highbd_jnt_convolve_sse4.c (16618B)
1 /* 2 * Copyright (c) 2018, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include <smmintrin.h> 13 #include <assert.h> 14 15 #include "config/av1_rtcd.h" 16 17 #include "aom_dsp/x86/convolve_sse2.h" 18 #include "aom_dsp/x86/convolve_sse4_1.h" 19 20 void av1_highbd_dist_wtd_convolve_y_sse4_1( 21 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, 22 int h, const InterpFilterParams *filter_params_y, const int subpel_y_qn, 23 ConvolveParams *conv_params, int bd) { 24 CONV_BUF_TYPE *dst = conv_params->dst; 25 int dst_stride = conv_params->dst_stride; 26 const int fo_vert = filter_params_y->taps / 2 - 1; 27 const uint16_t *const src_ptr = src - fo_vert * src_stride; 28 const int bits = FILTER_BITS - conv_params->round_0; 29 30 assert(bits >= 0); 31 int i, j; 32 const int do_average = conv_params->do_average; 33 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; 34 35 const int w0 = conv_params->fwd_offset; 36 const int w1 = conv_params->bck_offset; 37 const __m128i wt0 = _mm_set1_epi32(w0); 38 const __m128i wt1 = _mm_set1_epi32(w1); 39 const __m128i round_const_y = 40 _mm_set1_epi32(((1 << conv_params->round_1) >> 1)); 41 const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1); 42 const __m128i round_shift_bits = _mm_cvtsi32_si128(bits); 43 44 const int offset_0 = 45 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 46 const int offset = (1 << offset_0) + (1 << (offset_0 - 1)); 47 const __m128i offset_const = _mm_set1_epi32(offset); 48 const int rounding_shift = 49 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 50 const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1); 51 const __m128i clip_pixel_to_bd = 52 _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); 53 const __m128i zero = _mm_setzero_si128(); 54 __m128i s[16], coeffs_y[4]; 55 56 prepare_coeffs(filter_params_y, subpel_y_qn, coeffs_y); 57 58 for (j = 0; j < w; j += 8) { 59 const uint16_t *data = &src_ptr[j]; 60 /* Vertical filter */ 61 { 62 __m128i s0 = _mm_loadu_si128((__m128i *)(data + 0 * src_stride)); 63 __m128i s1 = _mm_loadu_si128((__m128i *)(data + 1 * src_stride)); 64 __m128i s2 = _mm_loadu_si128((__m128i *)(data + 2 * src_stride)); 65 __m128i s3 = _mm_loadu_si128((__m128i *)(data + 3 * src_stride)); 66 __m128i s4 = _mm_loadu_si128((__m128i *)(data + 4 * src_stride)); 67 __m128i s5 = _mm_loadu_si128((__m128i *)(data + 5 * src_stride)); 68 __m128i s6 = _mm_loadu_si128((__m128i *)(data + 6 * src_stride)); 69 70 s[0] = _mm_unpacklo_epi16(s0, s1); 71 s[1] = _mm_unpacklo_epi16(s2, s3); 72 s[2] = _mm_unpacklo_epi16(s4, s5); 73 74 s[4] = _mm_unpackhi_epi16(s0, s1); 75 s[5] = _mm_unpackhi_epi16(s2, s3); 76 s[6] = _mm_unpackhi_epi16(s4, s5); 77 78 s[0 + 8] = _mm_unpacklo_epi16(s1, s2); 79 s[1 + 8] = _mm_unpacklo_epi16(s3, s4); 80 s[2 + 8] = _mm_unpacklo_epi16(s5, s6); 81 82 s[4 + 8] = _mm_unpackhi_epi16(s1, s2); 83 s[5 + 8] = _mm_unpackhi_epi16(s3, s4); 84 s[6 + 8] = _mm_unpackhi_epi16(s5, s6); 85 86 for (i = 0; i < h; i += 2) { 87 data = &src_ptr[i * src_stride + j]; 88 89 __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * src_stride)); 90 __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * src_stride)); 91 92 s[3] = _mm_unpacklo_epi16(s6, s7); 93 s[7] = _mm_unpackhi_epi16(s6, s7); 94 95 s[3 + 8] = _mm_unpacklo_epi16(s7, s8); 96 s[7 + 8] = _mm_unpackhi_epi16(s7, s8); 97 98 const __m128i res_a0 = convolve(s, coeffs_y); 99 __m128i res_a_round0 = _mm_sll_epi32(res_a0, round_shift_bits); 100 res_a_round0 = _mm_sra_epi32(_mm_add_epi32(res_a_round0, round_const_y), 101 round_shift_y); 102 103 const __m128i res_a1 = convolve(s + 8, coeffs_y); 104 __m128i res_a_round1 = _mm_sll_epi32(res_a1, round_shift_bits); 105 res_a_round1 = _mm_sra_epi32(_mm_add_epi32(res_a_round1, round_const_y), 106 round_shift_y); 107 108 __m128i res_unsigned_lo_0 = _mm_add_epi32(res_a_round0, offset_const); 109 __m128i res_unsigned_lo_1 = _mm_add_epi32(res_a_round1, offset_const); 110 111 if (w - j < 8) { 112 if (do_average) { 113 const __m128i data_0 = 114 _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])); 115 const __m128i data_1 = _mm_loadl_epi64( 116 (__m128i *)(&dst[i * dst_stride + j + dst_stride])); 117 118 const __m128i data_ref_0 = _mm_unpacklo_epi16(data_0, zero); 119 const __m128i data_ref_1 = _mm_unpacklo_epi16(data_1, zero); 120 121 const __m128i comp_avg_res_0 = 122 highbd_comp_avg_sse4_1(&data_ref_0, &res_unsigned_lo_0, &wt0, 123 &wt1, use_dist_wtd_comp_avg); 124 const __m128i comp_avg_res_1 = 125 highbd_comp_avg_sse4_1(&data_ref_1, &res_unsigned_lo_1, &wt0, 126 &wt1, use_dist_wtd_comp_avg); 127 128 const __m128i round_result_0 = 129 highbd_convolve_rounding_sse2(&comp_avg_res_0, &offset_const, 130 &rounding_const, rounding_shift); 131 const __m128i round_result_1 = 132 highbd_convolve_rounding_sse2(&comp_avg_res_1, &offset_const, 133 &rounding_const, rounding_shift); 134 135 const __m128i res_16b_0 = 136 _mm_packus_epi32(round_result_0, round_result_0); 137 const __m128i res_clip_0 = 138 _mm_min_epi16(res_16b_0, clip_pixel_to_bd); 139 const __m128i res_16b_1 = 140 _mm_packus_epi32(round_result_1, round_result_1); 141 const __m128i res_clip_1 = 142 _mm_min_epi16(res_16b_1, clip_pixel_to_bd); 143 144 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), 145 res_clip_0); 146 _mm_storel_epi64( 147 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), 148 res_clip_1); 149 150 } else { 151 __m128i res_16b_0 = 152 _mm_packus_epi32(res_unsigned_lo_0, res_unsigned_lo_0); 153 154 __m128i res_16b_1 = 155 _mm_packus_epi32(res_unsigned_lo_1, res_unsigned_lo_1); 156 157 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_16b_0); 158 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride], 159 res_16b_1); 160 } 161 } else { 162 const __m128i res_b0 = convolve(s + 4, coeffs_y); 163 __m128i res_b_round0 = _mm_sll_epi32(res_b0, round_shift_bits); 164 res_b_round0 = _mm_sra_epi32( 165 _mm_add_epi32(res_b_round0, round_const_y), round_shift_y); 166 167 const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y); 168 __m128i res_b_round1 = _mm_sll_epi32(res_b1, round_shift_bits); 169 res_b_round1 = _mm_sra_epi32( 170 _mm_add_epi32(res_b_round1, round_const_y), round_shift_y); 171 172 __m128i res_unsigned_hi_0 = _mm_add_epi32(res_b_round0, offset_const); 173 __m128i res_unsigned_hi_1 = _mm_add_epi32(res_b_round1, offset_const); 174 175 if (do_average) { 176 const __m128i data_0 = 177 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])); 178 const __m128i data_1 = _mm_loadu_si128( 179 (__m128i *)(&dst[i * dst_stride + j + dst_stride])); 180 const __m128i data_ref_0_lo_0 = _mm_unpacklo_epi16(data_0, zero); 181 const __m128i data_ref_0_lo_1 = _mm_unpacklo_epi16(data_1, zero); 182 183 const __m128i data_ref_0_hi_0 = _mm_unpackhi_epi16(data_0, zero); 184 const __m128i data_ref_0_hi_1 = _mm_unpackhi_epi16(data_1, zero); 185 186 const __m128i comp_avg_res_lo_0 = 187 highbd_comp_avg_sse4_1(&data_ref_0_lo_0, &res_unsigned_lo_0, 188 &wt0, &wt1, use_dist_wtd_comp_avg); 189 const __m128i comp_avg_res_lo_1 = 190 highbd_comp_avg_sse4_1(&data_ref_0_lo_1, &res_unsigned_lo_1, 191 &wt0, &wt1, use_dist_wtd_comp_avg); 192 const __m128i comp_avg_res_hi_0 = 193 highbd_comp_avg_sse4_1(&data_ref_0_hi_0, &res_unsigned_hi_0, 194 &wt0, &wt1, use_dist_wtd_comp_avg); 195 const __m128i comp_avg_res_hi_1 = 196 highbd_comp_avg_sse4_1(&data_ref_0_hi_1, &res_unsigned_hi_1, 197 &wt0, &wt1, use_dist_wtd_comp_avg); 198 199 const __m128i round_result_lo_0 = 200 highbd_convolve_rounding_sse2(&comp_avg_res_lo_0, &offset_const, 201 &rounding_const, rounding_shift); 202 const __m128i round_result_lo_1 = 203 highbd_convolve_rounding_sse2(&comp_avg_res_lo_1, &offset_const, 204 &rounding_const, rounding_shift); 205 const __m128i round_result_hi_0 = 206 highbd_convolve_rounding_sse2(&comp_avg_res_hi_0, &offset_const, 207 &rounding_const, rounding_shift); 208 const __m128i round_result_hi_1 = 209 highbd_convolve_rounding_sse2(&comp_avg_res_hi_1, &offset_const, 210 &rounding_const, rounding_shift); 211 212 const __m128i res_16b_0 = 213 _mm_packus_epi32(round_result_lo_0, round_result_hi_0); 214 const __m128i res_clip_0 = 215 _mm_min_epi16(res_16b_0, clip_pixel_to_bd); 216 217 const __m128i res_16b_1 = 218 _mm_packus_epi32(round_result_lo_1, round_result_hi_1); 219 const __m128i res_clip_1 = 220 _mm_min_epi16(res_16b_1, clip_pixel_to_bd); 221 222 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), 223 res_clip_0); 224 _mm_store_si128( 225 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), 226 res_clip_1); 227 } else { 228 __m128i res_16bit0 = 229 _mm_packus_epi32(res_unsigned_lo_0, res_unsigned_hi_0); 230 __m128i res_16bit1 = 231 _mm_packus_epi32(res_unsigned_lo_1, res_unsigned_hi_1); 232 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_16bit0); 233 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]), 234 res_16bit1); 235 } 236 } 237 s[0] = s[1]; 238 s[1] = s[2]; 239 s[2] = s[3]; 240 241 s[4] = s[5]; 242 s[5] = s[6]; 243 s[6] = s[7]; 244 245 s[0 + 8] = s[1 + 8]; 246 s[1 + 8] = s[2 + 8]; 247 s[2 + 8] = s[3 + 8]; 248 249 s[4 + 8] = s[5 + 8]; 250 s[5 + 8] = s[6 + 8]; 251 s[6 + 8] = s[7 + 8]; 252 253 s6 = s8; 254 } 255 } 256 } 257 } 258 259 void av1_highbd_dist_wtd_convolve_x_sse4_1( 260 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, 261 int h, const InterpFilterParams *filter_params_x, const int subpel_x_qn, 262 ConvolveParams *conv_params, int bd) { 263 CONV_BUF_TYPE *dst = conv_params->dst; 264 int dst_stride = conv_params->dst_stride; 265 const int fo_horiz = filter_params_x->taps / 2 - 1; 266 const uint16_t *const src_ptr = src - fo_horiz; 267 const int bits = FILTER_BITS - conv_params->round_1; 268 269 int i, j; 270 __m128i s[4], coeffs_x[4]; 271 272 const int do_average = conv_params->do_average; 273 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; 274 const int w0 = conv_params->fwd_offset; 275 const int w1 = conv_params->bck_offset; 276 const __m128i wt0 = _mm_set1_epi32(w0); 277 const __m128i wt1 = _mm_set1_epi32(w1); 278 const __m128i zero = _mm_setzero_si128(); 279 280 const __m128i round_const_x = 281 _mm_set1_epi32(((1 << conv_params->round_0) >> 1)); 282 const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0); 283 const __m128i round_shift_bits = _mm_cvtsi32_si128(bits); 284 285 const int offset_0 = 286 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 287 const int offset = (1 << offset_0) + (1 << (offset_0 - 1)); 288 const __m128i offset_const = _mm_set1_epi32(offset); 289 const int rounding_shift = 290 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 291 const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1); 292 const __m128i clip_pixel_to_bd = 293 _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); 294 295 assert(bits >= 0); 296 prepare_coeffs(filter_params_x, subpel_x_qn, coeffs_x); 297 298 for (j = 0; j < w; j += 8) { 299 /* Horizontal filter */ 300 for (i = 0; i < h; i += 1) { 301 const __m128i row00 = 302 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]); 303 const __m128i row01 = 304 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]); 305 306 // even pixels 307 s[0] = _mm_alignr_epi8(row01, row00, 0); 308 s[1] = _mm_alignr_epi8(row01, row00, 4); 309 s[2] = _mm_alignr_epi8(row01, row00, 8); 310 s[3] = _mm_alignr_epi8(row01, row00, 12); 311 312 __m128i res_even = convolve(s, coeffs_x); 313 res_even = 314 _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x), round_shift_x); 315 316 // odd pixels 317 s[0] = _mm_alignr_epi8(row01, row00, 2); 318 s[1] = _mm_alignr_epi8(row01, row00, 6); 319 s[2] = _mm_alignr_epi8(row01, row00, 10); 320 s[3] = _mm_alignr_epi8(row01, row00, 14); 321 322 __m128i res_odd = convolve(s, coeffs_x); 323 res_odd = 324 _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x), round_shift_x); 325 326 res_even = _mm_sll_epi32(res_even, round_shift_bits); 327 res_odd = _mm_sll_epi32(res_odd, round_shift_bits); 328 329 __m128i res1 = _mm_unpacklo_epi32(res_even, res_odd); 330 __m128i res_unsigned_lo = _mm_add_epi32(res1, offset_const); 331 if (w - j < 8) { 332 if (do_average) { 333 const __m128i data_0 = 334 _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])); 335 const __m128i data_ref_0 = _mm_unpacklo_epi16(data_0, zero); 336 337 const __m128i comp_avg_res = highbd_comp_avg_sse4_1( 338 &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_dist_wtd_comp_avg); 339 const __m128i round_result = highbd_convolve_rounding_sse2( 340 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 341 342 const __m128i res_16b = _mm_packus_epi32(round_result, round_result); 343 const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd); 344 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); 345 } else { 346 __m128i res_16b = _mm_packus_epi32(res_unsigned_lo, res_unsigned_lo); 347 _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_16b); 348 } 349 } else { 350 __m128i res2 = _mm_unpackhi_epi32(res_even, res_odd); 351 __m128i res_unsigned_hi = _mm_add_epi32(res2, offset_const); 352 if (do_average) { 353 const __m128i data_0 = 354 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])); 355 const __m128i data_ref_0_lo = _mm_unpacklo_epi16(data_0, zero); 356 const __m128i data_ref_0_hi = _mm_unpackhi_epi16(data_0, zero); 357 358 const __m128i comp_avg_res_lo = 359 highbd_comp_avg_sse4_1(&data_ref_0_lo, &res_unsigned_lo, &wt0, 360 &wt1, use_dist_wtd_comp_avg); 361 const __m128i comp_avg_res_hi = 362 highbd_comp_avg_sse4_1(&data_ref_0_hi, &res_unsigned_hi, &wt0, 363 &wt1, use_dist_wtd_comp_avg); 364 365 const __m128i round_result_lo = highbd_convolve_rounding_sse2( 366 &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift); 367 const __m128i round_result_hi = highbd_convolve_rounding_sse2( 368 &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift); 369 370 const __m128i res_16b = 371 _mm_packus_epi32(round_result_lo, round_result_hi); 372 const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd); 373 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); 374 } else { 375 __m128i res_16b = _mm_packus_epi32(res_unsigned_lo, res_unsigned_hi); 376 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_16b); 377 } 378 } 379 } 380 } 381 }