1static inline __m256i
2dec_reshuffle (const __m256i in)
3{
4 // in, lower lane, bits, upper case are most significant bits, lower
5 // case are least significant bits:
6 // 00llllll 00kkkkLL 00jjKKKK 00JJJJJJ
7 // 00iiiiii 00hhhhII 00ggHHHH 00GGGGGG
8 // 00ffffff 00eeeeFF 00ddEEEE 00DDDDDD
9 // 00cccccc 00bbbbCC 00aaBBBB 00AAAAAA
10
11 const __m256i merge_ab_and_bc = _mm256_maddubs_epi16(in, _mm256_set1_epi32(0x01400140));
12 // 0000kkkk LLllllll 0000JJJJ JJjjKKKK
13 // 0000hhhh IIiiiiii 0000GGGG GGggHHHH
14 // 0000eeee FFffffff 0000DDDD DDddEEEE
15 // 0000bbbb CCcccccc 0000AAAA AAaaBBBB
16
17 __m256i out = _mm256_madd_epi16(merge_ab_and_bc, _mm256_set1_epi32(0x00011000));
18 // 00000000 JJJJJJjj KKKKkkkk LLllllll
19 // 00000000 GGGGGGgg HHHHhhhh IIiiiiii
20 // 00000000 DDDDDDdd EEEEeeee FFffffff
21 // 00000000 AAAAAAaa BBBBbbbb CCcccccc
22
23 // Pack bytes together in each lane:
24 out = _mm256_shuffle_epi8(out, _mm256_setr_epi8(
25 2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1,
26 2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1));
27 // 00000000 00000000 00000000 00000000
28 // LLllllll KKKKkkkk JJJJJJjj IIiiiiii
29 // HHHHhhhh GGGGGGgg FFffffff EEEEeeee
30 // DDDDDDdd CCcccccc BBBBbbbb AAAAAAaa
31
32 // Pack lanes:
33 return _mm256_permutevar8x32_epi32(out, _mm256_setr_epi32(0, 1, 2, 4, 5, 6, -1, -1));
34}
35