From b43102440489df9d0175c88e602dfa425b574a94 Mon Sep 17 00:00:00 2001 From: Rasmus Munk Larsen Date: Wed, 7 Oct 2020 19:05:18 +0000 Subject: Don't make assumptions about NaN-propagation for pmin/pmax - it various across platforms. Change test to only test for NaN-propagation for pfmin/pfmax. --- Eigen/src/Core/GenericPacketMath.h | 55 +++++++++++++++++++++++++++++--------- 1 file changed, 43 insertions(+), 12 deletions(-) (limited to 'Eigen/src/Core/GenericPacketMath.h') diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 075d18aa6..b5eb1cf99 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -216,12 +216,12 @@ template EIGEN_DEVICE_FUNC inline Packet pdiv(const Packet& a, const Packet& b) { return a/b; } /** \internal \returns the min of \a a and \a b (coeff-wise). -Equivalent to std::min(a, b), so if either a or b is NaN, a is returned. */ + If \a a or \b b is NaN, the return value is implementation defined. */ template EIGEN_DEVICE_FUNC inline Packet pmin(const Packet& a, const Packet& b) { return numext::mini(a, b); } /** \internal \returns the max of \a a and \a b (coeff-wise) -Equivalent to std::max(a, b), so if either a or b is NaN, a is returned.*/ + If \a a or \b b is NaN, the return value is implementation defined. */ template EIGEN_DEVICE_FUNC inline Packet pmax(const Packet& a, const Packet& b) { return numext::maxi(a, b); } @@ -635,23 +635,54 @@ Packet print(const Packet& a) { using numext::rint; return rint(a); } template EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); } -/** \internal \returns the min of \a a and \a b (coeff-wise) - Equivalent to std::fmin(a, b). Only if both a and b are NaN is NaN returned. -*/ + +/** \internal \returns the max of \a a and \a b (coeff-wise) + If both \a a and \a b are NaN, NaN is returned. + Equivalent to std::fmax(a, b). */ +template EIGEN_DEVICE_FUNC inline Packet +pfmax(const Packet& a, const Packet& b) { + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmax(a, b), a), + b); +} + +/** \internal \returns the min of \a a and \a b (coeff-wise) + If both \a a and \a b are NaN, NaN is returned. + Equivalent to std::fmin(a, b). */ template EIGEN_DEVICE_FUNC inline Packet pfmin(const Packet& a, const Packet& b) { - Packet not_nan_mask = pcmp_eq(a, a); - return pselect(not_nan_mask, pmin(a, b), b); + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmin(a, b), a), + b); } /** \internal \returns the max of \a a and \a b (coeff-wise) - Equivalent to std::fmax(a, b). Only if both a and b are NaN is NaN returned.*/ -template EIGEN_DEVICE_FUNC inline Packet -pfmax(const Packet& a, const Packet& b) { - Packet not_nan_mask = pcmp_eq(a, a); - return pselect(not_nan_mask, pmax(a, b), b); + If either \a a or \a b are NaN, NaN is returned. */ +template EIGEN_DEVICE_FUNC inline Packet +pfmax_nan(const Packet& a, const Packet& b) { + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmax(a, b), b), + a); +} + +/** \internal \returns the min of \a a and \a b (coeff-wise) + If either \a a or \a b are NaN, NaN is returned. */ +template EIGEN_DEVICE_FUNC inline Packet +pfmin_nan(const Packet& a, const Packet& b) { + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmin(a, b), b), + a); } + /*************************************************************************** * The following functions might not have to be overwritten for vectorized types ***************************************************************************/ -- cgit v1.2.3