用于检查两个双精度值是否足够不同的 simd 代码

Question

假设我有两个双精度值，旧的和新的。我想实现一个矢量化函数如果 abs(x-y) < p，则 returns 旧，否则为新。

这是代码 (test.cpp):

#include <emmintrin.h>
#include <iostream>

#define ARRAY_LENGTH 2

int main(void) {
    // x = old value, y = new value, res = result
    double *x, *y, *res;
    posix_memalign((void **)&x, 16, sizeof(double) * ARRAY_LENGTH);
    posix_memalign((void **)&y, 16, sizeof(double) * ARRAY_LENGTH);
    posix_memalign((void **)&res, 16, sizeof(double) * ARRAY_LENGTH);

    double p = 1e-4; // precision
    __m128d sp = _mm_set1_pd(p);
    x[0] = 1.5; y[0] = 1.50011; // x - old value, y - new value
    x[1] = 2.; y[1] = 2.0000001;

    __m128d sx = _mm_load_pd(x);
    __m128d sy = _mm_load_pd(y);

    // sign mask to compute fabs()
    __m128d sign_mask = _mm_set1_pd(-0.);
    // |x-y|
    __m128d absval = _mm_andnot_pd(sign_mask, _mm_sub_pd(sx, sy) );
    // mask of |x-y| < p
    __m128d mask = _mm_cmplt_pd(absval, sp);
    // sres = |x-y| < p ? x : y;
    __m128d sres = _mm_or_pd(
            _mm_and_pd(mask, sx), _mm_andnot_pd(mask, sy) );
    _mm_store_pd(res, sres);
    std::cerr << "res=" << res[0] << "," << res[1] << std::endl;
    return 0;
}

建造：

g++ -std=c++11 -msse4 test.cpp

我们首先计算 fabs(x-y)，与 p 进行比较，然后使用获得面具

有没有人看到更有效的编码方式？谢谢

Answer 1

有一种方法可以使该算法更快一些，但会降低准确性：

// d = x - y;
__m128d diff = _mm_sub_pd(sx, sy);
// mask of |y - x| < p
__m128d mask = _mm_cmplt_pd(_mm_andnot_pd(sign_mask, diff), sp);
// sres = y + (|y - x| < p) ? (x - y) : 0;
__m128d sres = _mm_add_pd(sy, _mm_and_pd(mask, diff));

另一种方式 - 使用 AVX or/and 单精度。

用于检查两个双精度值是否足够不同的 simd 代码

simd code to check if two double values are sufficiently different

c++

simd

absolute-value