); 323 324 q4p4 = _mm_andnot_si128(flat2, q4p4); 325 flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4); 326 q4p4 = _mm_or_si128(q4p4, flat2_q4p4); 327 _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4); 328 _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4)); 329 330 q3p3 = _mm_andnot_si128(flat2, q3p3); 331 flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3); 332 q3p3 = _mm_or_si128(q3p3, flat2_q3p3); 333 _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3); 334 _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3)); 335 336 q2p2 = _mm_andnot_si128(flat2, q2p2); 337 flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2); 338 q2p2 = _mm_or_si128(q2p2, flat2_q2p2); 339 _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2); 340 _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2)); 341 342 q1p1 = _mm_andnot_si128(flat2, q1p1); 343 flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1); 344 q1p1 = _mm_or_si128(q1p1, flat2_q1p1); 345 _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1); 346 _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1)); 347 348 q0p0 = _mm_andnot_si128(flat2, q0p0); 349 flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0); 350 q0p0 = _mm_or_si128(q0p0, flat2_q0p0); 351 _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0); 352 _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0)); 353 } 354} 355 356static INLINE __m128i filter_add2_sub2(const __m128i *const total, 357 const __m128i *const a1, 358 const __m128i *const a2, 359 const __m128i *const s1, 360 const __m128i *const s2) { 361 __m128i x = _mm_add_epi16(*a1, *total); 362 x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2); 363 return x; 364} 365 366static INLINE __m128i filter8_mask(const __m128i *const flat, 367 const __m128i *const other_filt, 368 const __m128i *const f8_lo, 369 const __m128i *const f8_hi) { 370 const __m128i f8 = _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3), 371 _mm_srli_epi16(*f8_hi, 3)); 372 const __m128i result = _mm_and_si128(*flat, f8); 373 return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result); 374} 375 376static INLINE __m128i filter16_mask(const __m128i *const flat, 377 const __m128i *const other_filt, 378 const __m128i *const f_lo, 379 const __m128i *const f_hi) { 380 const __m128i f = _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4), 381 _mm_srli_epi16(*f_hi, 4)); 382 const __m128i result = _mm_and_si128(*flat, f); 383 return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result); 384} 385 386static void mb_lpf_horizontal_edge_w_sse2_16(unsigned char *s, 387 int p, 388 const unsigned char *_blimit, 389 const unsigned char *_limit, 390 const unsigned char *_thresh) { 391 const __m128i zero = _mm_set1_epi16(0); 392 const __m128i one = _mm_set1_epi8(1); 393 const __m128i blimit = _mm_load_si128((const __m128i *)_blimit); 394 const __m128i limit = _mm_load_si128((const __m128i *)_limit); 395 const __m128i thresh = _mm_load_si128((const __m128i *)_thresh); 396 __m128i mask, hev, flat, flat2; 397 __m128i p7, p6, p5; 398 __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4; 399 __m128i q5, q6, q7; 400 401 __m128i op2, op1, op0, oq0, oq1, oq2; 402 403 __m128i max_abs_p1p0q1q0; 404 405 p7 = _mm_loadu_si128((__m128i *)(s - 8 * p)); 406 p6 = _mm_loadu_si128((__m128i *)(s - 7 * p)); 407 p5 = _mm_loadu_si128((__m128i *)(s - 6 * p)); 408 p4 = _mm_loadu_si128((__m128i *)(s - 5 * p)); 409 p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); 410 p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); 411 p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); 412 p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); 413 q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); 414 q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); 415 q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); 416 q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); 417 q4 = _mm_loadu_si128((__m128i *)(s + 4 * p)); 418 q5 = _mm_loadu_si128((__m128i *)(s + 5 * p)); 419 q6 = _mm_loadu_si128((__m128i *)(s + 6 * p)); 420 q7 = _mm_loadu_si128((__m128i *)(s + 7 * p)); 421 422 { 423 const __m128i abs_p1p0 = abs_diff(p1, p0); 424 const __m128i abs_q1q0 = abs_diff(q1, q0); 425 const __m128i fe = _mm_set1_epi8(0xfe); 426 const __m128i ff = _mm_cmpeq_epi8(zero, zero); 427 __m128i abs_p0q0 = abs_diff(p0, q0); 428 __m128i abs_p1q1 = abs_diff(p1, q1); 429 __m128i work; 430 max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0); 431 432 abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0); 433 abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); 434 mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); 435 mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); 436 // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 437 mask = _mm_max_epu8(max_abs_p1p0q1q0, mask); 438 // mask |= (abs(p1 - p0) > limit) * -1; 439 // mask |= (abs(q1 - q0) > limit) * -1; 440 work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2)); 441 mask = _mm_max_epu8(work, mask); 442 work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2)); 443 mask = _mm_max_epu8(work, mask); 444 mask = _mm_subs_epu8(mask, limit); 445 mask = _mm_cmpeq_epi8(mask, zero); 446 } 447 448 { 449 __m128i work; 450 work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0)); 451 flat = _mm_max_epu8(work, max_abs_p1p0q1q0); 452 work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0)); 453 flat = _mm_max_epu8(work, flat); 454 work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0)); 455 flat = _mm_subs_epu8(flat, one); 456 flat = _mm_cmpeq_epi8(flat, zero); 457 flat = _mm_and_si128(flat, mask); 458 flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0)); 459 flat2 = _mm_max_epu8(work, flat2); 460 work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0)); 461 flat2 = _mm_max_epu8(work, flat2); 462 work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0)); 463 flat2 = _mm_max_epu8(work, flat2); 464 flat2 = _mm_subs_epu8(flat2, one); 465 flat2 = _mm_cmpeq_epi8(flat2, zero); 466 flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask 467 } 468 469 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 470 // filter4 471 { 472 const __m128i t4 = _mm_set1_epi8(4); 473 const __m128i t3 = _mm_set1_epi8(3); 474 const __m128i t80 = _mm_set1_epi8(0x80); 475 const __m128i te0 = _mm_set1_epi8(0xe0); 476 const __m128i t1f = _mm_set1_epi8(0x1f); 477 const __m128i t1 = _mm_set1_epi8(0x1); 478 const __m128i t7f = _mm_set1_epi8(0x7f); 479 const __m128i ff = _mm_cmpeq_epi8(t4, t4); 480 481 __m128i filt; 482 __m128i work_a; 483 __m128i filter1, filter2; 484 485 op1 = _mm_xor_si128(p1, t80); 486 op0 = _mm_xor_si128(p0, t80); 487 oq0 = _mm_xor_si128(q0, t80); 488 oq1 = _mm_xor_si128(q1, t80); 489 490 hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh); 491 hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); 492 filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev); 493 494 work_a = _mm_subs_epi8(oq0, op0); 495 filt = _mm_adds_epi8(filt, work_a); 496 filt = _mm_adds_epi8(filt, work_a); 497 filt = _mm_adds_epi8(filt, work_a); 498 // (vpx_filter + 3 * (qs0 - ps0)) & mask 499 filt = _mm_and_si128(filt, mask); 500 filter1 = _mm_adds_epi8(filt, t4); 501 filter2 = _mm_adds_epi8(filt, t3); 502 503 // Filter1 >> 3 504 work_a = _mm_cmpgt_epi8(zero, filter1); 505 filter1 = _mm_srli_epi16(filter1, 3); 506 work_a = _mm_and_si128(work_a, te0); 507 filter1 = _mm_and_si128(filter1, t1f); 508 filter1 = _mm_or_si128(filter1, work_a); 509 oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80); 510 511 // Filter2 >> 3 512 work_a = _mm_cmpgt_epi8(zero, filter2); 513 filter2 = _mm_srli_epi16(filter2, 3); 514 work_a = _mm_and_si128(work_a, te0); 515 filter2 = _mm_and_si128(filter2, t1f); 516 filter2 = _mm_or_si128(filter2, work_a); 517 op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80); 518 519 // filt >> 1 520 filt = _mm_adds_epi8(filter1, t1); 521 work_a = _mm_cmpgt_epi8(zero, filt); 522 filt = _mm_srli_epi16(filt, 1); 523 work_a = _mm_and_si128(work_a, t80); 524 filt = _mm_and_si128(filt, t7f); 525 filt = _mm_or_si128(filt, work_a); 526 filt = _mm_andnot_si128(hev, filt); 527 op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80); 528 oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80); 529 // loopfilter done 530 531 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 532 // filter8 533 { 534 const __m128i four = _mm_set1_epi16(4); 535 const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero); 536 const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero); 537 const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero); 538 const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero); 539 const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero); 540 const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero); 541 const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero); 542 const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero); 543 544 const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero); 545 const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero); 546 const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero); 547 const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero); 548 const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero); 549 const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero); 550 const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero); 551 const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero); 552 __m128i f8_lo, f8_hi; 553 554 f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four), 555 _mm_add_epi16(p3_lo, p2_lo)); 556 f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo), 557 _mm_add_epi16(p2_lo, p1_lo)); 558 f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo); 559 560 f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four), 561 _mm_add_epi16(p3_hi, p2_hi)); 562 f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi), 563 _mm_add_epi16(p2_hi, p1_hi)); 564 f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi); 565 566 op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi); 567 568 f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo); 569 f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi); 570 op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi); 571 572 f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo); 573 f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi); 574 op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi); 575 576 f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo); 577 f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi); 578 oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi); 579 580 f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo); 581 f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi); 582 oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi); 583 584 f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo); 585 f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi); 586 oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi); 587 } 588 589 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 590 // wide flat calculations 591 { 592 const __m128i eight = _mm_set1_epi16(8); 593 const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero); 594 const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero); 595 const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero); 596 const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero); 597 const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero); 598 const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero); 599 const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero); 600 const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero); 601 const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero); 602 const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero); 603 const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero); 604 const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero); 605 const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero); 606 const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero); 607 const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero); 608 const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero); 609 610 const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero); 611 const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero); 612 const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero); 613 const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero); 614 const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero); 615 const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero); 616 const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero); 617 const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero); 618 const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero); 619 const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero); 620 const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero); 621 const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero); 622 const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero); 623 const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero); 624 const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero); 625 const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero); 626 627 __m128i f_lo; 628 __m128i f_hi; 629 630 f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo); // p7 * 7 631 f_lo = _mm_add_epi16(_mm_slli_epi16(p6_lo, 1), 632 _mm_add_epi16(p4_lo, f_lo)); 633 f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo), 634 _mm_add_epi16(p2_lo, p1_lo)); 635 f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo); 636 f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo); 637 638 f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi); // p7 * 7 639 f_hi = _mm_add_epi16(_mm_slli_epi16(p6_hi, 1), 640 _mm_add_epi16(p4_hi, f_hi)); 641 f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi), 642 _mm_add_epi16(p2_hi, p1_hi)); 643 f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi); 644 f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi); 645 646 p6 = filter16_mask(&flat2, &p6, &f_lo, &f_hi); 647 _mm_storeu_si128((__m128i *)(s - 7 * p), p6); 648 649 f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p5_lo, &p6_lo, &p7_lo); 650 f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p5_hi, &p6_hi, &p7_hi); 651 p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi); 652 _mm_storeu_si128((__m128i *)(s - 6 * p), p5); 653 654 f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p4_lo, &p5_lo, &p7_lo); 655 f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p4_hi, &p5_hi, &p7_hi); 656 p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi); 657 _mm_storeu_si128((__m128i *)(s - 5 * p), p4); 658 659 f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p3_lo, &p4_lo, &p7_lo); 660 f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p3_hi, &p4_hi, &p7_hi); 661 p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi); 662 _mm_storeu_si128((__m128i *)(s - 4 * p), p3); 663 664 f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p2_lo, &p3_lo, &p7_lo); 665 f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p2_hi, &p3_hi, &p7_hi); 666 op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi); 667 _mm_storeu_si128((__m128i *)(s - 3 * p), op2); 668 669 f_lo = filter_add2_sub2(&f_lo, &q5_lo, &p1_lo, &p2_lo, &p7_lo); 670 f_hi = filter_add2_sub2(&f_hi, &q5_hi, &p1_hi, &p2_hi, &p7_hi); 671 op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi); 672 _mm_storeu_si128((__m128i *)(s - 2 * p), op1); 673 674 f_lo = filter_add2_sub2(&f_lo, &q6_lo, &p0_lo, &p1_lo, &p7_lo); 675 f_hi = filter_add2_sub2(&f_hi, &q6_hi, &p0_hi, &p1_hi, &p7_hi); 676 op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi); 677 _mm_storeu_si128((__m128i *)(s - 1 * p), op0); 678 679 f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q0_lo, &p0_lo, &p7_lo); 680 f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q0_hi, &p0_hi, &p7_hi); 681 oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi); 682 _mm_storeu_si128((__m128i *)(s - 0 * p), oq0); 683 684 f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q1_lo, &p6_lo, &q0_lo); 685 f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q1_hi, &p6_hi, &q0_hi); 686 oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi); 687 _mm_storeu_si128((__m128i *)(s + 1 * p), oq1); 688 689 f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q2_lo, &p5_lo, &q1_lo); 690 f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q2_hi, &p5_hi, &q1_hi); 691 oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi); 692 _mm_storeu_si128((__m128i *)(s + 2 * p), oq2); 693 694 f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q3_lo, &p4_lo, &q2_lo); 695 f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q3_hi, &p4_hi, &q2_hi); 696 q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi); 697 _mm_storeu_si128((__m128i *)(s + 3 * p), q3); 698 699 f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q4_lo, &p3_lo, &q3_lo); 700 f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q4_hi, &p3_hi, &q3_hi); 701 q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi); 702 _mm_storeu_si128((__m128i *)(s + 4 * p), q4); 703 704 f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q5_lo, &p2_lo, &q4_lo); 705 f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q5_hi, &p2_hi, &q4_hi); 706 q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi); 707 _mm_storeu_si128((__m128i *)(s + 5 * p), q5); 708 709 f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q6_lo, &p1_lo, &q5_lo); 710 f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q6_hi, &p1_hi, &q5_hi); 711 q6 = filter16_mask(&flat2, &q6, &f_lo, &f_hi); 712 _mm_storeu_si128((__m128i *)(s + 6 * p), q6); 713 } 714 // wide flat 715 // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 716 } 717} 718 719// TODO(yunqingwang): remove count and call these 2 functions(8 or 16) directly. 720void vpx_lpf_horizontal_16_sse2(unsigned char *s, int p, 721 const unsigned char *_blimit, 722 const unsigned char *_limit, 723 const unsigned char *_thresh, int count) { 724 if (count == 1) 725 mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh); 726 else 727 mb_lpf_horizontal_edge_w_sse2_16(s, p, _blimit, _limit, _thresh); 728} 729 730void vpx_lpf_horizontal_8_sse2(unsigned char *s, int p, 731 const unsigned char *_blimit, 732 const unsigned char *_limit, 733 const unsigned char *_thresh, int count) { 734 DECLARE_ALIGNED(16, unsigned char, flat_op2[16]); 735 DECLARE_ALIGNED(16, unsigned char, flat_op1[16]); 736 DECLARE_ALIGNED(16, unsigned char, flat_op0[16]); 737 DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]); 738 DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]); 739 DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]); 740 const __m128i zero = _mm_set1_epi16(0); 741 const __m128i blimit = _mm_load_si128((const __m128i *)_blimit); 742 const __m128i limit = _mm_load_si128((const __m128i *)_limit); 743 const __m128i thresh = _mm_load_si128((const __m128i *)_thresh); 744 __m128i mask, hev, flat; 745 __m128i p3, p2, p1, p0, q0, q1, q2, q3; 746 __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0; 747 748 (void)count; 749 750 q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)), 751 _mm_loadl_epi64((__m128i *)(s + 3 * p))); 752 q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)), 753 _mm_loadl_epi64((__m128i *)(s + 2 * p))); 754 q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)), 755 _mm_loadl_epi64((__m128i *)(s + 1 * p))); 756 q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)), 757 _mm_loadl_epi64((__m128i *)(s - 0 * p))); 758 p1q1 = _mm_shuffle_epi32(q1p1, 78); 759 p0q0 = _mm_shuffle_epi32(q0p0, 78); 760 761 { 762 // filter_mask and hev_mask 763 const __m128i one = _mm_set1_epi8(1); 764 const __m128i fe = _mm_set1_epi8(0xfe); 765 const __m128i ff = _mm_cmpeq_epi8(fe, fe); 766 __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work; 767 abs_p1p0 = abs_diff(q1p1, q0p0); 768 abs_q1q0 = _mm_srli_si128(abs_p1p0, 8); 769 770 abs_p0q0 = abs_diff(q0p0, p0q0); 771 abs_p1q1 = abs_diff(q1p1, p1q1); 772 flat = _mm_max_epu8(abs_p1p0, abs_q1q0); 773 hev = _mm_subs_epu8(flat, thresh); 774 hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); 775 776 abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0); 777 abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); 778 mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); 779 mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); 780 // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 781 mask = _mm_max_epu8(abs_p1p0, mask); 782 // mask |= (abs(p1 - p0) > limit) * -1; 783 // mask |= (abs(q1 - q0) > limit) * -1; 784 785 work = _mm_max_epu8(abs_diff(q2p2, q1p1), 786 abs_diff(q3p3, q2p2)); 787 mask = _mm_max_epu8(work, mask); 788 mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8)); 789 mask = _mm_subs_epu8(mask, limit); 790 mask = _mm_cmpeq_epi8(mask, zero); 791 792 // flat_mask4 793 794 flat = _mm_max_epu8(abs_diff(q2p2, q0p0), 795 abs_diff(q3p3, q0p0)); 796 flat = _mm_max_epu8(abs_p1p0, flat); 797 flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8)); 798 flat = _mm_subs_epu8(flat, one); 799 flat = _mm_cmpeq_epi8(flat, zero); 800 flat = _mm_and_si128(flat, mask); 801 } 802 803 { 804 const __m128i four = _mm_set1_epi16(4); 805 unsigned char *src = s; 806 { 807 __m128i workp_a, workp_b, workp_shft; 808 p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero); 809 p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero); 810 p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero); 811 p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero); 812 q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero); 813 q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero); 814 q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero); 815 q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero); 816 817 workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1)); 818 workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0); 819 workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3); 820 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 821 _mm_storel_epi64((__m128i *)&flat_op2[0], 822 _mm_packus_epi16(workp_shft, workp_shft)); 823 824 workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1); 825 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 826 _mm_storel_epi64((__m128i *)&flat_op1[0], 827 _mm_packus_epi16(workp_shft, workp_shft)); 828 829 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2); 830 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0); 831 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 832 _mm_storel_epi64((__m128i *)&flat_op0[0], 833 _mm_packus_epi16(workp_shft, workp_shft)); 834 835 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3); 836 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0); 837 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 838 _mm_storel_epi64((__m128i *)&flat_oq0[0], 839 _mm_packus_epi16(workp_shft, workp_shft)); 840 841 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3); 842 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1); 843 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 844 _mm_storel_epi64((__m128i *)&flat_oq1[0], 845 _mm_packus_epi16(workp_shft, workp_shft)); 846 847 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3); 848 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2); 849 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 850 _mm_storel_epi64((__m128i *)&flat_oq2[0], 851 _mm_packus_epi16(workp_shft, workp_shft)); 852 } 853 } 854 // lp filter 855 { 856 const __m128i t4 = _mm_set1_epi8(4); 857 const __m128i t3 = _mm_set1_epi8(3); 858 const __m128i t80 = _mm_set1_epi8(0x80); 859 const __m128i t1 = _mm_set1_epi8(0x1); 860 const __m128i ps1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)), 861 t80); 862 const __m128i ps0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)), 863 t80); 864 const __m128i qs0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)), 865 t80); 866 const __m128i qs1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)), 867 t80); 868 __m128i filt; 869 __m128i work_a; 870 __m128i filter1, filter2; 871 872 filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev); 873 work_a = _mm_subs_epi8(qs0, ps0); 874 filt = _mm_adds_epi8(filt, work_a); 875 filt = _mm_adds_epi8(filt, work_a); 876 filt = _mm_adds_epi8(filt, work_a); 877 // (vpx_filter + 3 * (qs0 - ps0)) & mask 878 filt = _mm_and_si128(filt, mask); 879 880 filter1 = _mm_adds_epi8(filt, t4); 881 filter2 = _mm_adds_epi8(filt, t3); 882 883 // Filter1 >> 3 884 filter1 = _mm_unpacklo_epi8(zero, filter1); 885 filter1 = _mm_srai_epi16(filter1, 11); 886 filter1 = _mm_packs_epi16(filter1, filter1); 887 888 // Filter2 >> 3 889 filter2 = _mm_unpacklo_epi8(zero, filter2); 890 filter2 = _mm_srai_epi16(filter2, 11); 891 filter2 = _mm_packs_epi16(filter2, zero); 892 893 // filt >> 1 894 filt = _mm_adds_epi8(filter1, t1); 895 filt = _mm_unpacklo_epi8(zero, filt); 896 filt = _mm_srai_epi16(filt, 9); 897 filt = _mm_packs_epi16(filt, zero); 898 899 filt = _mm_andnot_si128(hev, filt); 900 901 work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80); 902 q0 = _mm_loadl_epi64((__m128i *)flat_oq0); 903 work_a = _mm_andnot_si128(flat, work_a); 904 q0 = _mm_and_si128(flat, q0); 905 q0 = _mm_or_si128(work_a, q0); 906 907 work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80); 908 q1 = _mm_loadl_epi64((__m128i *)flat_oq1); 909 work_a = _mm_andnot_si128(flat, work_a); 910 q1 = _mm_and_si128(flat, q1); 911 q1 = _mm_or_si128(work_a, q1); 912 913 work_a = _mm_loadu_si128((__m128i *)(s + 2 * p)); 914 q2 = _mm_loadl_epi64((__m128i *)flat_oq2); 915 work_a = _mm_andnot_si128(flat, work_a); 916 q2 = _mm_and_si128(flat, q2); 917 q2 = _mm_or_si128(work_a, q2); 918 919 work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80); 920 p0 = _mm_loadl_epi64((__m128i *)flat_op0); 921 work_a = _mm_andnot_si128(flat, work_a); 922 p0 = _mm_and_si128(flat, p0); 923 p0 = _mm_or_si128(work_a, p0); 924 925 work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80); 926 p1 = _mm_loadl_epi64((__m128i *)flat_op1); 927 work_a = _mm_andnot_si128(flat, work_a); 928 p1 = _mm_and_si128(flat, p1); 929 p1 = _mm_or_si128(work_a, p1); 930 931 work_a = _mm_loadu_si128((__m128i *)(s - 3 * p)); 932 p2 = _mm_loadl_epi64((__m128i *)flat_op2); 933 work_a = _mm_andnot_si128(flat, work_a); 934 p2 = _mm_and_si128(flat, p2); 935 p2 = _mm_or_si128(work_a, p2); 936 937 _mm_storel_epi64((__m128i *)(s - 3 * p), p2); 938 _mm_storel_epi64((__m128i *)(s - 2 * p), p1); 939 _mm_storel_epi64((__m128i *)(s - 1 * p), p0); 940 _mm_storel_epi64((__m128i *)(s + 0 * p), q0); 941 _mm_storel_epi64((__m128i *)(s + 1 * p), q1); 942 _mm_storel_epi64((__m128i *)(s + 2 * p), q2); 943 } 944} 945 946void vpx_lpf_horizontal_8_dual_sse2(uint8_t *s, int p, 947 const uint8_t *_blimit0, 948 const uint8_t *_limit0, 949 const uint8_t *_thresh0, 950 const uint8_t *_blimit1, 951 const uint8_t *_limit1, 952 const uint8_t *_thresh1) { 953 DECLARE_ALIGNED(16, unsigned char, flat_op2[16]); 954 DECLARE_ALIGNED(16, unsigned char, flat_op1[16]); 955 DECLARE_ALIGNED(16, unsigned char, flat_op0[16]); 956 DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]); 957 DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]); 958 DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]); 959 const __m128i zero = _mm_set1_epi16(0); 960 const __m128i blimit = 961 _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0), 962 _mm_load_si128((const __m128i *)_blimit1)); 963 const __m128i limit = 964 _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0), 965 _mm_load_si128((const __m128i *)_limit1)); 966 const __m128i thresh = 967 _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0), 968 _mm_load_si128((const __m128i *)_thresh1)); 969 970 __m128i mask, hev, flat; 971 __m128i p3, p2, p1, p0, q0, q1, q2, q3; 972 973 p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); 974 p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); 975 p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); 976 p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); 977 q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); 978 q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); 979 q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); 980 q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); 981 { 982 const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0), 983 _mm_subs_epu8(p0, p1)); 984 const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0), 985 _mm_subs_epu8(q0, q1)); 986 const __m128i one = _mm_set1_epi8(1); 987 const __m128i fe = _mm_set1_epi8(0xfe); 988 const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); 989 __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0), 990 _mm_subs_epu8(q0, p0)); 991 __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1), 992 _mm_subs_epu8(q1, p1)); 993 __m128i work; 994 995 // filter_mask and hev_mask 996 flat = _mm_max_epu8(abs_p1p0, abs_q1q0); 997 hev = _mm_subs_epu8(flat, thresh); 998 hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); 999 1000 abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0); 1001 abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); 1002 mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); 1003 mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); 1004 // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 1005 mask = _mm_max_epu8(flat, mask); 1006 // mask |= (abs(p1 - p0) > limit) * -1; 1007 // mask |= (abs(q1 - q0) > limit) * -1; 1008 work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1), 1009 _mm_subs_epu8(p1, p2)), 1010 _mm_or_si128(_mm_subs_epu8(p3, p2), 1011 _mm_subs_epu8(p2, p3))); 1012 mask = _mm_max_epu8(work, mask); 1013 work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1), 1014 _mm_subs_epu8(q1, q2)), 1015 _mm_or_si128(_mm_subs_epu8(q3, q2), 1016 _mm_subs_epu8(q2, q3))); 1017 mask = _mm_max_epu8(work, mask); 1018 mask = _mm_subs_epu8(mask, limit); 1019 mask = _mm_cmpeq_epi8(mask, zero); 1020 1021 // flat_mask4 1022 work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0), 1023 _mm_subs_epu8(p0, p2)), 1024 _mm_or_si128(_mm_subs_epu8(q2, q0), 1025 _mm_subs_epu8(q0, q2))); 1026 flat = _mm_max_epu8(work, flat); 1027 work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0), 1028 _mm_subs_epu8(p0, p3)), 1029 _mm_or_si128(_mm_subs_epu8(q3, q0), 1030 _mm_subs_epu8(q0, q3))); 1031 flat = _mm_max_epu8(work, flat); 1032 flat = _mm_subs_epu8(flat, one); 1033 flat = _mm_cmpeq_epi8(flat, zero); 1034 flat = _mm_and_si128(flat, mask); 1035 } 1036 { 1037 const __m128i four = _mm_set1_epi16(4); 1038 unsigned char *src = s; 1039 int i = 0; 1040 1041 do { 1042 __m128i workp_a, workp_b, workp_shft; 1043 p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero); 1044 p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero); 1045 p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero); 1046 p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero); 1047 q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero); 1048 q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero); 1049 q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero); 1050 q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero); 1051 1052 workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1)); 1053 workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0); 1054 workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3); 1055 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 1056 _mm_storel_epi64((__m128i *)&flat_op2[i * 8], 1057 _mm_packus_epi16(workp_shft, workp_shft)); 1058 1059 workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1); 1060 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 1061 _mm_storel_epi64((__m128i *)&flat_op1[i * 8], 1062 _mm_packus_epi16(workp_shft, workp_shft)); 1063 1064 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2); 1065 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0); 1066 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 1067 _mm_storel_epi64((__m128i *)&flat_op0[i * 8], 1068 _mm_packus_epi16(workp_shft, workp_shft)); 1069 1070 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3); 1071 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0); 1072 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 1073 _mm_storel_epi64((__m128i *)&flat_oq0[i * 8], 1074 _mm_packus_epi16(workp_shft, workp_shft)); 1075 1076 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3); 1077 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1); 1078 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 1079 _mm_storel_epi64((__m128i *)&flat_oq1[i * 8], 1080 _mm_packus_epi16(workp_shft, workp_shft)); 1081 1082 workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3); 1083 workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2); 1084 workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3); 1085 _mm_storel_epi64((__m128i *)&flat_oq2[i * 8], 1086 _mm_packus_epi16(workp_shft, workp_shft)); 1087 1088 src += 8; 1089 } while (++i < 2); 1090 } 1091 // lp filter 1092 { 1093 const __m128i t4 = _mm_set1_epi8(4); 1094 const __m128i t3 = _mm_set1_epi8(3); 1095 const __m128i t80 = _mm_set1_epi8(0x80); 1096 const __m128i te0 = _mm_set1_epi8(0xe0); 1097 const __m128i t1f = _mm_set1_epi8(0x1f); 1098 const __m128i t1 = _mm_set1_epi8(0x1); 1099 const __m128i t7f = _mm_set1_epi8(0x7f); 1100 1101 const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)), 1102 t80); 1103 const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)), 1104 t80); 1105 const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)), 1106 t80); 1107 const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)), 1108 t80); 1109 __m128i filt; 1110 __m128i work_a; 1111 __m128i filter1, filter2; 1112 1113 filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev); 1114 work_a = _mm_subs_epi8(qs0, ps0); 1115 filt = _mm_adds_epi8(filt, work_a); 1116 filt = _mm_adds_epi8(filt, work_a); 1117 filt = _mm_adds_epi8(filt, work_a); 1118 // (vpx_filter + 3 * (qs0 - ps0)) & mask 1119 filt = _mm_and_si128(filt, mask); 1120 1121 filter1 = _mm_adds_epi8(filt, t4); 1122 filter2 = _mm_adds_epi8(filt, t3); 1123 1124 // Filter1 >> 3 1125 work_a = _mm_cmpgt_epi8(zero, filter1); 1126 filter1 = _mm_srli_epi16(filter1, 3); 1127 work_a = _mm_and_si128(work_a, te0); 1128 filter1 = _mm_and_si128(filter1, t1f); 1129 filter1 = _mm_or_si128(filter1, work_a); 1130 1131 // Filter2 >> 3 1132 work_a = _mm_cmpgt_epi8(zero, filter2); 1133 filter2 = _mm_srli_epi16(filter2, 3); 1134 work_a = _mm_and_si128(work_a, te0); 1135 filter2 = _mm_and_si128(filter2, t1f); 1136 filter2 = _mm_or_si128(filter2, work_a); 1137 1138 // filt >> 1 1139 filt = _mm_adds_epi8(filter1, t1); 1140 work_a = _mm_cmpgt_epi8(zero, filt); 1141 filt = _mm_srli_epi16(filt, 1); 1142 work_a = _mm_and_si128(work_a, t80); 1143 filt = _mm_and_si128(filt, t7f); 1144 filt = _mm_or_si128(filt, work_a); 1145 1146 filt = _mm_andnot_si128(hev, filt); 1147 1148 work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80); 1149 q0 = _mm_load_si128((__m128i *)flat_oq0); 1150 work_a = _mm_andnot_si128(flat, work_a); 1151 q0 = _mm_and_si128(flat, q0); 1152 q0 = _mm_or_si128(work_a, q0); 1153 1154 work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80); 1155 q1 = _mm_load_si128((__m128i *)flat_oq1); 1156 work_a = _mm_andnot_si128(flat, work_a); 1157 q1 = _mm_and_si128(flat, q1); 1158 q1 = _mm_or_si128(work_a, q1); 1159 1160 work_a = _mm_loadu_si128((__m128i *)(s + 2 * p)); 1161 q2 = _mm_load_si128((__m128i *)flat_oq2); 1162 work_a = _mm_andnot_si128(flat, work_a); 1163 q2 = _mm_and_si128(flat, q2); 1164 q2 = _mm_or_si128(work_a, q2); 1165 1166 work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80); 1167 p0 = _mm_load_si128((__m128i *)flat_op0); 1168 work_a = _mm_andnot_si128(flat, work_a); 1169 p0 = _mm_and_si128(flat, p0); 1170 p0 = _mm_or_si128(work_a, p0); 1171 1172 work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80); 1173 p1 = _mm_load_si128((__m128i *)flat_op1); 1174 work_a = _mm_andnot_si128(flat, work_a); 1175 p1 = _mm_and_si128(flat, p1); 1176 p1 = _mm_or_si128(work_a, p1); 1177 1178 work_a = _mm_loadu_si128((__m128i *)(s - 3 * p)); 1179 p2 = _mm_load_si128((__m128i *)flat_op2); 1180 work_a = _mm_andnot_si128(flat, work_a); 1181 p2 = _mm_and_si128(flat, p2); 1182 p2 = _mm_or_si128(work_a, p2); 1183 1184 _mm_storeu_si128((__m128i *)(s - 3 * p), p2); 1185 _mm_storeu_si128((__m128i *)(s - 2 * p), p1); 1186 _mm_storeu_si128((__m128i *)(s - 1 * p), p0); 1187 _mm_storeu_si128((__m128i *)(s + 0 * p), q0); 1188 _mm_storeu_si128((__m128i *)(s + 1 * p), q1); 1189 _mm_storeu_si128((__m128i *)(s + 2 * p), q2); 1190 } 1191} 1192 1193void vpx_lpf_horizontal_4_dual_sse2(unsigned char *s, int p, 1194 const unsigned char *_blimit0, 1195 const unsigned char *_limit0, 1196 const unsigned char *_thresh0, 1197 const unsigned char *_blimit1, 1198 const unsigned char *_limit1, 1199 const unsigned char *_thresh1) { 1200 const __m128i blimit = 1201 _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0), 1202 _mm_load_si128((const __m128i *)_blimit1)); 1203 const __m128i limit = 1204 _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0), 1205 _mm_load_si128((const __m128i *)_limit1)); 1206 const __m128i thresh = 1207 _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0), 1208 _mm_load_si128((const __m128i *)_thresh1)); 1209 const __m128i zero = _mm_set1_epi16(0); 1210 __m128i p3, p2, p1, p0, q0, q1, q2, q3; 1211 __m128i mask, hev, flat; 1212 1213 p3 = _mm_loadu_si128((__m128i *)(s - 4 * p)); 1214 p2 = _mm_loadu_si128((__m128i *)(s - 3 * p)); 1215 p1 = _mm_loadu_si128((__m128i *)(s - 2 * p)); 1216 p0 = _mm_loadu_si128((__m128i *)(s - 1 * p)); 1217 q0 = _mm_loadu_si128((__m128i *)(s - 0 * p)); 1218 q1 = _mm_loadu_si128((__m128i *)(s + 1 * p)); 1219 q2 = _mm_loadu_si128((__m128i *)(s + 2 * p)); 1220 q3 = _mm_loadu_si128((__m128i *)(s + 3 * p)); 1221 1222 // filter_mask and hev_mask 1223 { 1224 const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0), 1225 _mm_subs_epu8(p0, p1)); 1226 const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0), 1227 _mm_subs_epu8(q0, q1)); 1228 const __m128i fe = _mm_set1_epi8(0xfe); 1229 const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0); 1230 __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0), 1231 _mm_subs_epu8(q0, p0)); 1232 __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1), 1233 _mm_subs_epu8(q1, p1)); 1234 __m128i work; 1235 1236 flat = _mm_max_epu8(abs_p1p0, abs_q1q0); 1237 hev = _mm_subs_epu8(flat, thresh); 1238 hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff); 1239 1240 abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0); 1241 abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1); 1242 mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit); 1243 mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff); 1244 // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 1245 mask = _mm_max_epu8(flat, mask); 1246 // mask |= (abs(p1 - p0) > limit) * -1; 1247 // mask |= (abs(q1 - q0) > limit) * -1; 1248 work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1), 1249 _mm_subs_epu8(p1, p2)), 1250 _mm_or_si128(_mm_subs_epu8(p3, p2), 1251 _mm_subs_epu8(p2, p3))); 1252 mask = _mm_max_epu8(work, mask); 1253 work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1), 1254 _mm_subs_epu8(q1, q2)), 1255 _mm_or_si128(_mm_subs_epu8(q3, q2), 1256 _mm_subs_epu8(q2, q3))); 1257 mask = _mm_max_epu8(work, mask); 1258 mask = _mm_subs_epu8(mask, limit); 1259 mask = _mm_cmpeq_epi8(mask, zero); 1260 } 1261 1262 // filter4 1263 { 1264 const __m128i t4 = _mm_set1_epi8(4); 1265 const __m128i t3 = _mm_set1_epi8(3); 1266 const __m128i t80 = _mm_set1_epi8(0x80); 1267 const __m128i te0 = _mm_set1_epi8(0xe0); 1268 const __m128i t1f = _mm_set1_epi8(0x1f); 1269 const __m128i t1 = _mm_set1_epi8(0x1); 1270 const __m128i t7f = _mm_set1_epi8(0x7f); 1271 1272 const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)), 1273 t80); 1274 const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)), 1275 t80); 1276 const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)), 1277 t80); 1278 const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)), 1279 t80); 1280 __m128i filt; 1281 __m128i work_a; 1282 __m128i filter1, filter2; 1283 1284 filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev); 1285 work_a = _mm_subs_epi8(qs0, ps0); 1286 filt = _mm_adds_epi8(filt, work_a); 1287 filt = _mm_adds_epi8(filt, work_a); 1288 filt = _mm_adds_epi8(filt, work_a); 1289 // (vpx_filter + 3 * (qs0 - ps0)) & mask 1290 filt = _mm_and_si128(filt, mask); 1291 1292 filter1 = _mm_adds_epi8(filt, t4); 1293 filter2 = _mm_adds_epi8(filt, t3); 1294 1295 // Filter1 >> 3 1296 work_a = _mm_cmpgt_epi8(zero, filter1); 1297 filter1 = _mm_srli_epi16(filter1, 3); 1298 work_a = _mm_and_si128(work_a, te0); 1299 filter1 = _mm_and_si128(filter1, t1f); 1300 filter1 = _mm_or_si128(filter1, work_a); 1301 1302 // Filter2 >> 3 1303 work_a = _mm_cmpgt_epi8(zero, filter2); 1304 filter2 = _mm_srli_epi16(filter2, 3); 1305 work_a = _mm_and_si128(work_a, te0); 1306 filter2 = _mm_and_si128(filter2, t1f); 1307 filter2 = _mm_or_si128(filter2, work_a); 1308 1309 // filt >> 1 1310 filt = _mm_adds_epi8(filter1, t1); 1311 work_a = _mm_cmpgt_epi8(zero, filt); 1312 filt = _mm_srli_epi16(filt, 1); 1313 work_a = _mm_and_si128(work_a, t80); 1314 filt = _mm_and_si128(filt, t7f); 1315 filt = _mm_or_si128(filt, work_a); 1316 1317 filt = _mm_andnot_si128(hev, filt); 1318 1319 q0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80); 1320 q1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80); 1321 p0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80); 1322 p1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80); 1323 1324 _mm_storeu_si128((__m128i *)(s - 2 * p), p1); 1325 _mm_storeu_si128((__m128i *)(s - 1 * p), p0); 1326 _mm_storeu_si128((__m128i *)(s + 0 * p), q0); 1327 _mm_storeu_si128((__m128i *)(s + 1 * p), q1); 1328 } 1329} 1330 1331static INLINE void transpose8x16(unsigned char *in0, unsigned char *in1, 1332 int in_p, unsigned char *out, int out_p) { 1333 __m128i x0, x1, x2, x3, x4, x5, x6, x7; 1334 __m128i x8, x9, x10, x11, x12, x13, x14, x15; 1335 1336 // 2-way interleave w/hoisting of unpacks 1337 x0 = _mm_loadl_epi64((__m128i *)in0); // 1 1338 x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p)); // 3 1339 x0 = _mm_unpacklo_epi8(x0, x1); // 1 1340 1341 x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p)); // 5 1342 x3 = _mm_loadl_epi64((__m128i *)(in0 + 3*in_p)); // 7 1343 x1 = _mm_unpacklo_epi8(x2, x3); // 2 1344 1345 x4 = _mm_loadl_epi64((__m128i *)(in0 + 4*in_p)); // 9 1346 x5 = _mm_loadl_epi64((__m128i *)(in0 + 5*in_p)); // 11 1347 x2 = _mm_unpacklo_epi8(x4, x5); // 3 1348 1349 x6 = _mm_loadl_epi64((__m128i *)(in0 + 6*in_p)); // 13 1350 x7 = _mm_loadl_epi64((__m128i *)(in0 + 7*in_p)); // 15 1351 x3 = _mm_unpacklo_epi8(x6, x7); // 4 1352 x4 = _mm_unpacklo_epi16(x0, x1); // 9 1353 1354 x8 = _mm_loadl_epi64((__m128i *)in1); // 2 1355 x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p)); // 4 1356 x8 = _mm_unpacklo_epi8(x8, x9); // 5 1357 x5 = _mm_unpacklo_epi16(x2, x3); // 10 1358 1359 x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p)); // 6 1360 x11 = _mm_loadl_epi64((__m128i *)(in1 + 3*in_p)); // 8 1361 x9 = _mm_unpacklo_epi8(x10, x11); // 6 1362 1363 x12 = _mm_loadl_epi64((__m128i *)(in1 + 4*in_p)); // 10 1364 x13 = _mm_loadl_epi64((__m128i *)(in1 + 5*in_p)); // 12 1365 x10 = _mm_unpacklo_epi8(x12, x13); // 7 1366 x12 = _mm_unpacklo_epi16(x8, x9); // 11 1367 1368 x14 = _mm_loadl_epi64((__m128i *)(in1 + 6*in_p)); // 14 1369 x15 = _mm_loadl_epi64((__m128i *)(in1 + 7*in_p)); // 16 1370 x11 = _mm_unpacklo_epi8(x14, x15); // 8 1371 x13 = _mm_unpacklo_epi16(x10, x11); // 12 1372 1373 x6 = _mm_unpacklo_epi32(x4, x5); // 13 1374 x7 = _mm_unpackhi_epi32(x4, x5); // 14 1375 x14 = _mm_unpacklo_epi32(x12, x13); // 15 1376 x15 = _mm_unpackhi_epi32(x12, x13); // 16 1377 1378 // Store first 4-line result 1379 _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14)); 1380 _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14)); 1381 _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15)); 1382 _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15)); 1383 1384 x4 = _mm_unpackhi_epi16(x0, x1); 1385 x5 = _mm_unpackhi_epi16(x2, x3); 1386 x12 = _mm_unpackhi_epi16(x8, x9); 1387 x13 = _mm_unpackhi_epi16(x10, x11); 1388 1389 x6 = _mm_unpacklo_epi32(x4, x5); 1390 x7 = _mm_unpackhi_epi32(x4, x5); 1391 x14 = _mm_unpacklo_epi32(x12, x13); 1392 x15 = _mm_unpackhi_epi32(x12, x13); 1393 1394 // Store second 4-line result 1395 _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14)); 1396 _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14)); 1397 _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15)); 1398 _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15)); 1399} 1400 1401static INLINE void transpose(unsigned char *src[], int in_p, 1402 unsigned char *dst[], int out_p, 1403 int num_8x8_to_transpose) { 1404 int idx8x8 = 0; 1405 __m128i x0, x1, x2, x3, x4, x5, x6, x7; 1406 do { 1407 unsigned char *in = src[idx8x8]; 1408 unsigned char *out = dst[idx8x8]; 1409 1410 x0 = _mm_loadl_epi64((__m128i *)(in + 0*in_p)); // 00 01 02 03 04 05 06 07 1411 x1 = _mm_loadl_epi64((__m128i *)(in + 1*in_p)); // 10 11 12 13 14 15 16 17 1412 // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17 1413 x0 = _mm_unpacklo_epi8(x0, x1); 1414 1415 x2 = _mm_loadl_epi64((__m128i *)(in + 2*in_p)); // 20 21 22 23 24 25 26 27 1416 x3 = _mm_loadl_epi64((__m128i *)(in + 3*in_p)); // 30 31 32 33 34 35 36 37 1417 // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37 1418 x1 = _mm_unpacklo_epi8(x2, x3); 1419 1420 x4 = _mm_loadl_epi64((__m128i *)(in + 4*in_p)); // 40 41 42 43 44 45 46 47 1421 x5 = _mm_loadl_epi64((__m128i *)(in + 5*in_p)); // 50 51 52 53 54 55 56 57 1422 // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57 1423 x2 = _mm_unpacklo_epi8(x4, x5); 1424 1425 x6 = _mm_loadl_epi64((__m128i *)(in + 6*in_p)); // 60 61 62 63 64 65 66 67 1426 x7 = _mm_loadl_epi64((__m128i *)(in + 7*in_p)); // 70 71 72 73 74 75 76 77 1427 // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77 1428 x3 = _mm_unpacklo_epi8(x6, x7); 1429 1430 // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 1431 x4 = _mm_unpacklo_epi16(x0, x1); 1432 // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73 1433 x5 = _mm_unpacklo_epi16(x2, x3); 1434 // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71 1435 x6 = _mm_unpacklo_epi32(x4, x5); 1436 _mm_storel_pd((double *)(out + 0*out_p), 1437 _mm_castsi128_pd(x6)); // 00 10 20 30 40 50 60 70 1438 _mm_storeh_pd((double *)(out + 1*out_p), 1439 _mm_castsi128_pd(x6)); // 01 11 21 31 41 51 61 71 1440 // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73 1441 x7 = _mm_unpackhi_epi32(x4, x5); 1442 _mm_storel_pd((double *)(out + 2*out_p), 1443 _mm_castsi128_pd(x7)); // 02 12 22 32 42 52 62 72 1444 _mm_storeh_pd((double *)(out + 3*out_p), 1445 _mm_castsi128_pd(x7)); // 03 13 23 33 43 53 63 73 1446 1447 // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 1448 x4 = _mm_unpackhi_epi16(x0, x1); 1449 // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77 1450 x5 = _mm_unpackhi_epi16(x2, x3); 1451 // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75 1452 x6 = _mm_unpacklo_epi32(x4, x5); 1453 _mm_storel_pd((double *)(out + 4*out_p), 1454 _mm_castsi128_pd(x6)); // 04 14 24 34 44 54 64 74 1455 _mm_storeh_pd((double *)(out + 5*out_p), 1456 _mm_castsi128_pd(x6)); // 05 15 25 35 45 55 65 75 1457 // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77 1458 x7 = _mm_unpackhi_epi32(x4, x5); 1459 1460 _mm_storel_pd((double *)(out + 6*out_p), 1461 _mm_castsi128_pd(x7)); // 06 16 26 36 46 56 66 76 1462 _mm_storeh_pd((double *)(out + 7*out_p), 1463 _mm_castsi128_pd(x7)); // 07 17 27 37 47 57 67 77 1464 } while (++idx8x8 < num_8x8_to_transpose); 1465} 1466 1467void vpx_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0, 1468 const uint8_t *limit0, 1469 const uint8_t *thresh0, 1470 const uint8_t *blimit1, 1471 const uint8_t *limit1, 1472 const uint8_t *thresh1) { 1473 DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]); 1474 unsigned char *src[2]; 1475 unsigned char *dst[2]; 1476 1477 // Transpose 8x16 1478 transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); 1479 1480 // Loop filtering 1481 vpx_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0, 1482 blimit1, limit1, thresh1); 1483 src[0] = t_dst; 1484 src[1] = t_dst + 8; 1485 dst[0] = s - 4; 1486 dst[1] = s - 4 + p * 8; 1487 1488 // Transpose back 1489 transpose(src, 16, dst, p, 2); 1490} 1491 1492void vpx_lpf_vertical_8_sse2(unsigned char *s, int p, 1493 const unsigned char *blimit, 1494 const unsigned char *limit, 1495 const unsigned char *thresh, int count) { 1496 DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 8]); 1497 unsigned char *src[1]; 1498 unsigned char *dst[1]; 1499 (void)count; 1500 1501 // Transpose 8x8 1502 src[0] = s - 4; 1503 dst[0] = t_dst; 1504 1505 transpose(src, p, dst, 8, 1); 1506 1507 // Loop filtering 1508 vpx_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1); 1509 1510 src[0] = t_dst; 1511 dst[0] = s - 4; 1512 1513 // Transpose back 1514 transpose(src, 8, dst, p, 1); 1515} 1516 1517void vpx_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0, 1518 const uint8_t *limit0, 1519 const uint8_t *thresh0, 1520 const uint8_t *blimit1, 1521 const uint8_t *limit1, 1522 const uint8_t *thresh1) { 1523 DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]); 1524 unsigned char *src[2]; 1525 unsigned char *dst[2]; 1526 1527 // Transpose 8x16 1528 transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16); 1529 1530 // Loop filtering 1531 vpx_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0, 1532 blimit1, limit1, thresh1); 1533 src[0] = t_dst; 1534 src[1] = t_dst + 8; 1535 1536 dst[0] = s - 4; 1537 dst[1] = s - 4 + p * 8; 1538 1539 // Transpose back 1540 transpose(src, 16, dst, p, 2); 1541} 1542 1543void vpx_lpf_vertical_16_sse2(unsigned char *s, int p, 1544 const unsigned char *blimit, 1545 const unsigned char *limit, 1546 const unsigned char *thresh) { 1547 DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 16]); 1548 unsigned char *src[2]; 1549 unsigned char *dst[2]; 1550 1551 src[0] = s - 8; 1552 src[1] = s; 1553 dst[0] = t_dst; 1554 dst[1] = t_dst + 8 * 8; 1555 1556 // Transpose 16x8 1557 transpose(src, p, dst, 8, 2); 1558 1559 // Loop filtering 1560 mb_lpf_horizontal_edge_w_sse2_8(t_dst + 8 * 8, 8, blimit, limit, thresh); 1561 1562 src[0] = t_dst; 1563 src[1] = t_dst + 8 * 8; 1564 dst[0] = s - 8; 1565 dst[1] = s; 1566 1567 // Transpose back 1568 transpose(src, 8, dst, p, 2); 1569} 1570 1571void vpx_lpf_vertical_16_dual_sse2(unsigned char *s, int p, 1572 const uint8_t *blimit, const uint8_t *limit, 1573 const uint8_t *thresh) { 1574 DECLARE_ALIGNED(16, unsigned char, t_dst[256]); 1575 1576 // Transpose 16x16 1577 transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16); 1578 transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16); 1579 1580 // Loop filtering 1581 mb_lpf_horizontal_edge_w_sse2_16(t_dst + 8 * 16, 16, blimit, limit, 1582 thresh); 1583 1584 // Transpose back 1585 transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p); 1586 transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p); 1587} 1588