devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merged latest updates from the Eigen trunk.

Browse Source
This commit is contained in:
Benoit Steiner
2014-09-15 09:18:16 -07:00
parent efdff15749 9452eb38f8
commit 10a79ca3a3
66 changed files with 882 additions and 1343 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
Eigen/src/Core/ArrayWrapper.h
View File

@@ -29,6 +29,11 @@ struct traits<ArrayWrapper<ExpressionType> >
: public traits<typename remove_all<typename ExpressionType::Nested>::type >
{
typedef ArrayXpr XprKind;
// Let's remove NestByRefBit
enum {
Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
Flags = Flags0 & ~NestByRefBit
};
};
}
@@ -166,6 +171,11 @@ struct traits<MatrixWrapper<ExpressionType> >
: public traits<typename remove_all<typename ExpressionType::Nested>::type >
{
typedef MatrixXpr XprKind;
// Let's remove NestByRefBit
enum {
Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
Flags = Flags0 & ~NestByRefBit
};
};
}

4
Eigen/src/Core/GeneralProduct.h
View File

@@ -231,7 +231,7 @@ EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest&
// FIXME not very good if rhs is real and lhs complex while alpha is real too
const Index cols = dest.cols();
for (Index j=0; j<cols; ++j)
func(dest.col(j), prod.rhs().coeff(j) * prod.lhs());
func(dest.col(j), prod.rhs().coeff(0,j) * prod.lhs());
}
// Row major
@@ -242,7 +242,7 @@ EIGEN_DONT_INLINE void outer_product_selector_run(const ProductType& prod, Dest&
// FIXME not very good if lhs is real and rhs complex while alpha is real too
const Index rows = dest.rows();
for (Index i=0; i<rows; ++i)
func(dest.row(i), prod.lhs().coeff(i) * prod.rhs());
func(dest.row(i), prod.lhs().coeff(i,0) * prod.rhs());
}
template<typename Lhs, typename Rhs>

24
Eigen/src/Core/MathFunctions.h
View File

@@ -12,6 +12,15 @@
namespace Eigen {
// On WINCE, std::abs is defined for int only, so let's defined our own overloads:
// This issue has been confirmed with MSVC 2008 only, but the issue might exist for more recent versions too.
#if defined(_WIN32_WCE) && defined(_MSC_VER) && _MSC_VER<=1500
long abs(long x) { return (labs(x)); }
double abs(double x) { return (fabs(x)); }
float abs(float x) { return (fabsf(x)); }
long double abs(long double x) { return (fabsl(x)); }
#endif
namespace internal {
/** \internal \struct global_math_functions_filtering_base
@@ -308,10 +317,17 @@ struct hypot_impl
using std::sqrt;
RealScalar _x = abs(x);
RealScalar _y = abs(y);
RealScalar p = (max)(_x, _y);
if(p==RealScalar(0)) return 0;
RealScalar q = (min)(_x, _y);
RealScalar qp = q/p;
Scalar p, qp;
if(_x>_y)
{
p = _x;
qp = _y / p;
}
else
{
p = _y;
qp = _x / p;
}
return p * sqrt(RealScalar(1) + qp*qp);
}
};

1
Eigen/src/Core/MatrixBase.h
View File

@@ -81,6 +81,7 @@ template<typename Derived> class MatrixBase
using Base::operator-=;
using Base::operator*=;
using Base::operator/=;
using Base::operator*;
typedef typename Base::CoeffReturnType CoeffReturnType;
typedef typename Base::ConstTransposeReturnType ConstTransposeReturnType;

2
Eigen/src/Core/Redux.h
View File

@@ -207,7 +207,7 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
const Index packetSize = packet_traits<Scalar>::size;
const Index alignedStart = internal::first_aligned(mat);
enum {
alignment = bool(Derived::Flags & DirectAccessBit) || bool(Derived::Flags & AlignedBit)
alignment = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) || bool(Derived::Flags & AlignedBit)
? Aligned : Unaligned
};
const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);

10
Eigen/src/Core/SelfCwiseBinaryOp.h
View File

@@ -209,15 +209,9 @@ inline Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
template<typename Derived>
inline Derived& DenseBase<Derived>::operator/=(const Scalar& other)
{
typedef typename internal::conditional<NumTraits<Scalar>::IsInteger,
internal::scalar_quotient_op<Scalar>,
internal::scalar_product_op<Scalar> >::type BinOp;
typedef typename Derived::PlainObject PlainObject;
SelfCwiseBinaryOp<BinOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
Scalar actual_other;
if(NumTraits<Scalar>::IsInteger) actual_other = other;
else actual_other = Scalar(1)/other;
tmp = PlainObject::Constant(rows(),cols(), actual_other);
SelfCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
tmp = PlainObject::Constant(rows(),cols(), other);
return derived();
}

22
Eigen/src/Core/StableNorm.h
View File

@@ -20,7 +20,7 @@ inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& sc
using std::max;
Scalar maxCoeff = bl.cwiseAbs().maxCoeff();
if (maxCoeff>scale)
if(maxCoeff>scale)
{
ssq = ssq * numext::abs2(scale/maxCoeff);
Scalar tmp = Scalar(1)/maxCoeff;
@@ -29,12 +29,21 @@ inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& sc
invScale = NumTraits<Scalar>::highest();
scale = Scalar(1)/invScale;
}
else if(maxCoeff>NumTraits<Scalar>::highest()) // we got a INF
{
invScale = Scalar(1);
scale = maxCoeff;
}
else
{
scale = maxCoeff;
invScale = tmp;
}
}
else if(maxCoeff!=maxCoeff) // we got a NaN
{
scale = maxCoeff;
}
// TODO if the maxCoeff is much much smaller than the current scale,
// then we can neglect this sub vector
@@ -55,7 +64,7 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
using std::abs;
const Derived& vec(_vec.derived());
static bool initialized = false;
static RealScalar b1, b2, s1m, s2m, overfl, rbig, relerr;
static RealScalar b1, b2, s1m, s2m, rbig, relerr;
if(!initialized)
{
int ibeta, it, iemin, iemax, iexp;
@@ -84,7 +93,6 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
iexp = - ((iemax+it)/2);
s2m = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp))); // scaling factor for upper range
overfl = rbig*s2m; // overflow boundary for abig
eps = RealScalar(pow(double(ibeta), 1-it));
relerr = sqrt(eps); // tolerance for neglecting asml
initialized = true;
@@ -101,13 +109,13 @@ blueNorm_impl(const EigenBase<Derived>& _vec)
else if(ax < b1) asml += numext::abs2(ax*s1m);
else amed += numext::abs2(ax);
}
if(amed!=amed)
return amed; // we got a NaN
if(abig > RealScalar(0))
{
abig = sqrt(abig);
if(abig > overfl)
{
return rbig;
}
if(abig > rbig) // overflow, or *this contains INF values
return abig; // return INF
if(amed > RealScalar(0))
{
abig = abig/s2m;

4
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@@ -143,7 +143,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f&
// so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate
// the result of the product.
Packet8f res = c;
asm("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
__asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
return res;
#else
return _mm256_fmadd_ps(a,b,c);
@@ -153,7 +153,7 @@ template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d&
#if defined(__clang__) || defined(__GNUC__)
// see above
Packet4d res = c;
asm("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
__asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
return res;
#else
return _mm256_fmadd_pd(a,b,c);

4
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@@ -52,12 +52,12 @@ typedef uint32x4_t Packet4ui;
// arm64 does have the pld instruction. If available, let's trust the __builtin_prefetch built-in function
// which available on LLVM and GCC (at least)
#if (defined(__has_builtin) && __has_builtin(__builtin_prefetch)) || defined(__GNUC__)
#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || defined(__GNUC__)
#define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR);
#elif defined __pld
#define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR)
#elif !defined(__aarch64__)
#define EIGEN_ARM_PREFETCH(ADDR) asm volatile ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
#define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ( " pld [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
#else
// by default no explicit prefetching
#define EIGEN_ARM_PREFETCH(ADDR)

14
Eigen/src/Core/functors/BinaryFunctors.h
View File

@@ -167,9 +167,17 @@ template<typename Scalar> struct scalar_hypot_op {
EIGEN_USING_STD_MATH(max);
EIGEN_USING_STD_MATH(min);
using std::sqrt;
Scalar p = (max)(_x, _y);
Scalar q = (min)(_x, _y);
Scalar qp = q/p;
Scalar p, qp;
if(_x>_y)
{
p = _x;
qp = _y / p;
}
else
{
p = _y;
qp = _x / p;
}
return p * sqrt(Scalar(1) + qp*qp);
}
};

2
Eigen/src/Core/products/GeneralMatrixMatrix_MKL.h
View File

@@ -53,6 +53,8 @@ template< \
int RhsStorageOrder, bool ConjugateRhs> \
struct general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor> \
{ \
typedef gebp_traits<EIGTYPE,EIGTYPE> Traits; \
\
static void run(Index rows, Index cols, Index depth, \
const EIGTYPE* _lhs, Index lhsStride, \
const EIGTYPE* _rhs, Index rhsStride, \

4
Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h
View File

@@ -109,7 +109,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
/* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
if (rows != depth) { \
\
int nthr = mkl_domain_get_max_threads(MKL_BLAS); \
int nthr = mkl_domain_get_max_threads(EIGEN_MKL_DOMAIN_BLAS); \
\
if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \
/* Most likely no benefit to call TRMM or GEMM from MKL*/ \
@@ -223,7 +223,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
/* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
if (cols != depth) { \
\
int nthr = mkl_domain_get_max_threads(MKL_BLAS); \
int nthr = mkl_domain_get_max_threads(EIGEN_MKL_DOMAIN_BLAS); \
\
if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \
/* Most likely no benefit to call TRMM or GEMM from MKL*/ \

32
Eigen/src/Core/util/MKL_support.h
View File

@@ -76,6 +76,38 @@
#include <mkl_lapacke.h>
#define EIGEN_MKL_VML_THRESHOLD 128
/* MKL_DOMAIN_BLAS, etc are defined only in 10.3 update 7 */
/* MKL_BLAS, etc are not defined in 11.2 */
#ifdef MKL_DOMAIN_ALL
#define EIGEN_MKL_DOMAIN_ALL MKL_DOMAIN_ALL
#else
#define EIGEN_MKL_DOMAIN_ALL MKL_ALL
#endif
#ifdef MKL_DOMAIN_BLAS
#define EIGEN_MKL_DOMAIN_BLAS MKL_DOMAIN_BLAS
#else
#define EIGEN_MKL_DOMAIN_BLAS MKL_BLAS
#endif
#ifdef MKL_DOMAIN_FFT
#define EIGEN_MKL_DOMAIN_FFT MKL_DOMAIN_FFT
#else
#define EIGEN_MKL_DOMAIN_FFT MKL_FFT
#endif
#ifdef MKL_DOMAIN_VML
#define EIGEN_MKL_DOMAIN_VML MKL_DOMAIN_VML
#else
#define EIGEN_MKL_DOMAIN_VML MKL_VML
#endif
#ifdef MKL_DOMAIN_PARDISO
#define EIGEN_MKL_DOMAIN_PARDISO MKL_DOMAIN_PARDISO
#else
#define EIGEN_MKL_DOMAIN_PARDISO MKL_PARDISO
#endif
namespace Eigen {
typedef std::complex<double> dcomplex;

9
Eigen/src/Core/util/Macros.h
View File

@@ -107,6 +107,13 @@
#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
#endif
// Cross compiler wrapper around LLVM's __has_builtin
#ifdef __has_builtin
# define EIGEN_HAS_BUILTIN(x) __has_builtin(x)
#else
# define EIGEN_HAS_BUILTIN(x) 0
#endif
// A Clang feature extension to determine compiler features.
// We use it to determine 'cxx_rvalue_references'
#ifndef __has_feature
@@ -277,7 +284,7 @@ namespace Eigen {
#if !defined(EIGEN_ASM_COMMENT)
#if (defined __GNUC__) && ( defined(__i386__) || defined(__x86_64__) )
#define EIGEN_ASM_COMMENT(X) asm("#" X)
#define EIGEN_ASM_COMMENT(X) __asm__("#" X)
#else
#define EIGEN_ASM_COMMENT(X)
#endif

2
Eigen/src/Core/util/Memory.h
View File

@@ -64,7 +64,7 @@
// Currently, let's include it only on unix systems:
#if defined(__unix__) || defined(__unix)
#include <unistd.h>
#if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
#if ((defined __QNXNTO__) || (defined _GNU_SOURCE) || (defined __PGI) || ((defined _XOPEN_SOURCE) && (_XOPEN_SOURCE >= 600))) && (defined _POSIX_ADVISORY_INFO) && (_POSIX_ADVISORY_INFO > 0)
#define EIGEN_HAS_POSIX_MEMALIGN 1
#endif
#endif

10
Eigen/src/Core/util/XprHelper.h
View File

@@ -383,13 +383,21 @@ struct special_scalar_op_base<Derived,Scalar,OtherScalar,true> : public DenseCo
const CwiseUnaryOp<scalar_multiple2_op<Scalar,OtherScalar>, Derived>
operator*(const OtherScalar& scalar) const
{
#ifdef EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN
EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN
#endif
return CwiseUnaryOp<scalar_multiple2_op<Scalar,OtherScalar>, Derived>
(*static_cast<const Derived*>(this), scalar_multiple2_op<Scalar,OtherScalar>(scalar));
}
inline friend const CwiseUnaryOp<scalar_multiple2_op<Scalar,OtherScalar>, Derived>
operator*(const OtherScalar& scalar, const Derived& matrix)
{ return static_cast<const special_scalar_op_base&>(matrix).operator*(scalar); }
{
#ifdef EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN
EIGEN_SPECIAL_SCALAR_MULTIPLE_PLUGIN
#endif
return static_cast<const special_scalar_op_base&>(matrix).operator*(scalar);
}
};
template<typename XprType, typename CastType> struct cast_return_type