jnt_convolve_sse2.c (15331B)
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 <emmintrin.h> 13 14 #include "config/av1_rtcd.h" 15 16 #include "aom_dsp/aom_filter.h" 17 #include "aom_dsp/x86/convolve_sse2.h" 18 #include "aom_dsp/x86/synonyms.h" 19 20 void av1_dist_wtd_convolve_x_sse2(const uint8_t *src, int src_stride, 21 uint8_t *dst0, int dst_stride0, int w, int h, 22 const InterpFilterParams *filter_params_x, 23 const int subpel_x_qn, 24 ConvolveParams *conv_params) { 25 const int bd = 8; 26 CONV_BUF_TYPE *dst = conv_params->dst; 27 const int dst_stride = conv_params->dst_stride; 28 const int fo_horiz = filter_params_x->taps / 2 - 1; 29 const uint8_t *src_ptr = src - fo_horiz; 30 const int bits = FILTER_BITS - conv_params->round_1; 31 const __m128i left_shift = _mm_cvtsi32_si128(bits); 32 const __m128i round_const = _mm_set1_epi32((1 << conv_params->round_0) >> 1); 33 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0); 34 const int w0 = conv_params->fwd_offset; 35 const int w1 = conv_params->bck_offset; 36 const __m128i wt0 = _mm_set1_epi16(w0); 37 const __m128i wt1 = _mm_set1_epi16(w1); 38 const __m128i wt = _mm_unpacklo_epi16(wt0, wt1); 39 const int do_average = conv_params->do_average; 40 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; 41 const int offset_0 = 42 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 43 const int offset = (1 << offset_0) + (1 << (offset_0 - 1)); 44 const __m128i offset_const = _mm_set1_epi16(offset); 45 const int rounding_shift = 46 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 47 const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1); 48 __m128i coeffs[4]; 49 50 prepare_coeffs(filter_params_x, subpel_x_qn, coeffs); 51 52 if (w == 4) { 53 do { 54 const __m128i data = _mm_loadu_si128((__m128i *)src_ptr); 55 __m128i s[4]; 56 57 s[0] = _mm_unpacklo_epi8(data, _mm_srli_si128(data, 1)); 58 s[1] = 59 _mm_unpacklo_epi8(_mm_srli_si128(data, 2), _mm_srli_si128(data, 3)); 60 s[2] = 61 _mm_unpacklo_epi8(_mm_srli_si128(data, 4), _mm_srli_si128(data, 5)); 62 s[3] = 63 _mm_unpacklo_epi8(_mm_srli_si128(data, 6), _mm_srli_si128(data, 7)); 64 const __m128i res_lo = convolve_lo_x(s, coeffs); 65 const __m128i res_lo_round = 66 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); 67 const __m128i res_lo_shift = _mm_sll_epi32(res_lo_round, left_shift); 68 69 const __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_lo_shift); 70 const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const); 71 72 // Accumulate values into the destination buffer 73 if (do_average) { 74 const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst); 75 76 const __m128i comp_avg_res = 77 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); 78 79 const __m128i round_result = convolve_rounding( 80 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 81 82 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); 83 *(int *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); 84 } else { 85 _mm_store_si128((__m128i *)(&dst[0]), res_unsigned); 86 } 87 src_ptr += src_stride; 88 dst += dst_stride; 89 dst0 += dst_stride0; 90 } while (--h); 91 } else { 92 assert(!(w % 8)); 93 int i = 0; 94 do { 95 int j = 0; 96 do { 97 const __m128i data = 98 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]); 99 __m128i s[4]; 100 101 // Filter even-index pixels 102 s[0] = data; 103 s[1] = _mm_srli_si128(data, 2); 104 s[2] = _mm_srli_si128(data, 4); 105 s[3] = _mm_srli_si128(data, 6); 106 const __m128i res_even = convolve_lo_x(s, coeffs); 107 108 // Filter odd-index pixels 109 s[0] = _mm_srli_si128(data, 1); 110 s[1] = _mm_srli_si128(data, 3); 111 s[2] = _mm_srli_si128(data, 5); 112 s[3] = _mm_srli_si128(data, 7); 113 const __m128i res_odd = convolve_lo_x(s, coeffs); 114 115 // Rearrange pixels back into the order 0 ... 7 116 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); 117 const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd); 118 const __m128i res_lo_round = 119 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); 120 const __m128i res_hi_round = 121 _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift); 122 const __m128i res_lo_shift = _mm_sll_epi32(res_lo_round, left_shift); 123 const __m128i res_hi_shift = _mm_sll_epi32(res_hi_round, left_shift); 124 125 const __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift); 126 const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const); 127 128 // Accumulate values into the destination buffer 129 if (do_average) { 130 const __m128i data_ref_0 = 131 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])); 132 133 const __m128i comp_avg_res = 134 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); 135 136 const __m128i round_result = convolve_rounding( 137 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 138 139 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); 140 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); 141 } else { 142 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned); 143 } 144 j += 8; 145 } while (j < w); 146 } while (++i < h); 147 } 148 } 149 150 void av1_dist_wtd_convolve_y_sse2(const uint8_t *src, int src_stride, 151 uint8_t *dst0, int dst_stride0, int w, int h, 152 const InterpFilterParams *filter_params_y, 153 const int subpel_y_qn, 154 ConvolveParams *conv_params) { 155 const int bd = 8; 156 CONV_BUF_TYPE *dst = conv_params->dst; 157 const int dst_stride = conv_params->dst_stride; 158 const int fo_vert = filter_params_y->taps / 2 - 1; 159 const uint8_t *src_ptr = src - fo_vert * src_stride; 160 const int bits = FILTER_BITS - conv_params->round_0; 161 const __m128i left_shift = _mm_cvtsi32_si128(bits); 162 const __m128i wt0 = _mm_set1_epi16(conv_params->fwd_offset); 163 const __m128i wt1 = _mm_set1_epi16(conv_params->bck_offset); 164 const __m128i wt = _mm_unpacklo_epi16(wt0, wt1); 165 const int do_average = conv_params->do_average; 166 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; 167 const int offset_0 = 168 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 169 const int offset = (1 << offset_0) + (1 << (offset_0 - 1)); 170 const __m128i offset_const = _mm_set1_epi16(offset); 171 const int rounding_shift = 172 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 173 const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1); 174 const __m128i round_const = _mm_set1_epi32((1 << conv_params->round_1) >> 1); 175 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1); 176 __m128i coeffs[4]; 177 178 prepare_coeffs(filter_params_y, subpel_y_qn, coeffs); 179 180 if (w == 4) { 181 __m128i s[8], src6, res, res_shift; 182 s[0] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 0 * src_stride), 183 xx_loadl_32(src_ptr + 1 * src_stride)); 184 s[1] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 1 * src_stride), 185 xx_loadl_32(src_ptr + 2 * src_stride)); 186 s[2] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 2 * src_stride), 187 xx_loadl_32(src_ptr + 3 * src_stride)); 188 s[3] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 3 * src_stride), 189 xx_loadl_32(src_ptr + 4 * src_stride)); 190 s[4] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 4 * src_stride), 191 xx_loadl_32(src_ptr + 5 * src_stride)); 192 src6 = xx_loadl_32(src_ptr + 6 * src_stride); 193 s[5] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 5 * src_stride), src6); 194 195 do { 196 s[6] = _mm_unpacklo_epi8(src6, xx_loadl_32(src_ptr + 7 * src_stride)); 197 src6 = xx_loadl_32(src_ptr + 8 * src_stride); 198 s[7] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 7 * src_stride), src6); 199 200 res = convolve_lo_y(s + 0, coeffs); 201 res_shift = _mm_sll_epi32(res, left_shift); 202 res_shift = 203 _mm_sra_epi32(_mm_add_epi32(res_shift, round_const), round_shift); 204 205 __m128i res_16b = _mm_packs_epi32(res_shift, res_shift); 206 __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const); 207 208 // Accumulate values into the destination buffer 209 if (do_average) { 210 const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst); 211 212 const __m128i comp_avg_res = 213 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); 214 215 const __m128i round_result = convolve_rounding( 216 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 217 218 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); 219 *(int *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); 220 221 } else { 222 _mm_store_si128((__m128i *)dst, res_unsigned); 223 } 224 225 src_ptr += src_stride; 226 dst += dst_stride; 227 dst0 += dst_stride0; 228 229 res = convolve_lo_y(s + 1, coeffs); 230 res_shift = _mm_sll_epi32(res, left_shift); 231 res_shift = 232 _mm_sra_epi32(_mm_add_epi32(res_shift, round_const), round_shift); 233 234 res_16b = _mm_packs_epi32(res_shift, res_shift); 235 res_unsigned = _mm_add_epi16(res_16b, offset_const); 236 237 // Accumulate values into the destination buffer 238 if (do_average) { 239 const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst); 240 241 const __m128i comp_avg_res = 242 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); 243 244 const __m128i round_result = convolve_rounding( 245 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 246 247 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); 248 *(int *)(&dst0[0]) = _mm_cvtsi128_si32(res_8); 249 250 } else { 251 _mm_store_si128((__m128i *)dst, res_unsigned); 252 } 253 254 src_ptr += src_stride; 255 dst += dst_stride; 256 dst0 += dst_stride0; 257 258 s[0] = s[2]; 259 s[1] = s[3]; 260 s[2] = s[4]; 261 s[3] = s[5]; 262 s[4] = s[6]; 263 s[5] = s[7]; 264 h -= 2; 265 } while (h); 266 } else { 267 assert(!(w % 8)); 268 int j = 0; 269 do { 270 __m128i s[8], src6, res_lo, res_hi, res_lo_shift, res_hi_shift; 271 const uint8_t *data = &src_ptr[j]; 272 273 src6 = _mm_loadl_epi64((__m128i *)(data + 6 * src_stride)); 274 s[0] = _mm_unpacklo_epi8( 275 _mm_loadl_epi64((__m128i *)(data + 0 * src_stride)), 276 _mm_loadl_epi64((__m128i *)(data + 1 * src_stride))); 277 s[1] = _mm_unpacklo_epi8( 278 _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)), 279 _mm_loadl_epi64((__m128i *)(data + 2 * src_stride))); 280 s[2] = _mm_unpacklo_epi8( 281 _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)), 282 _mm_loadl_epi64((__m128i *)(data + 3 * src_stride))); 283 s[3] = _mm_unpacklo_epi8( 284 _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)), 285 _mm_loadl_epi64((__m128i *)(data + 4 * src_stride))); 286 s[4] = _mm_unpacklo_epi8( 287 _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)), 288 _mm_loadl_epi64((__m128i *)(data + 5 * src_stride))); 289 s[5] = _mm_unpacklo_epi8( 290 _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)), src6); 291 292 int i = 0; 293 do { 294 data = &src_ptr[i * src_stride + j]; 295 s[6] = _mm_unpacklo_epi8( 296 src6, _mm_loadl_epi64((__m128i *)(data + 7 * src_stride))); 297 src6 = _mm_loadl_epi64((__m128i *)(data + 8 * src_stride)); 298 s[7] = _mm_unpacklo_epi8( 299 _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)), src6); 300 301 res_lo = convolve_lo_y(s, coeffs); // Filter low index pixels 302 res_hi = convolve_hi_y(s, coeffs); // Filter high index pixels 303 res_lo_shift = _mm_sll_epi32(res_lo, left_shift); 304 res_hi_shift = _mm_sll_epi32(res_hi, left_shift); 305 res_lo_shift = _mm_sra_epi32(_mm_add_epi32(res_lo_shift, round_const), 306 round_shift); 307 res_hi_shift = _mm_sra_epi32(_mm_add_epi32(res_hi_shift, round_const), 308 round_shift); 309 310 __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift); 311 __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const); 312 313 // Accumulate values into the destination buffer 314 if (do_average) { 315 const __m128i data_ref_0 = 316 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])); 317 318 const __m128i comp_avg_res = 319 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); 320 321 const __m128i round_result = convolve_rounding( 322 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 323 324 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); 325 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); 326 } else { 327 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned); 328 } 329 i++; 330 331 res_lo = convolve_lo_y(s + 1, coeffs); // Filter low index pixels 332 res_hi = convolve_hi_y(s + 1, coeffs); // Filter high index pixels 333 res_lo_shift = _mm_sll_epi32(res_lo, left_shift); 334 res_hi_shift = _mm_sll_epi32(res_hi, left_shift); 335 res_lo_shift = _mm_sra_epi32(_mm_add_epi32(res_lo_shift, round_const), 336 round_shift); 337 res_hi_shift = _mm_sra_epi32(_mm_add_epi32(res_hi_shift, round_const), 338 round_shift); 339 res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift); 340 res_unsigned = _mm_add_epi16(res_16b, offset_const); 341 342 // Accumulate values into the destination buffer 343 if (do_average) { 344 __m128i data_ref_0 = 345 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])); 346 347 const __m128i comp_avg_res = 348 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg); 349 350 const __m128i round_result = convolve_rounding( 351 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 352 353 const __m128i res_8 = _mm_packus_epi16(round_result, round_result); 354 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8); 355 } else { 356 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned); 357 } 358 i++; 359 360 s[0] = s[2]; 361 s[1] = s[3]; 362 s[2] = s[4]; 363 s[3] = s[5]; 364 s[4] = s[6]; 365 s[5] = s[7]; 366 } while (i < h); 367 j += 8; 368 } while (j < w); 369 } 370 }