Question

I have a __m128i register (Vector A) with 16 bit values with the content:

{100,26,26,26,26,26,26,100} // A Vector

Now I subtract the vector

{82,82,82,82,82,82,82,82}

With the instruction

_mm_sub_epi16(a_vec,_mm_set1_epi16(82)) 

The expected result should be the following vector

{18,-56,-56,-56,-56,-56,-56,18}

But I get

{18,65480,65480,65480,65480,65480,65480,18}

How can I solve that the vector is treated as signed?

The A Vector was created by this instruction:

__m128i a_vec = _mm_srli_epi16(_mm_unpacklo_epi8(score_vec_8bit, score_vec_8bit), 8)

Answer 1

65480 is the same value as -56 (they are both 0xffc8 at the register level) – you’re just displaying it as if it were an unsigned short.

Note that for non-saturating addition and subtraction of binary values without carry/borrow flags it really is irrelevant whether the values are signed or unsigned – hence the same instruction can be used for adding both signed and unsigned shorts – the only difference is how you subsequently interpret (or display) the result.