12#ifndef EIGEN_MATHFUNCTIONSIMPL_H
13#define EIGEN_MATHFUNCTIONSIMPL_H
16#include "./InternalHeaderCheck.h"
39template <
typename Packet,
int Steps>
40struct generic_reciprocal_newton_step {
41 static_assert(Steps > 0,
"Steps must be at least 1.");
42 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Packet run(
const Packet& a,
const Packet& approx_a_recip) {
43 using Scalar =
typename unpacket_traits<Packet>::type;
44 const Packet one = pset1<Packet>(Scalar(1));
47 const Packet x = generic_reciprocal_newton_step<Packet, Steps - 1>::run(a, approx_a_recip);
48 const Packet tmp = pnmadd(a, x, one);
50 const Packet refined = pmadd(x, tmp, x);
53 const Packet redo = pcmp_lt_or_nan(pabs(refined), pset1<Packet>((std::numeric_limits<Scalar>::denorm_min)()));
54 return predux_any(redo) ? pselect(redo, pdiv(one, a), refined) : refined;
58template <
typename Packet>
59struct generic_reciprocal_newton_step<Packet, 0> {
60 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Packet run(
const Packet& ,
const Packet& approx_rsqrt) {
82template <
typename Packet,
int Steps>
83struct generic_rsqrt_newton_step {
84 static_assert(Steps > 0,
"Steps must be at least 1.");
85 using Scalar =
typename unpacket_traits<Packet>::type;
86 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Packet run(
const Packet& a,
const Packet& approx_rsqrt) {
87 const Scalar kMinusHalf = Scalar(-1) / Scalar(2);
88 const Packet cst_minus_half = pset1<Packet>(kMinusHalf);
89 const Packet cst_minus_one = pset1<Packet>(Scalar(-1));
91 Packet inv_sqrt = approx_rsqrt;
92 for (
int step = 0; step < Steps; ++step) {
96 Packet r2 = pmul(a, inv_sqrt);
97 Packet half_r = pmul(inv_sqrt, cst_minus_half);
98 Packet h_n = pmadd(r2, inv_sqrt, cst_minus_one);
99 inv_sqrt = pmadd(half_r, h_n, inv_sqrt);
106 return pselect(pisnan(inv_sqrt), approx_rsqrt, inv_sqrt);
110template <
typename Packet>
111struct generic_rsqrt_newton_step<Packet, 0> {
112 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Packet run(
const Packet& ,
const Packet& approx_rsqrt) {
135template <
typename Packet,
int Steps = 1>
136struct generic_sqrt_newton_step {
137 static_assert(Steps > 0,
"Steps must be at least 1.");
139 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE Packet run(
const Packet& a,
const Packet& approx_rsqrt) {
140 using Scalar =
typename unpacket_traits<Packet>::type;
141 const Packet one_point_five = pset1<Packet>(Scalar(1.5));
142 const Packet minus_half = pset1<Packet>(Scalar(-0.5));
144 const Packet inf_mask = pcmp_eq(a, pset1<Packet>(NumTraits<Scalar>::infinity()));
145 const Packet return_a = por(pcmp_eq(a, pzero(a)), inf_mask);
149 Packet rsqrt = pmul(approx_rsqrt, pmadd(pmul(minus_half, approx_rsqrt), pmul(a, approx_rsqrt), one_point_five));
150 for (
int step = 1; step < Steps; ++step) {
151 rsqrt = pmul(rsqrt, pmadd(pmul(minus_half, rsqrt), pmul(a, rsqrt), one_point_five));
156 return pselect(return_a, a, pmul(a, rsqrt));
162template <
typename RealScalar,
typename EnableIf =
void>
163struct positive_real_hypot_impl {
164 EIGEN_DEVICE_FUNC
static constexpr EIGEN_STRONG_INLINE RealScalar run(
const RealScalar& x,
const RealScalar& y) {
166 if ((numext::isinf)(x) || (numext::isinf)(y))
return NumTraits<RealScalar>::infinity();
167 if ((numext::isnan)(x) || (numext::isnan)(y))
return NumTraits<RealScalar>::quiet_NaN();
169 RealScalar p = numext::maxi(x, y);
170 if (numext::is_exactly_zero(p))
return RealScalar(0);
171 RealScalar qp = numext::mini(y, x) / p;
172 return p * numext::sqrt<RealScalar>(RealScalar(1) + qp * qp);
179#if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(EIGEN_SYCL_DEVICE_ONLY) && !defined(EIGEN_PARSED_BY_DOXYGEN)
183struct positive_real_hypot_impl<
185 std::enable_if_t<std::is_same<T, float>::value && std::numeric_limits<float>::is_iec559 &&
186 std::numeric_limits<double>::is_iec559 &&
187 (std::numeric_limits<double>::digits >= 2 * std::numeric_limits<float>::digits + 2) &&
188 (std::numeric_limits<double>::max_exponent > 2 * std::numeric_limits<float>::max_exponent + 1)>> {
189 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE
float run(
const float& x,
const float& y) {
193 if (EIGEN_PREDICT_FALSE((numext::isinf)(x) || (numext::isinf)(y)))
return NumTraits<float>::infinity();
194 const double dx =
static_cast<double>(x);
195 const double dy =
static_cast<double>(y);
196 const double xx = dx * dx;
197 return static_cast<float>(numext::sqrt(xx + dy * dy));
206struct positive_real_hypot_impl<
207 T, std::enable_if_t<std::is_same<T, double>::value && std::numeric_limits<double>::is_iec559>> {
208 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE
double run(
const double& x,
const double& y) {
211 const double p = x > y ? x : y;
212 const double q = y < x ? y : x;
216 const double sum = x + y;
217 if (EIGEN_PREDICT_TRUE(sum >= 4.004166190366202e-146 && sum <= 3.3519519824856493e+153)) {
218 return numext::sqrt(sum_of_squares(p, q, has_fast_fma<double>()));
220 return run_extreme(x, y);
227 EIGEN_DEVICE_FUNC
static EIGEN_DONT_INLINE
double run_extreme(
double x,
double y) {
229 if ((numext::isinf)(x) || (numext::isinf)(y))
return NumTraits<double>::infinity();
230 if ((numext::isnan)(x) || (numext::isnan)(y))
return NumTraits<double>::quiet_NaN();
231 const double p = numext::maxi(x, y);
232 const double q = numext::mini(x, y);
233 if (numext::is_exactly_zero(p))
return 0.0;
236 const bool large = p > 1.0;
237 const double scale = large ? 1.8645851828000517e-155 : 4.149515568880993e+180;
238 const double unscale = large ? 5.363123171977039e+154 : 2.409919865102884e-181;
239 return unscale * numext::sqrt(sum_of_squares(p * scale, q * scale, has_fast_fma<double>()));
241 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE
double sum_of_squares(
const double& p,
const double& q, std::true_type) {
242 const double qq = q * q;
243 return numext::fma(p, p, qq);
245 EIGEN_DEVICE_FUNC
static EIGEN_STRONG_INLINE
double sum_of_squares(
const double& p,
const double& q,
247 const double pp = p * p;
248 const double qq = q * q;
254template <
typename RealScalar>
255EIGEN_DEVICE_FUNC
constexpr EIGEN_STRONG_INLINE RealScalar positive_real_hypot(
const RealScalar& x,
256 const RealScalar& y) {
257 return positive_real_hypot_impl<RealScalar>::run(x, y);
260template <
typename Scalar>
262 using RealScalar =
typename NumTraits<Scalar>::Real;
263 static EIGEN_DEVICE_FUNC
inline RealScalar run(
const Scalar& x,
const Scalar& y) {
264 return positive_real_hypot<RealScalar>(numext::abs(x), numext::abs(y));
268template <
typename ComplexT,
bool Reciprocal>
269EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE ComplexT complex_sqrt_extreme(
const ComplexT& z) {
270 using T =
typename NumTraits<ComplexT>::Real;
271 const T x = numext::real(z);
272 const T y = numext::imag(z);
274 const T inf = NumTraits<T>::infinity();
275 EIGEN_IF_CONSTEXPR (Reciprocal) {
276 if ((numext::isinf)(x) || (numext::isinf)(y))
return ComplexT(zero, numext::copysign<T>(zero, -y));
278 if ((numext::isinf)(y))
return ComplexT(inf, y);
279 if ((numext::isinf)(x)) {
280 const T other = (numext::isnan)(y) ? y : zero;
281 return x > zero ? ComplexT(inf, numext::copysign(other, y))
282 : ComplexT(numext::abs(other), numext::copysign(inf, y));
285 if ((numext::isnan)(x) || (numext::isnan)(y)) {
286 return ComplexT(NumTraits<T>::quiet_NaN(), NumTraits<T>::quiet_NaN());
288 if (numext::is_exactly_zero(x) && numext::is_exactly_zero(y)) {
289 return Reciprocal ? ComplexT(inf, NumTraits<T>::quiet_NaN()) : ComplexT(zero, y);
292 T ax = numext::abs(x);
293 T ay = numext::abs(y);
294 const T p = numext::maxi(ax, ay);
295 const T r = numext::mini(ax, ay) / p;
296 const T h = numext::sqrt(T(1) + r * r);
300 const T scale = numext::sqrt(p);
301 const T sum = ax + h;
302 const T w = numext::sqrt(T(0.5) * sum);
305 EIGEN_IF_CONSTEXPR (Reciprocal) {
306 major = (w / h) / scale;
307 minor = (ay / sum) * major;
310 minor = numext::abs(y) / (T(2) * major);
312 if (numext::is_exactly_zero(x)) minor = major;
313 const T imag_sign = Reciprocal ? -y : y;
314 return x < zero ? ComplexT(minor, numext::copysign(major, imag_sign))
315 : ComplexT(major, numext::copysign(minor, imag_sign));
318template <
typename ComplexT,
bool Reciprocal>
319EIGEN_DEVICE_FUNC
constexpr ComplexT complex_sqrt_impl(
const ComplexT& z) {
320 using T =
typename NumTraits<ComplexT>::Real;
321 const T x = numext::real(z);
322 const T y = numext::imag(z);
323 const T ax = numext::abs(x);
324 const T ay = numext::abs(y);
325 const bool real_larger = ax > ay;
326 const T p = real_larger ? ax : ay;
327 const T q = real_larger ? ay : ax;
330 if (EIGEN_PREDICT_FALSE(!(p > T(2) * (numext::numeric_limits<T>::min)() && p <= NumTraits<T>::highest() / T(4)))) {
331 return complex_sqrt_extreme<ComplexT, Reciprocal>(z);
334 const T abs_z = p * numext::sqrt(T(1) + r * r);
335 const T sum = ax + abs_z;
336 const T w = numext::sqrt(T(0.5) * sum);
339 EIGEN_IF_CONSTEXPR (Reciprocal) {
342 minor = (ay / sum) * major;
344 minor = ay / (T(2) * w);
346 if (numext::is_exactly_zero(x)) minor = major;
347 const T imag_sign = Reciprocal ? -y : y;
348 return x < T(0) ? ComplexT(minor, numext::copysign(major, imag_sign))
349 : ComplexT(major, numext::copysign(minor, imag_sign));
353template <
typename ComplexT>
354EIGEN_DEVICE_FUNC
constexpr ComplexT complex_sqrt(
const ComplexT& z) {
355 return complex_sqrt_impl<ComplexT, false>(z);
358template <
typename ComplexT>
359EIGEN_DEVICE_FUNC
constexpr ComplexT complex_rsqrt(
const ComplexT& z) {
360 return complex_sqrt_impl<ComplexT, true>(z);
363template <
typename ComplexT>
364EIGEN_DEVICE_FUNC
constexpr ComplexT complex_log(
const ComplexT& z) {
366 using T =
typename NumTraits<ComplexT>::Real;
367 T a = numext::abs(z);
368 T b = numext::atan2(z.imag(), z.real());
370 if (EIGEN_PREDICT_FALSE((numext::isinf)(a) && (numext::isfinite)(z.real()) && (numext::isfinite)(z.imag()))) {
371 const T half = T(0.5);
372 return ComplexT(numext::log(numext::abs(ComplexT(z.real() * half, z.imag() * half))) + T(EIGEN_LN2), b);
374 if (EIGEN_PREDICT_FALSE(a < (numext::numeric_limits<T>::min)() && !numext::is_exactly_zero(a))) {
375 const T eps = NumTraits<T>::epsilon();
376 return ComplexT(numext::log(numext::abs(ComplexT(z.real() / eps, z.imag() / eps))) + numext::log(eps), b);
378 return ComplexT(numext::log(a), b);
383template <
typename ComplexT>
384EIGEN_DEVICE_FUNC ComplexT complex_log2(
const ComplexT& z) {
385 using T =
typename NumTraits<ComplexT>::Real;
386 const T a = numext::abs(z);
387 const T b = T(EIGEN_LOG2E) * numext::atan2(z.imag(), z.real());
388 if (EIGEN_PREDICT_FALSE((numext::isinf)(a) && (numext::isfinite)(z.real()) && (numext::isfinite)(z.imag()))) {
389 const T half = T(0.5);
390 return ComplexT(numext::log2(numext::abs(ComplexT(z.real() * half, z.imag() * half))) + T(1), b);
392 if (EIGEN_PREDICT_FALSE(a < (numext::numeric_limits<T>::min)() && !numext::is_exactly_zero(a))) {
393 const T eps = NumTraits<T>::epsilon();
394 return ComplexT(numext::log2(numext::abs(ComplexT(z.real() / eps, z.imag() / eps))) + numext::log2(eps), b);
396 return ComplexT(numext::log2(a), b);