SSE Instructions: Byte+Short

instructions, sse, x86

Solution

You need to unpack each vector of 8 bit values to two vectors of 16 bit values and then add those.

__m128i v = _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
__m128i vl = _mm_unpacklo_epi8(v, _mm_set1_epi8(0)); // vl = { 7, 6, 5, 4, 3, 2, 1, 0 }
__m128i vh = _mm_unpackhi_epi8(v, _mm_set1_epi8(0)); // vh = { 15, 14, 13, 12, 11, 10, 9, 8 }

where `v` is a vector of 16 x 8 bit values and `vl`, `vh` are the two unpacked vectors of 8 x 16 bit values.

Note that I'm assuming that the 8 bit values are unsigned so when unpacking to 16 bits the high byte is set to 0 (i.e. no sign extension).

If you want to sum a lot of these vectors and get a 32 bit result then a useful trick is to use `_mm_madd_epi16` with a multiplier of 1, e.g.

__m128i vsuml = _mm_set1_epi32(0);
__m128i vsumh = _mm_set1_epi32(0);
__m128i vsum;
int sum;

for (int i = 0; i < N; i += 16)
{
    __m128i v = _mm_load_si128(&x[i]);
    __m128i vl = _mm_unpacklo_epi8(v, _mm_set1_epi8(0));
    __m128i vh = _mm_unpackhi_epi8(v, _mm_set1_epi8(0));
    vsuml = _mm_add_epi32(vsuml, _mm_madd_epi16(vl, _mm_set1_epi16(1)));
    vsumh = _mm_add_epi32(vsumh, _mm_madd_epi16(vh, _mm_set1_epi16(1)));
}
// do horizontal sum of 4 partial sums and store in scalar int
vsum = _mm_add_epi32(vsuml, vsumh);
vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
sum = _mm_cvtsi128_si32(vsum);

Problem

I have very long byte arrays that need to be added to a destination array of type `short` (or `int`). Does such SSE instruction exist? Or maybe their set ?

Original source