Search code examples
csimdintrinsicsavxavx2

How can I extract a byte from __m256i AVX2 register into another __m256i register?

Given a __m256i register and an index i, I want to extract a single byte from each value stored in the register and save it in antoher __m256i register. Also after performing some calculation on the second register, I want to load the byte back to the first register without touching the other bytes.

Example:
index i = 2
__m256i a:
           3210
|AAAA|AAAA|AAAA|AAAA|AAAA|AAAA|AAAA|AAAA|

__m256i b:
|FAFF|FAFF|FAFF|FAFF|FAFF|FAFF|FAFF|FAFF|

// some calculation

__m256i a:
|A6AA|A6AA|A6AA|A6AA|A6AA|A6AA|A6AA|A6AA|

I am sorry, if this question was asked before, but since I am new to this topic it is quite hard to find answers for this topic. Thank you!

Solution

  • I try to generalize answers above:

    const int index = 2; // byte index
__m256i mask = _mm256_set1_epi32(0xFF << index*8); // bit mask |0F00|0F00|...|0F00|

__m256i a; // source vector |AAAA|AAAA|...|AAAA|
__m256i b = _mm256_blendv_epi8(_mm256_set1_epi8(-1), a, mask);// extract byte |FAFF|FAFF|...|FAFF|

b;  // after some manipulations |BBBB|BBBB|...|BBBB|
a = _mm256_blendv_epi8(a, b, mask); // store byte |ABAA|ABAA|...|ABAA|