| | 100 | __m128i temp = _mm_setzero_si128(); |
| | 101 | for( j = 0; j < 8; j++ ) { |
| | 102 | __m128i _top_left = unpack(top_left, j); |
| | 103 | __m128i _top = unpack( top, j ); |
| | 104 | __m128i _top_right = unpack( top_right, j ); |
| | 105 | __m128i _left = unpack( left, j ); |
| | 106 | __m128i _right = unpack ( right, j ); |
| | 107 | __m128i _bot = unpack( bot, j ); |
| | 108 | __m128i _bot_left = unpack( bot_right, j ); |
| | 109 | __m128i _bot_right = unpack( bot_right, j ); |
| 104 | | for( j = 0; j < 8; j++ ) { |
| 105 | | __m128i sum1 = _mm_add_epi8( unpack( top_left, j ), unpack( top, j ) ); |
| 106 | | __m128i sum2 = _mm_add_epi8( unpack( top_right, j ), unpack( right, j ) ); |
| 107 | | __m128i sum3 = _mm_add_epi8( unpack( left, j ), unpack( bot_left, j ) ); |
| 108 | | __m128i sum4 = _mm_add_epi8( unpack( bot, j ), unpack( bot_right, j ) ); |
| | 111 | __m128i sum1 = _mm_add_epi8( _top_left, _top ); |
| | 112 | __m128i sum2 = _mm_add_epi8( _top_right, _right ); |
| | 113 | __m128i sum3 = _mm_add_epi8( _left, _bot_left ); |
| | 114 | __m128i sum4 = _mm_add_epi8( _bot, _bot_right ); |
| | 115 | |
| | 116 | sum3 = _mm_add_epi8( sum3, sum4 ); |