Line data    Source code

       1             : /*
       2             : * Copyright(c) 2019 Intel Corporation
       3             : * SPDX - License - Identifier: BSD - 2 - Clause - Patent
       4             : */
       5             : 
       6             : #include "stdint.h"
       7             : #include "emmintrin.h"
       8             : #include "EbComputeSAD_SSE2.h"
       9             : 
      10           0 : uint32_t combined_averaging_4xm_sad_sse2_intrin(
      11             :     uint8_t  *src,
      12             :     uint32_t  src_stride,
      13             :     uint8_t  *ref1,
      14             :     uint32_t  ref1_stride,
      15             :     uint8_t  *ref2,
      16             :     uint32_t  ref2_stride,
      17             :     uint32_t  height,
      18             :     uint32_t  width)
      19             : {
      20             :     __m128i sad0, sad1;
      21             :     uint32_t y;
      22             :     (void)width;
      23           0 :     sad0 = sad1 = _mm_setzero_si128();
      24             : 
      25           0 :     for (y = 0; y < height; y += 2) {
      26           0 :         sad0 = _mm_add_epi32(sad0, _mm_sad_epu8(_mm_cvtsi32_si128(*(uint32_t *)src), _mm_avg_epu8(_mm_cvtsi32_si128(*(uint32_t *)ref1), _mm_cvtsi32_si128(*(uint32_t *)ref2))));
      27             : 
      28           0 :         sad1 = _mm_add_epi32(sad1, _mm_sad_epu8(_mm_cvtsi32_si128(*(uint32_t *)(src + src_stride)), _mm_avg_epu8(_mm_cvtsi32_si128(*(uint32_t *)(ref1 + ref1_stride)), _mm_cvtsi32_si128(*(uint32_t *)(ref2 + ref2_stride)))));
      29           0 :         src += src_stride << 1;
      30           0 :         ref1 += ref1_stride << 1;
      31           0 :         ref2 += ref2_stride << 1;
      32             :     }
      33           0 :     return _mm_cvtsi128_si32(_mm_add_epi32(sad0, sad1));
      34             : }