From f9fac1d5b044afb8104ecb5c1e5183c309bafca4 Mon Sep 17 00:00:00 2001 From: Rasmus Munk Larsen Date: Fri, 4 Dec 2020 21:45:09 +0000 Subject: Add log2() to Eigen. --- Eigen/src/Core/GenericPacketMath.h | 7 +++ Eigen/src/Core/GlobalFunctions.h | 3 +- Eigen/src/Core/arch/AVX/MathFunctions.h | 14 +++++ Eigen/src/Core/arch/AVX512/MathFunctions.h | 15 ++++++ Eigen/src/Core/arch/Default/BFloat16.h | 3 ++ .../Core/arch/Default/GenericPacketMathFunctions.h | 63 ++++++++++++++++++---- .../arch/Default/GenericPacketMathFunctionsFwd.h | 12 +++++ Eigen/src/Core/arch/Default/Half.h | 4 ++ Eigen/src/Core/arch/SSE/MathFunctions.h | 10 ++++ Eigen/src/Core/functors/UnaryFunctors.h | 16 ++++++ Eigen/src/plugins/ArrayCwiseUnaryOps.h | 13 +++++ test/array_cwise.cpp | 4 ++ test/packetmath.cpp | 8 +++ 13 files changed, 160 insertions(+), 12 deletions(-) diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index bdde62e1f..3f80e8033 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -650,6 +650,13 @@ Packet plog1p(const Packet& a) { return numext::log1p(a); } template EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet plog10(const Packet& a) { EIGEN_USING_STD(log10); return log10(a); } +/** \internal \returns the log10 of \a a (coeff-wise) */ +template EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +Packet plog2(const Packet& a) { + typedef typename internal::unpacket_traits::type Scalar; + return pmul(pset1(Scalar(M_LOG2E)), plog(a)); +} + /** \internal \returns the square-root of \a a (coeff-wise) */ template EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet psqrt(const Packet& a) { EIGEN_USING_STD(sqrt); return sqrt(a); } diff --git a/Eigen/src/Core/GlobalFunctions.h b/Eigen/src/Core/GlobalFunctions.h index 0a539ed8e..629af94b9 100644 --- a/Eigen/src/Core/GlobalFunctions.h +++ b/Eigen/src/Core/GlobalFunctions.h @@ -81,7 +81,8 @@ namespace Eigen EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(expm1,scalar_expm1_op,exponential of a value minus 1,\sa ArrayBase::expm1) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op,natural logarithm,\sa Eigen::log10 DOXCOMMA ArrayBase::log) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log1p,scalar_log1p_op,natural logarithm of 1 plus the value,\sa ArrayBase::log1p) - EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op,base 10 logarithm,\sa Eigen::log DOXCOMMA ArrayBase::log) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op,base 10 logarithm,\sa Eigen::log DOXCOMMA ArrayBase::log10) + EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log2,scalar_log2_op,base 2 logarithm,\sa Eigen::log DOXCOMMA ArrayBase::log2) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op,absolute value,\sa ArrayBase::abs DOXCOMMA MatrixBase::cwiseAbs) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs2,scalar_abs2_op,squared absolute value,\sa ArrayBase::abs2 DOXCOMMA MatrixBase::cwiseAbs2) EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(arg,scalar_arg_op,complex argument,\sa ArrayBase::arg DOXCOMMA MatrixBase::cwiseArg) diff --git a/Eigen/src/Core/arch/AVX/MathFunctions.h b/Eigen/src/Core/arch/AVX/MathFunctions.h index e2e704d82..5fe2cff32 100644 --- a/Eigen/src/Core/arch/AVX/MathFunctions.h +++ b/Eigen/src/Core/arch/AVX/MathFunctions.h @@ -42,6 +42,18 @@ plog(const Packet4d& _x) { return plog_double(_x); } +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f +plog2(const Packet8f& _x) { + return plog2_float(_x); +} + +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4d +plog2(const Packet4d& _x) { + return plog2_double(_x); +} + template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f plog1p(const Packet8f& _x) { return generic_plog1p(_x); @@ -161,6 +173,7 @@ Packet4d prsqrt(const Packet4d& _x) { F16_PACKET_FUNCTION(Packet8f, Packet8h, psin) F16_PACKET_FUNCTION(Packet8f, Packet8h, pcos) F16_PACKET_FUNCTION(Packet8f, Packet8h, plog) +F16_PACKET_FUNCTION(Packet8f, Packet8h, plog2) F16_PACKET_FUNCTION(Packet8f, Packet8h, plog1p) F16_PACKET_FUNCTION(Packet8f, Packet8h, pexpm1) F16_PACKET_FUNCTION(Packet8f, Packet8h, pexp) @@ -171,6 +184,7 @@ F16_PACKET_FUNCTION(Packet8f, Packet8h, prsqrt) BF16_PACKET_FUNCTION(Packet8f, Packet8bf, psin) BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pcos) BF16_PACKET_FUNCTION(Packet8f, Packet8bf, plog) +BF16_PACKET_FUNCTION(Packet8f, Packet8bf, plog2) BF16_PACKET_FUNCTION(Packet8f, Packet8bf, plog1p) BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pexpm1) BF16_PACKET_FUNCTION(Packet8f, Packet8bf, pexp) diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h index c04da6795..66f3252cd 100644 --- a/Eigen/src/Core/arch/AVX512/MathFunctions.h +++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h @@ -50,6 +50,21 @@ plog(const Packet8d& _x) { F16_PACKET_FUNCTION(Packet16f, Packet16h, plog) BF16_PACKET_FUNCTION(Packet16f, Packet16bf, plog) +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f +plog2(const Packet16f& _x) { + return plog2_float(_x); +} + +template <> +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d +plog2(const Packet8d& _x) { + return plog2_double(_x); +} + +F16_PACKET_FUNCTION(Packet16f, Packet16h, plog2) +BF16_PACKET_FUNCTION(Packet16f, Packet16bf, plog2) + // Exponential function. Works by writing "x = m*log(2) + r" where // "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then // "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1). diff --git a/Eigen/src/Core/arch/Default/BFloat16.h b/Eigen/src/Core/arch/Default/BFloat16.h index 351f451a3..616dcf667 100644 --- a/Eigen/src/Core/arch/Default/BFloat16.h +++ b/Eigen/src/Core/arch/Default/BFloat16.h @@ -512,6 +512,9 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 log1p(const bfloat16& a) { EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 log10(const bfloat16& a) { return bfloat16(::log10f(float(a))); } +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 log2(const bfloat16& a) { + return bfloat16(static_cast(M_LOG2E) * ::logf(float(a))); +} EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bfloat16 sqrt(const bfloat16& a) { return bfloat16(::sqrtf(float(a))); } diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h index 60db2e12f..c6bb89b05 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h @@ -59,16 +59,16 @@ pldexp_double(Packet a, Packet exponent) return pmul(a, preinterpret(plogical_shift_left<52>(ei))); } -// Natural logarithm +// Natural or base 2 logarithm. // Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2) // and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can // be easily approximated by a polynomial centered on m=1 for stability. // TODO(gonnet): Further reduce the interval allowing for lower-degree // polynomial interpolants -> ... -> profit! -template +template EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED -Packet plog_float(const Packet _x) +Packet plog_impl_float(const Packet _x) { Packet x = _x; @@ -131,8 +131,13 @@ Packet plog_float(const Packet _x) x = padd(x, y); // Add the logarithm of the exponent back to the result of the interpolation. - const Packet cst_ln2 = pset1(M_LN2); - x = pmadd(e, cst_ln2, x); + if (base2) { + const Packet cst_log2e = pset1(static_cast(M_LOG2E)); + x = pmadd(x, cst_log2e, e); + } else { + const Packet cst_ln2 = pset1(static_cast(M_LN2)); + x = pmadd(e, cst_ln2, x); + } Packet invalid_mask = pcmp_lt_or_nan(_x, pzero(_x)); Packet iszero_mask = pcmp_eq(_x,pzero(_x)); @@ -145,8 +150,23 @@ Packet plog_float(const Packet _x) por(pselect(pos_inf_mask,cst_pos_inf,x), invalid_mask)); } +template +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet plog_float(const Packet _x) +{ + return plog_impl_float(_x); +} + +template +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet plog2_float(const Packet _x) +{ + return plog_impl_float(_x); +} -/* Returns the base e (2.718...) logarithm of x. +/* Returns the base e (2.718...) or base 2 logarithm of x. * The argument is separated into its exponent and fractional parts. * The logarithm of the fraction in the interval [sqrt(1/2), sqrt(2)], * is approximated by @@ -155,16 +175,16 @@ Packet plog_float(const Packet _x) * * for more detail see: http://www.netlib.org/cephes/ */ -template +template EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED -Packet plog_double(const Packet _x) +Packet plog_impl_double(const Packet _x) { Packet x = _x; const Packet cst_1 = pset1(1.0); const Packet cst_neg_half = pset1(-0.5); - // The smallest non denormalized float number. + // The smallest non denormalized double. const Packet cst_min_norm_pos = pset1frombits( static_cast(0x0010000000000000ull)); const Packet cst_minus_inf = pset1frombits( static_cast(0xfff0000000000000ull)); const Packet cst_pos_inf = pset1frombits( static_cast(0x7ff0000000000000ull)); @@ -232,8 +252,13 @@ Packet plog_double(const Packet _x) x = padd(x, y); // Add the logarithm of the exponent back to the result of the interpolation. - const Packet cst_ln2 = pset1(M_LN2); - x = pmadd(e, cst_ln2, x); + if (base2) { + const Packet cst_log2e = pset1(M_LOG2E); + x = pmadd(x, cst_log2e, e); + } else { + const Packet cst_ln2 = pset1(M_LN2); + x = pmadd(e, cst_ln2, x); + } Packet invalid_mask = pcmp_lt_or_nan(_x, pzero(_x)); Packet iszero_mask = pcmp_eq(_x,pzero(_x)); @@ -246,6 +271,22 @@ Packet plog_double(const Packet _x) por(pselect(pos_inf_mask,cst_pos_inf,x), invalid_mask)); } +template +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet plog_double(const Packet _x) +{ + return plog_impl_double(_x); +} + +template +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet plog2_double(const Packet _x) +{ + return plog_impl_double(_x); +} + /** \internal \returns log(1 + x) computed using W. Kahan's formula. See: http://www.plunk.org/~hatch/rightway.php */ diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctionsFwd.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctionsFwd.h index 0e02a1b20..b0f0b78fc 100644 --- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctionsFwd.h +++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctionsFwd.h @@ -32,12 +32,24 @@ EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet plog_float(const Packet _x); +/** \internal \returns log2(x) for single precision float */ +template +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet plog2_float(const Packet _x); + /** \internal \returns log(x) for single precision float */ template EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet plog_double(const Packet _x); +/** \internal \returns log2(x) for single precision float */ +template +EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS +EIGEN_UNUSED +Packet plog2_double(const Packet _x); + /** \internal \returns log(1 + x) */ template Packet generic_plog1p(const Packet& x); diff --git a/Eigen/src/Core/arch/Default/Half.h b/Eigen/src/Core/arch/Default/Half.h index 76156c416..204076e25 100644 --- a/Eigen/src/Core/arch/Default/Half.h +++ b/Eigen/src/Core/arch/Default/Half.h @@ -622,6 +622,10 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) { EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) { return half(::log10f(float(a))); } +EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log2(const half& a) { + return half(static_cast(M_LOG2E) * ::logf(float(a))); +} + EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) { #if (EIGEN_CUDA_SDK_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 530) || \ defined(EIGEN_HIP_DEVICE_COMPILE) diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h index d5b62e86a..2d7df2f7b 100644 --- a/Eigen/src/Core/arch/SSE/MathFunctions.h +++ b/Eigen/src/Core/arch/SSE/MathFunctions.h @@ -29,6 +29,16 @@ Packet2d plog(const Packet2d& _x) { return plog_double(_x); } +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet4f plog2(const Packet4f& _x) { + return plog2_float(_x); +} + +template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED +Packet2d plog2(const Packet2d& _x) { + return plog2_double(_x); +} + template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f plog1p(const Packet4f& _x) { return generic_plog1p(_x); diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h index 8c0ed8481..c037576dc 100644 --- a/Eigen/src/Core/functors/UnaryFunctors.h +++ b/Eigen/src/Core/functors/UnaryFunctors.h @@ -395,6 +395,22 @@ template struct functor_traits > { enum { Cost = 5 * NumTraits::MulCost, PacketAccess = packet_traits::HasLog10 }; }; +/** \internal + * + * \brief Template functor to compute the base-2 logarithm of a scalar + * + * \sa class CwiseUnaryOp, Cwise::log2() + */ +template struct scalar_log2_op { + EIGEN_EMPTY_STRUCT_CTOR(scalar_log2_op) + EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return Scalar(M_LOG2E) * std::log(a); } + template + EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog2(a); } +}; +template +struct functor_traits > +{ enum { Cost = 5 * NumTraits::MulCost, PacketAccess = packet_traits::HasLog }; }; + /** \internal * \brief Template functor to compute the square root of a scalar * \sa class CwiseUnaryOp, Cwise::sqrt() diff --git a/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/Eigen/src/plugins/ArrayCwiseUnaryOps.h index 59a4ee6a0..b7ea22a9d 100644 --- a/Eigen/src/plugins/ArrayCwiseUnaryOps.h +++ b/Eigen/src/plugins/ArrayCwiseUnaryOps.h @@ -14,6 +14,7 @@ typedef CwiseUnaryOp, const Derived> Expm1Retu typedef CwiseUnaryOp, const Derived> LogReturnType; typedef CwiseUnaryOp, const Derived> Log1pReturnType; typedef CwiseUnaryOp, const Derived> Log10ReturnType; +typedef CwiseUnaryOp, const Derived> Log2ReturnType; typedef CwiseUnaryOp, const Derived> CosReturnType; typedef CwiseUnaryOp, const Derived> SinReturnType; typedef CwiseUnaryOp, const Derived> TanReturnType; @@ -159,6 +160,18 @@ log10() const return Log10ReturnType(derived()); } +/** \returns an expression of the coefficient-wise base-2 logarithm of *this. + * + * This function computes the coefficient-wise base-2 logarithm. + * + */ +EIGEN_DEVICE_FUNC +inline const Log2ReturnType +log2() const +{ + return Log2ReturnType(derived()); +} + /** \returns an expression of the coefficient-wise square root of *this. * * This function computes the coefficient-wise square root. The function MatrixBase::sqrt() in the diff --git a/test/array_cwise.cpp b/test/array_cwise.cpp index 48c0935a4..6910f0e1f 100644 --- a/test/array_cwise.cpp +++ b/test/array_cwise.cpp @@ -319,6 +319,7 @@ template void array_real(const ArrayType& m) VERIFY_IS_APPROX(m3.log(), log(m3)); VERIFY_IS_APPROX(m3.log1p(), log1p(m3)); VERIFY_IS_APPROX(m3.log10(), log10(m3)); + VERIFY_IS_APPROX(m3.log2(), log2(m3)); VERIFY((!(m1>m2) == (m1<=m2)).all()); @@ -372,6 +373,7 @@ template void array_real(const ArrayType& m) VERIFY_IS_APPROX(pow(m3,RealScalar(-0.5)), m3.rsqrt()); VERIFY_IS_APPROX(log10(m3), log(m3)/log(10)); + VERIFY_IS_APPROX(log2(m3), log(m3)/log(2)); // scalar by array division const RealScalar tiny = sqrt(std::numeric_limits::epsilon()); @@ -423,6 +425,7 @@ template void array_complex(const ArrayType& m) VERIFY_IS_APPROX(m1.inverse(), inverse(m1)); VERIFY_IS_APPROX(m1.log(), log(m1)); VERIFY_IS_APPROX(m1.log10(), log10(m1)); + VERIFY_IS_APPROX(m1.log2(), log2(m1)); VERIFY_IS_APPROX(m1.abs(), abs(m1)); VERIFY_IS_APPROX(m1.abs2(), abs2(m1)); VERIFY_IS_APPROX(m1.sqrt(), sqrt(m1)); @@ -478,6 +481,7 @@ template void array_complex(const ArrayType& m) VERIFY_IS_APPROX(abs(m1), sqrt(square(m1.real())+square(m1.imag()))); VERIFY_IS_APPROX(abs(m1), sqrt(abs2(m1))); VERIFY_IS_APPROX(log10(m1), log(m1)/log(10)); + VERIFY_IS_APPROX(log2(m1), log(m1)/log(2)); VERIFY_IS_APPROX( m1.sign(), -(-m1).sign() ); VERIFY_IS_APPROX( m1.sign() * m1.abs(), m1); diff --git a/test/packetmath.cpp b/test/packetmath.cpp index ae0ead820..d995e8b71 100644 --- a/test/packetmath.cpp +++ b/test/packetmath.cpp @@ -488,6 +488,13 @@ void packetmath() { packetmath_minus_zero_add(); } +// Notice that this definition works for complex types as well. +// c++11 has std::log2 for real, but not for complex types. +template +Scalar log2(Scalar x) { + return Scalar(M_LOG2E) * std::log(x); +} + template void packetmath_real() { typedef internal::packet_traits PacketTraits; @@ -506,6 +513,7 @@ void packetmath_real() { if (internal::random(0, 1) < 0.1f) data1[internal::random(0, PacketSize)] = Scalar(0); CHECK_CWISE1_IF(PacketTraits::HasLog, std::log, internal::plog); + CHECK_CWISE1_IF(PacketTraits::HasLog, log2, internal::plog2); CHECK_CWISE1_IF(PacketTraits::HasRsqrt, 1 / std::sqrt, internal::prsqrt); for (int i = 0; i < size; ++i) { -- cgit v1.2.3