Add optimized routines for pairwise long adds and _mm_mullo_epi32

[easyaspi314]

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);
}

[Zvictoria]

Hi, easyaspi314. Thanks for your input!!!! it helps indeed. The only place that have to be changed is the _mm_set1_epi64x function - unfortunately it is not available say in VS compiler (for 32 bit version). So if you change it to another "set" - _epi32 with the corresponding arguments, I will merge your commit with a great pleasure. Thanks in advance!

[easyaspi314]

I also added a version that uses _mm_blend_epi16 for SSE4.1 which only requires a pxor instead of movqda.

[Zvictoria]

The final thanks for this useful commit!