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 default into unary-array-cwise-functors

Browse Source
This commit is contained in:
Deanna Hood
2015-04-20 14:01:35 -04:00
parent 0339502a4f 0eb220c00d
commit 0250f4a9f2
102 changed files with 2595 additions and 946 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
Eigen/src/Cholesky/LDLT.h
View File

@@ -226,6 +226,11 @@ template<typename _MatrixType, int _UpLo> class LDLT
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
/** \internal
* Used to compute and store the Cholesky decomposition A = L D L^* = U^* D U.
@@ -424,6 +429,8 @@ template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
template<typename MatrixType, int _UpLo>
LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
{
check_template_parameters();
eigen_assert(a.rows()==a.cols());
const Index size = a.rows();

8
Eigen/src/Cholesky/LLT.h
View File

@@ -170,6 +170,12 @@ template<typename _MatrixType, int _UpLo> class LLT
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
/** \internal
* Used to compute and store L
* The strict upper part is not used and even not initialized.
@@ -377,6 +383,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
template<typename MatrixType, int _UpLo>
LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
{
check_template_parameters();
eigen_assert(a.rows()==a.cols());
const Index size = a.rows();
m_matrix.resize(size, size);

17
Eigen/src/Core/CoreEvaluators.h
View File

@@ -647,11 +647,15 @@ struct evaluator<Map<PlainObjectType, MapOptions, StrideType> >
HasNoStride = HasNoInnerStride && HasNoOuterStride,
IsAligned = bool(EIGEN_ALIGN) && ((int(MapOptions)&Aligned)==Aligned),
IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic,
// TODO: should check for smaller packet types once we can handle multi-sized packet types
AlignBytes = int(packet_traits<Scalar>::size) * sizeof(Scalar),
KeepsPacketAccess = bool(HasNoInnerStride)
&& ( bool(IsDynamicSize)
|| HasNoOuterStride
|| ( OuterStrideAtCompileTime!=Dynamic
&& ((static_cast<int>(sizeof(Scalar))*OuterStrideAtCompileTime)%EIGEN_ALIGN_BYTES)==0 ) ),
&& ((static_cast<int>(sizeof(Scalar))*OuterStrideAtCompileTime) % AlignBytes)==0 ) ),
Flags0 = evaluator<PlainObjectType>::Flags,
Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit),
Flags2 = (bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime))
@@ -717,7 +721,10 @@ struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
&& (InnerStrideAtCompileTime == 1)
? PacketAccessBit : 0,
MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0)) ? AlignedBit : 0,
// TODO: should check for smaller packet types once we can handle multi-sized packet types
AlignBytes = int(packet_traits<Scalar>::size) * sizeof(Scalar),
MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % AlignBytes) == 0)) ? AlignedBit : 0,
FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (evaluator<ArgType>::Flags&LinearAccessBit))) ? LinearAccessBit : 0,
FlagsRowMajorBit = XprType::Flags&RowMajorBit,
Flags0 = evaluator<ArgType>::Flags & ( (HereditaryBits & ~RowMajorBit) |
@@ -825,12 +832,16 @@ struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /* HasDirectAc
typename Block<ArgType, BlockRows, BlockCols, InnerPanel>::PlainObject>
{
typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
typedef typename XprType::Scalar Scalar;
EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block)
: mapbase_evaluator<XprType, typename XprType::PlainObject>(block)
{
// TODO: should check for smaller packet types once we can handle multi-sized packet types
const int AlignBytes = int(packet_traits<Scalar>::size) * sizeof(Scalar);
EIGEN_ONLY_USED_FOR_DEBUG(AlignBytes)
// FIXME this should be an internal assertion
eigen_assert(EIGEN_IMPLIES(evaluator<XprType>::Flags&AlignedBit, (size_t(block.data()) % EIGEN_ALIGN_BYTES) == 0) && "data is not aligned");
eigen_assert(EIGEN_IMPLIES(evaluator<XprType>::Flags&AlignedBit, (size_t(block.data()) % AlignBytes) == 0) && "data is not aligned");
}
};

11
Eigen/src/Core/CwiseNullaryOp.h
View File

@@ -300,9 +300,10 @@ template<typename Derived>
bool DenseBase<Derived>::isApproxToConstant
(const Scalar& val, const RealScalar& prec) const
{
typename internal::nested_eval<Derived,1>::type self(derived());
for(Index j = 0; j < cols(); ++j)
for(Index i = 0; i < rows(); ++i)
if(!internal::isApprox(this->coeff(i, j), val, prec))
if(!internal::isApprox(self.coeff(i, j), val, prec))
return false;
return true;
}
@@ -484,9 +485,10 @@ DenseBase<Derived>::Zero()
template<typename Derived>
bool DenseBase<Derived>::isZero(const RealScalar& prec) const
{
typename internal::nested_eval<Derived,1>::type self(derived());
for(Index j = 0; j < cols(); ++j)
for(Index i = 0; i < rows(); ++i)
if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<Scalar>(1), prec))
if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<Scalar>(1), prec))
return false;
return true;
}
@@ -719,18 +721,19 @@ template<typename Derived>
bool MatrixBase<Derived>::isIdentity
(const RealScalar& prec) const
{
typename internal::nested_eval<Derived,1>::type self(derived());
for(Index j = 0; j < cols(); ++j)
{
for(Index i = 0; i < rows(); ++i)
{
if(i == j)
{
if(!internal::isApprox(this->coeff(i, j), static_cast<Scalar>(1), prec))
if(!internal::isApprox(self.coeff(i, j), static_cast<Scalar>(1), prec))
return false;
}
else
{
if(!internal::isMuchSmallerThan(this->coeff(i, j), static_cast<RealScalar>(1), prec))
if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<RealScalar>(1), prec))
return false;
}
}

88
Eigen/src/Core/DenseStorage.h
View File

@@ -34,14 +34,35 @@ void check_static_allocation_size()
#endif
}
template<typename T, int Size, typename Packet = typename packet_traits<T>::type,
bool Match = bool((Size%unpacket_traits<Packet>::size)==0),
bool TryHalf = bool(int(unpacket_traits<Packet>::size) > 1)
&& bool(int(unpacket_traits<Packet>::size) > int(unpacket_traits<typename unpacket_traits<Packet>::half>::size)) >
struct compute_default_alignment
{
enum { value = 0 };
};
template<typename T, int Size, typename Packet, bool TryHalf>
struct compute_default_alignment<T, Size, Packet, true, TryHalf> // Match
{
enum { value = sizeof(T) * unpacket_traits<Packet>::size };
};
template<typename T, int Size, typename Packet>
struct compute_default_alignment<T, Size, Packet, false, true> // Try-half
{
// current packet too large, try with an half-packet
enum { value = compute_default_alignment<T, Size, typename unpacket_traits<Packet>::half>::value };
};
/** \internal
* Static array. If the MatrixOrArrayOptions require auto-alignment, the array will be automatically aligned:
* to 16 bytes boundary if the total size is a multiple of 16 bytes.
*/
template <typename T, int Size, int MatrixOrArrayOptions,
int Alignment = (MatrixOrArrayOptions&DontAlign) ? 0
: (((Size*sizeof(T))%EIGEN_ALIGN_BYTES)==0) ? EIGEN_ALIGN_BYTES
: 0 >
: compute_default_alignment<T,Size>::value >
struct plain_array
{
T array[Size];
@@ -81,14 +102,71 @@ struct plain_array
#endif
template <typename T, int Size, int MatrixOrArrayOptions>
struct plain_array<T, Size, MatrixOrArrayOptions, EIGEN_ALIGN_BYTES>
struct plain_array<T, Size, MatrixOrArrayOptions, 8>
{
EIGEN_USER_ALIGN_DEFAULT T array[Size];
EIGEN_ALIGN_TO_BOUNDARY(8) T array[Size];
EIGEN_DEVICE_FUNC
plain_array()
{
EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(EIGEN_ALIGN_BYTES-1);
EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(7);
check_static_allocation_size<T,Size>();
}
EIGEN_DEVICE_FUNC
plain_array(constructor_without_unaligned_array_assert)
{
check_static_allocation_size<T,Size>();
}
};
template <typename T, int Size, int MatrixOrArrayOptions>
struct plain_array<T, Size, MatrixOrArrayOptions, 16>
{
EIGEN_ALIGN_TO_BOUNDARY(16) T array[Size];
EIGEN_DEVICE_FUNC
plain_array()
{
EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(15);
check_static_allocation_size<T,Size>();
}
EIGEN_DEVICE_FUNC
plain_array(constructor_without_unaligned_array_assert)
{
check_static_allocation_size<T,Size>();
}
};
template <typename T, int Size, int MatrixOrArrayOptions>
struct plain_array<T, Size, MatrixOrArrayOptions, 32>
{
EIGEN_ALIGN_TO_BOUNDARY(32) T array[Size];
EIGEN_DEVICE_FUNC
plain_array()
{
EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(31);
check_static_allocation_size<T,Size>();
}
EIGEN_DEVICE_FUNC
plain_array(constructor_without_unaligned_array_assert)
{
check_static_allocation_size<T,Size>();
}
};
template <typename T, int Size, int MatrixOrArrayOptions>
struct plain_array<T, Size, MatrixOrArrayOptions, 64>
{
EIGEN_ALIGN_TO_BOUNDARY(64) T array[Size];
EIGEN_DEVICE_FUNC
plain_array()
{
EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(63);
check_static_allocation_size<T,Size>();
}

6
Eigen/src/Core/Dot.h
View File

@@ -224,13 +224,13 @@ bool MatrixBase<Derived>::isOrthogonal
template<typename Derived>
bool MatrixBase<Derived>::isUnitary(const RealScalar& prec) const
{
typename Derived::Nested nested(derived());
typename internal::nested_eval<Derived,1>::type self(derived());
for(Index i = 0; i < cols(); ++i)
{
if(!internal::isApprox(nested.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
if(!internal::isApprox(self.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
return false;
for(Index j = 0; j < i; ++j)
if(!internal::isMuchSmallerThan(nested.col(i).dot(nested.col(j)), static_cast<Scalar>(1), prec))
if(!internal::isMuchSmallerThan(self.col(i).dot(self.col(j)), static_cast<Scalar>(1), prec))
return false;
}
return true;

58
Eigen/src/Core/MathFunctions.h
View File

@@ -328,6 +328,7 @@ struct hypot_impl
p = _y;
qp = _x / p;
}
if(p==RealScalar(0)) return RealScalar(0);
return p * sqrt(RealScalar(1) + qp*qp);
}
};
@@ -560,48 +561,48 @@ struct random_default_impl<Scalar, false, false>
};
enum {
floor_log2_terminate,
floor_log2_move_up,
floor_log2_move_down,
floor_log2_bogus
meta_floor_log2_terminate,
meta_floor_log2_move_up,
meta_floor_log2_move_down,
meta_floor_log2_bogus
};
template<unsigned int n, int lower, int upper> struct floor_log2_selector
template<unsigned int n, int lower, int upper> struct meta_floor_log2_selector
{
enum { middle = (lower + upper) / 2,
value = (upper <= lower + 1) ? int(floor_log2_terminate)
: (n < (1 << middle)) ? int(floor_log2_move_down)
: (n==0) ? int(floor_log2_bogus)
: int(floor_log2_move_up)
value = (upper <= lower + 1) ? int(meta_floor_log2_terminate)
: (n < (1 << middle)) ? int(meta_floor_log2_move_down)
: (n==0) ? int(meta_floor_log2_bogus)
: int(meta_floor_log2_move_up)
};
};
template<unsigned int n,
int lower = 0,
int upper = sizeof(unsigned int) * CHAR_BIT - 1,
int selector = floor_log2_selector<n, lower, upper>::value>
struct floor_log2 {};
int selector = meta_floor_log2_selector<n, lower, upper>::value>
struct meta_floor_log2 {};
template<unsigned int n, int lower, int upper>
struct floor_log2<n, lower, upper, floor_log2_move_down>
struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_down>
{
enum { value = floor_log2<n, lower, floor_log2_selector<n, lower, upper>::middle>::value };
enum { value = meta_floor_log2<n, lower, meta_floor_log2_selector<n, lower, upper>::middle>::value };
};
template<unsigned int n, int lower, int upper>
struct floor_log2<n, lower, upper, floor_log2_move_up>
struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_up>
{
enum { value = floor_log2<n, floor_log2_selector<n, lower, upper>::middle, upper>::value };
enum { value = meta_floor_log2<n, meta_floor_log2_selector<n, lower, upper>::middle, upper>::value };
};
template<unsigned int n, int lower, int upper>
struct floor_log2<n, lower, upper, floor_log2_terminate>
struct meta_floor_log2<n, lower, upper, meta_floor_log2_terminate>
{
enum { value = (n >= ((unsigned int)(1) << (lower+1))) ? lower+1 : lower };
};
template<unsigned int n, int lower, int upper>
struct floor_log2<n, lower, upper, floor_log2_bogus>
struct meta_floor_log2<n, lower, upper, meta_floor_log2_bogus>
{
// no value, error at compile time
};
@@ -609,11 +610,24 @@ struct floor_log2<n, lower, upper, floor_log2_bogus>
template<typename Scalar>
struct random_default_impl<Scalar, false, true>
{
typedef typename NumTraits<Scalar>::NonInteger NonInteger;
static inline Scalar run(const Scalar& x, const Scalar& y)
{
return x + Scalar((NonInteger(y)-x+1) * std::rand() / (RAND_MAX + NonInteger(1)));
{
using std::max;
using std::min;
typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
if(y<x)
return x;
std::size_t range = ScalarX(y)-ScalarX(x);
std::size_t offset = 0;
// rejection sampling
std::size_t divisor = (range+RAND_MAX-1)/(range+1);
std::size_t multiplier = (range+RAND_MAX-1)/std::size_t(RAND_MAX);
do {
offset = ( (std::size_t(std::rand()) * multiplier) / divisor );
} while (offset > range);
return Scalar(ScalarX(x) + offset);
}
static inline Scalar run()
@@ -621,7 +635,7 @@ struct random_default_impl<Scalar, false, true>
#ifdef EIGEN_MAKING_DOCS
return run(Scalar(NumTraits<Scalar>::IsSigned ? -10 : 0), Scalar(10));
#else
enum { rand_bits = floor_log2<(unsigned int)(RAND_MAX)+1>::value,
enum { rand_bits = meta_floor_log2<(unsigned int)(RAND_MAX)+1>::value,
scalar_bits = sizeof(Scalar) * CHAR_BIT,
shift = EIGEN_PLAIN_ENUM_MAX(0, int(rand_bits) - int(scalar_bits)),
offset = NumTraits<Scalar>::IsSigned ? (1 << (EIGEN_PLAIN_ENUM_MIN(rand_bits,scalar_bits)-1)) : 0

44
Eigen/src/Core/ProductEvaluators.h
View File

@@ -409,7 +409,8 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
LhsCoeffReadCost = LhsEtorType::CoeffReadCost,
RhsCoeffReadCost = RhsEtorType::CoeffReadCost,
CoeffReadCost = (InnerSize == Dynamic || LhsCoeffReadCost==Dynamic || RhsCoeffReadCost==Dynamic || NumTraits<Scalar>::AddCost==Dynamic || NumTraits<Scalar>::MulCost==Dynamic) ? Dynamic
CoeffReadCost = InnerSize==0 ? NumTraits<Scalar>::ReadCost
: (InnerSize == Dynamic || LhsCoeffReadCost==Dynamic || RhsCoeffReadCost==Dynamic || NumTraits<Scalar>::AddCost==Dynamic || NumTraits<Scalar>::MulCost==Dynamic) ? Dynamic
: InnerSize * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost)
+ (InnerSize - 1) * NumTraits<Scalar>::AddCost,
@@ -484,7 +485,7 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
{
PacketScalar res;
typedef etor_product_packet_impl<Flags&RowMajorBit ? RowMajor : ColMajor,
Unroll ? InnerSize-1 : Dynamic,
Unroll ? InnerSize : Dynamic,
LhsEtorType, RhsEtorType, PacketScalar, LoadMode> PacketImpl;
PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res);
@@ -527,7 +528,7 @@ struct etor_product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, Load
static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
{
etor_product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
res = pmadd(pset1<Packet>(lhs.coeff(row, UnrollingIndex)), rhs.template packet<LoadMode>(UnrollingIndex, col), res);
res = pmadd(pset1<Packet>(lhs.coeff(row, UnrollingIndex-1)), rhs.template packet<LoadMode>(UnrollingIndex-1, col), res);
}
};
@@ -537,12 +538,12 @@ struct etor_product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, Load
static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
{
etor_product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
res = pmadd(lhs.template packet<LoadMode>(row, UnrollingIndex), pset1<Packet>(rhs.coeff(UnrollingIndex, col)), res);
res = pmadd(lhs.template packet<LoadMode>(row, UnrollingIndex-1), pset1<Packet>(rhs.coeff(UnrollingIndex-1, col)), res);
}
};
template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
struct etor_product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
struct etor_product_packet_impl<RowMajor, 1, Lhs, Rhs, Packet, LoadMode>
{
static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
{
@@ -551,7 +552,7 @@ struct etor_product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
};
template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
struct etor_product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
struct etor_product_packet_impl<ColMajor, 1, Lhs, Rhs, Packet, LoadMode>
{
static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
{
@@ -559,14 +560,31 @@ struct etor_product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
}
};
template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
struct etor_product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
{
static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
{
res = pset1<Packet>(0);
}
};
template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
struct etor_product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
{
static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
{
res = pset1<Packet>(0);
}
};
template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
struct etor_product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
{
static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
{
eigen_assert(innerDim>0 && "you are using a non initialized matrix");
res = pmul(pset1<Packet>(lhs.coeff(row, 0)),rhs.template packet<LoadMode>(0, col));
for(Index i = 1; i < innerDim; ++i)
res = pset1<Packet>(0);
for(Index i = 0; i < innerDim; ++i)
res = pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode>(i, col), res);
}
};
@@ -576,9 +594,8 @@ struct etor_product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
{
static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
{
eigen_assert(innerDim>0 && "you are using a non initialized matrix");
res = pmul(lhs.template packet<LoadMode>(row, 0), pset1<Packet>(rhs.coeff(0, col)));
for(Index i = 1; i < innerDim; ++i)
res = pset1<Packet>(0);
for(Index i = 0; i < innerDim; ++i)
res = pmadd(lhs.template packet<LoadMode>(row, i), pset1<Packet>(rhs.coeff(i, col)), res);
}
};
@@ -678,8 +695,7 @@ public:
//_Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagFlags)&PacketAccessBit))),
_Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))),
_LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0,
Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit
//(int(MatrixFlags)&int(DiagFlags)&AlignedBit),
Flags = ((HereditaryBits|_LinearAccessMask|AlignedBit) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0)
};
diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag)

73
Eigen/src/Core/Reverse.h
View File

@@ -200,17 +200,82 @@ DenseBase<Derived>::reverse() const
* In most cases it is probably better to simply use the reversed expression
* of a matrix. However, when reversing the matrix data itself is really needed,
* then this "in-place" version is probably the right choice because it provides
* the following additional features:
* the following additional benefits:
* - less error prone: doing the same operation with .reverse() requires special care:
* \code m = m.reverse().eval(); \endcode
* - this API allows to avoid creating a temporary (the current implementation creates a temporary, but that could be avoided using swap)
* - this API enables reverse operations without the need for a temporary
* - it allows future optimizations (cache friendliness, etc.)
*
* \sa reverse() */
* \sa VectorwiseOp::reverseInPlace(), reverse() */
template<typename Derived>
inline void DenseBase<Derived>::reverseInPlace()
{
derived() = derived().reverse().eval();
if(cols()>rows())
{
Index half = cols()/2;
leftCols(half).swap(rightCols(half).reverse());
if((cols()%2)==1)
{
Index half2 = rows()/2;
col(half).head(half2).swap(col(half).tail(half2).reverse());
}
}
else
{
Index half = rows()/2;
topRows(half).swap(bottomRows(half).reverse());
if((rows()%2)==1)
{
Index half2 = cols()/2;
row(half).head(half2).swap(row(half).tail(half2).reverse());
}
}
}
namespace internal {
template<int Direction>
struct vectorwise_reverse_inplace_impl;
template<>
struct vectorwise_reverse_inplace_impl<Vertical>
{
template<typename ExpressionType>
static void run(ExpressionType &xpr)
{
Index half = xpr.rows()/2;
xpr.topRows(half).swap(xpr.bottomRows(half).colwise().reverse());
}
};
template<>
struct vectorwise_reverse_inplace_impl<Horizontal>
{
template<typename ExpressionType>
static void run(ExpressionType &xpr)
{
Index half = xpr.cols()/2;
xpr.leftCols(half).swap(xpr.rightCols(half).rowwise().reverse());
}
};
} // end namespace internal
/** This is the "in place" version of VectorwiseOp::reverse: it reverses each column or row of \c *this.
*
* In most cases it is probably better to simply use the reversed expression
* of a matrix. However, when reversing the matrix data itself is really needed,
* then this "in-place" version is probably the right choice because it provides
* the following additional benefits:
* - less error prone: doing the same operation with .reverse() requires special care:
* \code m = m.reverse().eval(); \endcode
* - this API enables reverse operations without the need for a temporary
*
* \sa DenseBase::reverseInPlace(), reverse() */
template<typename ExpressionType, int Direction>
void VectorwiseOp<ExpressionType,Direction>::reverseInPlace()
{
internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived());
}
} // end namespace Eigen

8
Eigen/src/Core/Swap.h
View File

@@ -38,13 +38,17 @@ public:
template<int StoreMode, int LoadMode>
void assignPacket(Index row, Index col)
{
m_functor.template swapPacket<StoreMode,LoadMode,PacketScalar>(&m_dst.coeffRef(row,col), &const_cast<SrcEvaluatorTypeT&>(m_src).coeffRef(row,col));
PacketScalar tmp = m_src.template packet<LoadMode>(row,col);
const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(row,col, m_dst.template packet<StoreMode>(row,col));
m_dst.template writePacket<StoreMode>(row,col,tmp);
}
template<int StoreMode, int LoadMode>
void assignPacket(Index index)
{
m_functor.template swapPacket<StoreMode,LoadMode,PacketScalar>(&m_dst.coeffRef(index), &const_cast<SrcEvaluatorTypeT&>(m_src).coeffRef(index));
PacketScalar tmp = m_src.template packet<LoadMode>(index);
const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(index, m_dst.template packet<StoreMode>(index));
m_dst.template writePacket<StoreMode>(index,tmp);
}
// TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael)

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

@@ -562,6 +562,8 @@ template<typename ExpressionType, int Direction> class VectorwiseOp
void normalize() {
m_matrix = this->normalized();
}
inline void reverseInPlace();
/////////// Geometry module ///////////

16
Eigen/src/Core/arch/CUDA/PacketMath.h
View File

@@ -197,21 +197,21 @@ EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Unaligned>(cons
}
#endif
template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, int stride) {
template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, Index stride) {
return make_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]);
}
template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, int stride) {
template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, Index stride) {
return make_double2(from[0*stride], from[1*stride]);
}
template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, int stride) {
template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, Index stride) {
to[stride*0] = from.x;
to[stride*1] = from.y;
to[stride*2] = from.z;
to[stride*3] = from.w;
}
template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, int stride) {
template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, Index stride) {
to[stride*0] = from.x;
to[stride*1] = from.y;
}
@@ -245,14 +245,14 @@ template<> EIGEN_DEVICE_FUNC inline double predux_min<double2>(const double2& a)
}
template<> EIGEN_DEVICE_FUNC inline float4 pabs<float4>(const float4& a) {
return make_float4(fabs(a.x), fabs(a.y), fabs(a.z), fabs(a.w));
return make_float4(fabsf(a.x), fabsf(a.y), fabsf(a.z), fabsf(a.w));
}
template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) {
return make_double2(abs(a.x), abs(a.y));
return make_double2(fabs(a.x), fabs(a.y));
}
template<> EIGEN_DEVICE_FUNC inline void
EIGEN_DEVICE_FUNC inline void
ptranspose(PacketBlock<float4,4>& kernel) {
double tmp = kernel.packet[0].y;
kernel.packet[0].y = kernel.packet[1].x;
@@ -279,7 +279,7 @@ ptranspose(PacketBlock<float4,4>& kernel) {
kernel.packet[3].z = tmp;
}
template<> EIGEN_DEVICE_FUNC inline void
EIGEN_DEVICE_FUNC inline void
ptranspose(PacketBlock<double2,2>& kernel) {
double tmp = kernel.packet[0].y;
kernel.packet[0].y = kernel.packet[1].x;

110
Eigen/src/Core/arch/NEON/BlockingSizesLookupTables.h Normal file
View File

@@ -0,0 +1,110 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2015 Benoit Jacob <benoitjacob@google.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_NEON_BLOCKING_SIZES_LOOKUP_TABLES_H
#define EIGEN_NEON_BLOCKING_SIZES_LOOKUP_TABLES_H
namespace Eigen {
namespace internal {
/* The following lookup table was generated from measurements on a Nexus 5,
* which has a Qualcomm Krait 400 CPU. This is very representative of current
* 32bit (ARMv7) Android devices. On the other hand, I don't know how
* representative that is outside of these conditions. Accordingly,
* let's only use this lookup table on ARM 32bit on Android for now.
*
* Measurements were single-threaded, with Scalar=float, compiled with
* -mfpu=neon-vfpv4, so the pmadd instruction used was VFMA.F32.
*
* The device was cooled, allowing it to run a the max clock speed throughout.
* This may not be representative of real-world thermal conditions.
*
* The benchmark attempted to flush caches to test cold-cache performance.
*/
#if EIGEN_ARCH_ARM && EIGEN_OS_ANDROID
template<>
struct BlockingSizesLookupTable<float, float> {
static const size_t BaseSize = 16;
static const size_t NumSizes = 8;
static const unsigned short* Data() {
static const unsigned short data[512] = {
0x444, 0x445, 0x446, 0x447, 0x448, 0x449, 0x447, 0x447,
0x454, 0x455, 0x456, 0x457, 0x458, 0x459, 0x45a, 0x456,
0x464, 0x465, 0x466, 0x467, 0x468, 0x469, 0x46a, 0x467,
0x474, 0x475, 0x476, 0x467, 0x478, 0x479, 0x476, 0x478,
0x474, 0x475, 0x476, 0x477, 0x478, 0x479, 0x476, 0x476,
0x474, 0x475, 0x476, 0x477, 0x478, 0x479, 0x496, 0x488,
0x474, 0x475, 0x476, 0x4a6, 0x496, 0x496, 0x495, 0x4a6,
0x474, 0x475, 0x466, 0x4a6, 0x497, 0x4a5, 0x496, 0x4a5,
0x544, 0x545, 0x546, 0x547, 0x548, 0x549, 0x54a, 0x54b,
0x554, 0x555, 0x556, 0x557, 0x558, 0x559, 0x55a, 0x55b,
0x564, 0x565, 0x566, 0x567, 0x568, 0x569, 0x56a, 0x56b,
0x564, 0x565, 0x566, 0x567, 0x568, 0x569, 0x56a, 0x576,
0x564, 0x565, 0x566, 0x567, 0x568, 0x569, 0x56a, 0x587,
0x564, 0x565, 0x566, 0x567, 0x596, 0x596, 0x596, 0x597,
0x574, 0x565, 0x566, 0x596, 0x596, 0x5a6, 0x5a6, 0x5a6,
0x564, 0x565, 0x5a6, 0x596, 0x5a6, 0x5a6, 0x5a6, 0x5a6,
0x644, 0x645, 0x646, 0x647, 0x648, 0x649, 0x64a, 0x64b,
0x644, 0x655, 0x656, 0x657, 0x658, 0x659, 0x65a, 0x65b,
0x664, 0x665, 0x666, 0x667, 0x668, 0x669, 0x65a, 0x667,
0x654, 0x665, 0x676, 0x677, 0x678, 0x679, 0x67a, 0x675,
0x684, 0x675, 0x686, 0x687, 0x688, 0x688, 0x687, 0x686,
0x664, 0x685, 0x666, 0x677, 0x697, 0x696, 0x697, 0x697,
0x664, 0x665, 0x696, 0x696, 0x685, 0x6a6, 0x696, 0x696,
0x664, 0x675, 0x686, 0x696, 0x6a6, 0x696, 0x696, 0x696,
0x744, 0x745, 0x746, 0x747, 0x748, 0x749, 0x74a, 0x747,
0x754, 0x755, 0x756, 0x757, 0x758, 0x759, 0x75a, 0x757,
0x764, 0x765, 0x756, 0x767, 0x768, 0x759, 0x75a, 0x766,
0x744, 0x755, 0x766, 0x777, 0x768, 0x759, 0x778, 0x777,
0x744, 0x745, 0x766, 0x777, 0x788, 0x786, 0x786, 0x788,
0x754, 0x755, 0x766, 0x787, 0x796, 0x796, 0x787, 0x796,
0x684, 0x695, 0x696, 0x6a6, 0x795, 0x786, 0x795, 0x796,
0x684, 0x695, 0x696, 0x795, 0x786, 0x796, 0x795, 0x796,
0x844, 0x845, 0x846, 0x847, 0x848, 0x849, 0x848, 0x848,
0x844, 0x855, 0x846, 0x847, 0x848, 0x849, 0x855, 0x857,
0x844, 0x845, 0x846, 0x857, 0x848, 0x859, 0x866, 0x865,
0x844, 0x855, 0x846, 0x847, 0x878, 0x859, 0x877, 0x877,
0x844, 0x855, 0x846, 0x867, 0x886, 0x887, 0x885, 0x886,
0x784, 0x785, 0x786, 0x877, 0x897, 0x885, 0x896, 0x896,
0x684, 0x695, 0x686, 0x886, 0x885, 0x885, 0x886, 0x896,
0x694, 0x6a5, 0x6a6, 0x885, 0x885, 0x886, 0x896, 0x896,
0x944, 0x945, 0x946, 0x947, 0x948, 0x847, 0x847, 0x848,
0x954, 0x855, 0x856, 0x947, 0x858, 0x857, 0x858, 0x858,
0x944, 0x945, 0x946, 0x867, 0x948, 0x866, 0x867, 0x867,
0x944, 0x975, 0x976, 0x877, 0x877, 0x877, 0x877, 0x877,
0x784, 0x785, 0x886, 0x887, 0x886, 0x887, 0x887, 0x887,
0x784, 0x785, 0x786, 0x796, 0x887, 0x897, 0x896, 0x896,
0x684, 0x695, 0x6a6, 0x886, 0x886, 0x896, 0x896, 0x896,
0x6a4, 0x6a5, 0x696, 0x896, 0x886, 0x896, 0x896, 0x896,
0xa44, 0xa45, 0xa46, 0xa47, 0x847, 0x848, 0x847, 0x848,
0xa44, 0xa45, 0x856, 0x857, 0x857, 0x857, 0x857, 0x857,
0xa44, 0xa65, 0x866, 0x867, 0x867, 0x867, 0x867, 0x867,
0x774, 0x875, 0x876, 0x877, 0x877, 0x877, 0x877, 0x877,
0x784, 0x785, 0x886, 0x887, 0x887, 0x887, 0x887, 0x887,
0x784, 0x785, 0x786, 0x787, 0x887, 0x896, 0x897, 0x897,
0x684, 0x6a5, 0x696, 0x886, 0x886, 0x896, 0x896, 0x896,
0x684, 0x6a5, 0x6a5, 0x886, 0x886, 0x896, 0x896, 0x896,
0xb44, 0x845, 0x846, 0x847, 0x847, 0x945, 0x846, 0x946,
0xb54, 0x855, 0x856, 0x857, 0x857, 0x856, 0x857, 0x856,
0x864, 0x865, 0x866, 0x867, 0x867, 0x866, 0x866, 0x867,
0x864, 0x875, 0x876, 0x877, 0x877, 0x877, 0x877, 0x877,
0x784, 0x885, 0x886, 0x787, 0x887, 0x887, 0x887, 0x887,
0x784, 0x785, 0x786, 0x796, 0x886, 0x897, 0x897, 0x897,
0x684, 0x695, 0x696, 0x886, 0x896, 0x896, 0x896, 0x896,
0x684, 0x685, 0x696, 0xb57, 0x896, 0x896, 0x896, 0x896
};
return data;
}
};
#endif
}
}
#endif // EIGEN_NEON_BLOCKING_SIZES_LOOKUP_TABLES_H

8
Eigen/src/Core/functors/AssignmentFunctors.h
View File

@@ -150,14 +150,6 @@ template<typename Scalar> struct swap_assign_op {
swap(a,const_cast<Scalar&>(b));
#endif
}
template<int LhsAlignment, int RhsAlignment, typename Packet>
EIGEN_STRONG_INLINE void swapPacket(Scalar* a, Scalar* b) const
{
Packet tmp = internal::ploadt<Packet,RhsAlignment>(b);
internal::pstoret<Scalar,Packet,RhsAlignment>(b, internal::ploadt<Packet,LhsAlignment>(a));
internal::pstoret<Scalar,Packet,LhsAlignment>(a, tmp);
}
};
template<typename Scalar>
struct functor_traits<swap_assign_op<Scalar> > {

151
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@@ -25,21 +25,31 @@ inline std::ptrdiff_t manage_caching_sizes_helper(std::ptrdiff_t a, std::ptrdiff
return a<=0 ? b : a;
}
#if EIGEN_ARCH_i386_OR_x86_64
const std::ptrdiff_t defaultL1CacheSize = 32*1024;
const std::ptrdiff_t defaultL2CacheSize = 256*1024;
const std::ptrdiff_t defaultL3CacheSize = 2*1024*1024;
#else
const std::ptrdiff_t defaultL1CacheSize = 16*1024;
const std::ptrdiff_t defaultL2CacheSize = 512*1024;
const std::ptrdiff_t defaultL3CacheSize = 512*1024;
#endif
/** \internal */
inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff_t* l2, std::ptrdiff_t* l3)
{
static bool m_cache_sizes_initialized = false;
static std::ptrdiff_t m_l1CacheSize = 32*1024;
static std::ptrdiff_t m_l2CacheSize = 256*1024;
static std::ptrdiff_t m_l3CacheSize = 2*1024*1024;
static std::ptrdiff_t m_l1CacheSize = 0;
static std::ptrdiff_t m_l2CacheSize = 0;
static std::ptrdiff_t m_l3CacheSize = 0;
if(!m_cache_sizes_initialized)
{
int l1CacheSize, l2CacheSize, l3CacheSize;
queryCacheSizes(l1CacheSize, l2CacheSize, l3CacheSize);
m_l1CacheSize = manage_caching_sizes_helper(l1CacheSize, 8*1024);
m_l2CacheSize = manage_caching_sizes_helper(l2CacheSize, 256*1024);
m_l3CacheSize = manage_caching_sizes_helper(l3CacheSize, 8*1024*1024);
m_l1CacheSize = manage_caching_sizes_helper(l1CacheSize, defaultL1CacheSize);
m_l2CacheSize = manage_caching_sizes_helper(l2CacheSize, defaultL2CacheSize);
m_l3CacheSize = manage_caching_sizes_helper(l3CacheSize, defaultL3CacheSize);
m_cache_sizes_initialized = true;
}
@@ -64,45 +74,23 @@ inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff
}
}
/** \brief Computes the blocking parameters for a m x k times k x n matrix product
*
* \param[in,out] k Input: the third dimension of the product. Output: the blocking size along the same dimension.
* \param[in,out] m Input: the number of rows of the left hand side. Output: the blocking size along the same dimension.
* \param[in,out] n Input: the number of columns of the right hand side. Output: the blocking size along the same dimension.
*
* Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar,
* this function computes the blocking size parameters along the respective dimensions
* for matrix products and related algorithms. The blocking sizes depends on various
* parameters:
* - the L1 and L2 cache sizes,
* - the register level blocking sizes defined by gebp_traits,
* - the number of scalars that fit into a packet (when vectorization is enabled).
*
* \sa setCpuCacheSizes */
/* Helper for computeProductBlockingSizes.
*
* Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar,
* this function computes the blocking size parameters along the respective dimensions
* for matrix products and related algorithms. The blocking sizes depends on various
* parameters:
* - the L1 and L2 cache sizes,
* - the register level blocking sizes defined by gebp_traits,
* - the number of scalars that fit into a packet (when vectorization is enabled).
*
* \sa setCpuCacheSizes */
template<typename LhsScalar, typename RhsScalar, int KcFactor>
void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1)
void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index num_threads = 1)
{
typedef gebp_traits<LhsScalar,RhsScalar> Traits;
#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
if (EIGEN_TEST_SPECIFIC_BLOCKING_SIZES) {
EIGEN_UNUSED_VARIABLE(num_threads);
enum {
kr = 8,
mr = Traits::mr,
nr = Traits::nr
};
k = std::min<Index>(k, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_K);
if (k > kr) k -= k % kr;
m = std::min<Index>(m, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_M);
if (m > mr) m -= m % mr;
n = std::min<Index>(n, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_N);
if (n > nr) n -= n % nr;
return;
}
#endif
// Explanations:
// Let's recall that the product algorithms form mc x kc vertical panels A' on the lhs and
// kc x nc blocks B' on the rhs. B' has to fit into L2/L3 cache. Moreover, A' is processed
@@ -261,16 +249,69 @@ void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads
actual_lm = l2;
max_mc = 576;
}
Index mc = (std::min<Index>)(actual_lm/(3*k*sizeof(LhsScalar)), max_mc);
if (mc > Traits::mr) mc -= mc % Traits::mr;
else if (mc==0) return;
m = (m%mc)==0 ? mc
: (mc - Traits::mr * ((mc/*-1*/-(m%mc))/(Traits::mr*(m/mc+1))));
}
}
}
inline bool useSpecificBlockingSizes(Index& k, Index& m, Index& n)
{
#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
if (EIGEN_TEST_SPECIFIC_BLOCKING_SIZES) {
k = std::min<Index>(k, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_K);
m = std::min<Index>(m, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_M);
n = std::min<Index>(n, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_N);
return true;
}
#else
EIGEN_UNUSED_VARIABLE(k)
EIGEN_UNUSED_VARIABLE(m)
EIGEN_UNUSED_VARIABLE(n)
#endif
return false;
}
/** \brief Computes the blocking parameters for a m x k times k x n matrix product
*
* \param[in,out] k Input: the third dimension of the product. Output: the blocking size along the same dimension.
* \param[in,out] m Input: the number of rows of the left hand side. Output: the blocking size along the same dimension.
* \param[in,out] n Input: the number of columns of the right hand side. Output: the blocking size along the same dimension.
*
* Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar,
* this function computes the blocking size parameters along the respective dimensions
* for matrix products and related algorithms.
*
* The blocking size parameters may be evaluated:
* - either by a heuristic based on cache sizes;
* - or using a precomputed lookup table;
* - or using fixed prescribed values (for testing purposes).
*
* \sa setCpuCacheSizes */
template<typename LhsScalar, typename RhsScalar, int KcFactor>
void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1)
{
if (!useSpecificBlockingSizes(k, m, n)) {
if (!lookupBlockingSizesFromTable<LhsScalar, RhsScalar>(k, m, n, num_threads)) {
evaluateProductBlockingSizesHeuristic<LhsScalar, RhsScalar, KcFactor>(k, m, n, num_threads);
}
}
typedef gebp_traits<LhsScalar,RhsScalar> Traits;
enum {
kr = 8,
mr = Traits::mr,
nr = Traits::nr
};
if (k > kr) k -= k % kr;
if (m > mr) m -= m % mr;
if (n > nr) n -= n % nr;
}
template<typename LhsScalar, typename RhsScalar>
inline void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1)
{
@@ -339,11 +380,14 @@ public:
nr = 4,
// register block size along the M direction (currently, this one cannot be modified)
default_mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize,
#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
// we assume 16 registers
mr = 3*LhsPacketSize,
// See bug 992, if the scalar type is not vectorizable but that EIGEN_HAS_SINGLE_INSTRUCTION_MADD is defined,
// then using 3*LhsPacketSize triggers non-implemented paths in syrk.
mr = Vectorizable ? 3*LhsPacketSize : default_mr,
#else
mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize,
mr = default_mr,
#endif
LhsProgress = LhsPacketSize,
@@ -974,12 +1018,11 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
// Blocking sizes, i.e., 'depth' has been computed so that the micro horizontal panel of the lhs fit in L1.
// However, if depth is too small, we can extend the number of rows of these horizontal panels.
// This actual number of rows is computed as follow:
const Index l1 = 32*1024; // in Bytes, TODO, l1 should be passed to this function.
#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function.
// The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size
// suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess),
// or because we are testing specific blocking sizes.
const Index actual_panel_rows = (3*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 3*LhsProgress) ));
#else
const Index actual_panel_rows = (3*LhsProgress) * ( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 3*LhsProgress) );
#endif
for(Index i1=0; i1<peeled_mc3; i1+=actual_panel_rows)
{
const Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc3);
@@ -1211,12 +1254,12 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
//---------- Process 2 * LhsProgress rows at once ----------
if(mr>=2*Traits::LhsProgress)
{
const Index l1 = 32*1024; // in Bytes, TODO, l1 should be passed to this function.
#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function.
// The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size
// suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess),
// or because we are testing specific blocking sizes.
Index actual_panel_rows = (2*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 2*LhsProgress) ));
#else
Index actual_panel_rows = (2*LhsProgress) * ( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 2*LhsProgress) );
#endif
for(Index i1=peeled_mc3; i1<peeled_mc2; i1+=actual_panel_rows)
{
Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc2);

97
Eigen/src/Core/products/LookupBlockingSizesTable.h Normal file
View File

@@ -0,0 +1,97 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2015 Benoit Jacob <benoitjacob@google.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_LOOKUP_BLOCKING_SIZES_TABLE_H
#define EIGEN_LOOKUP_BLOCKING_SIZES_TABLE_H
namespace Eigen {
namespace internal {
template <typename LhsScalar,
typename RhsScalar,
bool HasLookupTable = BlockingSizesLookupTable<LhsScalar, RhsScalar>::NumSizes != 0 >
struct LookupBlockingSizesFromTableImpl
{
static bool run(Index&, Index&, Index&, Index)
{
return false;
}
};
inline size_t floor_log2_helper(unsigned short& x, size_t offset)
{
unsigned short y = x >> offset;
if (y) {
x = y;
return offset;
} else {
return 0;
}
}
inline size_t floor_log2(unsigned short x)
{
return floor_log2_helper(x, 8)
+ floor_log2_helper(x, 4)
+ floor_log2_helper(x, 2)
+ floor_log2_helper(x, 1);
}
inline size_t ceil_log2(unsigned short x)
{
return x > 1 ? floor_log2(x - 1) + 1 : 0;
}
template <typename LhsScalar,
typename RhsScalar>
struct LookupBlockingSizesFromTableImpl<LhsScalar, RhsScalar, true>
{
static bool run(Index& k, Index& m, Index& n, Index)
{
using std::min;
using std::max;
typedef BlockingSizesLookupTable<LhsScalar, RhsScalar> Table;
const unsigned short minsize = Table::BaseSize;
const unsigned short maxsize = minsize << (Table::NumSizes - 1);
const unsigned short k_clamped = max<unsigned short>(minsize, min<Index>(k, maxsize));
const unsigned short m_clamped = max<unsigned short>(minsize, min<Index>(m, maxsize));
const unsigned short n_clamped = max<unsigned short>(minsize, min<Index>(n, maxsize));
const size_t k_index = ceil_log2(k_clamped / minsize);
const size_t m_index = ceil_log2(m_clamped / minsize);
const size_t n_index = ceil_log2(n_clamped / minsize);
const size_t index = n_index + Table::NumSizes * (m_index + Table::NumSizes * k_index);
const unsigned short table_entry = Table::Data()[index];
k = min<Index>(k, 1 << ((table_entry & 0xf00) >> 8));
m = min<Index>(m, 1 << ((table_entry & 0x0f0) >> 4));
n = min<Index>(n, 1 << ((table_entry & 0x00f) >> 0));
return true;
}
};
template <typename LhsScalar,
typename RhsScalar>
bool lookupBlockingSizesFromTable(Index& k, Index& m, Index& n, Index num_threads)
{
if (num_threads > 1) {
// We don't currently have lookup tables recorded for multithread performance,
// and we have confirmed experimentally that our single-thread-recorded LUTs are
// poor for multithread performance, and our LUTs don't currently contain
// any annotation about multithread status (FIXME - we need that).
// So for now, we just early-return here.
return false;
}
return LookupBlockingSizesFromTableImpl<LhsScalar, RhsScalar>::run(k, m, n, num_threads);
}
}
}
#endif // EIGEN_LOOKUP_BLOCKING_SIZES_TABLE_H

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

@@ -214,7 +214,7 @@ class blas_data_mapper {
}
template<typename SubPacket>
EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, SubPacket p) const {
EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, const SubPacket &p) const {
pscatter<Scalar, SubPacket>(&operator()(i, j), p, m_stride);
}

8
Eigen/src/Core/util/ForwardDeclarations.h
View File

@@ -287,6 +287,14 @@ struct stem_function
typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
typedef ComplexScalar type(ComplexScalar, int);
};
template <typename LhsScalar,
typename RhsScalar>
struct BlockingSizesLookupTable
{
static const size_t NumSizes = 0;
};
}
} // end namespace Eigen

6
Eigen/src/Core/util/Macros.h
View File

@@ -213,7 +213,8 @@
#endif
/// \internal EIGEN_OS_ANDROID set to 1 if the OS is Android
#if defined(__ANDROID__)
// note: ANDROID is defined when using ndk_build, __ANDROID__ is defined when using a standalone toolchain.
#if defined(__ANDROID__) || defined(ANDROID)
#define EIGEN_OS_ANDROID 1
#else
#define EIGEN_OS_ANDROID 0
@@ -318,6 +319,9 @@
// Defined the boundary (in bytes) on which the data needs to be aligned. Note
// that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be
// aligned at all regardless of the value of this #define.
// TODO should be renamed EIGEN_MAXIMAL_ALIGN_BYTES,
// for instance with AVX 1 EIGEN_MAXIMAL_ALIGN_BYTES=32 while for 'int' 16 bytes alignment is always enough,
// and 16 bytes alignment is also enough for Vector4f.
#define EIGEN_ALIGN_BYTES 16
#ifdef EIGEN_DONT_ALIGN

5
Eigen/src/Core/util/XprHelper.h
View File

@@ -159,13 +159,16 @@ class compute_matrix_evaluator_flags
enum {
row_major_bit = Options&RowMajor ? RowMajorBit : 0,
is_dynamic_size_storage = MaxRows==Dynamic || MaxCols==Dynamic,
// TODO: should check for smaller packet types once we can handle multi-sized packet types
align_bytes = int(packet_traits<Scalar>::size) * sizeof(Scalar),
aligned_bit =
(
((Options&DontAlign)==0)
&& (
#if EIGEN_ALIGN_STATICALLY
((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % EIGEN_ALIGN_BYTES) == 0))
((!is_dynamic_size_storage) && (((MaxCols*MaxRows*int(sizeof(Scalar))) % align_bytes) == 0))
#else
0
#endif

8
Eigen/src/Eigenvalues/ComplexEigenSolver.h
View File

@@ -234,6 +234,12 @@ template<typename _MatrixType> class ComplexEigenSolver
}
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
EigenvectorType m_eivec;
EigenvalueType m_eivalues;
ComplexSchur<MatrixType> m_schur;
@@ -251,6 +257,8 @@ template<typename MatrixType>
ComplexEigenSolver<MatrixType>&
ComplexEigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvectors)
{
check_template_parameters();
// this code is inspired from Jampack
eigen_assert(matrix.cols() == matrix.rows());

13
Eigen/src/Eigenvalues/EigenSolver.h
View File

@@ -299,6 +299,13 @@ template<typename _MatrixType> class EigenSolver
void doComputeEigenvectors();
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL);
}
MatrixType m_eivec;
EigenvalueType m_eivalues;
bool m_isInitialized;
@@ -366,6 +373,8 @@ template<typename MatrixType>
EigenSolver<MatrixType>&
EigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvectors)
{
check_template_parameters();
using std::sqrt;
using std::abs;
using numext::isfinite;
@@ -408,7 +417,7 @@ EigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvect
{
Scalar t0 = m_matT.coeff(i+1, i);
Scalar t1 = m_matT.coeff(i, i+1);
Scalar maxval = numext::maxi(abs(p),numext::maxi(abs(t0),abs(t1)));
Scalar maxval = numext::maxi<Scalar>(abs(p),numext::maxi<Scalar>(abs(t0),abs(t1)));
t0 /= maxval;
t1 /= maxval;
Scalar p0 = p/maxval;
@@ -599,7 +608,7 @@ void EigenSolver<MatrixType>::doComputeEigenvectors()
}
// Overflow control
Scalar t = numext::maxi(abs(m_matT.coeff(i,n-1)),abs(m_matT.coeff(i,n)));
Scalar t = numext::maxi<Scalar>(abs(m_matT.coeff(i,n-1)),abs(m_matT.coeff(i,n)));
if ((eps * t) * t > Scalar(1))
m_matT.block(i, n-1, size-i, 2) /= t;

9
Eigen/src/Eigenvalues/GeneralizedEigenSolver.h
View File

@@ -263,6 +263,13 @@ template<typename _MatrixType> class GeneralizedEigenSolver
}
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL);
}
MatrixType m_eivec;
ComplexVectorType m_alphas;
VectorType m_betas;
@@ -290,6 +297,8 @@ template<typename MatrixType>
GeneralizedEigenSolver<MatrixType>&
GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixType& B, bool computeEigenvectors)
{
check_template_parameters();
using std::sqrt;
using std::abs;
eigen_assert(A.cols() == A.rows() && B.cols() == A.rows() && B.cols() == B.rows());

12
Eigen/src/Eigenvalues/RealQZ.h
View File

@@ -240,10 +240,10 @@ namespace Eigen {
m_S.coeffRef(i,j) = Scalar(0.0);
m_S.rightCols(dim-j-1).applyOnTheLeft(i-1,i,G.adjoint());
m_T.rightCols(dim-i+1).applyOnTheLeft(i-1,i,G.adjoint());
// update Q
if (m_computeQZ)
m_Q.applyOnTheRight(i-1,i,G);
}
// update Q
if (m_computeQZ)
m_Q.applyOnTheRight(i-1,i,G);
// kill T(i,i-1)
if(m_T.coeff(i,i-1)!=Scalar(0))
{
@@ -251,10 +251,10 @@ namespace Eigen {
m_T.coeffRef(i,i-1) = Scalar(0.0);
m_S.applyOnTheRight(i,i-1,G);
m_T.topRows(i).applyOnTheRight(i,i-1,G);
// update Z
if (m_computeQZ)
m_Z.applyOnTheLeft(i,i-1,G.adjoint());
}
// update Z
if (m_computeQZ)
m_Z.applyOnTheLeft(i,i-1,G.adjoint());
}
}
}

7
Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
View File

@@ -347,6 +347,11 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
static const int m_maxIterations = 30;
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
MatrixType m_eivec;
RealVectorType m_eivalues;
typename TridiagonalizationType::SubDiagonalType m_subdiag;
@@ -382,6 +387,8 @@ EIGEN_DEVICE_FUNC
SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
::compute(const MatrixType& matrix, int options)
{
check_template_parameters();
using std::abs;
eigen_assert(matrix.cols() == matrix.rows());
eigen_assert((options&~(EigVecMask|GenEigMask))==0

6
Eigen/src/Geometry/Quaternion.h
View File

@@ -161,8 +161,8 @@ class QuaternionBase : public RotationBase<Derived, 3>
bool isApprox(const QuaternionBase<OtherDerived>& other, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const
{ return coeffs().isApprox(other.coeffs(), prec); }
/** return the result vector of \a v through the rotation*/
EIGEN_STRONG_INLINE Vector3 _transformVector(Vector3 v) const;
/** return the result vector of \a v through the rotation*/
EIGEN_STRONG_INLINE Vector3 _transformVector(const Vector3& v) const;
/** \returns \c *this with scalar type casted to \a NewScalarType
*
@@ -462,7 +462,7 @@ EIGEN_STRONG_INLINE Derived& QuaternionBase<Derived>::operator*= (const Quaterni
*/
template <class Derived>
EIGEN_STRONG_INLINE typename QuaternionBase<Derived>::Vector3
QuaternionBase<Derived>::_transformVector(Vector3 v) const
QuaternionBase<Derived>::_transformVector(const Vector3& v) const
{
// Note that this algorithm comes from the optimization by hand
// of the conversion to a Matrix followed by a Matrix/Vector product.

8
Eigen/src/LU/FullPivLU.h
View File

@@ -390,6 +390,12 @@ template<typename _MatrixType> class FullPivLU
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
MatrixType m_lu;
PermutationPType m_p;
PermutationQType m_q;
@@ -434,6 +440,8 @@ FullPivLU<MatrixType>::FullPivLU(const MatrixType& matrix)
template<typename MatrixType>
FullPivLU<MatrixType>& FullPivLU<MatrixType>::compute(const MatrixType& matrix)
{
check_template_parameters();
// the permutations are stored as int indices, so just to be sure:
eigen_assert(matrix.rows()<=NumTraits<int>::highest() && matrix.cols()<=NumTraits<int>::highest());

8
Eigen/src/LU/PartialPivLU.h
View File

@@ -209,6 +209,12 @@ template<typename _MatrixType> class PartialPivLU
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
MatrixType m_lu;
PermutationType m_p;
TranspositionType m_rowsTranspositions;
@@ -425,6 +431,8 @@ void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, t
template<typename MatrixType>
PartialPivLU<MatrixType>& PartialPivLU<MatrixType>::compute(const MatrixType& matrix)
{
check_template_parameters();
// the row permutation is stored as int indices, so just to be sure:
eigen_assert(matrix.rows()<NumTraits<int>::highest());

19
Eigen/src/OrderingMethods/Amd.h
View File

@@ -137,22 +137,27 @@ void minimum_degree_ordering(SparseMatrix<Scalar,ColMajor,StorageIndex>& C, Perm
degree[i] = len[i]; // degree of node i
}
mark = internal::cs_wclear<StorageIndex>(0, 0, w, n); /* clear w */
elen[n] = -2; /* n is a dead element */
Cp[n] = -1; /* n is a root of assembly tree */
w[n] = 0; /* n is a dead element */
/* --- Initialize degree lists ------------------------------------------ */
for(i = 0; i < n; i++)
{
bool has_diag = false;
for(p = Cp[i]; p<Cp[i+1]; ++p)
if(Ci[p]==i)
{
has_diag = true;
break;
}
d = degree[i];
if(d == 0) /* node i is empty */
if(d == 1) /* node i is empty */
{
elen[i] = -2; /* element i is dead */
nel++;
Cp[i] = -1; /* i is a root of assembly tree */
w[i] = 0;
}
else if(d > dense) /* node i is dense */
else if(d > dense || !has_diag) /* node i is dense or has no structural diagonal element */
{
nv[i] = 0; /* absorb i into element n */
elen[i] = -1; /* node i is dead */
@@ -168,6 +173,10 @@ void minimum_degree_ordering(SparseMatrix<Scalar,ColMajor,StorageIndex>& C, Perm
}
}
elen[n] = -2; /* n is a dead element */
Cp[n] = -1; /* n is a root of assembly tree */
w[n] = 0; /* n is a dead element */
while (nel < n) /* while (selecting pivots) do */
{
/* --- Select node of minimum approximate degree -------------------- */

8
Eigen/src/QR/ColPivHouseholderQR.h
View File

@@ -398,6 +398,12 @@ template<typename _MatrixType> class ColPivHouseholderQR
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
MatrixType m_qr;
HCoeffsType m_hCoeffs;
PermutationType m_colsPermutation;
@@ -436,6 +442,8 @@ typename MatrixType::RealScalar ColPivHouseholderQR<MatrixType>::logAbsDetermina
template<typename MatrixType>
ColPivHouseholderQR<MatrixType>& ColPivHouseholderQR<MatrixType>::compute(const MatrixType& matrix)
{
check_template_parameters();
using std::abs;
Index rows = matrix.rows();
Index cols = matrix.cols();

8
Eigen/src/QR/FullPivHouseholderQR.h
View File

@@ -380,6 +380,12 @@ template<typename _MatrixType> class FullPivHouseholderQR
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
MatrixType m_qr;
HCoeffsType m_hCoeffs;
IntDiagSizeVectorType m_rows_transpositions;
@@ -419,6 +425,8 @@ typename MatrixType::RealScalar FullPivHouseholderQR<MatrixType>::logAbsDetermin
template<typename MatrixType>
FullPivHouseholderQR<MatrixType>& FullPivHouseholderQR<MatrixType>::compute(const MatrixType& matrix)
{
check_template_parameters();
using std::abs;
Index rows = matrix.rows();
Index cols = matrix.cols();

8
Eigen/src/QR/HouseholderQR.h
View File

@@ -196,6 +196,12 @@ template<typename _MatrixType> class HouseholderQR
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
MatrixType m_qr;
HCoeffsType m_hCoeffs;
RowVectorType m_temp;
@@ -348,6 +354,8 @@ void HouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) c
template<typename MatrixType>
HouseholderQR<MatrixType>& HouseholderQR<MatrixType>::compute(const MatrixType& matrix)
{
check_template_parameters();
Index rows = matrix.rows();
Index cols = matrix.cols();
Index size = (std::min)(rows,cols);

119
Eigen/src/SVD/BDCSVD.h
View File

@@ -84,6 +84,8 @@ public:
typedef Matrix<RealScalar, Dynamic, 1> VectorType;
typedef Array<RealScalar, Dynamic, 1> ArrayXr;
typedef Array<Index,1,Dynamic> ArrayXi;
typedef Ref<ArrayXr> ArrayRef;
typedef Ref<ArrayXi> IndicesRef;
/** \brief Default Constructor.
*
@@ -159,21 +161,23 @@ private:
void allocate(Index rows, Index cols, unsigned int computationOptions);
void divide(Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift);
void computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V);
void computeSingVals(const ArrayXr& col0, const ArrayXr& diag, const ArrayXi& perm, VectorType& singVals, ArrayXr& shifts, ArrayXr& mus);
void perturbCol0(const ArrayXr& col0, const ArrayXr& diag, const ArrayXi& perm, const VectorType& singVals, const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat);
void computeSingVecs(const ArrayXr& zhat, const ArrayXr& diag, const ArrayXi& perm, const VectorType& singVals, const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V);
void computeSingVals(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef& perm, VectorType& singVals, ArrayRef shifts, ArrayRef mus);
void perturbCol0(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef& perm, const VectorType& singVals, const ArrayRef& shifts, const ArrayRef& mus, ArrayRef zhat);
void computeSingVecs(const ArrayRef& zhat, const ArrayRef& diag, const IndicesRef& perm, const VectorType& singVals, const ArrayRef& shifts, const ArrayRef& mus, MatrixXr& U, MatrixXr& V);
void deflation43(Index firstCol, Index shift, Index i, Index size);
void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size);
void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift);
template<typename HouseholderU, typename HouseholderV, typename NaiveU, typename NaiveV>
void copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naivev);
static void structured_update(Block<MatrixXr,Dynamic,Dynamic> A, const MatrixXr &B, Index n1);
static RealScalar secularEq(RealScalar x, const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm, const ArrayXr& diagShifted, RealScalar shift);
void structured_update(Block<MatrixXr,Dynamic,Dynamic> A, const MatrixXr &B, Index n1);
static RealScalar secularEq(RealScalar x, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift);
protected:
MatrixXr m_naiveU, m_naiveV;
MatrixXr m_computed;
Index m_nRec;
ArrayXr m_workspace;
ArrayXi m_workspaceI;
int m_algoswap;
bool m_isTranspose, m_compU, m_compV;
@@ -212,6 +216,9 @@ void BDCSVD<MatrixType>::allocate(Index rows, Index cols, unsigned int computati
else m_naiveU = MatrixXr::Zero(2, m_diagSize + 1 );
if (m_compV) m_naiveV = MatrixXr::Zero(m_diagSize, m_diagSize);
m_workspace.resize((m_diagSize+1)*(m_diagSize+1)*3);
m_workspaceI.resize(3*m_diagSize);
}// end allocate
template<typename MatrixType>
@@ -223,6 +230,19 @@ BDCSVD<MatrixType>& BDCSVD<MatrixType>::compute(const MatrixType& matrix, unsign
allocate(matrix.rows(), matrix.cols(), computationOptions);
using std::abs;
//**** step -1 - If the problem is too small, directly falls back to JacobiSVD and return
if(matrix.cols() < m_algoswap)
{
// FIXME this line involves temporaries
JacobiSVD<MatrixType> jsvd(matrix,computationOptions);
if(computeU()) m_matrixU = jsvd.matrixU();
if(computeV()) m_matrixV = jsvd.matrixV();
m_singularValues = jsvd.singularValues();
m_nonzeroSingularValues = jsvd.nonzeroSingularValues();
m_isInitialized = true;
return *this;
}
//**** step 0 - Copy the input matrix and apply scaling to reduce over/under-flows
RealScalar scale = matrix.cwiseAbs().maxCoeff();
if(scale==RealScalar(0)) scale = RealScalar(1);
@@ -231,11 +251,13 @@ BDCSVD<MatrixType>& BDCSVD<MatrixType>::compute(const MatrixType& matrix, unsign
else copy = matrix/scale;
//**** step 1 - Bidiagonalization
// FIXME this line involves temporaries
internal::UpperBidiagonalization<MatrixX> bid(copy);
//**** step 2 - Divide & Conquer
m_naiveU.setZero();
m_naiveV.setZero();
// FIXME this line involves a temporary matrix
m_computed.topRows(m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose();
m_computed.template bottomRows<1>().setZero();
divide(0, m_diagSize - 1, 0, 0, 0);
@@ -257,6 +279,7 @@ BDCSVD<MatrixType>& BDCSVD<MatrixType>::compute(const MatrixType& matrix, unsign
break;
}
}
#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
// std::cout << "m_naiveU\n" << m_naiveU << "\n\n";
// std::cout << "m_naiveV\n" << m_naiveV << "\n\n";
@@ -279,14 +302,14 @@ void BDCSVD<MatrixType>::copyUV(const HouseholderU &householderU, const Househol
Index Ucols = m_computeThinU ? m_diagSize : householderU.cols();
m_matrixU = MatrixX::Identity(householderU.cols(), Ucols);
m_matrixU.topLeftCorner(m_diagSize, m_diagSize) = naiveV.template cast<Scalar>().topLeftCorner(m_diagSize, m_diagSize);
householderU.applyThisOnTheLeft(m_matrixU);
householderU.applyThisOnTheLeft(m_matrixU); // FIXME this line involves a temporary buffer
}
if (computeV())
{
Index Vcols = m_computeThinV ? m_diagSize : householderV.cols();
m_matrixV = MatrixX::Identity(householderV.cols(), Vcols);
m_matrixV.topLeftCorner(m_diagSize, m_diagSize) = naiveU.template cast<Scalar>().topLeftCorner(m_diagSize, m_diagSize);
householderV.applyThisOnTheLeft(m_matrixV);
householderV.applyThisOnTheLeft(m_matrixV); // FIXME this line involves a temporary buffer
}
}
@@ -307,7 +330,10 @@ void BDCSVD<MatrixType>::structured_update(Block<MatrixXr,Dynamic,Dynamic> A, co
// If the matrices are large enough, let's exploit the sparse structure of A by
// splitting it in half (wrt n1), and packing the non-zero columns.
Index n2 = n - n1;
MatrixXr A1(n1,n), A2(n2,n), B1(n,n), B2(n,n);
Map<MatrixXr> A1(m_workspace.data() , n1, n);
Map<MatrixXr> A2(m_workspace.data()+ n1*n, n2, n);
Map<MatrixXr> B1(m_workspace.data()+ n*n, n, n);
Map<MatrixXr> B2(m_workspace.data()+2*n*n, n, n);
Index k1=0, k2=0;
for(Index j=0; j<n; ++j)
{
@@ -329,7 +355,11 @@ void BDCSVD<MatrixType>::structured_update(Block<MatrixXr,Dynamic,Dynamic> A, co
A.bottomRows(n2).noalias() = A2.leftCols(k2) * B2.topRows(k2);
}
else
A *= B; // FIXME this requires a temporary
{
Map<MatrixXr,Aligned> tmp(m_workspace.data(),n,n);
tmp.noalias() = A*B;
A = tmp;
}
}
// The divide algorithm is done "in place", we are always working on subsets of the same matrix. The divide methods takes as argument the
@@ -360,7 +390,8 @@ void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW,
// matrices.
if (n < m_algoswap)
{
JacobiSVD<MatrixXr> b(m_computed.block(firstCol, firstCol, n + 1, n), ComputeFullU | (m_compV ? ComputeFullV : 0)) ;
// FIXME this line involves temporaries
JacobiSVD<MatrixXr> b(m_computed.block(firstCol, firstCol, n + 1, n), ComputeFullU | (m_compV ? ComputeFullV : 0));
if (m_compU)
m_naiveU.block(firstCol, firstCol, n + 1, n + 1).real() = b.matrixU();
else
@@ -438,7 +469,7 @@ void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW,
}
else
{
RealScalar q1 = (m_naiveU(0, firstCol + k));
RealScalar q1 = m_naiveU(0, firstCol + k);
// we shift Q1 to the right
for (Index i = firstCol + k - 1; i >= firstCol; i--)
m_naiveU(0, i + 1) = m_naiveU(0, i);
@@ -491,8 +522,14 @@ void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW,
assert(VofSVD.allFinite());
#endif
if (m_compU) structured_update(m_naiveU.block(firstCol, firstCol, n + 1, n + 1), UofSVD, (n+2)/2);
else m_naiveU.middleCols(firstCol, n + 1) *= UofSVD; // FIXME this requires a temporary, and exploit that there are 2 rows at compile time
if (m_compU)
structured_update(m_naiveU.block(firstCol, firstCol, n + 1, n + 1), UofSVD, (n+2)/2);
else
{
Map<Matrix<RealScalar,2,Dynamic>,Aligned> tmp(m_workspace.data(),2,n+1);
tmp.noalias() = m_naiveU.middleCols(firstCol, n+1) * UofSVD;
m_naiveU.middleCols(firstCol, n + 1) = tmp;
}
if (m_compV) structured_update(m_naiveV.block(firstRowW, firstColW, n, n), VofSVD, (n+1)/2);
@@ -517,10 +554,9 @@ void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW,
template <typename MatrixType>
void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V)
{
// TODO Get rid of these copies (?)
// FIXME at least preallocate them
ArrayXr col0 = m_computed.col(firstCol).segment(firstCol, n);
ArrayXr diag = m_computed.block(firstCol, firstCol, n, n).diagonal();
ArrayRef col0 = m_computed.col(firstCol).segment(firstCol, n);
m_workspace.head(n) = m_computed.block(firstCol, firstCol, n, n).diagonal();
ArrayRef diag = m_workspace.head(n);
diag(0) = 0;
// Allocate space for singular values and vectors
@@ -539,13 +575,14 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
Index actual_n = n;
while(actual_n>1 && diag(actual_n-1)==0) --actual_n;
Index m = 0; // size of the deflated problem
ArrayXi perm(actual_n);
for(Index k=0;k<actual_n;++k)
if(col0(k)!=0)
perm(m++) = k;
perm.conservativeResize(m);
m_workspaceI(m++) = k;
Map<ArrayXi> perm(m_workspaceI.data(),m);
ArrayXr shifts(n), mus(n), zhat(n);
Map<ArrayXr> shifts(m_workspace.data()+1*n, n);
Map<ArrayXr> mus(m_workspace.data()+2*n, n);
Map<ArrayXr> zhat(m_workspace.data()+3*n, n);
#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
std::cout << "computeSVDofM using:\n";
@@ -622,8 +659,8 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
// Reverse order so that singular values in increased order
// Because of deflation, the zeros singular-values are already at the end
singVals.head(actual_n).reverseInPlace();
U.leftCols(actual_n) = U.leftCols(actual_n).rowwise().reverse().eval(); // FIXME this requires a temporary
if (m_compV) V.leftCols(actual_n) = V.leftCols(actual_n).rowwise().reverse().eval(); // FIXME this requires a temporary
U.leftCols(actual_n).rowwise().reverseInPlace();
if (m_compV) V.leftCols(actual_n).rowwise().reverseInPlace();
#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
JacobiSVD<MatrixXr> jsvd(m_computed.block(firstCol, firstCol, n, n) );
@@ -634,7 +671,7 @@ void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, Vec
}
template <typename MatrixType>
typename BDCSVD<MatrixType>::RealScalar BDCSVD<MatrixType>::secularEq(RealScalar mu, const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm, const ArrayXr& diagShifted, RealScalar shift)
typename BDCSVD<MatrixType>::RealScalar BDCSVD<MatrixType>::secularEq(RealScalar mu, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift)
{
Index m = perm.size();
RealScalar res = 1;
@@ -647,8 +684,8 @@ typename BDCSVD<MatrixType>::RealScalar BDCSVD<MatrixType>::secularEq(RealScalar
}
template <typename MatrixType>
void BDCSVD<MatrixType>::computeSingVals(const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm,
VectorType& singVals, ArrayXr& shifts, ArrayXr& mus)
void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm,
VectorType& singVals, ArrayRef shifts, ArrayRef mus)
{
using std::abs;
using std::swap;
@@ -703,7 +740,8 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayXr& col0, const ArrayXr& dia
RealScalar shift = (k == actual_n-1 || fMid > 0) ? left : right;
// measure everything relative to shift
ArrayXr diagShifted = diag - shift;
Map<ArrayXr> diagShifted(m_workspace.data()+4*n, n);
diagShifted = diag - shift;
// initial guess
RealScalar muPrev, muCur;
@@ -730,7 +768,7 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayXr& col0, const ArrayXr& dia
// rational interpolation: fit a function of the form a / mu + b through the two previous
// iterates and use its zero to compute the next iterate
bool useBisection = fPrev*fCur>0;
while (fCur!=0 && abs(muCur - muPrev) > 8 * NumTraits<RealScalar>::epsilon() * numext::maxi(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits<RealScalar>::epsilon() && !useBisection)
while (fCur!=0 && abs(muCur - muPrev) > 8 * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits<RealScalar>::epsilon() && !useBisection)
{
++m_numIters;
@@ -773,7 +811,10 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayXr& col0, const ArrayXr& dia
}
RealScalar fLeft = secularEq(leftShifted, col0, diag, perm, diagShifted, shift);
#if defined EIGEN_INTERNAL_DEBUGGING || defined EIGEN_BDCSVD_DEBUG_VERBOSE
RealScalar fRight = secularEq(rightShifted, col0, diag, perm, diagShifted, shift);
#endif
#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
if(!(fLeft * fRight<0))
@@ -781,14 +822,13 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayXr& col0, const ArrayXr& dia
#endif
eigen_internal_assert(fLeft * fRight < 0);
while (rightShifted - leftShifted > 2 * NumTraits<RealScalar>::epsilon() * numext::maxi(abs(leftShifted), abs(rightShifted)))
while (rightShifted - leftShifted > 2 * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(leftShifted), abs(rightShifted)))
{
RealScalar midShifted = (leftShifted + rightShifted) / 2;
RealScalar fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift);
if (fLeft * fMid < 0)
{
rightShifted = midShifted;
fRight = fMid;
}
else
{
@@ -816,8 +856,8 @@ void BDCSVD<MatrixType>::computeSingVals(const ArrayXr& col0, const ArrayXr& dia
// zhat is perturbation of col0 for which singular vectors can be computed stably (see Section 3.1)
template <typename MatrixType>
void BDCSVD<MatrixType>::perturbCol0
(const ArrayXr& col0, const ArrayXr& diag, const ArrayXi &perm, const VectorType& singVals,
const ArrayXr& shifts, const ArrayXr& mus, ArrayXr& zhat)
(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const VectorType& singVals,
const ArrayRef& shifts, const ArrayRef& mus, ArrayRef zhat)
{
using std::sqrt;
Index n = col0.size();
@@ -865,8 +905,8 @@ void BDCSVD<MatrixType>::perturbCol0
// compute singular vectors
template <typename MatrixType>
void BDCSVD<MatrixType>::computeSingVecs
(const ArrayXr& zhat, const ArrayXr& diag, const ArrayXi &perm, const VectorType& singVals,
const ArrayXr& shifts, const ArrayXr& mus, MatrixXr& U, MatrixXr& V)
(const ArrayRef& zhat, const ArrayRef& diag, const IndicesRef &perm, const VectorType& singVals,
const ArrayRef& shifts, const ArrayRef& mus, MatrixXr& U, MatrixXr& V)
{
Index n = zhat.size();
Index m = perm.size();
@@ -991,7 +1031,7 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
RealScalar maxDiag = diag.tail((std::max)(Index(1),length-1)).cwiseAbs().maxCoeff();
RealScalar epsilon_strict = NumTraits<RealScalar>::epsilon() * maxDiag;
RealScalar epsilon_coarse = 8 * NumTraits<RealScalar>::epsilon() * numext::maxi(col0.cwiseAbs().maxCoeff(), maxDiag);
RealScalar epsilon_coarse = 8 * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(col0.cwiseAbs().maxCoeff(), maxDiag);
#ifdef EIGEN_BDCSVD_SANITY_CHECKS
assert(m_naiveU.allFinite());
@@ -1047,7 +1087,7 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
// Sort the diagonal entries, since diag(1:k-1) and diag(k:length) are already sorted, let's do a sorted merge.
// First, compute the respective permutation.
Index *permutation = new Index[length]; // FIXME avoid repeated dynamic memory allocation
Index *permutation = m_workspaceI.data();
{
permutation[0] = 0;
Index p = 1;
@@ -1084,8 +1124,8 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
}
// Current index of each col, and current column of each index
Index *realInd = new Index[length]; // FIXME avoid repeated dynamic memory allocation
Index *realCol = new Index[length]; // FIXME avoid repeated dynamic memory allocation
Index *realInd = m_workspaceI.data()+length;
Index *realCol = m_workspaceI.data()+2*length;
for(int pos = 0; pos< length; pos++)
{
@@ -1115,9 +1155,6 @@ void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index
realInd[J] = realI;
realInd[i] = pi;
}
delete[] permutation;
delete[] realInd;
delete[] realCol;
}
#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
std::cout << "sorted: " << diag.transpose().format(bdcsvdfmt) << "\n";

14
Eigen/src/SVD/JacobiSVD.h
View File

@@ -425,12 +425,13 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q,
// If d!=0, then t/d cannot overflow because the magnitude of the
// entries forming d are not too small compared to the ones forming t.
RealScalar u = t / d;
rot1.s() = RealScalar(1) / sqrt(RealScalar(1) + numext::abs2(u));
rot1.c() = rot1.s() * u;
RealScalar tmp = sqrt(RealScalar(1) + numext::abs2(u));
rot1.s() = RealScalar(1) / tmp;
rot1.c() = u / tmp;
}
m.applyOnTheLeft(0,1,rot1);
j_right->makeJacobi(m,0,1);
*j_left = rot1 * j_right->transpose();
*j_left = rot1 * j_right->transpose();
}
template<typename _MatrixType, int QRPreconditioner>
@@ -680,6 +681,8 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
const RealScalar precision = RealScalar(2) * NumTraits<Scalar>::epsilon();
// limit for very small denormal numbers to be considered zero in order to avoid infinite loops (see bug 286)
// FIXME What about considerering any denormal numbers as zero, using:
// const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
const RealScalar considerAsZero = RealScalar(2) * std::numeric_limits<RealScalar>::denorm_min();
// Scaling factor to reduce over/under-flows
@@ -719,8 +722,9 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
// if this 2x2 sub-matrix is not diagonal already...
// notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't
// keep us iterating forever. Similarly, small denormal numbers are considered zero.
RealScalar threshold = numext::maxi(considerAsZero, precision * numext::maxi(abs(m_workMatrix.coeff(p,p)),
abs(m_workMatrix.coeff(q,q))));
RealScalar threshold = numext::maxi<RealScalar>(considerAsZero,
precision * numext::maxi<RealScalar>(abs(m_workMatrix.coeff(p,p)),
abs(m_workMatrix.coeff(q,q))));
// We compare both values to threshold instead of calling max to be robust to NaN (See bug 791)
if(abs(m_workMatrix.coeff(p,q))>threshold || abs(m_workMatrix.coeff(q,p)) > threshold)
{

13
Eigen/src/SVD/SVDBase.h
View File

@@ -130,9 +130,10 @@ public:
inline Index rank() const
{
using std::abs;
using std::max;
eigen_assert(m_isInitialized && "JacobiSVD is not initialized.");
if(m_singularValues.size()==0) return 0;
RealScalar premultiplied_threshold = m_singularValues.coeff(0) * threshold();
RealScalar premultiplied_threshold = (max)(m_singularValues.coeff(0) * threshold(), (std::numeric_limits<RealScalar>::min)());
Index i = m_nonzeroSingularValues-1;
while(i>=0 && m_singularValues.coeff(i) < premultiplied_threshold) --i;
return i+1;
@@ -217,6 +218,12 @@ public:
#endif
protected:
static void check_template_parameters()
{
EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
}
// return true if already allocated
bool allocate(Index rows, Index cols, unsigned int computationOptions) ;
@@ -240,7 +247,9 @@ protected:
m_usePrescribedThreshold(false),
m_computationOptions(0),
m_rows(-1), m_cols(-1), m_diagSize(0)
{}
{
check_template_parameters();
}
};

7
Eigen/src/SparseCore/CompressedStorage.h
View File

@@ -86,7 +86,12 @@ class CompressedStorage
void resize(Index size, double reserveSizeFactor = 0)
{
if (m_allocatedSize<size)
reallocate(size + Index(reserveSizeFactor*double(size)));
{
Index realloc_size = (std::min<Index>)(NumTraits<StorageIndex>::highest(), size + Index(reserveSizeFactor*double(size)));
if(realloc_size<size)
internal::throw_std_bad_alloc();
reallocate(realloc_size);
}
m_size = size;
}

8
Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
View File

@@ -30,16 +30,16 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
std::memset(mask,0,sizeof(bool)*rows);
typename evaluator<Lhs>::type lhsEval(lhs);
typename evaluator<Rhs>::type rhsEval(rhs);
// estimate the number of non zero entries
// given a rhs column containing Y non zeros, we assume that the respective Y columns
// of the lhs differs in average of one non zeros, thus the number of non zeros for
// the product of a rhs column with the lhs is X+Y where X is the average number of non zero
// per column of the lhs.
// Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs)
Index estimated_nnz_prod = lhs.nonZeros() + rhs.nonZeros();
typename evaluator<Lhs>::type lhsEval(lhs);
typename evaluator<Rhs>::type rhsEval(rhs);
Index estimated_nnz_prod = lhsEval.nonZerosEstimate() + rhsEval.nonZerosEstimate();
res.setZero();
res.reserve(Index(estimated_nnz_prod));

159
Eigen/src/SparseCore/SparseBlock.h
View File

@@ -49,6 +49,16 @@ public:
return nnz;
}
inline const Scalar coeff(Index row, Index col) const
{
return m_matrix.coeff(row + (IsRowMajor ? m_outerStart : 0), col + (IsRowMajor ? 0 : m_outerStart));
}
inline const Scalar coeff(Index index) const
{
return m_matrix.coeff(IsRowMajor ? m_outerStart : index, IsRowMajor ? index : m_outerStart);
}
inline const _MatrixTypeNested& nestedExpression() const { return m_matrix; }
Index startRow() const { return IsRowMajor ? m_outerStart : 0; }
Index startCol() const { return IsRowMajor ? 0 : m_outerStart; }
@@ -80,7 +90,8 @@ class sparse_matrix_block_impl
typedef Block<SparseMatrixType, BlockRows, BlockCols, true> BlockType;
public:
enum { IsRowMajor = internal::traits<BlockType>::IsRowMajor };
EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType)
typedef SparseCompressedBase<Block<SparseMatrixType,BlockRows,BlockCols,true> > Base;
_EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType)
protected:
typedef typename Base::IndexVector IndexVector;
enum { OuterSize = IsRowMajor ? BlockRows : BlockCols };
@@ -188,27 +199,31 @@ public:
{ return m_matrix.const_cast_derived().outerIndexPtr() + m_outerStart; }
inline const StorageIndex* innerNonZeroPtr() const
{ return isCompressed() ? 0 : m_matrix.innerNonZeroPtr(); }
{ return isCompressed() ? 0 : (m_matrix.innerNonZeroPtr()+m_outerStart); }
inline StorageIndex* innerNonZeroPtr()
{ return isCompressed() ? 0 : m_matrix.const_cast_derived().innerNonZeroPtr(); }
Index nonZeros() const
{
if(m_matrix.isCompressed())
return ( (m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()])
- (m_matrix.outerIndexPtr()[m_outerStart]));
else if(m_outerSize.value()==0)
return 0;
else
return Map<const IndexVector>(m_matrix.innerNonZeroPtr()+m_outerStart, m_outerSize.value()).sum();
}
{ return isCompressed() ? 0 : (m_matrix.const_cast_derived().innerNonZeroPtr()+m_outerStart); }
bool isCompressed() const { return m_matrix.innerNonZeroPtr()==0; }
inline Scalar& coeffRef(Index row, Index col)
{
return m_matrix.const_cast_derived().coeffRef(row + (IsRowMajor ? m_outerStart : 0), col + (IsRowMajor ? 0 : m_outerStart));
}
inline const Scalar coeff(Index row, Index col) const
{
return m_matrix.coeff(row + (IsRowMajor ? m_outerStart : 0), col + (IsRowMajor ? 0 : m_outerStart));
}
inline const Scalar coeff(Index index) const
{
return m_matrix.coeff(IsRowMajor ? m_outerStart : index, IsRowMajor ? index : m_outerStart);
}
const Scalar& lastCoeff() const
{
EIGEN_STATIC_ASSERT_VECTOR_ONLY(sparse_matrix_block_impl);
eigen_assert(nonZeros()>0);
eigen_assert(Base::nonZeros()>0);
if(m_matrix.isCompressed())
return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart+1]-1];
else
@@ -314,17 +329,6 @@ SparseMatrixBase<Derived>::innerVectors(Index outerStart, Index outerSize) const
}
namespace internal {
template< typename XprType, int BlockRows, int BlockCols, bool InnerPanel,
bool OuterVector = (BlockCols==1 && XprType::IsRowMajor)
| // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&".
// revert to || as soon as not needed anymore.
(BlockRows==1 && !XprType::IsRowMajor)>
class GenericSparseBlockInnerIteratorImpl;
}
/** Generic implementation of sparse Block expression.
* Real-only.
*/
@@ -390,8 +394,11 @@ public:
Index blockCols() const { return m_blockCols.value(); }
protected:
friend class internal::GenericSparseBlockInnerIteratorImpl<XprType,BlockRows,BlockCols,InnerPanel>;
// friend class internal::GenericSparseBlockInnerIteratorImpl<XprType,BlockRows,BlockCols,InnerPanel>;
friend class ReverseInnerIterator;
friend struct internal::unary_evaluator<Block<XprType,BlockRows,BlockCols,InnerPanel>, internal::IteratorBased, Scalar >;
Index nonZeros() const { return Dynamic; }
EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl)
@@ -404,94 +411,6 @@ public:
};
namespace internal {
template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
class GenericSparseBlockInnerIteratorImpl<XprType,BlockRows,BlockCols,InnerPanel,false> : public Block<XprType, BlockRows, BlockCols, InnerPanel>::_MatrixTypeNested::InnerIterator
{
typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
enum {
IsRowMajor = BlockType::IsRowMajor
};
typedef typename BlockType::_MatrixTypeNested _MatrixTypeNested;
typedef typename BlockType::StorageIndex StorageIndex;
typedef typename _MatrixTypeNested::InnerIterator Base;
const BlockType& m_block;
Index m_end;
public:
EIGEN_STRONG_INLINE GenericSparseBlockInnerIteratorImpl(const BlockType& block, Index outer)
: Base(block.derived().nestedExpression(), outer + (IsRowMajor ? block.m_startRow.value() : block.m_startCol.value())),
m_block(block),
m_end(IsRowMajor ? block.m_startCol.value()+block.m_blockCols.value() : block.m_startRow.value()+block.m_blockRows.value())
{
while( (Base::operator bool()) && (Base::index() < (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value())) )
Base::operator++();
}
inline Index index() const { return Base::index() - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); }
inline Index outer() const { return Base::outer() - (IsRowMajor ? m_block.m_startRow.value() : m_block.m_startCol.value()); }
inline Index row() const { return Base::row() - m_block.m_startRow.value(); }
inline Index col() const { return Base::col() - m_block.m_startCol.value(); }
inline operator bool() const { return Base::operator bool() && Base::index() < m_end; }
};
// Row vector of a column-major sparse matrix or column of a row-major one.
template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
class GenericSparseBlockInnerIteratorImpl<XprType,BlockRows,BlockCols,InnerPanel,true>
{
typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
enum {
IsRowMajor = BlockType::IsRowMajor
};
typedef typename BlockType::_MatrixTypeNested _MatrixTypeNested;
typedef typename BlockType::StorageIndex StorageIndex;
typedef typename BlockType::Scalar Scalar;
const BlockType& m_block;
Index m_outerPos;
Index m_innerIndex;
Scalar m_value;
Index m_end;
public:
explicit EIGEN_STRONG_INLINE GenericSparseBlockInnerIteratorImpl(const BlockType& block, Index outer = 0)
:
m_block(block),
m_outerPos( (IsRowMajor ? block.m_startCol.value() : block.m_startRow.value()) - 1), // -1 so that operator++ finds the first non-zero entry
m_innerIndex(IsRowMajor ? block.m_startRow.value() : block.m_startCol.value()),
m_end(IsRowMajor ? block.m_startCol.value()+block.m_blockCols.value() : block.m_startRow.value()+block.m_blockRows.value())
{
EIGEN_UNUSED_VARIABLE(outer);
eigen_assert(outer==0);
++(*this);
}
inline Index index() const { return m_outerPos - (IsRowMajor ? m_block.m_startCol.value() : m_block.m_startRow.value()); }
inline Index outer() const { return 0; }
inline Index row() const { return IsRowMajor ? 0 : index(); }
inline Index col() const { return IsRowMajor ? index() : 0; }
inline Scalar value() const { return m_value; }
inline GenericSparseBlockInnerIteratorImpl& operator++()
{
// search next non-zero entry
while(++m_outerPos<m_end)
{
typename XprType::InnerIterator it(m_block.m_matrix, m_outerPos);
// search for the key m_innerIndex in the current outer-vector
while(it && it.index() < m_innerIndex) ++it;
if(it && it.index()==m_innerIndex)
{
m_value = it.value();
break;
}
}
return *this;
}
inline operator bool() const { return m_outerPos < m_end; }
};
template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
struct unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBased >
@@ -523,9 +442,16 @@ struct unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBa
explicit unary_evaluator(const XprType& op)
: m_argImpl(op.nestedExpression()), m_block(op)
{}
inline Index nonZerosEstimate() const {
Index nnz = m_block.nonZeros();
if(nnz<0)
return m_argImpl.nonZerosEstimate() * m_block.size() / m_block.nestedExpression().size();
return nnz;
}
protected:
typedef typename evaluator<ArgType>::InnerIterator EvalIterator;
typedef typename evaluator<ArgType>::InnerIterator EvalIterator;
typename evaluator<ArgType>::nestedType m_argImpl;
const XprType &m_block;
@@ -570,6 +496,7 @@ public:
: m_eval(aEval),
m_outerPos( (IsRowMajor ? aEval.m_block.startCol() : aEval.m_block.startRow()) - 1), // -1 so that operator++ finds the first non-zero entry
m_innerIndex(IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol()),
m_value(0),
m_end(IsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows())
{
EIGEN_UNUSED_VARIABLE(outer);

23
Eigen/src/SparseCore/SparseCompressedBase.h
View File

@@ -35,6 +35,25 @@ class SparseCompressedBase
class InnerIterator;
class ReverseInnerIterator;
protected:
typedef typename Base::IndexVector IndexVector;
Eigen::Map<IndexVector> innerNonZeros() { return Eigen::Map<IndexVector>(innerNonZeroPtr(), isCompressed()?0:derived().outerSize()); }
const Eigen::Map<const IndexVector> innerNonZeros() const { return Eigen::Map<const IndexVector>(innerNonZeroPtr(), isCompressed()?0:derived().outerSize()); }
public:
/** \returns the number of non zero coefficients */
inline Index nonZeros() const
{
if(isCompressed())
return outerIndexPtr()[derived().outerSize()]-outerIndexPtr()[0];
else if(derived().outerSize()==0)
return 0;
else
return innerNonZeros().sum();
}
/** \returns a const pointer to the array of values.
* This function is aimed at interoperability with other libraries.
* \sa innerIndexPtr(), outerIndexPtr() */
@@ -165,6 +184,10 @@ struct evaluator<SparseCompressedBase<Derived> >
evaluator() : m_matrix(0) {}
explicit evaluator(const Derived &mat) : m_matrix(&mat) {}
inline Index nonZerosEstimate() const {
return m_matrix->nonZeros();
}
operator Derived&() { return m_matrix->const_cast_derived(); }
operator const Derived&() const { return *m_matrix; }

20
Eigen/src/SparseCore/SparseCwiseBinaryOp.h
View File

@@ -121,6 +121,10 @@ public:
m_lhsImpl(xpr.lhs()),
m_rhsImpl(xpr.rhs())
{ }
inline Index nonZerosEstimate() const {
return m_lhsImpl.nonZerosEstimate() + m_rhsImpl.nonZerosEstimate();
}
protected:
const BinaryOp m_functor;
@@ -198,6 +202,10 @@ public:
m_lhsImpl(xpr.lhs()),
m_rhsImpl(xpr.rhs())
{ }
inline Index nonZerosEstimate() const {
return (std::min)(m_lhsImpl.nonZerosEstimate(), m_rhsImpl.nonZerosEstimate());
}
protected:
const BinaryOp m_functor;
@@ -243,7 +251,7 @@ public:
EIGEN_STRONG_INLINE Index col() const { return m_rhsIter.col(); }
EIGEN_STRONG_INLINE operator bool() const { return m_rhsIter; }
protected:
const LhsEvaluator &m_lhsEval;
RhsIterator m_rhsIter;
@@ -262,6 +270,10 @@ public:
m_lhsImpl(xpr.lhs()),
m_rhsImpl(xpr.rhs())
{ }
inline Index nonZerosEstimate() const {
return m_rhsImpl.nonZerosEstimate();
}
protected:
const BinaryOp m_functor;
@@ -308,7 +320,7 @@ public:
EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); }
EIGEN_STRONG_INLINE operator bool() const { return m_lhsIter; }
protected:
LhsIterator m_lhsIter;
const RhsEvaluator &m_rhsEval;
@@ -327,6 +339,10 @@ public:
m_lhsImpl(xpr.lhs()),
m_rhsImpl(xpr.rhs())
{ }
inline Index nonZerosEstimate() const {
return m_lhsImpl.nonZerosEstimate();
}
protected:
const BinaryOp m_functor;

4
Eigen/src/SparseCore/SparseCwiseUnaryOp.h
View File

@@ -30,6 +30,10 @@ struct unary_evaluator<CwiseUnaryOp<UnaryOp,ArgType>, IteratorBased>
};
explicit unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression()) {}
inline Index nonZerosEstimate() const {
return m_argImpl.nonZerosEstimate();
}
protected:
typedef typename evaluator<ArgType>::InnerIterator EvalIterator;

3
Eigen/src/SparseCore/SparseMap.h
View File

@@ -105,9 +105,6 @@ class SparseMapBase<Derived,ReadOnlyAccessors>
return ((*r==inner) && (id<end)) ? m_values[id] : Scalar(0);
}
/** \returns the number of non zero coefficients */
inline Index nonZeros() const { return m_nnz; }
inline SparseMapBase(Index rows, Index cols, Index nnz, IndexPointer outerIndexPtr, IndexPointer innerIndexPtr,
ScalarPointer valuePtr, IndexPointer innerNonZerosPtr = 0)
: m_outerSize(IsRowMajor?rows:cols), m_innerSize(IsRowMajor?cols:rows), m_nnz(nnz), m_outerIndex(outerIndexPtr),

20
Eigen/src/SparseCore/SparseMatrix.h
View File

@@ -95,6 +95,7 @@ class SparseMatrix
public:
typedef SparseCompressedBase<SparseMatrix> Base;
using Base::isCompressed;
using Base::nonZeros;
_EIGEN_SPARSE_PUBLIC_INTERFACE(SparseMatrix)
EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(SparseMatrix, +=)
EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(SparseMatrix, -=)
@@ -122,9 +123,6 @@ class SparseMatrix
StorageIndex* m_outerIndex;
StorageIndex* m_innerNonZeros; // optional, if null then the data is compressed
Storage m_data;
Eigen::Map<IndexVector> innerNonZeros() { return Eigen::Map<IndexVector>(m_innerNonZeros, m_innerNonZeros?m_outerSize:0); }
const Eigen::Map<const IndexVector> innerNonZeros() const { return Eigen::Map<const IndexVector>(m_innerNonZeros, m_innerNonZeros?m_outerSize:0); }
public:
@@ -252,14 +250,6 @@ class SparseMatrix
memset(m_innerNonZeros, 0, (m_outerSize)*sizeof(StorageIndex));
}
/** \returns the number of non zero coefficients */
inline Index nonZeros() const
{
if(m_innerNonZeros)
return innerNonZeros().sum();
return convert_index(Index(m_data.size()));
}
/** Preallocates \a reserveSize non zeros.
*
* Precondition: the matrix must be in compressed mode. */
@@ -1172,8 +1162,12 @@ typename SparseMatrix<_Scalar,_Options,_Index>::Scalar& SparseMatrix<_Scalar,_Op
return (m_data.value(p) = 0);
}
// make sure the matrix is compatible to random un-compressed insertion:
m_data.resize(m_data.allocatedSize());
if(m_data.size() != m_data.allocatedSize())
{
// make sure the matrix is compatible to random un-compressed insertion:
m_data.resize(m_data.allocatedSize());
this->reserveInnerVectors(Array<StorageIndex,Dynamic,1>::Constant(2*m_outerSize, convert_index(m_outerSize)));
}
return insertUncompressed(row,col);
}

3
Eigen/src/SparseCore/SparseMatrixBase.h
View File

@@ -149,9 +149,6 @@ template<typename Derived> class SparseMatrixBase : public EigenBase<Derived>
/** \returns the number of coefficients, which is \a rows()*cols().
* \sa rows(), cols(). */
inline Index size() const { return rows() * cols(); }
/** \returns the number of nonzero coefficients which is in practice the number
* of stored coefficients. */
inline Index nonZeros() const { return derived().nonZeros(); }
/** \returns true if either the number of rows or the number of columns is equal to 1.
* In other words, this function returns
* \code rows()==1 || cols()==1 \endcode

16
Eigen/src/SparseCore/SparseSparseProductWithPruning.h
View File

@@ -33,14 +33,6 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
// allocate a temporary buffer
AmbiVector<Scalar,StorageIndex> tempVector(rows);
// estimate the number of non zero entries
// given a rhs column containing Y non zeros, we assume that the respective Y columns
// of the lhs differs in average of one non zeros, thus the number of non zeros for
// the product of a rhs column with the lhs is X+Y where X is the average number of non zero
// per column of the lhs.
// Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs)
Index estimated_nnz_prod = lhs.nonZeros() + rhs.nonZeros();
// mimics a resizeByInnerOuter:
if(ResultType::IsRowMajor)
res.resize(cols, rows);
@@ -49,6 +41,14 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
typename evaluator<Lhs>::type lhsEval(lhs);
typename evaluator<Rhs>::type rhsEval(rhs);
// estimate the number of non zero entries
// given a rhs column containing Y non zeros, we assume that the respective Y columns
// of the lhs differs in average of one non zeros, thus the number of non zeros for
// the product of a rhs column with the lhs is X+Y where X is the average number of non zero
// per column of the lhs.
// Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs)
Index estimated_nnz_prod = lhsEval.nonZerosEstimate() + rhsEval.nonZerosEstimate();
res.reserve(estimated_nnz_prod);
double ratioColRes = double(estimated_nnz_prod)/double(lhs.rows()*rhs.cols());

8
Eigen/src/SparseCore/SparseTranspose.h
View File

@@ -40,15 +40,11 @@ namespace internal {
};
}
// Implement nonZeros() for transpose. I'm not sure that's the best approach for that.
// Perhaps it should be implemented in Transpose<> itself.
template<typename MatrixType> class TransposeImpl<MatrixType,Sparse>
: public internal::SparseTransposeImpl<MatrixType>
{
protected:
typedef internal::SparseTransposeImpl<MatrixType> Base;
public:
inline Index nonZeros() const { return Base::derived().nestedExpression().nonZeros(); }
};
namespace internal {
@@ -61,6 +57,10 @@ struct unary_evaluator<Transpose<ArgType>, IteratorBased>
typedef typename evaluator<ArgType>::ReverseInnerIterator EvalReverseIterator;
public:
typedef Transpose<ArgType> XprType;
inline Index nonZerosEstimate() const {
return m_argImpl.nonZerosEstimate();
}
class InnerIterator : public EvalIterator
{

11
Eigen/src/SparseCore/SparseTriangularView.h
View File

@@ -50,13 +50,6 @@ protected:
template<typename OtherDerived> void solveInPlace(MatrixBase<OtherDerived>& other) const;
template<typename OtherDerived> void solveInPlace(SparseMatrixBase<OtherDerived>& other) const;
inline Index nonZeros() const {
// FIXME HACK number of nonZeros is required for product logic
// this returns only an upper bound (but should be OK for most purposes)
return derived().nestedExpression().nonZeros();
}
};
@@ -191,6 +184,10 @@ public:
explicit unary_evaluator(const XprType &xpr) : m_argImpl(xpr.nestedExpression()) {}
inline Index nonZerosEstimate() const {
return m_argImpl.nonZerosEstimate();
}
class InnerIterator : public EvalIterator
{
typedef EvalIterator Base;

4
Eigen/src/SparseCore/SparseVector.h
View File

@@ -442,6 +442,10 @@ struct evaluator<SparseVector<_Scalar,_Options,_Index> >
explicit evaluator(const SparseVectorType &mat) : m_matrix(mat) {}
inline Index nonZerosEstimate() const {
return m_matrix.nonZeros();
}
operator SparseVectorType&() { return m_matrix.const_cast_derived(); }
operator const SparseVectorType&() const { return m_matrix; }

17
Eigen/src/SuperLUSupport/SuperLUSupport.h
View File

@@ -165,8 +165,9 @@ struct SluMatrix : SuperMatrix
}
template<typename MatrixType>
static SluMatrix Map(SparseMatrixBase<MatrixType>& mat)
static SluMatrix Map(SparseMatrixBase<MatrixType>& a_mat)
{
MatrixType &mat(a_mat.derived());
SluMatrix res;
if ((MatrixType::Flags&RowMajorBit)==RowMajorBit)
{
@@ -184,9 +185,9 @@ struct SluMatrix : SuperMatrix
res.Mtype = SLU_GE;
res.storage.nnz = internal::convert_index<int>(mat.nonZeros());
res.storage.values = mat.derived().valuePtr();
res.storage.innerInd = mat.derived().innerIndexPtr();
res.storage.outerInd = mat.derived().outerIndexPtr();
res.storage.values = mat.valuePtr();
res.storage.innerInd = mat.innerIndexPtr();
res.storage.outerInd = mat.outerIndexPtr();
res.setScalarType<typename MatrixType::Scalar>();
@@ -302,6 +303,7 @@ class SuperLUBase : public SparseSolverBase<Derived>
typedef Matrix<Scalar,Dynamic,1> Vector;
typedef Matrix<int, 1, MatrixType::ColsAtCompileTime> IntRowVectorType;
typedef Matrix<int, MatrixType::RowsAtCompileTime, 1> IntColVectorType;
typedef Map<PermutationMatrix<Dynamic,Dynamic,int> > PermutationMap;
typedef SparseMatrix<Scalar> LUMatrixType;
public:
@@ -459,10 +461,11 @@ class SuperLU : public SuperLUBase<_MatrixType,SuperLU<_MatrixType> >
typedef typename Base::RealScalar RealScalar;
typedef typename Base::StorageIndex StorageIndex;
typedef typename Base::IntRowVectorType IntRowVectorType;
typedef typename Base::IntColVectorType IntColVectorType;
typedef typename Base::IntColVectorType IntColVectorType;
typedef typename Base::PermutationMap PermutationMap;
typedef typename Base::LUMatrixType LUMatrixType;
typedef TriangularView<LUMatrixType, Lower|UnitDiag> LMatrixType;
typedef TriangularView<LUMatrixType, Upper> UMatrixType;
typedef TriangularView<LUMatrixType, Upper> UMatrixType;
public:
using Base::_solve_impl;
@@ -774,6 +777,8 @@ typename SuperLU<MatrixType>::Scalar SuperLU<MatrixType>::determinant() const
det *= m_u.valuePtr()[lastId];
}
}
if(PermutationMap(m_p.data(),m_p.size()).determinant()*PermutationMap(m_q.data(),m_q.size()).determinant()<0)
det = -det;
if(m_sluEqued!='N')
return det/m_sluRscale.prod()/m_sluCscale.prod();
else