1 /* 2 * Copyright (c) 2012 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 <immintrin.h> // AVX2 12 13 #include "./vpx_dsp_rtcd.h" 14 15 /* clang-format off */ 16 DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = { 17 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 18 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 19 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 20 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 21 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 22 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 23 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 24 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 25 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 26 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 27 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 28 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 29 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 30 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 31 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 32 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 33 }; 34 35 DECLARE_ALIGNED(32, static const int8_t, adjacent_sub_avx2[32]) = { 36 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 37 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1 38 }; 39 /* clang-format on */ 40 41 void vpx_get16x16var_avx2(const unsigned char *src_ptr, int source_stride, 42 const unsigned char *ref_ptr, int recon_stride, 43 unsigned int *sse, int *sum) { 44 unsigned int i, src_2strides, ref_2strides; 45 __m256i sum_reg = _mm256_setzero_si256(); 46 __m256i sse_reg = _mm256_setzero_si256(); 47 // process two 16 byte locations in a 256 bit register 48 src_2strides = source_stride << 1; 49 ref_2strides = recon_stride << 1; 50 for (i = 0; i < 8; ++i) { 51 // convert up values in 128 bit registers across lanes 52 const __m256i src0 = 53 _mm256_cvtepu8_epi16(_mm_loadu_si128((__m128i const *)(src_ptr))); 54 const __m256i src1 = _mm256_cvtepu8_epi16( 55 _mm_loadu_si128((__m128i const *)(src_ptr + source_stride))); 56 const __m256i ref0 = 57 _mm256_cvtepu8_epi16(_mm_loadu_si128((__m128i const *)(ref_ptr))); 58 const __m256i ref1 = _mm256_cvtepu8_epi16( 59 _mm_loadu_si128((__m128i const *)(ref_ptr + recon_stride))); 60 const __m256i diff0 = _mm256_sub_epi16(src0, ref0); 61 const __m256i diff1 = _mm256_sub_epi16(src1, ref1); 62 const __m256i madd0 = _mm256_madd_epi16(diff0, diff0); 63 const __m256i madd1 = _mm256_madd_epi16(diff1, diff1); 64 65 // add to the running totals 66 sum_reg = _mm256_add_epi16(sum_reg, _mm256_add_epi16(diff0, diff1)); 67 sse_reg = _mm256_add_epi32(sse_reg, _mm256_add_epi32(madd0, madd1)); 68 69 src_ptr += src_2strides; 70 ref_ptr += ref_2strides; 71 } 72 { 73 // extract the low lane and add it to the high lane 74 const __m128i sum_reg_128 = _mm_add_epi16( 75 _mm256_castsi256_si128(sum_reg), _mm256_extractf128_si256(sum_reg, 1)); 76 const __m128i sse_reg_128 = _mm_add_epi32( 77 _mm256_castsi256_si128(sse_reg), _mm256_extractf128_si256(sse_reg, 1)); 78 79 // sum upper and lower 64 bits together and convert up to 32 bit values 80 const __m128i sum_reg_64 = 81 _mm_add_epi16(sum_reg_128, _mm_srli_si128(sum_reg_128, 8)); 82 const __m128i sum_int32 = _mm_cvtepi16_epi32(sum_reg_64); 83 84 // unpack sse and sum registers and add 85 const __m128i sse_sum_lo = _mm_unpacklo_epi32(sse_reg_128, sum_int32); 86 const __m128i sse_sum_hi = _mm_unpackhi_epi32(sse_reg_128, sum_int32); 87 const __m128i sse_sum = _mm_add_epi32(sse_sum_lo, sse_sum_hi); 88 89 // perform the final summation and extract the results 90 const __m128i res = _mm_add_epi32(sse_sum, _mm_srli_si128(sse_sum, 8)); 91 *((int *)sse) = _mm_cvtsi128_si32(res); 92 *((int *)sum) = _mm_extract_epi32(res, 1); 93 } 94 } 95 96 static void get32x16var_avx2(const unsigned char *src_ptr, int source_stride, 97 const unsigned char *ref_ptr, int recon_stride, 98 unsigned int *sse, int *sum) { 99 unsigned int i, src_2strides, ref_2strides; 100 const __m256i adj_sub = _mm256_load_si256((__m256i const *)adjacent_sub_avx2); 101 __m256i sum_reg = _mm256_setzero_si256(); 102 __m256i sse_reg = _mm256_setzero_si256(); 103 104 // process 64 elements in an iteration 105 src_2strides = source_stride << 1; 106 ref_2strides = recon_stride << 1; 107 for (i = 0; i < 8; i++) { 108 const __m256i src0 = _mm256_loadu_si256((__m256i const *)(src_ptr)); 109 const __m256i src1 = 110 _mm256_loadu_si256((__m256i const *)(src_ptr + source_stride)); 111 const __m256i ref0 = _mm256_loadu_si256((__m256i const *)(ref_ptr)); 112 const __m256i ref1 = 113 _mm256_loadu_si256((__m256i const *)(ref_ptr + recon_stride)); 114 115 // unpack into pairs of source and reference values 116 const __m256i src_ref0 = _mm256_unpacklo_epi8(src0, ref0); 117 const __m256i src_ref1 = _mm256_unpackhi_epi8(src0, ref0); 118 const __m256i src_ref2 = _mm256_unpacklo_epi8(src1, ref1); 119 const __m256i src_ref3 = _mm256_unpackhi_epi8(src1, ref1); 120 121 // subtract adjacent elements using src*1 + ref*-1 122 const __m256i diff0 = _mm256_maddubs_epi16(src_ref0, adj_sub); 123 const __m256i diff1 = _mm256_maddubs_epi16(src_ref1, adj_sub); 124 const __m256i diff2 = _mm256_maddubs_epi16(src_ref2, adj_sub); 125 const __m256i diff3 = _mm256_maddubs_epi16(src_ref3, adj_sub); 126 const __m256i madd0 = _mm256_madd_epi16(diff0, diff0); 127 const __m256i madd1 = _mm256_madd_epi16(diff1, diff1); 128 const __m256i madd2 = _mm256_madd_epi16(diff2, diff2); 129 const __m256i madd3 = _mm256_madd_epi16(diff3, diff3); 130 131 // add to the running totals 132 sum_reg = _mm256_add_epi16(sum_reg, _mm256_add_epi16(diff0, diff1)); 133 sum_reg = _mm256_add_epi16(sum_reg, _mm256_add_epi16(diff2, diff3)); 134 sse_reg = _mm256_add_epi32(sse_reg, _mm256_add_epi32(madd0, madd1)); 135 sse_reg = _mm256_add_epi32(sse_reg, _mm256_add_epi32(madd2, madd3)); 136 137 src_ptr += src_2strides; 138 ref_ptr += ref_2strides; 139 } 140 141 { 142 // extract the low lane and add it to the high lane 143 const __m128i sum_reg_128 = _mm_add_epi16( 144 _mm256_castsi256_si128(sum_reg), _mm256_extractf128_si256(sum_reg, 1)); 145 const __m128i sse_reg_128 = _mm_add_epi32( 146 _mm256_castsi256_si128(sse_reg), _mm256_extractf128_si256(sse_reg, 1)); 147 148 // sum upper and lower 64 bits together and convert up to 32 bit values 149 const __m128i sum_reg_64 = 150 _mm_add_epi16(sum_reg_128, _mm_srli_si128(sum_reg_128, 8)); 151 const __m128i sum_int32 = _mm_cvtepi16_epi32(sum_reg_64); 152 153 // unpack sse and sum registers and add 154 const __m128i sse_sum_lo = _mm_unpacklo_epi32(sse_reg_128, sum_int32); 155 const __m128i sse_sum_hi = _mm_unpackhi_epi32(sse_reg_128, sum_int32); 156 const __m128i sse_sum = _mm_add_epi32(sse_sum_lo, sse_sum_hi); 157 158 // perform the final summation and extract the results 159 const __m128i res = _mm_add_epi32(sse_sum, _mm_srli_si128(sse_sum, 8)); 160 *((int *)sse) = _mm_cvtsi128_si32(res); 161 *((int *)sum) = _mm_extract_epi32(res, 1); 162 } 163 } 164 165 #define FILTER_SRC(filter) \ 166 /* filter the source */ \ 167 exp_src_lo = _mm256_maddubs_epi16(exp_src_lo, filter); \ 168 exp_src_hi = _mm256_maddubs_epi16(exp_src_hi, filter); \ 169 \ 170 /* add 8 to source */ \ 171 exp_src_lo = _mm256_add_epi16(exp_src_lo, pw8); \ 172 exp_src_hi = _mm256_add_epi16(exp_src_hi, pw8); \ 173 \ 174 /* divide source by 16 */ \ 175 exp_src_lo = _mm256_srai_epi16(exp_src_lo, 4); \ 176 exp_src_hi = _mm256_srai_epi16(exp_src_hi, 4); 177 178 #define CALC_SUM_SSE_INSIDE_LOOP \ 179 /* expand each byte to 2 bytes */ \ 180 exp_dst_lo = _mm256_unpacklo_epi8(dst_reg, zero_reg); \ 181 exp_dst_hi = _mm256_unpackhi_epi8(dst_reg, zero_reg); \ 182 /* source - dest */ \ 183 exp_src_lo = _mm256_sub_epi16(exp_src_lo, exp_dst_lo); \ 184 exp_src_hi = _mm256_sub_epi16(exp_src_hi, exp_dst_hi); \ 185 /* caculate sum */ \ 186 *sum_reg = _mm256_add_epi16(*sum_reg, exp_src_lo); \ 187 exp_src_lo = _mm256_madd_epi16(exp_src_lo, exp_src_lo); \ 188 *sum_reg = _mm256_add_epi16(*sum_reg, exp_src_hi); \ 189 exp_src_hi = _mm256_madd_epi16(exp_src_hi, exp_src_hi); \ 190 /* calculate sse */ \ 191 *sse_reg = _mm256_add_epi32(*sse_reg, exp_src_lo); \ 192 *sse_reg = _mm256_add_epi32(*sse_reg, exp_src_hi); 193 194 // final calculation to sum and sse 195 #define CALC_SUM_AND_SSE \ 196 res_cmp = _mm256_cmpgt_epi16(zero_reg, sum_reg); \ 197 sse_reg_hi = _mm256_srli_si256(sse_reg, 8); \ 198 sum_reg_lo = _mm256_unpacklo_epi16(sum_reg, res_cmp); \ 199 sum_reg_hi = _mm256_unpackhi_epi16(sum_reg, res_cmp); \ 200 sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \ 201 sum_reg = _mm256_add_epi32(sum_reg_lo, sum_reg_hi); \ 202 \ 203 sse_reg_hi = _mm256_srli_si256(sse_reg, 4); \ 204 sum_reg_hi = _mm256_srli_si256(sum_reg, 8); \ 205 \ 206 sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \ 207 sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \ 208 *((int *)sse) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) + \ 209 _mm_cvtsi128_si32(_mm256_extractf128_si256(sse_reg, 1)); \ 210 sum_reg_hi = _mm256_srli_si256(sum_reg, 4); \ 211 sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \ 212 sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_reg)) + \ 213 _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_reg, 1)); 214 215 static INLINE void spv32_x0_y0(const uint8_t *src, int src_stride, 216 const uint8_t *dst, int dst_stride, 217 const uint8_t *sec, int sec_stride, int do_sec, 218 int height, __m256i *sum_reg, __m256i *sse_reg) { 219 const __m256i zero_reg = _mm256_setzero_si256(); 220 __m256i exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; 221 int i; 222 for (i = 0; i < height; i++) { 223 const __m256i dst_reg = _mm256_loadu_si256((__m256i const *)dst); 224 const __m256i src_reg = _mm256_loadu_si256((__m256i const *)src); 225 if (do_sec) { 226 const __m256i sec_reg = _mm256_loadu_si256((__m256i const *)sec); 227 const __m256i avg_reg = _mm256_avg_epu8(src_reg, sec_reg); 228 exp_src_lo = _mm256_unpacklo_epi8(avg_reg, zero_reg); 229 exp_src_hi = _mm256_unpackhi_epi8(avg_reg, zero_reg); 230 sec += sec_stride; 231 } else { 232 exp_src_lo = _mm256_unpacklo_epi8(src_reg, zero_reg); 233 exp_src_hi = _mm256_unpackhi_epi8(src_reg, zero_reg); 234 } 235 CALC_SUM_SSE_INSIDE_LOOP 236 src += src_stride; 237 dst += dst_stride; 238 } 239 } 240 241 // (x == 0, y == 4) or (x == 4, y == 0). sstep determines the direction. 242 static INLINE void spv32_half_zero(const uint8_t *src, int src_stride, 243 const uint8_t *dst, int dst_stride, 244 const uint8_t *sec, int sec_stride, 245 int do_sec, int height, __m256i *sum_reg, 246 __m256i *sse_reg, int sstep) { 247 const __m256i zero_reg = _mm256_setzero_si256(); 248 __m256i exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; 249 int i; 250 for (i = 0; i < height; i++) { 251 const __m256i dst_reg = _mm256_loadu_si256((__m256i const *)dst); 252 const __m256i src_0 = _mm256_loadu_si256((__m256i const *)src); 253 const __m256i src_1 = _mm256_loadu_si256((__m256i const *)(src + sstep)); 254 const __m256i src_avg = _mm256_avg_epu8(src_0, src_1); 255 if (do_sec) { 256 const __m256i sec_reg = _mm256_loadu_si256((__m256i const *)sec); 257 const __m256i avg_reg = _mm256_avg_epu8(src_avg, sec_reg); 258 exp_src_lo = _mm256_unpacklo_epi8(avg_reg, zero_reg); 259 exp_src_hi = _mm256_unpackhi_epi8(avg_reg, zero_reg); 260 sec += sec_stride; 261 } else { 262 exp_src_lo = _mm256_unpacklo_epi8(src_avg, zero_reg); 263 exp_src_hi = _mm256_unpackhi_epi8(src_avg, zero_reg); 264 } 265 CALC_SUM_SSE_INSIDE_LOOP 266 src += src_stride; 267 dst += dst_stride; 268 } 269 } 270 271 static INLINE void spv32_x0_y4(const uint8_t *src, int src_stride, 272 const uint8_t *dst, int dst_stride, 273 const uint8_t *sec, int sec_stride, int do_sec, 274 int height, __m256i *sum_reg, __m256i *sse_reg) { 275 spv32_half_zero(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 276 height, sum_reg, sse_reg, src_stride); 277 } 278 279 static INLINE void spv32_x4_y0(const uint8_t *src, int src_stride, 280 const uint8_t *dst, int dst_stride, 281 const uint8_t *sec, int sec_stride, int do_sec, 282 int height, __m256i *sum_reg, __m256i *sse_reg) { 283 spv32_half_zero(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 284 height, sum_reg, sse_reg, 1); 285 } 286 287 static INLINE void spv32_x4_y4(const uint8_t *src, int src_stride, 288 const uint8_t *dst, int dst_stride, 289 const uint8_t *sec, int sec_stride, int do_sec, 290 int height, __m256i *sum_reg, __m256i *sse_reg) { 291 const __m256i zero_reg = _mm256_setzero_si256(); 292 const __m256i src_a = _mm256_loadu_si256((__m256i const *)src); 293 const __m256i src_b = _mm256_loadu_si256((__m256i const *)(src + 1)); 294 __m256i prev_src_avg = _mm256_avg_epu8(src_a, src_b); 295 __m256i exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; 296 int i; 297 src += src_stride; 298 for (i = 0; i < height; i++) { 299 const __m256i dst_reg = _mm256_loadu_si256((__m256i const *)dst); 300 const __m256i src_0 = _mm256_loadu_si256((__m256i const *)(src)); 301 const __m256i src_1 = _mm256_loadu_si256((__m256i const *)(src + 1)); 302 const __m256i src_avg = _mm256_avg_epu8(src_0, src_1); 303 const __m256i current_avg = _mm256_avg_epu8(prev_src_avg, src_avg); 304 prev_src_avg = src_avg; 305 306 if (do_sec) { 307 const __m256i sec_reg = _mm256_loadu_si256((__m256i const *)sec); 308 const __m256i avg_reg = _mm256_avg_epu8(current_avg, sec_reg); 309 exp_src_lo = _mm256_unpacklo_epi8(avg_reg, zero_reg); 310 exp_src_hi = _mm256_unpackhi_epi8(avg_reg, zero_reg); 311 sec += sec_stride; 312 } else { 313 exp_src_lo = _mm256_unpacklo_epi8(current_avg, zero_reg); 314 exp_src_hi = _mm256_unpackhi_epi8(current_avg, zero_reg); 315 } 316 // save current source average 317 CALC_SUM_SSE_INSIDE_LOOP 318 dst += dst_stride; 319 src += src_stride; 320 } 321 } 322 323 // (x == 0, y == bil) or (x == 4, y == bil). sstep determines the direction. 324 static INLINE void spv32_bilin_zero(const uint8_t *src, int src_stride, 325 const uint8_t *dst, int dst_stride, 326 const uint8_t *sec, int sec_stride, 327 int do_sec, int height, __m256i *sum_reg, 328 __m256i *sse_reg, int offset, int sstep) { 329 const __m256i zero_reg = _mm256_setzero_si256(); 330 const __m256i pw8 = _mm256_set1_epi16(8); 331 const __m256i filter = _mm256_load_si256( 332 (__m256i const *)(bilinear_filters_avx2 + (offset << 5))); 333 __m256i exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; 334 int i; 335 for (i = 0; i < height; i++) { 336 const __m256i dst_reg = _mm256_loadu_si256((__m256i const *)dst); 337 const __m256i src_0 = _mm256_loadu_si256((__m256i const *)src); 338 const __m256i src_1 = _mm256_loadu_si256((__m256i const *)(src + sstep)); 339 exp_src_lo = _mm256_unpacklo_epi8(src_0, src_1); 340 exp_src_hi = _mm256_unpackhi_epi8(src_0, src_1); 341 342 FILTER_SRC(filter) 343 if (do_sec) { 344 const __m256i sec_reg = _mm256_loadu_si256((__m256i const *)sec); 345 const __m256i exp_src = _mm256_packus_epi16(exp_src_lo, exp_src_hi); 346 const __m256i avg_reg = _mm256_avg_epu8(exp_src, sec_reg); 347 sec += sec_stride; 348 exp_src_lo = _mm256_unpacklo_epi8(avg_reg, zero_reg); 349 exp_src_hi = _mm256_unpackhi_epi8(avg_reg, zero_reg); 350 } 351 CALC_SUM_SSE_INSIDE_LOOP 352 src += src_stride; 353 dst += dst_stride; 354 } 355 } 356 357 static INLINE void spv32_x0_yb(const uint8_t *src, int src_stride, 358 const uint8_t *dst, int dst_stride, 359 const uint8_t *sec, int sec_stride, int do_sec, 360 int height, __m256i *sum_reg, __m256i *sse_reg, 361 int y_offset) { 362 spv32_bilin_zero(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 363 height, sum_reg, sse_reg, y_offset, src_stride); 364 } 365 366 static INLINE void spv32_xb_y0(const uint8_t *src, int src_stride, 367 const uint8_t *dst, int dst_stride, 368 const uint8_t *sec, int sec_stride, int do_sec, 369 int height, __m256i *sum_reg, __m256i *sse_reg, 370 int x_offset) { 371 spv32_bilin_zero(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 372 height, sum_reg, sse_reg, x_offset, 1); 373 } 374 375 static INLINE void spv32_x4_yb(const uint8_t *src, int src_stride, 376 const uint8_t *dst, int dst_stride, 377 const uint8_t *sec, int sec_stride, int do_sec, 378 int height, __m256i *sum_reg, __m256i *sse_reg, 379 int y_offset) { 380 const __m256i zero_reg = _mm256_setzero_si256(); 381 const __m256i pw8 = _mm256_set1_epi16(8); 382 const __m256i filter = _mm256_load_si256( 383 (__m256i const *)(bilinear_filters_avx2 + (y_offset << 5))); 384 const __m256i src_a = _mm256_loadu_si256((__m256i const *)src); 385 const __m256i src_b = _mm256_loadu_si256((__m256i const *)(src + 1)); 386 __m256i prev_src_avg = _mm256_avg_epu8(src_a, src_b); 387 __m256i exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; 388 int i; 389 src += src_stride; 390 for (i = 0; i < height; i++) { 391 const __m256i dst_reg = _mm256_loadu_si256((__m256i const *)dst); 392 const __m256i src_0 = _mm256_loadu_si256((__m256i const *)src); 393 const __m256i src_1 = _mm256_loadu_si256((__m256i const *)(src + 1)); 394 const __m256i src_avg = _mm256_avg_epu8(src_0, src_1); 395 exp_src_lo = _mm256_unpacklo_epi8(prev_src_avg, src_avg); 396 exp_src_hi = _mm256_unpackhi_epi8(prev_src_avg, src_avg); 397 prev_src_avg = src_avg; 398 399 FILTER_SRC(filter) 400 if (do_sec) { 401 const __m256i sec_reg = _mm256_loadu_si256((__m256i const *)sec); 402 const __m256i exp_src_avg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); 403 const __m256i avg_reg = _mm256_avg_epu8(exp_src_avg, sec_reg); 404 exp_src_lo = _mm256_unpacklo_epi8(avg_reg, zero_reg); 405 exp_src_hi = _mm256_unpackhi_epi8(avg_reg, zero_reg); 406 sec += sec_stride; 407 } 408 CALC_SUM_SSE_INSIDE_LOOP 409 dst += dst_stride; 410 src += src_stride; 411 } 412 } 413 414 static INLINE void spv32_xb_y4(const uint8_t *src, int src_stride, 415 const uint8_t *dst, int dst_stride, 416 const uint8_t *sec, int sec_stride, int do_sec, 417 int height, __m256i *sum_reg, __m256i *sse_reg, 418 int x_offset) { 419 const __m256i zero_reg = _mm256_setzero_si256(); 420 const __m256i pw8 = _mm256_set1_epi16(8); 421 const __m256i filter = _mm256_load_si256( 422 (__m256i const *)(bilinear_filters_avx2 + (x_offset << 5))); 423 const __m256i src_a = _mm256_loadu_si256((__m256i const *)src); 424 const __m256i src_b = _mm256_loadu_si256((__m256i const *)(src + 1)); 425 __m256i exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; 426 __m256i src_reg, src_pack; 427 int i; 428 exp_src_lo = _mm256_unpacklo_epi8(src_a, src_b); 429 exp_src_hi = _mm256_unpackhi_epi8(src_a, src_b); 430 FILTER_SRC(filter) 431 // convert each 16 bit to 8 bit to each low and high lane source 432 src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); 433 434 src += src_stride; 435 for (i = 0; i < height; i++) { 436 const __m256i dst_reg = _mm256_loadu_si256((__m256i const *)dst); 437 const __m256i src_0 = _mm256_loadu_si256((__m256i const *)src); 438 const __m256i src_1 = _mm256_loadu_si256((__m256i const *)(src + 1)); 439 exp_src_lo = _mm256_unpacklo_epi8(src_0, src_1); 440 exp_src_hi = _mm256_unpackhi_epi8(src_0, src_1); 441 442 FILTER_SRC(filter) 443 444 src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi); 445 // average between previous pack to the current 446 src_pack = _mm256_avg_epu8(src_pack, src_reg); 447 448 if (do_sec) { 449 const __m256i sec_reg = _mm256_loadu_si256((__m256i const *)sec); 450 const __m256i avg_pack = _mm256_avg_epu8(src_pack, sec_reg); 451 exp_src_lo = _mm256_unpacklo_epi8(avg_pack, zero_reg); 452 exp_src_hi = _mm256_unpackhi_epi8(avg_pack, zero_reg); 453 sec += sec_stride; 454 } else { 455 exp_src_lo = _mm256_unpacklo_epi8(src_pack, zero_reg); 456 exp_src_hi = _mm256_unpackhi_epi8(src_pack, zero_reg); 457 } 458 CALC_SUM_SSE_INSIDE_LOOP 459 src_pack = src_reg; 460 dst += dst_stride; 461 src += src_stride; 462 } 463 } 464 465 static INLINE void spv32_xb_yb(const uint8_t *src, int src_stride, 466 const uint8_t *dst, int dst_stride, 467 const uint8_t *sec, int sec_stride, int do_sec, 468 int height, __m256i *sum_reg, __m256i *sse_reg, 469 int x_offset, int y_offset) { 470 const __m256i zero_reg = _mm256_setzero_si256(); 471 const __m256i pw8 = _mm256_set1_epi16(8); 472 const __m256i xfilter = _mm256_load_si256( 473 (__m256i const *)(bilinear_filters_avx2 + (x_offset << 5))); 474 const __m256i yfilter = _mm256_load_si256( 475 (__m256i const *)(bilinear_filters_avx2 + (y_offset << 5))); 476 const __m256i src_a = _mm256_loadu_si256((__m256i const *)src); 477 const __m256i src_b = _mm256_loadu_si256((__m256i const *)(src + 1)); 478 __m256i exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; 479 __m256i prev_src_pack, src_pack; 480 int i; 481 exp_src_lo = _mm256_unpacklo_epi8(src_a, src_b); 482 exp_src_hi = _mm256_unpackhi_epi8(src_a, src_b); 483 FILTER_SRC(xfilter) 484 // convert each 16 bit to 8 bit to each low and high lane source 485 prev_src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); 486 src += src_stride; 487 488 for (i = 0; i < height; i++) { 489 const __m256i dst_reg = _mm256_loadu_si256((__m256i const *)dst); 490 const __m256i src_0 = _mm256_loadu_si256((__m256i const *)src); 491 const __m256i src_1 = _mm256_loadu_si256((__m256i const *)(src + 1)); 492 exp_src_lo = _mm256_unpacklo_epi8(src_0, src_1); 493 exp_src_hi = _mm256_unpackhi_epi8(src_0, src_1); 494 495 FILTER_SRC(xfilter) 496 src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi); 497 498 // merge previous pack to current pack source 499 exp_src_lo = _mm256_unpacklo_epi8(prev_src_pack, src_pack); 500 exp_src_hi = _mm256_unpackhi_epi8(prev_src_pack, src_pack); 501 502 FILTER_SRC(yfilter) 503 if (do_sec) { 504 const __m256i sec_reg = _mm256_loadu_si256((__m256i const *)sec); 505 const __m256i exp_src = _mm256_packus_epi16(exp_src_lo, exp_src_hi); 506 const __m256i avg_reg = _mm256_avg_epu8(exp_src, sec_reg); 507 exp_src_lo = _mm256_unpacklo_epi8(avg_reg, zero_reg); 508 exp_src_hi = _mm256_unpackhi_epi8(avg_reg, zero_reg); 509 sec += sec_stride; 510 } 511 512 prev_src_pack = src_pack; 513 514 CALC_SUM_SSE_INSIDE_LOOP 515 dst += dst_stride; 516 src += src_stride; 517 } 518 } 519 520 static INLINE int sub_pix_var32xh(const uint8_t *src, int src_stride, 521 int x_offset, int y_offset, 522 const uint8_t *dst, int dst_stride, 523 const uint8_t *sec, int sec_stride, 524 int do_sec, int height, unsigned int *sse) { 525 const __m256i zero_reg = _mm256_setzero_si256(); 526 __m256i sum_reg = _mm256_setzero_si256(); 527 __m256i sse_reg = _mm256_setzero_si256(); 528 __m256i sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi; 529 int sum; 530 // x_offset = 0 and y_offset = 0 531 if (x_offset == 0) { 532 if (y_offset == 0) { 533 spv32_x0_y0(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 534 height, &sum_reg, &sse_reg); 535 // x_offset = 0 and y_offset = 4 536 } else if (y_offset == 4) { 537 spv32_x0_y4(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 538 height, &sum_reg, &sse_reg); 539 // x_offset = 0 and y_offset = bilin interpolation 540 } else { 541 spv32_x0_yb(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 542 height, &sum_reg, &sse_reg, y_offset); 543 } 544 // x_offset = 4 and y_offset = 0 545 } else if (x_offset == 4) { 546 if (y_offset == 0) { 547 spv32_x4_y0(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 548 height, &sum_reg, &sse_reg); 549 // x_offset = 4 and y_offset = 4 550 } else if (y_offset == 4) { 551 spv32_x4_y4(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 552 height, &sum_reg, &sse_reg); 553 // x_offset = 4 and y_offset = bilin interpolation 554 } else { 555 spv32_x4_yb(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 556 height, &sum_reg, &sse_reg, y_offset); 557 } 558 // x_offset = bilin interpolation and y_offset = 0 559 } else { 560 if (y_offset == 0) { 561 spv32_xb_y0(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 562 height, &sum_reg, &sse_reg, x_offset); 563 // x_offset = bilin interpolation and y_offset = 4 564 } else if (y_offset == 4) { 565 spv32_xb_y4(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 566 height, &sum_reg, &sse_reg, x_offset); 567 // x_offset = bilin interpolation and y_offset = bilin interpolation 568 } else { 569 spv32_xb_yb(src, src_stride, dst, dst_stride, sec, sec_stride, do_sec, 570 height, &sum_reg, &sse_reg, x_offset, y_offset); 571 } 572 } 573 CALC_SUM_AND_SSE 574 return sum; 575 } 576 577 static unsigned int sub_pixel_variance32xh_avx2( 578 const uint8_t *src, int src_stride, int x_offset, int y_offset, 579 const uint8_t *dst, int dst_stride, int height, unsigned int *sse) { 580 return sub_pix_var32xh(src, src_stride, x_offset, y_offset, dst, dst_stride, 581 NULL, 0, 0, height, sse); 582 } 583 584 static unsigned int sub_pixel_avg_variance32xh_avx2( 585 const uint8_t *src, int src_stride, int x_offset, int y_offset, 586 const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride, 587 int height, unsigned int *sse) { 588 return sub_pix_var32xh(src, src_stride, x_offset, y_offset, dst, dst_stride, 589 sec, sec_stride, 1, height, sse); 590 } 591 592 typedef void (*get_var_avx2)(const uint8_t *src, int src_stride, 593 const uint8_t *ref, int ref_stride, 594 unsigned int *sse, int *sum); 595 596 static void variance_avx2(const uint8_t *src, int src_stride, 597 const uint8_t *ref, int ref_stride, int w, int h, 598 unsigned int *sse, int *sum, get_var_avx2 var_fn, 599 int block_size) { 600 int i, j; 601 602 *sse = 0; 603 *sum = 0; 604 605 for (i = 0; i < h; i += 16) { 606 for (j = 0; j < w; j += block_size) { 607 unsigned int sse0; 608 int sum0; 609 var_fn(&src[src_stride * i + j], src_stride, &ref[ref_stride * i + j], 610 ref_stride, &sse0, &sum0); 611 *sse += sse0; 612 *sum += sum0; 613 } 614 } 615 } 616 617 unsigned int vpx_variance16x16_avx2(const uint8_t *src, int src_stride, 618 const uint8_t *ref, int ref_stride, 619 unsigned int *sse) { 620 int sum; 621 variance_avx2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum, 622 vpx_get16x16var_avx2, 16); 623 return *sse - (uint32_t)(((int64_t)sum * sum) >> 8); 624 } 625 626 unsigned int vpx_mse16x16_avx2(const uint8_t *src, int src_stride, 627 const uint8_t *ref, int ref_stride, 628 unsigned int *sse) { 629 int sum; 630 vpx_get16x16var_avx2(src, src_stride, ref, ref_stride, sse, &sum); 631 return *sse; 632 } 633 634 unsigned int vpx_variance32x16_avx2(const uint8_t *src, int src_stride, 635 const uint8_t *ref, int ref_stride, 636 unsigned int *sse) { 637 int sum; 638 variance_avx2(src, src_stride, ref, ref_stride, 32, 16, sse, &sum, 639 get32x16var_avx2, 32); 640 return *sse - (uint32_t)(((int64_t)sum * sum) >> 9); 641 } 642 643 unsigned int vpx_variance32x32_avx2(const uint8_t *src, int src_stride, 644 const uint8_t *ref, int ref_stride, 645 unsigned int *sse) { 646 int sum; 647 variance_avx2(src, src_stride, ref, ref_stride, 32, 32, sse, &sum, 648 get32x16var_avx2, 32); 649 return *sse - (uint32_t)(((int64_t)sum * sum) >> 10); 650 } 651 652 unsigned int vpx_variance64x64_avx2(const uint8_t *src, int src_stride, 653 const uint8_t *ref, int ref_stride, 654 unsigned int *sse) { 655 int sum; 656 variance_avx2(src, src_stride, ref, ref_stride, 64, 64, sse, &sum, 657 get32x16var_avx2, 32); 658 return *sse - (uint32_t)(((int64_t)sum * sum) >> 12); 659 } 660 661 unsigned int vpx_variance64x32_avx2(const uint8_t *src, int src_stride, 662 const uint8_t *ref, int ref_stride, 663 unsigned int *sse) { 664 int sum; 665 variance_avx2(src, src_stride, ref, ref_stride, 64, 32, sse, &sum, 666 get32x16var_avx2, 32); 667 return *sse - (uint32_t)(((int64_t)sum * sum) >> 11); 668 } 669 670 unsigned int vpx_sub_pixel_variance64x64_avx2(const uint8_t *src, 671 int src_stride, int x_offset, 672 int y_offset, const uint8_t *dst, 673 int dst_stride, 674 unsigned int *sse) { 675 unsigned int sse1; 676 const int se1 = sub_pixel_variance32xh_avx2( 677 src, src_stride, x_offset, y_offset, dst, dst_stride, 64, &sse1); 678 unsigned int sse2; 679 const int se2 = 680 sub_pixel_variance32xh_avx2(src + 32, src_stride, x_offset, y_offset, 681 dst + 32, dst_stride, 64, &sse2); 682 const int se = se1 + se2; 683 *sse = sse1 + sse2; 684 return *sse - (uint32_t)(((int64_t)se * se) >> 12); 685 } 686 687 unsigned int vpx_sub_pixel_variance32x32_avx2(const uint8_t *src, 688 int src_stride, int x_offset, 689 int y_offset, const uint8_t *dst, 690 int dst_stride, 691 unsigned int *sse) { 692 const int se = sub_pixel_variance32xh_avx2( 693 src, src_stride, x_offset, y_offset, dst, dst_stride, 32, sse); 694 return *sse - (uint32_t)(((int64_t)se * se) >> 10); 695 } 696 697 unsigned int vpx_sub_pixel_avg_variance64x64_avx2( 698 const uint8_t *src, int src_stride, int x_offset, int y_offset, 699 const uint8_t *dst, int dst_stride, unsigned int *sse, const uint8_t *sec) { 700 unsigned int sse1; 701 const int se1 = sub_pixel_avg_variance32xh_avx2( 702 src, src_stride, x_offset, y_offset, dst, dst_stride, sec, 64, 64, &sse1); 703 unsigned int sse2; 704 const int se2 = sub_pixel_avg_variance32xh_avx2( 705 src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, sec + 32, 706 64, 64, &sse2); 707 const int se = se1 + se2; 708 709 *sse = sse1 + sse2; 710 711 return *sse - (uint32_t)(((int64_t)se * se) >> 12); 712 } 713 714 unsigned int vpx_sub_pixel_avg_variance32x32_avx2( 715 const uint8_t *src, int src_stride, int x_offset, int y_offset, 716 const uint8_t *dst, int dst_stride, unsigned int *sse, const uint8_t *sec) { 717 // Process 32 elements in parallel. 718 const int se = sub_pixel_avg_variance32xh_avx2( 719 src, src_stride, x_offset, y_offset, dst, dst_stride, sec, 32, 32, sse); 720 return *sse - (uint32_t)(((int64_t)se * se) >> 10); 721 } 722