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 eigen/eigen into default

Browse Source
This commit is contained in:
Konstantinos Margaritis
2017-10-12 22:23:13 +03:00
parent 98e52cc770 0e85a677e3
commit df7644aec3
36 changed files with 517 additions and 274 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -396,6 +396,7 @@ template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
*/
template<typename Derived>
template<typename OtherDerived>
EIGEN_DEVICE_FUNC
inline const Product<Derived, OtherDerived>
MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
{

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

@@ -114,7 +114,7 @@ template<typename PlainObjectType, int MapOptions, typename StrideType> class Ma
inline Index outerStride() const
{
return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
: internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? internal::traits<Map>::OuterStrideAtCompileTime
: internal::traits<Map>::OuterStrideAtCompileTime != Dynamic ? Index(internal::traits<Map>::OuterStrideAtCompileTime)
: IsVectorAtCompileTime ? (this->size() * innerStride())
: int(Flags)&RowMajorBit ? (this->cols() * innerStride())
: (this->rows() * innerStride());

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

@@ -99,7 +99,7 @@ class NoAlias
* \sa class NoAlias
*/
template<typename Derived>
NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias()
NoAlias<Derived,MatrixBase> EIGEN_DEVICE_FUNC MatrixBase<Derived>::noalias()
{
return NoAlias<Derived, Eigen::MatrixBase >(derived());
}

62
Eigen/src/Core/arch/CUDA/Half.h
View File

@@ -50,38 +50,45 @@ struct half;
namespace half_impl {
#if !defined(EIGEN_HAS_CUDA_FP16)
// Make our own __half definition that is similar to CUDA's.
struct __half {
EIGEN_DEVICE_FUNC __half() : x(0) {}
explicit EIGEN_DEVICE_FUNC __half(unsigned short raw) : x(raw) {}
// Make our own __half_raw definition that is similar to CUDA's.
struct __half_raw {
EIGEN_DEVICE_FUNC __half_raw() : x(0) {}
explicit EIGEN_DEVICE_FUNC __half_raw(unsigned short raw) : x(raw) {}
unsigned short x;
};
#elif defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000
// In CUDA < 9.0, __half is the equivalent of CUDA 9's __half_raw
typedef __half __half_raw;
#endif
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x);
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff);
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h);
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x);
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff);
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h);
struct half_base : public __half {
struct half_base : public __half_raw {
EIGEN_DEVICE_FUNC half_base() {}
EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {}
EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {}
EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half_raw(h) {}
EIGEN_DEVICE_FUNC half_base(const __half_raw& h) : __half_raw(h) {}
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
EIGEN_DEVICE_FUNC half_base(const __half& h) : __half_raw(*(__half_raw*)&h) {}
#endif
};
} // namespace half_impl
// Class definition.
struct half : public half_impl::half_base {
#if !defined(EIGEN_HAS_CUDA_FP16)
typedef half_impl::__half __half;
#if !defined(EIGEN_HAS_CUDA_FP16) || (defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER < 90000)
typedef half_impl::__half_raw __half_raw;
#endif
EIGEN_DEVICE_FUNC half() {}
EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
EIGEN_DEVICE_FUNC half(const __half_raw& h) : half_impl::half_base(h) {}
EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDACC_VER) && EIGEN_CUDACC_VER >= 90000
EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
#endif
explicit EIGEN_DEVICE_FUNC half(bool b)
: half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
@@ -269,8 +276,8 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
// these in hardware. If we need more performance on older/other CPUs, they are
// also possible to vectorize directly.
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x) {
__half h;
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw raw_uint16_to_half(unsigned short x) {
__half_raw h;
h.x = x;
return h;
}
@@ -280,12 +287,13 @@ union FP32 {
float f;
};
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half_raw float_to_half_rtne(float ff) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
return __float2half(ff);
__half tmp_ff = __float2half(ff);
return *(__half_raw*)&tmp_ff;
#elif defined(EIGEN_HAS_FP16_C)
__half h;
__half_raw h;
h.x = _cvtss_sh(ff, 0);
return h;
@@ -296,7 +304,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
const FP32 f16max = { (127 + 16) << 23 };
const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
unsigned int sign_mask = 0x80000000u;
__half o;
__half_raw o;
o.x = static_cast<unsigned short>(0x0u);
unsigned int sign = f.u & sign_mask;
@@ -335,7 +343,7 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
#endif
}
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half_raw h) {
#if defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
return __half2float(h);
@@ -512,8 +520,8 @@ struct numeric_limits<Eigen::half> {
static const bool is_bounded = false;
static const bool is_modulo = false;
static const int digits = 11;
static const int digits10 = 2;
//static const int max_digits10 = ;
static const int digits10 = 3; // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
static const int max_digits10 = 5; // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
static const int radix = 2;
static const int min_exponent = -13;
static const int min_exponent10 = -4;
@@ -612,11 +620,15 @@ struct hash<Eigen::half> {
// Add the missing shfl_xor intrinsic
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300
__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
#if EIGEN_CUDACC_VER < 90000
return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
#else
return static_cast<Eigen::half>(__shfl_xor_sync(0xFFFFFFFF, static_cast<float>(var), laneMask, width));
#endif
}
#endif
// ldg() has an overload for __half, but we also need one for Eigen::half.
// ldg() has an overload for __half_raw, but we also need one for Eigen::half.
#if defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 350
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
return Eigen::half_impl::raw_uint16_to_half(

3
Eigen/src/Core/arch/CUDA/PacketMathHalf.h
View File

@@ -100,7 +100,8 @@ template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2&
template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
half2 result;
result.x = a.x & 0x7FFF7FFF;
unsigned temp = *(reinterpret_cast<const unsigned*>(&(a)));
*(reinterpret_cast<unsigned*>(&(result))) = temp & 0x7FFF7FFF;
return result;
}

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

@@ -410,6 +410,16 @@
#endif
#endif
// Does the compiler support type_trais?
#ifndef EIGEN_HAS_TYPE_TRAITS
#if EIGEN_MAX_CPP_VER>=11 && (EIGEN_HAS_CXX11 || EIGEN_COMP_MSVC >= 1700)
#define EIGEN_HAS_TYPE_TRAITS 1
#define EIGEN_INCLUDE_TYPE_TRAITS
#else
#define EIGEN_HAS_TYPE_TRAITS 0
#endif
#endif
// Does the compiler support variadic templates?
#ifndef EIGEN_HAS_VARIADIC_TEMPLATES
#if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \

12
Eigen/src/Core/util/Memory.h
View File

@@ -493,7 +493,7 @@ template<typename T> struct smart_copy_helper<T,true> {
IntPtr size = IntPtr(end)-IntPtr(start);
if(size==0) return;
eigen_internal_assert(start!=0 && end!=0 && target!=0);
memcpy(target, start, size);
std::memcpy(target, start, size);
}
};
@@ -696,7 +696,15 @@ template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
/** \class aligned_allocator
* \ingroup Core_Module
*
* \brief STL compatible allocator to use with with 16 byte aligned types
* \brief STL compatible allocator to use with types requiring a non standrad alignment.
*
* The memory is aligned as for dynamically aligned matrix/array types such as MatrixXd.
* By default, it will thus provide at least 16 bytes alignment and more in following cases:
* - 32 bytes alignment if AVX is enabled.
* - 64 bytes alignment if AVX512 is enabled.
*
* This can be controled using the \c EIGEN_MAX_ALIGN_BYTES macro as documented
* \link TopicPreprocessorDirectivesPerformance there \endlink.
*
* Example:
* \code

12
Eigen/src/Core/util/XprHelper.h
View File

@@ -34,6 +34,18 @@ inline IndexDest convert_index(const IndexSrc& idx) {
return IndexDest(idx);
}
// true if T can be considered as an integral index (i.e., and integral type or enum)
template<typename T> struct is_valid_index_type
{
enum { value =
#if EIGEN_HAS_TYPE_TRAITS
internal::is_integral<T>::value || std::is_enum<T>::value
#else
// without C++11, we use is_convertible to Index instead of is_integral in order to treat enums as Index.
internal::is_convertible<T,Index>::value
#endif
};
};
// promote_scalar_arg is an helper used in operation between an expression and a scalar, like:
// expression * scalar

256
Eigen/src/Jacobi/Jacobi.h
View File

@@ -309,16 +309,144 @@ inline void MatrixBase<Derived>::applyOnTheRight(Index p, Index q, const JacobiR
}
namespace internal {
template<typename Scalar, typename OtherScalar,
int SizeAtCompileTime, int MinAlignment, bool Vectorizable>
struct apply_rotation_in_the_plane_selector
{
static inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s)
{
for(Index i=0; i<size; ++i)
{
Scalar xi = *x;
Scalar yi = *y;
*x = c * xi + numext::conj(s) * yi;
*y = -s * xi + numext::conj(c) * yi;
x += incrx;
y += incry;
}
}
};
template<typename Scalar, typename OtherScalar,
int SizeAtCompileTime, int MinAlignment>
struct apply_rotation_in_the_plane_selector<Scalar,OtherScalar,SizeAtCompileTime,MinAlignment,true /* vectorizable */>
{
static inline void run(Scalar *x, Index incrx, Scalar *y, Index incry, Index size, OtherScalar c, OtherScalar s)
{
enum {
PacketSize = packet_traits<Scalar>::size,
OtherPacketSize = packet_traits<OtherScalar>::size
};
typedef typename packet_traits<Scalar>::type Packet;
typedef typename packet_traits<OtherScalar>::type OtherPacket;
/*** dynamic-size vectorized paths ***/
if(SizeAtCompileTime == Dynamic && ((incrx==1 && incry==1) || PacketSize == 1))
{
// both vectors are sequentially stored in memory => vectorization
enum { Peeling = 2 };
Index alignedStart = internal::first_default_aligned(y, size);
Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize;
const OtherPacket pc = pset1<OtherPacket>(c);
const OtherPacket ps = pset1<OtherPacket>(s);
conj_helper<OtherPacket,Packet,NumTraits<OtherScalar>::IsComplex,false> pcj;
conj_helper<OtherPacket,Packet,false,false> pm;
for(Index i=0; i<alignedStart; ++i)
{
Scalar xi = x[i];
Scalar yi = y[i];
x[i] = c * xi + numext::conj(s) * yi;
y[i] = -s * xi + numext::conj(c) * yi;
}
Scalar* EIGEN_RESTRICT px = x + alignedStart;
Scalar* EIGEN_RESTRICT py = y + alignedStart;
if(internal::first_default_aligned(x, size)==alignedStart)
{
for(Index i=alignedStart; i<alignedEnd; i+=PacketSize)
{
Packet xi = pload<Packet>(px);
Packet yi = pload<Packet>(py);
pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
px += PacketSize;
py += PacketSize;
}
}
else
{
Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize);
for(Index i=alignedStart; i<peelingEnd; i+=Peeling*PacketSize)
{
Packet xi = ploadu<Packet>(px);
Packet xi1 = ploadu<Packet>(px+PacketSize);
Packet yi = pload <Packet>(py);
Packet yi1 = pload <Packet>(py+PacketSize);
pstoreu(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstoreu(px+PacketSize, padd(pm.pmul(pc,xi1),pcj.pmul(ps,yi1)));
pstore (py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pm.pmul(ps,xi1)));
px += Peeling*PacketSize;
py += Peeling*PacketSize;
}
if(alignedEnd!=peelingEnd)
{
Packet xi = ploadu<Packet>(x+peelingEnd);
Packet yi = pload <Packet>(y+peelingEnd);
pstoreu(x+peelingEnd, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
}
}
for(Index i=alignedEnd; i<size; ++i)
{
Scalar xi = x[i];
Scalar yi = y[i];
x[i] = c * xi + numext::conj(s) * yi;
y[i] = -s * xi + numext::conj(c) * yi;
}
}
/*** fixed-size vectorized path ***/
else if(SizeAtCompileTime != Dynamic && MinAlignment>0) // FIXME should be compared to the required alignment
{
const OtherPacket pc = pset1<OtherPacket>(c);
const OtherPacket ps = pset1<OtherPacket>(s);
conj_helper<OtherPacket,Packet,NumTraits<OtherPacket>::IsComplex,false> pcj;
conj_helper<OtherPacket,Packet,false,false> pm;
Scalar* EIGEN_RESTRICT px = x;
Scalar* EIGEN_RESTRICT py = y;
for(Index i=0; i<size; i+=PacketSize)
{
Packet xi = pload<Packet>(px);
Packet yi = pload<Packet>(py);
pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
px += PacketSize;
py += PacketSize;
}
}
/*** non-vectorized path ***/
else
{
apply_rotation_in_the_plane_selector<Scalar,OtherScalar,SizeAtCompileTime,MinAlignment,false>::run(x,incrx,y,incry,size,c,s);
}
}
};
template<typename VectorX, typename VectorY, typename OtherScalar>
void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>& xpr_y, const JacobiRotation<OtherScalar>& j)
{
typedef typename VectorX::Scalar Scalar;
enum {
PacketSize = packet_traits<Scalar>::size,
OtherPacketSize = packet_traits<OtherScalar>::size
};
typedef typename packet_traits<Scalar>::type Packet;
typedef typename packet_traits<OtherScalar>::type OtherPacket;
const bool Vectorizable = (VectorX::Flags & VectorY::Flags & PacketAccessBit)
&& (int(packet_traits<Scalar>::size) == int(packet_traits<OtherScalar>::size));
eigen_assert(xpr_x.size() == xpr_y.size());
Index size = xpr_x.size();
Index incrx = xpr_x.derived().innerStride();
@@ -332,117 +460,11 @@ void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x
if (c==OtherScalar(1) && s==OtherScalar(0))
return;
/*** dynamic-size vectorized paths ***/
if(VectorX::SizeAtCompileTime == Dynamic &&
(VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
(PacketSize == OtherPacketSize) &&
((incrx==1 && incry==1) || PacketSize == 1))
{
// both vectors are sequentially stored in memory => vectorization
enum { Peeling = 2 };
Index alignedStart = internal::first_default_aligned(y, size);
Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize;
const OtherPacket pc = pset1<OtherPacket>(c);
const OtherPacket ps = pset1<OtherPacket>(s);
conj_helper<OtherPacket,Packet,NumTraits<OtherScalar>::IsComplex,false> pcj;
conj_helper<OtherPacket,Packet,false,false> pm;
for(Index i=0; i<alignedStart; ++i)
{
Scalar xi = x[i];
Scalar yi = y[i];
x[i] = c * xi + numext::conj(s) * yi;
y[i] = -s * xi + numext::conj(c) * yi;
}
Scalar* EIGEN_RESTRICT px = x + alignedStart;
Scalar* EIGEN_RESTRICT py = y + alignedStart;
if(internal::first_default_aligned(x, size)==alignedStart)
{
for(Index i=alignedStart; i<alignedEnd; i+=PacketSize)
{
Packet xi = pload<Packet>(px);
Packet yi = pload<Packet>(py);
pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
px += PacketSize;
py += PacketSize;
}
}
else
{
Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize);
for(Index i=alignedStart; i<peelingEnd; i+=Peeling*PacketSize)
{
Packet xi = ploadu<Packet>(px);
Packet xi1 = ploadu<Packet>(px+PacketSize);
Packet yi = pload <Packet>(py);
Packet yi1 = pload <Packet>(py+PacketSize);
pstoreu(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstoreu(px+PacketSize, padd(pm.pmul(pc,xi1),pcj.pmul(ps,yi1)));
pstore (py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pm.pmul(ps,xi1)));
px += Peeling*PacketSize;
py += Peeling*PacketSize;
}
if(alignedEnd!=peelingEnd)
{
Packet xi = ploadu<Packet>(x+peelingEnd);
Packet yi = pload <Packet>(y+peelingEnd);
pstoreu(x+peelingEnd, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
}
}
for(Index i=alignedEnd; i<size; ++i)
{
Scalar xi = x[i];
Scalar yi = y[i];
x[i] = c * xi + numext::conj(s) * yi;
y[i] = -s * xi + numext::conj(c) * yi;
}
}
/*** fixed-size vectorized path ***/
else if(VectorX::SizeAtCompileTime != Dynamic &&
(VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
(PacketSize == OtherPacketSize) &&
(EIGEN_PLAIN_ENUM_MIN(evaluator<VectorX>::Alignment, evaluator<VectorY>::Alignment)>0)) // FIXME should be compared to the required alignment
{
const OtherPacket pc = pset1<OtherPacket>(c);
const OtherPacket ps = pset1<OtherPacket>(s);
conj_helper<OtherPacket,Packet,NumTraits<OtherPacket>::IsComplex,false> pcj;
conj_helper<OtherPacket,Packet,false,false> pm;
Scalar* EIGEN_RESTRICT px = x;
Scalar* EIGEN_RESTRICT py = y;
for(Index i=0; i<size; i+=PacketSize)
{
Packet xi = pload<Packet>(px);
Packet yi = pload<Packet>(py);
pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
px += PacketSize;
py += PacketSize;
}
}
/*** non-vectorized path ***/
else
{
for(Index i=0; i<size; ++i)
{
Scalar xi = *x;
Scalar yi = *y;
*x = c * xi + numext::conj(s) * yi;
*y = -s * xi + numext::conj(c) * yi;
x += incrx;
y += incry;
}
}
apply_rotation_in_the_plane_selector<
Scalar,OtherScalar,
VectorX::SizeAtCompileTime,
EIGEN_PLAIN_ENUM_MIN(evaluator<VectorX>::Alignment, evaluator<VectorY>::Alignment),
Vectorizable>::run(x,incrx,y,incry,size,c,s);
}
} // end namespace internal

2
Eigen/src/SparseCore/AmbiVector.h
View File

@@ -94,7 +94,7 @@ class AmbiVector
Index allocSize = m_allocatedElements * sizeof(ListEl);
allocSize = (allocSize + sizeof(Scalar) - 1)/sizeof(Scalar);
Scalar* newBuffer = new Scalar[allocSize];
memcpy(newBuffer, m_buffer, copyElements * sizeof(ListEl));
std::memcpy(newBuffer, m_buffer, copyElements * sizeof(ListEl));
delete[] m_buffer;
m_buffer = newBuffer;
}

67
Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
View File

@@ -17,7 +17,9 @@ namespace internal {
template<typename Lhs, typename Rhs, typename ResultType>
static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res, bool sortedInsertion = false)
{
typedef typename remove_all<Lhs>::type::Scalar Scalar;
typedef typename remove_all<Lhs>::type::Scalar LhsScalar;
typedef typename remove_all<Rhs>::type::Scalar RhsScalar;
typedef typename remove_all<ResultType>::type::Scalar ResScalar;
// make sure to call innerSize/outerSize since we fake the storage order.
Index rows = lhs.innerSize();
@@ -25,7 +27,7 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
eigen_assert(lhs.outerSize() == rhs.innerSize());
ei_declare_aligned_stack_constructed_variable(bool, mask, rows, 0);
ei_declare_aligned_stack_constructed_variable(Scalar, values, rows, 0);
ei_declare_aligned_stack_constructed_variable(ResScalar, values, rows, 0);
ei_declare_aligned_stack_constructed_variable(Index, indices, rows, 0);
std::memset(mask,0,sizeof(bool)*rows);
@@ -51,12 +53,12 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
Index nnz = 0;
for (typename evaluator<Rhs>::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt)
{
Scalar y = rhsIt.value();
RhsScalar y = rhsIt.value();
Index k = rhsIt.index();
for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt)
{
Index i = lhsIt.index();
Scalar x = lhsIt.value();
LhsScalar x = lhsIt.value();
if(!mask[i])
{
mask[i] = true;
@@ -166,11 +168,12 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,C
{
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
{
typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
RowMajorMatrix rhsRow = rhs;
RowMajorMatrix resRow(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<RowMajorMatrix,Lhs,RowMajorMatrix>(rhsRow, lhs, resRow);
res = resRow;
typedef SparseMatrix<typename Rhs::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorRhs;
typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorRes;
RowMajorRhs rhsRow = rhs;
RowMajorRes resRow(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<RowMajorRhs,Lhs,RowMajorRes>(rhsRow, lhs, resRow);
res = resRow;
}
};
@@ -179,10 +182,11 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,R
{
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
{
typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
RowMajorMatrix lhsRow = lhs;
RowMajorMatrix resRow(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<Rhs,RowMajorMatrix,RowMajorMatrix>(rhs, lhsRow, resRow);
typedef SparseMatrix<typename Lhs::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorLhs;
typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorRes;
RowMajorLhs lhsRow = lhs;
RowMajorRes resRow(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<Rhs,RowMajorLhs,RowMajorRes>(rhs, lhsRow, resRow);
res = resRow;
}
};
@@ -219,10 +223,11 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,C
{
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
{
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
ColMajorMatrix lhsCol = lhs;
ColMajorMatrix resCol(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<ColMajorMatrix,Rhs,ColMajorMatrix>(lhsCol, rhs, resCol);
typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorLhs;
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRes;
ColMajorLhs lhsCol = lhs;
ColMajorRes resCol(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<ColMajorLhs,Rhs,ColMajorRes>(lhsCol, rhs, resCol);
res = resCol;
}
};
@@ -232,10 +237,11 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,R
{
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
{
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
ColMajorMatrix rhsCol = rhs;
ColMajorMatrix resCol(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<Lhs,ColMajorMatrix,ColMajorMatrix>(lhs, rhsCol, resCol);
typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRhs;
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRes;
ColMajorRhs rhsCol = rhs;
ColMajorRes resCol(lhs.rows(), rhs.cols());
internal::conservative_sparse_sparse_product_impl<Lhs,ColMajorRhs,ColMajorRes>(lhs, rhsCol, resCol);
res = resCol;
}
};
@@ -263,7 +269,8 @@ namespace internal {
template<typename Lhs, typename Rhs, typename ResultType>
static void sparse_sparse_to_dense_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res)
{
typedef typename remove_all<Lhs>::type::Scalar Scalar;
typedef typename remove_all<Lhs>::type::Scalar LhsScalar;
typedef typename remove_all<Rhs>::type::Scalar RhsScalar;
Index cols = rhs.outerSize();
eigen_assert(lhs.outerSize() == rhs.innerSize());
@@ -274,12 +281,12 @@ static void sparse_sparse_to_dense_product_impl(const Lhs& lhs, const Rhs& rhs,
{
for (typename evaluator<Rhs>::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt)
{
Scalar y = rhsIt.value();
RhsScalar y = rhsIt.value();
Index k = rhsIt.index();
for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt)
{
Index i = lhsIt.index();
Scalar x = lhsIt.value();
LhsScalar x = lhsIt.value();
res.coeffRef(i,j) += x * y;
}
}
@@ -310,9 +317,9 @@ struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,RowMajor,ColMa
{
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
{
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
ColMajorMatrix lhsCol(lhs);
internal::sparse_sparse_to_dense_product_impl<ColMajorMatrix,Rhs,ResultType>(lhsCol, rhs, res);
typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorLhs;
ColMajorLhs lhsCol(lhs);
internal::sparse_sparse_to_dense_product_impl<ColMajorLhs,Rhs,ResultType>(lhsCol, rhs, res);
}
};
@@ -321,9 +328,9 @@ struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,ColMajor,RowMa
{
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
{
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
ColMajorMatrix rhsCol(rhs);
internal::sparse_sparse_to_dense_product_impl<Lhs,ColMajorMatrix,ResultType>(lhs, rhsCol, res);
typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorRhs;
ColMajorRhs rhsCol(rhs);
internal::sparse_sparse_to_dense_product_impl<Lhs,ColMajorRhs,ResultType>(lhs, rhsCol, res);
}
};

22
Eigen/src/SparseCore/SparseSparseProductWithPruning.h
View File

@@ -21,7 +21,8 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
{
// return sparse_sparse_product_with_pruning_impl2(lhs,rhs,res);
typedef typename remove_all<Lhs>::type::Scalar Scalar;
typedef typename remove_all<Rhs>::type::Scalar RhsScalar;
typedef typename remove_all<ResultType>::type::Scalar ResScalar;
typedef typename remove_all<Lhs>::type::StorageIndex StorageIndex;
// make sure to call innerSize/outerSize since we fake the storage order.
@@ -31,7 +32,7 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
eigen_assert(lhs.outerSize() == rhs.innerSize());
// allocate a temporary buffer
AmbiVector<Scalar,StorageIndex> tempVector(rows);
AmbiVector<ResScalar,StorageIndex> tempVector(rows);
// mimics a resizeByInnerOuter:
if(ResultType::IsRowMajor)
@@ -63,14 +64,14 @@ static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& r
{
// FIXME should be written like this: tmp += rhsIt.value() * lhs.col(rhsIt.index())
tempVector.restart();
Scalar x = rhsIt.value();
RhsScalar x = rhsIt.value();
for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, rhsIt.index()); lhsIt; ++lhsIt)
{
tempVector.coeffRef(lhsIt.index()) += lhsIt.value() * x;
}
}
res.startVec(j);
for (typename AmbiVector<Scalar,StorageIndex>::Iterator it(tempVector,tolerance); it; ++it)
for (typename AmbiVector<ResScalar,StorageIndex>::Iterator it(tempVector,tolerance); it; ++it)
res.insertBackByOuterInner(j,it.index()) = it.value();
}
res.finalize();
@@ -85,7 +86,6 @@ struct sparse_sparse_product_with_pruning_selector;
template<typename Lhs, typename Rhs, typename ResultType>
struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,ColMajor,ColMajor>
{
typedef typename traits<typename remove_all<Lhs>::type>::Scalar Scalar;
typedef typename ResultType::RealScalar RealScalar;
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
@@ -129,8 +129,8 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,R
typedef typename ResultType::RealScalar RealScalar;
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
{
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
ColMajorMatrixLhs colLhs(lhs);
ColMajorMatrixRhs colRhs(rhs);
internal::sparse_sparse_product_with_pruning_impl<ColMajorMatrixLhs,ColMajorMatrixRhs,ResultType>(colLhs, colRhs, res, tolerance);
@@ -149,7 +149,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,R
typedef typename ResultType::RealScalar RealScalar;
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
{
typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixLhs;
typedef SparseMatrix<typename Lhs::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixLhs;
RowMajorMatrixLhs rowLhs(lhs);
sparse_sparse_product_with_pruning_selector<RowMajorMatrixLhs,Rhs,ResultType,RowMajor,RowMajor>(rowLhs,rhs,res,tolerance);
}
@@ -161,7 +161,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,C
typedef typename ResultType::RealScalar RealScalar;
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
{
typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixRhs;
typedef SparseMatrix<typename Rhs::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixRhs;
RowMajorMatrixRhs rowRhs(rhs);
sparse_sparse_product_with_pruning_selector<Lhs,RowMajorMatrixRhs,ResultType,RowMajor,RowMajor,RowMajor>(lhs,rowRhs,res,tolerance);
}
@@ -173,7 +173,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,R
typedef typename ResultType::RealScalar RealScalar;
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
{
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
typedef SparseMatrix<typename Rhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
ColMajorMatrixRhs colRhs(rhs);
internal::sparse_sparse_product_with_pruning_impl<Lhs,ColMajorMatrixRhs,ResultType>(lhs, colRhs, res, tolerance);
}
@@ -185,7 +185,7 @@ struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,C
typedef typename ResultType::RealScalar RealScalar;
static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
{
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
typedef SparseMatrix<typename Lhs::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
ColMajorMatrixLhs colLhs(lhs);
internal::sparse_sparse_product_with_pruning_impl<ColMajorMatrixLhs,Rhs,ResultType>(colLhs, rhs, res, tolerance);
}

12
Eigen/src/plugins/IndexedViewMethods.h
View File

@@ -55,9 +55,7 @@ ivcSize(const Indices& indices) const {
template<typename RowIndices, typename ColIndices>
struct valid_indexed_view_overload {
// Here we use is_convertible to Index instead of is_integral in order to treat enums as Index.
// In c++11 we could use is_integral<T> && is_enum<T> if is_convertible appears to be too permissive.
enum { value = !(internal::is_convertible<RowIndices,Index>::value && internal::is_convertible<ColIndices,Index>::value) };
enum { value = !(internal::is_valid_index_type<RowIndices>::value && internal::is_valid_index_type<ColIndices>::value) };
};
public:
@@ -146,7 +144,7 @@ operator()(const RowIndicesT (&rowIndices)[RowIndicesN], const ColIndicesT (&col
template<typename Indices>
typename internal::enable_if<
IsRowMajor && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_integral<Indices>::value)),
IsRowMajor && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_valid_index_type<Indices>::value)),
IndexedView<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,IvcIndex,typename IvcType<Indices>::type> >::type
operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
{
@@ -157,7 +155,7 @@ operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
template<typename Indices>
typename internal::enable_if<
(!IsRowMajor) && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_integral<Indices>::value)),
(!IsRowMajor) && (!(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1 || internal::is_valid_index_type<Indices>::value)),
IndexedView<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,typename IvcType<Indices>::type,IvcIndex> >::type
operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
{
@@ -168,7 +166,7 @@ operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
template<typename Indices>
typename internal::enable_if<
(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1) && (!internal::is_integral<Indices>::value) && (!Symbolic::is_symbolic<Indices>::value),
(internal::get_compile_time_incr<typename IvcType<Indices>::type>::value==1) && (!internal::is_valid_index_type<Indices>::value) && (!Symbolic::is_symbolic<Indices>::value),
VectorBlock<EIGEN_INDEXED_VIEW_METHOD_CONST Derived,internal::array_size<Indices>::value> >::type
operator()(const Indices& indices) EIGEN_INDEXED_VIEW_METHOD_CONST
{
@@ -250,6 +248,8 @@ operator()(const IndicesT (&indices)[IndicesN]) EIGEN_INDEXED_VIEW_METHOD_CONST
*
* For 1D vectors and arrays, you better use the operator()(const Indices&) overload, which behave the same way but taking a single parameter.
*
* See also this <a href="https://stackoverflow.com/questions/46110917/eigen-replicate-items-along-one-dimension-without-useless-allocations">question</a> and its answer for an example of how to duplicate coefficients.
*
* \sa operator()(const Indices&), class Block, class IndexedView, DenseBase::block(Index,Index,Index,Index)
*/
template<typename RowIndices, typename ColIndices>