How to achieve 8bit madd using SSE2

Question

Reading from the official Intel C++ Intrinsic Reference,

SSE 2 has the following command

__m128i _mm_madd_epi16(__m128i a, __m128i b)

Multiplies the 8 signed 16-bit integers from a by the 8 signed 16-bit integers from b. Adds the signed 32-bit integer results pairwise and packs the 4 signed 32-bit integer results.

while SSE 3 has

__m128i _mm_maddubs_epi16 (__m128i a, __m128i b)

Multiply signed and unsigned bytes, add horizontal pair of signed words, pack saturated signed words.

Since Im working with 8bit pixels and I must only use SSE 2(old architecture is the target) I need an 8bit madd instruction. How would I proceed with that?

Answer 1

Hope this works - I don't have a compiler here. But even if I missed something, you should get the overall idea.

EDIT: Thanks to @Peter Cordes for pointing out that _mm_setzero_si128 should better be used directly.

inline __m128i _mm_madd_epi8_SSE2(const __m128i & a, const __m128i & b)
{
    // a = 0x00 0x01 0xFE 0x04 ...
    // b = 0x00 0x02 0x80 0x84 ...

    // To extend signed 8-bit value, MSB has to be set to 0xFF
    __m128i sign_mask_a  = _mm_cmplt_epi8(a, _mm_setzero_si128());
    __m128i sign_mask_b  = _mm_cmplt_epi8(b, _mm_setzero_si128());

    // sign_mask_a = 0x00 0x00 0xFF 0x00 ...
    // sign_mask_b = 0x00 0x00 0xFF 0xFF ...

    // Unpack positives with 0x00, negatives with 0xFF
    __m128i a_epi16_l    = _mm_unpacklo_epi8(a, sign_mask_a);
    __m128i a_epi16_h    = _mm_unpackhi_epi8(a, sign_mask_a);
    __m128i b_epi16_l    = _mm_unpacklo_epi8(b, sign_mask_b);
    __m128i b_epi16_h    = _mm_unpackhi_epi8(b, sign_mask_b);

    // Here - valid 16-bit signed integers corresponding to the 8-bit input
    // a_epi16_l = 0x00 0x00 0x01 0x00 0xFE 0xFF 0x04 0x00 ... 

    // Get the a[i] * b[i] + a[i+1] * b[i+1] for both low and high parts
    __m128i madd_epi32_l = _mm_madd_epi16(a_epi16_l, b_epi16_l);
    __m128i madd_epi32_h = _mm_madd_epi16(a_epi16_h, b_epi16_h);

    // Now go back from 32-bit values to 16-bit values & signed saturate
    return _mm_packs_epi16(madd_epi32_l, madd_epi32_h);
}