Optimize extraction of 64 bit value from AVX2 register

Question

I try to extract 64 bit from an __m256i register. Example of my current extraction function:

             byte     31                    16 15                    0
byte_result_vec        000D  000C  000B  000A   000H  000G  000F  000E

_mm256_packs_epi32 ->  0D0C  0B0A  0D0C  0B0A   0H0G  0F0E  0H0G  0F0E

_mm256_packus_epi16 -> DCBA  DCBA  DCBA  DCBA   HGFE  HGFE  HGFE  HGFE
                                         ^^^^                     ^^^^
_mm256_castsi256_si128   -> HGFE  HGFE  HGFE  HGFE

_mm256_extracti128_si256 -> DCBA  DCBA  DCBA  DCBA

_mm_cvtsi128_si32(byte_result_vec1) ->  ABCD

_mm_cvtsi128_si32(byte_result_vec2) ->  EFGH

The following code is shifting 4x8 bites to the register position 0-3 and is than extracting 32 bit.

        byte_result_vec = _mm256_packs_epi32(byte_result_vec, byte_result_vec);
        byte_result_vec = _mm256_packus_epi16(byte_result_vec, byte_result_vec);
        __m128i byte_result_vec1 = _mm256_castsi256_si128(byte_result_vec);
        __m128i byte_result_vec2 = _mm256_extracti128_si256(byte_result_vec,1);
        const int res1 = _mm_cvtsi128_si32(byte_result_vec1);
        const int res2 = _mm_cvtsi128_si32(byte_result_vec2);
        result_array[j]       = res1;
        result_array[j+1]     = res2;

The code is working correct but it is slow. It looks like that copying of res1 and res2 to the result_array takes the most time. Is there a way to optimize it?

Answer 1

Probably this variant will be faster

/* byte_result_vec        000H  000G  000F  000E   000D  000C  000B  000A */
const __m256i shuffle_mask = _mm256_setr_epi8(0,  4,  8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1);
/* abcdefgh               0000  0000  HGFE  0000   0000  0000  0000  DCBA */
const __m256i abcdefgh = _mm256_shuffle_epi8(byte_result_vec, shuffle_mask);
/* abcd                                            0000  0000  0000  DCBA */
const __m128i abcd = _mm256_castsi256_si128(abcdefgh);
/* efgh                                            0000  0000  HGFE  0000 */
const __m128i efgh = _mm256_extracti128_si256(abcdefgh, 1);
_mm_storel_epi64((__m128i*)&result_array[j], _mm_or_si128(abcd, efgh));