highbd_convolve_2d_sse4.c (18237B)
1 /* 2 * Copyright (c) 2016, 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 <tmmintrin.h> 13 #include <smmintrin.h> 14 #include <assert.h> 15 16 #include "config/av1_rtcd.h" 17 18 #include "aom_dsp/aom_dsp_common.h" 19 #include "aom_dsp/aom_filter.h" 20 #include "aom_dsp/x86/convolve_sse2.h" 21 #include "aom_dsp/x86/convolve_sse4_1.h" 22 #include "av1/common/convolve.h" 23 24 void av1_highbd_dist_wtd_convolve_2d_copy_sse4_1(const uint16_t *src, 25 int src_stride, uint16_t *dst0, 26 int dst_stride0, int w, int h, 27 ConvolveParams *conv_params, 28 int bd) { 29 CONV_BUF_TYPE *dst = conv_params->dst; 30 int dst_stride = conv_params->dst_stride; 31 32 const int bits = 33 FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0; 34 const __m128i left_shift = _mm_cvtsi32_si128(bits); 35 const int do_average = conv_params->do_average; 36 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; 37 const int w0 = conv_params->fwd_offset; 38 const int w1 = conv_params->bck_offset; 39 const __m128i wt0 = _mm_set1_epi32(w0); 40 const __m128i wt1 = _mm_set1_epi32(w1); 41 const __m128i zero = _mm_setzero_si128(); 42 int i, j; 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 __m128i offset_const_16b = _mm_set1_epi16(offset); 49 const int rounding_shift = 50 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 51 const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1); 52 const __m128i clip_pixel_to_bd = 53 _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); 54 55 assert(bits <= 4); 56 57 if (!(w % 8)) { 58 for (i = 0; i < h; i += 1) { 59 for (j = 0; j < w; j += 8) { 60 const __m128i src_16bit = 61 _mm_loadu_si128((__m128i *)(&src[i * src_stride + j])); 62 const __m128i res = _mm_sll_epi16(src_16bit, left_shift); 63 if (do_average) { 64 const __m128i data_0 = 65 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])); 66 67 const __m128i data_ref_0_lo = _mm_unpacklo_epi16(data_0, zero); 68 const __m128i data_ref_0_hi = _mm_unpackhi_epi16(data_0, zero); 69 70 const __m128i res_32b_lo = _mm_unpacklo_epi16(res, zero); 71 const __m128i res_unsigned_lo = 72 _mm_add_epi32(res_32b_lo, offset_const); 73 74 const __m128i comp_avg_res_lo = 75 highbd_comp_avg_sse4_1(&data_ref_0_lo, &res_unsigned_lo, &wt0, 76 &wt1, use_dist_wtd_comp_avg); 77 78 const __m128i res_32b_hi = _mm_unpackhi_epi16(res, zero); 79 const __m128i res_unsigned_hi = 80 _mm_add_epi32(res_32b_hi, offset_const); 81 82 const __m128i comp_avg_res_hi = 83 highbd_comp_avg_sse4_1(&data_ref_0_hi, &res_unsigned_hi, &wt0, 84 &wt1, use_dist_wtd_comp_avg); 85 86 const __m128i round_result_lo = highbd_convolve_rounding_sse2( 87 &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift); 88 const __m128i round_result_hi = highbd_convolve_rounding_sse2( 89 &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift); 90 91 const __m128i res_16b = 92 _mm_packus_epi32(round_result_lo, round_result_hi); 93 const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd); 94 95 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); 96 } else { 97 const __m128i res_unsigned_16b = 98 _mm_adds_epu16(res, offset_const_16b); 99 100 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), 101 res_unsigned_16b); 102 } 103 } 104 } 105 } else if (!(w % 4)) { 106 for (i = 0; i < h; i += 2) { 107 for (j = 0; j < w; j += 4) { 108 const __m128i src_row_0 = 109 _mm_loadl_epi64((__m128i *)(&src[i * src_stride + j])); 110 const __m128i src_row_1 = 111 _mm_loadl_epi64((__m128i *)(&src[i * src_stride + j + src_stride])); 112 const __m128i src_10 = _mm_unpacklo_epi64(src_row_0, src_row_1); 113 114 const __m128i res = _mm_sll_epi16(src_10, left_shift); 115 116 if (do_average) { 117 const __m128i data_0 = 118 _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])); 119 const __m128i data_1 = _mm_loadl_epi64( 120 (__m128i *)(&dst[i * dst_stride + j + dst_stride])); 121 122 const __m128i data_ref_0 = _mm_unpacklo_epi16(data_0, zero); 123 const __m128i data_ref_1 = _mm_unpacklo_epi16(data_1, zero); 124 125 const __m128i res_32b = _mm_unpacklo_epi16(res, zero); 126 const __m128i res_unsigned_lo = _mm_add_epi32(res_32b, offset_const); 127 128 const __m128i res_32b_hi = _mm_unpackhi_epi16(res, zero); 129 const __m128i res_unsigned_hi = 130 _mm_add_epi32(res_32b_hi, offset_const); 131 132 const __m128i comp_avg_res_lo = highbd_comp_avg_sse4_1( 133 &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_dist_wtd_comp_avg); 134 const __m128i comp_avg_res_hi = highbd_comp_avg_sse4_1( 135 &data_ref_1, &res_unsigned_hi, &wt0, &wt1, use_dist_wtd_comp_avg); 136 137 const __m128i round_result_lo = highbd_convolve_rounding_sse2( 138 &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift); 139 const __m128i round_result_hi = highbd_convolve_rounding_sse2( 140 &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift); 141 142 const __m128i res_16b = 143 _mm_packus_epi32(round_result_lo, round_result_hi); 144 const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd); 145 146 const __m128i res_1 = _mm_srli_si128(res_clip, 8); 147 148 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); 149 _mm_storel_epi64( 150 (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1); 151 } else { 152 const __m128i res_unsigned_16b = 153 _mm_adds_epu16(res, offset_const_16b); 154 155 const __m128i res_1 = _mm_srli_si128(res_unsigned_16b, 8); 156 157 _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), 158 res_unsigned_16b); 159 _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j + dst_stride]), 160 res_1); 161 } 162 } 163 } 164 } 165 } 166 167 void av1_highbd_dist_wtd_convolve_2d_sse4_1( 168 const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w, 169 int h, const InterpFilterParams *filter_params_x, 170 const InterpFilterParams *filter_params_y, const int subpel_x_qn, 171 const int subpel_y_qn, ConvolveParams *conv_params, int bd) { 172 DECLARE_ALIGNED(16, int16_t, 173 im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]); 174 CONV_BUF_TYPE *dst = conv_params->dst; 175 int dst_stride = conv_params->dst_stride; 176 int im_h = h + filter_params_y->taps - 1; 177 int im_stride = MAX_SB_SIZE; 178 int i, j; 179 const int do_average = conv_params->do_average; 180 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg; 181 const int fo_vert = filter_params_y->taps / 2 - 1; 182 const int fo_horiz = filter_params_x->taps / 2 - 1; 183 const uint16_t *const src_ptr = src - fo_vert * src_stride - fo_horiz; 184 185 const int w0 = conv_params->fwd_offset; 186 const int w1 = conv_params->bck_offset; 187 const __m128i wt0 = _mm_set1_epi32(w0); 188 const __m128i wt1 = _mm_set1_epi32(w1); 189 190 const int offset_0 = 191 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 192 const int offset = (1 << offset_0) + (1 << (offset_0 - 1)); 193 const __m128i offset_const = _mm_set1_epi32(offset); 194 const int rounding_shift = 195 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1; 196 const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1); 197 const __m128i clip_pixel_to_bd = 198 _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255)); 199 200 // Check that, even with 12-bit input, the intermediate values will fit 201 // into an unsigned 16-bit intermediate array. 202 assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16); 203 204 /* Horizontal filter */ 205 { 206 const int16_t *x_filter = av1_get_interp_filter_subpel_kernel( 207 filter_params_x, subpel_x_qn & SUBPEL_MASK); 208 const __m128i coeffs_x = _mm_loadu_si128((__m128i *)x_filter); 209 210 // coeffs 0 1 0 1 2 3 2 3 211 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x); 212 // coeffs 4 5 4 5 6 7 6 7 213 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x); 214 215 // coeffs 0 1 0 1 0 1 0 1 216 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); 217 // coeffs 2 3 2 3 2 3 2 3 218 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); 219 // coeffs 4 5 4 5 4 5 4 5 220 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); 221 // coeffs 6 7 6 7 6 7 6 7 222 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); 223 224 const __m128i round_const = _mm_set1_epi32( 225 ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1))); 226 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0); 227 228 for (i = 0; i < im_h; ++i) { 229 for (j = 0; j < w; j += 8) { 230 const __m128i data = 231 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]); 232 const __m128i data2 = 233 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j + 8]); 234 235 // Filter even-index pixels 236 const __m128i res_0 = _mm_madd_epi16(data, coeff_01); 237 const __m128i res_2 = 238 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 4), coeff_23); 239 const __m128i res_4 = 240 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 8), coeff_45); 241 const __m128i res_6 = 242 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 12), coeff_67); 243 244 __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4), 245 _mm_add_epi32(res_2, res_6)); 246 res_even = 247 _mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift); 248 249 // Filter odd-index pixels 250 const __m128i res_1 = 251 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 2), coeff_01); 252 const __m128i res_3 = 253 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 6), coeff_23); 254 const __m128i res_5 = 255 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 10), coeff_45); 256 const __m128i res_7 = 257 _mm_madd_epi16(_mm_alignr_epi8(data2, data, 14), coeff_67); 258 259 __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5), 260 _mm_add_epi32(res_3, res_7)); 261 res_odd = 262 _mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift); 263 264 // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7 265 __m128i res = _mm_packs_epi32(res_even, res_odd); 266 _mm_storeu_si128((__m128i *)&im_block[i * im_stride + j], res); 267 } 268 } 269 } 270 271 /* Vertical filter */ 272 { 273 const int16_t *y_filter = av1_get_interp_filter_subpel_kernel( 274 filter_params_y, subpel_y_qn & SUBPEL_MASK); 275 const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter); 276 277 // coeffs 0 1 0 1 2 3 2 3 278 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y); 279 // coeffs 4 5 4 5 6 7 6 7 280 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y); 281 282 // coeffs 0 1 0 1 0 1 0 1 283 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0); 284 // coeffs 2 3 2 3 2 3 2 3 285 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0); 286 // coeffs 4 5 4 5 4 5 4 5 287 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1); 288 // coeffs 6 7 6 7 6 7 6 7 289 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1); 290 291 const __m128i round_const = _mm_set1_epi32( 292 ((1 << conv_params->round_1) >> 1) - 293 (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1))); 294 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1); 295 296 for (i = 0; i < h; ++i) { 297 for (j = 0; j < w; j += 8) { 298 // Filter even-index pixels 299 const int16_t *data = &im_block[i * im_stride + j]; 300 const __m128i src_0 = 301 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride), 302 *(__m128i *)(data + 1 * im_stride)); 303 const __m128i src_2 = 304 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride), 305 *(__m128i *)(data + 3 * im_stride)); 306 const __m128i src_4 = 307 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride), 308 *(__m128i *)(data + 5 * im_stride)); 309 const __m128i src_6 = 310 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride), 311 *(__m128i *)(data + 7 * im_stride)); 312 313 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01); 314 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23); 315 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45); 316 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67); 317 318 const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2), 319 _mm_add_epi32(res_4, res_6)); 320 321 // Filter odd-index pixels 322 const __m128i src_1 = 323 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride), 324 *(__m128i *)(data + 1 * im_stride)); 325 const __m128i src_3 = 326 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride), 327 *(__m128i *)(data + 3 * im_stride)); 328 const __m128i src_5 = 329 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride), 330 *(__m128i *)(data + 5 * im_stride)); 331 const __m128i src_7 = 332 _mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride), 333 *(__m128i *)(data + 7 * im_stride)); 334 335 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01); 336 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23); 337 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45); 338 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67); 339 340 const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3), 341 _mm_add_epi32(res_5, res_7)); 342 343 // Rearrange pixels back into the order 0 ... 7 344 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd); 345 const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd); 346 347 const __m128i res_lo_round = 348 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift); 349 350 const __m128i res_unsigned_lo = 351 _mm_add_epi32(res_lo_round, offset_const); 352 353 if (w < 8) { 354 if (do_average) { 355 const __m128i data_0 = 356 _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])); 357 358 const __m128i data_ref_0 = _mm_cvtepu16_epi32(data_0); 359 360 const __m128i comp_avg_res = 361 highbd_comp_avg_sse4_1(&data_ref_0, &res_unsigned_lo, &wt0, 362 &wt1, use_dist_wtd_comp_avg); 363 364 const __m128i round_result = highbd_convolve_rounding_sse2( 365 &comp_avg_res, &offset_const, &rounding_const, rounding_shift); 366 367 const __m128i res_16b = 368 _mm_packus_epi32(round_result, round_result); 369 const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd); 370 371 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); 372 } else { 373 const __m128i res_16b = 374 _mm_packus_epi32(res_unsigned_lo, res_unsigned_lo); 375 _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), res_16b); 376 } 377 } else { 378 const __m128i res_hi_round = 379 _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift); 380 381 const __m128i res_unsigned_hi = 382 _mm_add_epi32(res_hi_round, offset_const); 383 384 if (do_average) { 385 const __m128i data_lo = 386 _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])); 387 const __m128i data_hi = 388 _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j + 4])); 389 390 const __m128i data_ref_0_lo = _mm_cvtepu16_epi32(data_lo); 391 const __m128i data_ref_0_hi = _mm_cvtepu16_epi32(data_hi); 392 393 const __m128i comp_avg_res_lo = 394 highbd_comp_avg_sse4_1(&data_ref_0_lo, &res_unsigned_lo, &wt0, 395 &wt1, use_dist_wtd_comp_avg); 396 const __m128i comp_avg_res_hi = 397 highbd_comp_avg_sse4_1(&data_ref_0_hi, &res_unsigned_hi, &wt0, 398 &wt1, use_dist_wtd_comp_avg); 399 400 const __m128i round_result_lo = 401 highbd_convolve_rounding_sse2(&comp_avg_res_lo, &offset_const, 402 &rounding_const, rounding_shift); 403 const __m128i round_result_hi = 404 highbd_convolve_rounding_sse2(&comp_avg_res_hi, &offset_const, 405 &rounding_const, rounding_shift); 406 407 const __m128i res_16b = 408 _mm_packus_epi32(round_result_lo, round_result_hi); 409 const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd); 410 411 _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip); 412 } else { 413 const __m128i res_16b = 414 _mm_packus_epi32(res_unsigned_lo, res_unsigned_hi); 415 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_16b); 416 } 417 } 418 } 419 } 420 } 421 }