commit 6516a619fa09f6d70712328e4099b4b456577ad4
author easyaspi314 (Devin) <easyaspi314@users.noreply.github.com> Wed Feb 06 06:37:16 2019 -0500
committer Victoria <victoria.zhislina@intel.com> Wed Feb 06 14:37:16 2019 +0300
tree c03283b3425c6330bfdf8f5ffdb5a0a522f2f3e9
parent 1a99e5a4870db6fcdb12bd1f642fdab07894a402

Add optimized routines for pairwise long adds and _mm_mullo_epi32 (#25)

* Add optimized routines for pairwise long adds and _mm_mullo_epi32

vpaddlq_uN can be implemented as so:
{
    const __m128i ff = _mm_set1_epi2N((1 << N) - 1);
    __m128i low = _mm_and_si128(a, ff);
    __m128i high = _mm_srli_epi2N(a, N);
    return _mm_add_epi2N(low, high);
|

and the other unsigned pairwise adds are the same.

vpaddlq_s32 can be implemented like so:
{
    __m128i top, bot;
    bot = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
    bot = _MM_CVTEPI32_EPI64(bot);
    top = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 3, 1));
    top = _MM_CVTEPI32_EPI64(top);
    return _mm_add_epi64(top, bot);
}

And _mm_mullo_epi32 uses the same routine that GCC uses with vector
extensions (Clang uses a similar method, but it uses pshufd which is
slow on pre-Penryn chips):

{
    __m128i a_high = _mm_srli_epi64(a, 32);
    __m128i low = _mm_mul_epu32(a, b);
    __m128i b_high = _mm_srli_epi64(b, 32);
    __m128i high = _mm_mul_epu32(a_high, b_high);
    low = _mm_shuffle_epi32(low, _MM_SHUFFLE(0, 0, 2, 0));
    high = _mm_shuffle_epi32(high, _MM_SHUFFLE(0, 0, 2, 0));
    return _mm_unpacklo_epi32(low, high);
}