/*
* Compute: s = sqrt( t**2 - x**2 - y**2 - z**2 ), with s, t, x, y, z
* member variables of the st_coords structure arr.
*
* Traverse elements randomly
*/
void
comp_s(st_coords arr, int L)
{
for(int j=0; j<L; j+=8) {
int i = (rand() % (L/8)) * 8;
__m256 x = _mm256_load_ps(&arr.x[i]);
__m256 y = _mm256_load_ps(&arr.y[i]);
__m256 z = _mm256_load_ps(&arr.z[i]);
__m256 t = _mm256_load_ps(&arr.t[i]);
#ifdef FMA
register __m256 s0;
s0 = _mm256_mul_ps(x, x);
s0 = _mm256_fmadd_ps(y, y, s0);
s0 = _mm256_fmadd_ps(z, z, s0);
s0 = _mm256_fmsub_ps(t, t, s0);
s0 = _mm256_sqrt_ps(s0);
#else
register __m256 s0, s1;
s1 = _mm256_mul_ps(x, x);
s0 = _mm256_mul_ps(y, y);
s1 = _mm256_add_ps(s0, s1);
s0 = _mm256_mul_ps(z, z);
s1 = _mm256_add_ps(s0, s1);
s0 = _mm256_mul_ps(t, t);
s1 = _mm256_sub_ps(s0, s1);
s0 = _mm256_sqrt_ps(s1);
#endif
_mm256_store_ps(&arr.s[i], s0);
}
return;
}
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);
}
void
check_mm256_fmsub_ps (__m256 __A, __m256 __B, __m256 __C)
{
union256 a, b, c, e;
a.x = __A;
b.x = __B;
c.x = __C;
float d[8];
int i;
e.x = _mm256_fmsub_ps (__A, __B, __C);
for (i = 0; i < 8; i++)
{
d[i] = a.a[i] * b.a[i] - c.a[i];
}
if (check_union256 (e, d))
abort ();
}
inline void avx2_xy_trans_to_uv_f(__m256 x, __m256 y, __m256i& u, __m256i& v,
const float crot, const float srot, const float scale, const float dx = 0, const float dy = 0)
{
// x = (x - dx)/scale;
// y = (y - dy)/scale;
// double xx = x*crot + y*srot;
// y = y*crot - x*srot;
// xy_to_uv(xx,y,u,v);
const __m256 vsrot = _mm256_set1_ps(srot);
const __m256 vcrot = _mm256_set1_ps(crot);
const __m256 vscale = _mm256_set1_ps(1.0f/scale);
x = _mm256_mul_ps(_mm256_sub_ps(x, _mm256_set1_ps(dx)), vscale);
y = _mm256_mul_ps(_mm256_sub_ps(y, _mm256_set1_ps(dy)), vscale);
__m256 yy = _mm256_mul_ps(x, vsrot);
yy = _mm256_fmsub_ps(y, vcrot, yy);
x = _mm256_mul_ps(x, vcrot);
x = _mm256_fmadd_ps(y, vsrot, x);
avx2_xy_to_uv_f(x, yy, u, v);
}
inline void avx2_uv_to_xy_trans_f(const __m256i u, const __m256i v, __m256& x, __m256& y,
const float crot, const float srot, const float scale, const float dx = 0, const float dy = 0)
{
// uv_to_xy(u,v,x,y);
// double xx = x*crot - y*srot;
// y = scale * (y*crot + x*srot) + dy;
// x = scale * xx + dx;
avx2_uv_to_xy_f(u,v,x,y);
const __m256 vsrot = _mm256_set1_ps(srot);
const __m256 vcrot = _mm256_set1_ps(crot);
const __m256 vscale = _mm256_set1_ps(scale);
__m256 xx = _mm256_mul_ps(y, vsrot);
xx = _mm256_fmsub_ps(x, vcrot, xx);
y = _mm256_mul_ps(y, vcrot);
y = _mm256_fmadd_ps(x, vsrot, y);
x = xx;
x = _mm256_fmadd_ps(x, vscale, _mm256_set1_ps(dx));
y = _mm256_fmadd_ps(y, vscale, _mm256_set1_ps(dy));
}
inline void avx2_xy_trans_to_uv_with_remainder_f(
__m256& x_in_dx_out, __m256& y_in_dy_out, __m256i& u, __m256i& v,
const float crot, const float srot, const float scale, const float dx = 0, const float dy = 0)
{
const __m256 vsrot = _mm256_set1_ps(srot);
const __m256 vcrot = _mm256_set1_ps(crot);
__m256 vscale = _mm256_set1_ps(1.0f/scale);
x_in_dx_out = _mm256_mul_ps(_mm256_sub_ps(x_in_dx_out, _mm256_set1_ps(dx)), vscale);
y_in_dy_out = _mm256_mul_ps(_mm256_sub_ps(y_in_dy_out, _mm256_set1_ps(dy)), vscale);
__m256 yy = _mm256_mul_ps(x_in_dx_out, vsrot);
yy = _mm256_fmsub_ps(y_in_dy_out, vcrot, yy);
x_in_dx_out = _mm256_mul_ps(x_in_dx_out, vcrot);
x_in_dx_out = _mm256_fmadd_ps(y_in_dy_out, vsrot, x_in_dx_out);
avx2_xy_to_uv_with_remainder_f(x_in_dx_out, yy, u, v);
vscale = _mm256_set1_ps(scale);
y_in_dy_out = _mm256_mul_ps(yy, vcrot);
y_in_dy_out = _mm256_fmadd_ps(x_in_dx_out, vsrot, y_in_dy_out);
y_in_dy_out = _mm256_mul_ps(y_in_dy_out, vscale);
x_in_dx_out = _mm256_mul_ps(x_in_dx_out, vcrot);
x_in_dx_out = _mm256_fnmadd_ps(yy, vsrot, x_in_dx_out);
x_in_dx_out = _mm256_mul_ps(x_in_dx_out, vscale);
}
__m256 test_mm256_fmsub_ps(__m256 a, __m256 b, __m256 c) {
// CHECK-LABEL: test_mm256_fmsub_ps
// CHECK: [[NEG:%.+]] = fsub <8 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %{{.*}}
// CHECK: @llvm.x86.fma.vfmadd.ps.256(<8 x float> %{{.+}}, <8 x float> %{{.+}}, <8 x float> [[NEG]])
return _mm256_fmsub_ps(a, b, c);
}
__m256 test_mm256_fmsub_ps(__m256 a, __m256 b, __m256 c) {
// CHECK: @llvm.x86.fma.vfmsub.ps.256
return _mm256_fmsub_ps(a, b, c);
}
__m256
check_mm256_fmsub_ps (__m256 a, __m256 b, __m256 c)
{
return _mm256_fmsub_ps (a, b, c);
}