Is there a good way to find the module of two variables using SSE? (without SVML)

Question

I am trying to learn how to use SSE, one of the programs that I required to use modular division, and so I wrote this to do this (sorry this exceeded):

__m128i SSEModDiv(__m128i input, __m128i divisors)
{
    //Error Checking (div by zero)
    /*__m128i zeros = _mm_set1_epi32(0);
    __m128i error = _mm_set1_epi32(-1);
    __m128i zerocheck = _mm_cmpeq_epi32(zeros, divisors);
    if (_mm_extract_epi16(zerocheck, 0) != 0)
        return error;
    if (_mm_extract_epi16(zerocheck, 2) != 0)
        return error;
    if (_mm_extract_epi16(zerocheck, 4) != 0)
        return error;
    if (_mm_extract_epi16(zerocheck, 6) != 0)
        return error;*/

    //Now for the real work
    __m128 inputf = _mm_cvtepi32_ps(input);
    __m128 divisorsf = _mm_cvtepi32_ps(divisors);

    /*__m128 recip = _mm_rcp_ps(divisorsf); //Takes reciprocal
    __m128 divided = _mm_mul_ps(inputf, recip); //multiplies by reciprical values*/
    __m128 divided = _mm_div_ps(inputf, divisorsf);
    __m128i intermediateint = _mm_cvttps_epi32(divided); //makes an integer version truncated
    __m128 intermediate = _mm_cvtepi32_ps(intermediateint);
    __m128 multiplied = _mm_mul_ps(intermediate, divisorsf); //multiplies the intermediate with the divisors
    __m128 mods = _mm_sub_ps(inputf, multiplied); //subtracts to get moduli
    return _mm_cvtps_epi32(mods);
}

The problem is that it is about as fast as the module of each of the four 32-bit integers individually in the release and about 10 times slower in debugging (found by profiling).

Can someone tell me how to make this function faster?

-I cannot use SVML because I use Visual Studio -

Answer 1

input divisors SIMD x86 , . , , SIMD .

, (a + b)% c a b (.. a<c b<c), :

z = a + b
if(z>=c) z-=c;

vectorizing-modular-arithmetic

, , , . :

float fact = 1.0/x;
for(int i=0; i<n; i++) {
    z[i] = fact*y[i];  //z[i] = y[i]/x;
}

, .

y / x ≈ y * (2 n / x) >> n

(aka ) (2 n / x) n. . , , x/7 , GCC MSVC, , , , http://www.hackersdelight.org/magic.htm.

, Agner Fog Vector Class Library . SSE AVX.

, ,

for(int i=0; i<n; i++) {
    z[i] = y[i]%x[i];
}

x[i] , /.