v8sf exp256_ps(v8sf x) {
v8sf tmp = _mm256_setzero_ps(), fx;
v8si imm0;
v8sf one = *(v8sf*)_ps256_1;
x = _mm256_min_ps(x, *(v8sf*)_ps256_exp_hi);
x = _mm256_max_ps(x, *(v8sf*)_ps256_exp_lo);
/* express exp(x) as exp(g + n*log(2)) */
fx = _mm256_mul_ps(x, *(v8sf*)_ps256_cephes_LOG2EF);
fx = _mm256_add_ps(fx, *(v8sf*)_ps256_0p5);
/* how to perform a floorf with SSE: just below */
//imm0 = _mm256_cvttps_epi32(fx);
//tmp = _mm256_cvtepi32_ps(imm0);
tmp = _mm256_floor_ps(fx);
/* if greater, substract 1 */
//v8sf mask = _mm256_cmpgt_ps(tmp, fx);
v8sf mask = _mm256_cmp_ps(tmp, fx, _CMP_GT_OS);
mask = _mm256_and_ps(mask, one);
fx = _mm256_sub_ps(tmp, mask);
tmp = _mm256_mul_ps(fx, *(v8sf*)_ps256_cephes_exp_C1);
v8sf z = _mm256_mul_ps(fx, *(v8sf*)_ps256_cephes_exp_C2);
x = _mm256_sub_ps(x, tmp);
x = _mm256_sub_ps(x, z);
z = _mm256_mul_ps(x,x);
v8sf y = *(v8sf*)_ps256_cephes_exp_p0;
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p1);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p2);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p3);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p4);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_exp_p5);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, x);
y = _mm256_add_ps(y, one);
/* build 2^n */
imm0 = _mm256_cvttps_epi32(fx);
// another two AVX2 instructions
imm0 = _mm256_add_epi32(imm0, *(v8si*)_pi32_256_0x7f);
imm0 = _mm256_slli_epi32(imm0, 23);
v8sf pow2n = _mm256_castsi256_ps(imm0);
y = _mm256_mul_ps(y, pow2n);
return y;
}
void extern
avx512vl_test (void)
{
x1 = _mm256_roundscale_ps (x1, 0x42);
x1 = _mm256_ceil_ps (x1);
x1 = _mm256_floor_ps (x1);
x1 = _mm256_mask_roundscale_ps (x1, 2, x1, 0x42);
x1 = _mm256_maskz_roundscale_ps (2, x1, 0x42);
x2 = _mm_roundscale_ps (x2, 0x42);
x2 = _mm_ceil_ps (x2);
x2 = _mm_floor_ps (x2);
x2 = _mm_mask_roundscale_ps (x2, 2, x2, 0x42);
x2 = _mm_maskz_roundscale_ps (2, x2, 0x42);
}
inline __m256i avx2_positive_hexid_to_ringid_root(const __m256i hexid)
{
// The following algorithm works until hexid=12,589,056
// const unsigned iarg = 1+4*(hexid-1)/3;
// return (unsigned(std::sqrt(float(iarg)))+1)/2;
__m256 arg = _mm256_cvtepi32_ps(hexid);
arg = _mm256_fmsub_ps(arg, calin::math::simd::c_m256(_c_m256_four_thirds),
calin::math::simd::c_m256(_c_m256_one_third));
arg = _mm256_sqrt_ps(arg);
arg = _mm256_fmadd_ps(arg, calin::math::simd::c_m256(_c_m256_one_half),
calin::math::simd::c_m256(_c_m256_one_half));
arg = _mm256_floor_ps(arg);
return _mm256_cvtps_epi32(arg);
}
