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 13void vp9_get16x16var_avx2(const unsigned char *src_ptr, 14 int source_stride, 15 const unsigned char *ref_ptr, 16 int recon_stride, 17 unsigned int *SSE, 18 int *Sum) { 19 __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low; 20 __m256i ref_expand_high, madd_low, madd_high; 21 unsigned int i, src_2strides, ref_2strides; 22 __m256i zero_reg = _mm256_set1_epi16(0); 23 __m256i sum_ref_src = _mm256_set1_epi16(0); 24 __m256i madd_ref_src = _mm256_set1_epi16(0); 25 26 // processing two strides in a 256 bit register reducing the number 27 // of loop stride by half (comparing to the sse2 code) 28 src_2strides = source_stride << 1; 29 ref_2strides = recon_stride << 1; 30 for (i = 0; i < 8; i++) { 31 src = _mm256_castsi128_si256( 32 _mm_loadu_si128((__m128i const *) (src_ptr))); 33 src = _mm256_inserti128_si256(src, 34 _mm_loadu_si128((__m128i const *)(src_ptr+source_stride)), 1); 35 36 ref =_mm256_castsi128_si256( 37 _mm_loadu_si128((__m128i const *) (ref_ptr))); 38 ref = _mm256_inserti128_si256(ref, 39 _mm_loadu_si128((__m128i const *)(ref_ptr+recon_stride)), 1); 40 41 // expanding to 16 bit each lane 42 src_expand_low = _mm256_unpacklo_epi8(src, zero_reg); 43 src_expand_high = _mm256_unpackhi_epi8(src, zero_reg); 44 45 ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg); 46 ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg); 47 48 // src-ref 49 src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low); 50 src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high); 51 52 // madd low (src - ref) 53 madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low); 54 55 // add high to low 56 src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high); 57 58 // madd high (src - ref) 59 madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high); 60 61 sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low); 62 63 // add high to low 64 madd_ref_src = _mm256_add_epi32(madd_ref_src, 65 _mm256_add_epi32(madd_low, madd_high)); 66 67 src_ptr+= src_2strides; 68 ref_ptr+= ref_2strides; 69 } 70 71 { 72 __m128i sum_res, madd_res; 73 __m128i expand_sum_low, expand_sum_high, expand_sum; 74 __m128i expand_madd_low, expand_madd_high, expand_madd; 75 __m128i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum; 76 77 // extract the low lane and add it to the high lane 78 sum_res = _mm_add_epi16(_mm256_castsi256_si128(sum_ref_src), 79 _mm256_extractf128_si256(sum_ref_src, 1)); 80 81 madd_res = _mm_add_epi32(_mm256_castsi256_si128(madd_ref_src), 82 _mm256_extractf128_si256(madd_ref_src, 1)); 83 84 // padding each 2 bytes with another 2 zeroed bytes 85 expand_sum_low = _mm_unpacklo_epi16(_mm256_castsi256_si128(zero_reg), 86 sum_res); 87 expand_sum_high = _mm_unpackhi_epi16(_mm256_castsi256_si128(zero_reg), 88 sum_res); 89 90 // shifting the sign 16 bits right 91 expand_sum_low = _mm_srai_epi32(expand_sum_low, 16); 92 expand_sum_high = _mm_srai_epi32(expand_sum_high, 16); 93 94 expand_sum = _mm_add_epi32(expand_sum_low, expand_sum_high); 95 96 // expand each 32 bits of the madd result to 64 bits 97 expand_madd_low = _mm_unpacklo_epi32(madd_res, 98 _mm256_castsi256_si128(zero_reg)); 99 expand_madd_high = _mm_unpackhi_epi32(madd_res, 100 _mm256_castsi256_si128(zero_reg)); 101 102 expand_madd = _mm_add_epi32(expand_madd_low, expand_madd_high); 103 104 ex_expand_sum_low = _mm_unpacklo_epi32(expand_sum, 105 _mm256_castsi256_si128(zero_reg)); 106 ex_expand_sum_high = _mm_unpackhi_epi32(expand_sum, 107 _mm256_castsi256_si128(zero_reg)); 108 109 ex_expand_sum = _mm_add_epi32(ex_expand_sum_low, ex_expand_sum_high); 110 111 // shift 8 bytes eight 112 madd_res = _mm_srli_si128(expand_madd, 8); 113 sum_res = _mm_srli_si128(ex_expand_sum, 8); 114 115 madd_res = _mm_add_epi32(madd_res, expand_madd); 116 sum_res = _mm_add_epi32(sum_res, ex_expand_sum); 117 118 *((int*)SSE)= _mm_cvtsi128_si32(madd_res); 119 120 *((int*)Sum)= _mm_cvtsi128_si32(sum_res); 121 } 122} 123 124void vp9_get32x32var_avx2(const unsigned char *src_ptr, 125 int source_stride, 126 const unsigned char *ref_ptr, 127 int recon_stride, 128 unsigned int *SSE, 129 int *Sum) { 130 __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low; 131 __m256i ref_expand_high, madd_low, madd_high; 132 unsigned int i; 133 __m256i zero_reg = _mm256_set1_epi16(0); 134 __m256i sum_ref_src = _mm256_set1_epi16(0); 135 __m256i madd_ref_src = _mm256_set1_epi16(0); 136 137 // processing 32 elements in parallel 138 for (i = 0; i < 16; i++) { 139 src = _mm256_loadu_si256((__m256i const *) (src_ptr)); 140 141 ref = _mm256_loadu_si256((__m256i const *) (ref_ptr)); 142 143 // expanding to 16 bit each lane 144 src_expand_low = _mm256_unpacklo_epi8(src, zero_reg); 145 src_expand_high = _mm256_unpackhi_epi8(src, zero_reg); 146 147 ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg); 148 ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg); 149 150 // src-ref 151 src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low); 152 src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high); 153 154 // madd low (src - ref) 155 madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low); 156 157 // add high to low 158 src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high); 159 160 // madd high (src - ref) 161 madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high); 162 163 sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low); 164 165 // add high to low 166 madd_ref_src = _mm256_add_epi32(madd_ref_src, 167 _mm256_add_epi32(madd_low, madd_high)); 168 169 src_ptr+= source_stride; 170 ref_ptr+= recon_stride; 171 } 172 173 { 174 __m256i expand_sum_low, expand_sum_high, expand_sum; 175 __m256i expand_madd_low, expand_madd_high, expand_madd; 176 __m256i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum; 177 178 // padding each 2 bytes with another 2 zeroed bytes 179 expand_sum_low = _mm256_unpacklo_epi16(zero_reg, sum_ref_src); 180 expand_sum_high = _mm256_unpackhi_epi16(zero_reg, sum_ref_src); 181 182 // shifting the sign 16 bits right 183 expand_sum_low = _mm256_srai_epi32(expand_sum_low, 16); 184 expand_sum_high = _mm256_srai_epi32(expand_sum_high, 16); 185 186 expand_sum = _mm256_add_epi32(expand_sum_low, expand_sum_high); 187 188 // expand each 32 bits of the madd result to 64 bits 189 expand_madd_low = _mm256_unpacklo_epi32(madd_ref_src, zero_reg); 190 expand_madd_high = _mm256_unpackhi_epi32(madd_ref_src, zero_reg); 191 192 expand_madd = _mm256_add_epi32(expand_madd_low, expand_madd_high); 193 194 ex_expand_sum_low = _mm256_unpacklo_epi32(expand_sum, zero_reg); 195 ex_expand_sum_high = _mm256_unpackhi_epi32(expand_sum, zero_reg); 196 197 ex_expand_sum = _mm256_add_epi32(ex_expand_sum_low, ex_expand_sum_high); 198 199 // shift 8 bytes eight 200 madd_ref_src = _mm256_srli_si256(expand_madd, 8); 201 sum_ref_src = _mm256_srli_si256(ex_expand_sum, 8); 202 203 madd_ref_src = _mm256_add_epi32(madd_ref_src, expand_madd); 204 sum_ref_src = _mm256_add_epi32(sum_ref_src, ex_expand_sum); 205 206 // extract the low lane and the high lane and add the results 207 *((int*)SSE)= _mm_cvtsi128_si32(_mm256_castsi256_si128(madd_ref_src)) + 208 _mm_cvtsi128_si32(_mm256_extractf128_si256(madd_ref_src, 1)); 209 210 *((int*)Sum)= _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_ref_src)) + 211 _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_ref_src, 1)); 212 } 213} 214