sum_squares_avx2.c (12712B)
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 <immintrin.h> 13 #include <smmintrin.h> 14 15 #include "aom_dsp/x86/synonyms.h" 16 #include "aom_dsp/x86/synonyms_avx2.h" 17 #include "aom_dsp/x86/sum_squares_sse2.h" 18 #include "config/aom_config.h" 19 #include "config/aom_dsp_rtcd.h" 20 21 static uint64_t aom_sum_squares_2d_i16_nxn_avx2(const int16_t *src, int stride, 22 int width, int height) { 23 uint64_t result; 24 __m256i v_acc_q = _mm256_setzero_si256(); 25 const __m256i v_zext_mask_q = _mm256_set1_epi64x(~0u); 26 for (int col = 0; col < height; col += 4) { 27 __m256i v_acc_d = _mm256_setzero_si256(); 28 for (int row = 0; row < width; row += 16) { 29 const int16_t *tempsrc = src + row; 30 const __m256i v_val_0_w = 31 _mm256_loadu_si256((const __m256i *)(tempsrc + 0 * stride)); 32 const __m256i v_val_1_w = 33 _mm256_loadu_si256((const __m256i *)(tempsrc + 1 * stride)); 34 const __m256i v_val_2_w = 35 _mm256_loadu_si256((const __m256i *)(tempsrc + 2 * stride)); 36 const __m256i v_val_3_w = 37 _mm256_loadu_si256((const __m256i *)(tempsrc + 3 * stride)); 38 39 const __m256i v_sq_0_d = _mm256_madd_epi16(v_val_0_w, v_val_0_w); 40 const __m256i v_sq_1_d = _mm256_madd_epi16(v_val_1_w, v_val_1_w); 41 const __m256i v_sq_2_d = _mm256_madd_epi16(v_val_2_w, v_val_2_w); 42 const __m256i v_sq_3_d = _mm256_madd_epi16(v_val_3_w, v_val_3_w); 43 44 const __m256i v_sum_01_d = _mm256_add_epi32(v_sq_0_d, v_sq_1_d); 45 const __m256i v_sum_23_d = _mm256_add_epi32(v_sq_2_d, v_sq_3_d); 46 const __m256i v_sum_0123_d = _mm256_add_epi32(v_sum_01_d, v_sum_23_d); 47 48 v_acc_d = _mm256_add_epi32(v_acc_d, v_sum_0123_d); 49 } 50 v_acc_q = 51 _mm256_add_epi64(v_acc_q, _mm256_and_si256(v_acc_d, v_zext_mask_q)); 52 v_acc_q = _mm256_add_epi64(v_acc_q, _mm256_srli_epi64(v_acc_d, 32)); 53 src += 4 * stride; 54 } 55 __m128i lower_64_2_Value = _mm256_castsi256_si128(v_acc_q); 56 __m128i higher_64_2_Value = _mm256_extracti128_si256(v_acc_q, 1); 57 __m128i result_64_2_int = _mm_add_epi64(lower_64_2_Value, higher_64_2_Value); 58 59 result_64_2_int = _mm_add_epi64( 60 result_64_2_int, _mm_unpackhi_epi64(result_64_2_int, result_64_2_int)); 61 62 xx_storel_64(&result, result_64_2_int); 63 64 return result; 65 } 66 67 uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width, 68 int height) { 69 if (LIKELY(width == 4 && height == 4)) { 70 return aom_sum_squares_2d_i16_4x4_sse2(src, stride); 71 } else if (LIKELY(width == 4 && (height & 3) == 0)) { 72 return aom_sum_squares_2d_i16_4xn_sse2(src, stride, height); 73 } else if (LIKELY(width == 8 && (height & 3) == 0)) { 74 return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height); 75 } else if (LIKELY(((width & 15) == 0) && ((height & 3) == 0))) { 76 return aom_sum_squares_2d_i16_nxn_avx2(src, stride, width, height); 77 } else { 78 return aom_sum_squares_2d_i16_c(src, stride, width, height); 79 } 80 } 81 82 static uint64_t aom_sum_sse_2d_i16_nxn_avx2(const int16_t *src, int stride, 83 int width, int height, int *sum) { 84 uint64_t result; 85 const __m256i zero_reg = _mm256_setzero_si256(); 86 const __m256i one_reg = _mm256_set1_epi16(1); 87 88 __m256i v_sse_total = zero_reg; 89 __m256i v_sum_total = zero_reg; 90 91 for (int col = 0; col < height; col += 4) { 92 __m256i v_sse_row = zero_reg; 93 for (int row = 0; row < width; row += 16) { 94 const int16_t *tempsrc = src + row; 95 const __m256i v_val_0_w = 96 _mm256_loadu_si256((const __m256i *)(tempsrc + 0 * stride)); 97 const __m256i v_val_1_w = 98 _mm256_loadu_si256((const __m256i *)(tempsrc + 1 * stride)); 99 const __m256i v_val_2_w = 100 _mm256_loadu_si256((const __m256i *)(tempsrc + 2 * stride)); 101 const __m256i v_val_3_w = 102 _mm256_loadu_si256((const __m256i *)(tempsrc + 3 * stride)); 103 104 const __m256i v_sum_01 = _mm256_add_epi16(v_val_0_w, v_val_1_w); 105 const __m256i v_sum_23 = _mm256_add_epi16(v_val_2_w, v_val_3_w); 106 __m256i v_sum_0123 = _mm256_add_epi16(v_sum_01, v_sum_23); 107 v_sum_0123 = _mm256_madd_epi16(v_sum_0123, one_reg); 108 v_sum_total = _mm256_add_epi32(v_sum_total, v_sum_0123); 109 110 const __m256i v_sq_0_d = _mm256_madd_epi16(v_val_0_w, v_val_0_w); 111 const __m256i v_sq_1_d = _mm256_madd_epi16(v_val_1_w, v_val_1_w); 112 const __m256i v_sq_2_d = _mm256_madd_epi16(v_val_2_w, v_val_2_w); 113 const __m256i v_sq_3_d = _mm256_madd_epi16(v_val_3_w, v_val_3_w); 114 const __m256i v_sq_01_d = _mm256_add_epi32(v_sq_0_d, v_sq_1_d); 115 const __m256i v_sq_23_d = _mm256_add_epi32(v_sq_2_d, v_sq_3_d); 116 const __m256i v_sq_0123_d = _mm256_add_epi32(v_sq_01_d, v_sq_23_d); 117 v_sse_row = _mm256_add_epi32(v_sse_row, v_sq_0123_d); 118 } 119 const __m256i v_sse_row_low = _mm256_unpacklo_epi32(v_sse_row, zero_reg); 120 const __m256i v_sse_row_hi = _mm256_unpackhi_epi32(v_sse_row, zero_reg); 121 v_sse_row = _mm256_add_epi64(v_sse_row_low, v_sse_row_hi); 122 v_sse_total = _mm256_add_epi64(v_sse_total, v_sse_row); 123 src += 4 * stride; 124 } 125 126 const __m128i v_sum_total_low = _mm256_castsi256_si128(v_sum_total); 127 const __m128i v_sum_total_hi = _mm256_extracti128_si256(v_sum_total, 1); 128 __m128i sum_128bit = _mm_add_epi32(v_sum_total_hi, v_sum_total_low); 129 sum_128bit = _mm_add_epi32(sum_128bit, _mm_srli_si128(sum_128bit, 8)); 130 sum_128bit = _mm_add_epi32(sum_128bit, _mm_srli_si128(sum_128bit, 4)); 131 *sum += _mm_cvtsi128_si32(sum_128bit); 132 133 __m128i v_sse_total_lo = _mm256_castsi256_si128(v_sse_total); 134 __m128i v_sse_total_hi = _mm256_extracti128_si256(v_sse_total, 1); 135 __m128i sse_128bit = _mm_add_epi64(v_sse_total_lo, v_sse_total_hi); 136 137 sse_128bit = 138 _mm_add_epi64(sse_128bit, _mm_unpackhi_epi64(sse_128bit, sse_128bit)); 139 140 xx_storel_64(&result, sse_128bit); 141 142 return result; 143 } 144 145 uint64_t aom_sum_sse_2d_i16_avx2(const int16_t *src, int src_stride, int width, 146 int height, int *sum) { 147 if (LIKELY(width == 4 && height == 4)) { 148 return aom_sum_sse_2d_i16_4x4_sse2(src, src_stride, sum); 149 } else if (LIKELY(width == 4 && (height & 3) == 0)) { 150 return aom_sum_sse_2d_i16_4xn_sse2(src, src_stride, height, sum); 151 } else if (LIKELY(width == 8 && (height & 3) == 0)) { 152 return aom_sum_sse_2d_i16_nxn_sse2(src, src_stride, width, height, sum); 153 } else if (LIKELY(((width & 15) == 0) && ((height & 3) == 0))) { 154 return aom_sum_sse_2d_i16_nxn_avx2(src, src_stride, width, height, sum); 155 } else { 156 return aom_sum_sse_2d_i16_c(src, src_stride, width, height, sum); 157 } 158 } 159 160 // Accumulate sum of 16-bit elements in the vector 161 static inline int32_t mm256_accumulate_epi16(__m256i vec_a) { 162 __m128i vtmp1 = _mm256_extracti128_si256(vec_a, 1); 163 __m128i vtmp2 = _mm256_castsi256_si128(vec_a); 164 vtmp1 = _mm_add_epi16(vtmp1, vtmp2); 165 vtmp2 = _mm_srli_si128(vtmp1, 8); 166 vtmp1 = _mm_add_epi16(vtmp1, vtmp2); 167 vtmp2 = _mm_srli_si128(vtmp1, 4); 168 vtmp1 = _mm_add_epi16(vtmp1, vtmp2); 169 vtmp2 = _mm_srli_si128(vtmp1, 2); 170 vtmp1 = _mm_add_epi16(vtmp1, vtmp2); 171 return _mm_extract_epi16(vtmp1, 0); 172 } 173 174 // Accumulate sum of 32-bit elements in the vector 175 static inline int32_t mm256_accumulate_epi32(__m256i vec_a) { 176 __m128i vtmp1 = _mm256_extracti128_si256(vec_a, 1); 177 __m128i vtmp2 = _mm256_castsi256_si128(vec_a); 178 vtmp1 = _mm_add_epi32(vtmp1, vtmp2); 179 vtmp2 = _mm_srli_si128(vtmp1, 8); 180 vtmp1 = _mm_add_epi32(vtmp1, vtmp2); 181 vtmp2 = _mm_srli_si128(vtmp1, 4); 182 vtmp1 = _mm_add_epi32(vtmp1, vtmp2); 183 return _mm_cvtsi128_si32(vtmp1); 184 } 185 186 uint64_t aom_var_2d_u8_avx2(uint8_t *src, int src_stride, int width, 187 int height) { 188 uint8_t *srcp; 189 uint64_t s = 0, ss = 0; 190 __m256i vzero = _mm256_setzero_si256(); 191 __m256i v_acc_sum = vzero; 192 __m256i v_acc_sqs = vzero; 193 int i, j; 194 195 // Process 32 elements in a row 196 for (i = 0; i < width - 31; i += 32) { 197 srcp = src + i; 198 // Process 8 columns at a time 199 for (j = 0; j < height - 7; j += 8) { 200 __m256i vsrc[8]; 201 for (int k = 0; k < 8; k++) { 202 vsrc[k] = _mm256_loadu_si256((__m256i *)srcp); 203 srcp += src_stride; 204 } 205 for (int k = 0; k < 8; k++) { 206 __m256i vsrc0 = _mm256_unpacklo_epi8(vsrc[k], vzero); 207 __m256i vsrc1 = _mm256_unpackhi_epi8(vsrc[k], vzero); 208 v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc0); 209 v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc1); 210 211 __m256i vsqs0 = _mm256_madd_epi16(vsrc0, vsrc0); 212 __m256i vsqs1 = _mm256_madd_epi16(vsrc1, vsrc1); 213 v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0); 214 v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs1); 215 } 216 217 // Update total sum and clear the vectors 218 s += mm256_accumulate_epi16(v_acc_sum); 219 ss += mm256_accumulate_epi32(v_acc_sqs); 220 v_acc_sum = vzero; 221 v_acc_sqs = vzero; 222 } 223 224 // Process remaining rows (height not a multiple of 8) 225 for (; j < height; j++) { 226 __m256i vsrc = _mm256_loadu_si256((__m256i *)srcp); 227 __m256i vsrc0 = _mm256_unpacklo_epi8(vsrc, vzero); 228 __m256i vsrc1 = _mm256_unpackhi_epi8(vsrc, vzero); 229 v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc0); 230 v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc1); 231 232 __m256i vsqs0 = _mm256_madd_epi16(vsrc0, vsrc0); 233 __m256i vsqs1 = _mm256_madd_epi16(vsrc1, vsrc1); 234 v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0); 235 v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs1); 236 237 srcp += src_stride; 238 } 239 240 // Update total sum and clear the vectors 241 s += mm256_accumulate_epi16(v_acc_sum); 242 ss += mm256_accumulate_epi32(v_acc_sqs); 243 v_acc_sum = vzero; 244 v_acc_sqs = vzero; 245 } 246 247 // Process the remaining area using C 248 srcp = src; 249 for (int k = 0; k < height; k++) { 250 for (int m = i; m < width; m++) { 251 uint8_t val = srcp[m]; 252 s += val; 253 ss += val * val; 254 } 255 srcp += src_stride; 256 } 257 return (ss - s * s / (width * height)); 258 } 259 260 #if CONFIG_AV1_HIGHBITDEPTH 261 uint64_t aom_var_2d_u16_avx2(uint8_t *src, int src_stride, int width, 262 int height) { 263 uint16_t *srcp1 = CONVERT_TO_SHORTPTR(src), *srcp; 264 uint64_t s = 0, ss = 0; 265 __m256i vzero = _mm256_setzero_si256(); 266 __m256i v_acc_sum = vzero; 267 __m256i v_acc_sqs = vzero; 268 int i, j; 269 270 // Process 16 elements in a row 271 for (i = 0; i < width - 15; i += 16) { 272 srcp = srcp1 + i; 273 // Process 8 columns at a time 274 for (j = 0; j < height - 8; j += 8) { 275 __m256i vsrc[8]; 276 for (int k = 0; k < 8; k++) { 277 vsrc[k] = _mm256_loadu_si256((__m256i *)srcp); 278 srcp += src_stride; 279 } 280 for (int k = 0; k < 8; k++) { 281 __m256i vsrc0 = _mm256_unpacklo_epi16(vsrc[k], vzero); 282 __m256i vsrc1 = _mm256_unpackhi_epi16(vsrc[k], vzero); 283 v_acc_sum = _mm256_add_epi32(vsrc0, v_acc_sum); 284 v_acc_sum = _mm256_add_epi32(vsrc1, v_acc_sum); 285 286 __m256i vsqs0 = _mm256_madd_epi16(vsrc[k], vsrc[k]); 287 v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0); 288 } 289 290 // Update total sum and clear the vectors 291 s += mm256_accumulate_epi32(v_acc_sum); 292 ss += mm256_accumulate_epi32(v_acc_sqs); 293 v_acc_sum = vzero; 294 v_acc_sqs = vzero; 295 } 296 297 // Process remaining rows (height not a multiple of 8) 298 for (; j < height; j++) { 299 __m256i vsrc = _mm256_loadu_si256((__m256i *)srcp); 300 __m256i vsrc0 = _mm256_unpacklo_epi16(vsrc, vzero); 301 __m256i vsrc1 = _mm256_unpackhi_epi16(vsrc, vzero); 302 v_acc_sum = _mm256_add_epi32(vsrc0, v_acc_sum); 303 v_acc_sum = _mm256_add_epi32(vsrc1, v_acc_sum); 304 305 __m256i vsqs0 = _mm256_madd_epi16(vsrc, vsrc); 306 v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0); 307 srcp += src_stride; 308 } 309 310 // Update total sum and clear the vectors 311 s += mm256_accumulate_epi32(v_acc_sum); 312 ss += mm256_accumulate_epi32(v_acc_sqs); 313 v_acc_sum = vzero; 314 v_acc_sqs = vzero; 315 } 316 317 // Process the remaining area using C 318 srcp = srcp1; 319 for (int k = 0; k < height; k++) { 320 for (int m = i; m < width; m++) { 321 uint16_t val = srcp[m]; 322 s += val; 323 ss += val * val; 324 } 325 srcp += src_stride; 326 } 327 return (ss - s * s / (width * height)); 328 } 329 #endif // CONFIG_AV1_HIGHBITDEPTH