mm512 exp KNC implementation

exp512_ps.h

#include <immintrin.h>

static const __m512d c256 = _mm512_set1_pd(1.0 / 256.0);
static const __m512d c0 = _mm512_set1_pd(1);
static const __m512d c1 = _mm512_set1_pd(1.00000000006177459);
static const __m512d c2 = _mm512_set1_pd(0.49999988007542528);
static const __m512d c3 = _mm512_set1_pd(0.16666663108604157);
static const __m512d c4 = _mm512_set1_pd(0.041694294620381676);
static const __m512d c5 = _mm512_set1_pd(0.0083383426505236529);

__m512 exp512_ps(__m512 x)
{
	__m512d t0 = _mm512_cvtpslo_pd(x);
	__m512d t1 = _mm512_cvtpslo_pd(_mm512_permute4f128_ps(x, _MM_PERM_BADC));
	__m512d y0, y1;

	t0 = _mm512_mul_pd(t0, c256);
	t1 = _mm512_mul_pd(t1, c256);

	y0 = _mm512_fmadd_pd(c5, t0, c4);
	y1 = _mm512_fmadd_pd(c5, t1, c4);

	y0 = _mm512_fmadd_pd(y0, t0, c3);
	y1 = _mm512_fmadd_pd(y1, t1, c3);

	y0 = _mm512_fmadd_pd(y0, t0, c2);
	y1 = _mm512_fmadd_pd(y1, t1, c2);

	y0 = _mm512_fmadd_pd(y0, t0, c1);
	y1 = _mm512_fmadd_pd(y1, t1, c1);

	y0 = _mm512_fmadd_pd(y0, t0, c0);
	y1 = _mm512_fmadd_pd(y1, t1, c0);

	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);
	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);
	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);
	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);
	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);
	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);
	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);
	y0 = _mm512_mul_pd(y0, y0); y1 = _mm512_mul_pd(y1, y1);

	__m512 out = _mm512_cvtpd_pslo(y0);
	__m512 out2 = _mm512_cvtpd_pslo(y1);
	out = _mm512_mask_permute4f128_ps(out, 0xff00, out2, _MM_PERM_BADC);

	return m512f(out);
}