#include <deal.II/base/vectorization.h>

Public Member Functions
VectorizedArray &	operator= (const double x)

double &	operator[] (const unsigned int comp)

const double &	operator[] (const unsigned int comp) const

VectorizedArray &	operator+= (const VectorizedArray &vec)

VectorizedArray &	operator-= (const VectorizedArray &vec)

VectorizedArray &	operator*= (const VectorizedArray &vec)

VectorizedArray &	operator/= (const VectorizedArray &vec)

void	load (const double *ptr)

void	store (double *ptr) const

void	streaming_store (double *ptr) const

void	gather (const double base_ptr, const unsigned int offsets)

void	scatter (const unsigned int offsets, double base_ptr) const

Public Attributes
__m128d	data

Static Public Attributes
static const unsigned int	n_array_elements = 2

Private Member Functions
VectorizedArray	get_sqrt () const

VectorizedArray	get_abs () const

VectorizedArray	get_max (const VectorizedArray &other) const

VectorizedArray	get_min (const VectorizedArray &other) const

Friends
template<typename Number2 >
VectorizedArray< Number2 >	std::sqrt (const VectorizedArray< Number2 > &)

Detailed Description

template<>
class VectorizedArray< double >

Specialization for double and SSE2.

Definition at line 2141 of file vectorization.h.

Member Function Documentation

◆ operator=()

VectorizedArray& VectorizedArray< double >::operator= ( const double x )

inline

This function can be used to set all data fields to a given scalar.

Definition at line 2154 of file vectorization.h.

◆ operator[]() [1/2]

double& VectorizedArray< double >::operator[] ( const unsigned int comp )

inline

Access operator.

Definition at line 2165 of file vectorization.h.

◆ operator[]() [2/2]

const double& VectorizedArray< double >::operator[] ( const unsigned int comp ) const

inline

Constant access operator.

Definition at line 2176 of file vectorization.h.

◆ operator+=()

VectorizedArray& VectorizedArray< double >::operator+= ( const VectorizedArray< double > & vec )

inline

Addition.

Definition at line 2187 of file vectorization.h.

◆ operator-=()

VectorizedArray& VectorizedArray< double >::operator-= ( const VectorizedArray< double > & vec )

inline

Subtraction.

Definition at line 2202 of file vectorization.h.

◆ operator*=()

VectorizedArray& VectorizedArray< double >::operator*= ( const VectorizedArray< double > & vec )

inline

Multiplication.

Definition at line 2217 of file vectorization.h.

◆ operator/=()

VectorizedArray& VectorizedArray< double >::operator/= ( const VectorizedArray< double > & vec )

inline

Division.

Definition at line 2232 of file vectorization.h.

◆ load()

void VectorizedArray< double >::load ( const double * ptr )

inline

Load n_array_elements from memory into the calling class, starting at the given address. The memory need not be aligned by 16 bytes, as opposed to casting a double address to VectorizedArray<double>*.

Definition at line 2248 of file vectorization.h.

◆ store()

void VectorizedArray< double >::store ( double * ptr ) const

inline

Write the content of the calling class into memory in form of n_array_elements to the given address. The memory need not be aligned by 16 bytes, as opposed to casting a double address to VectorizedArray<double>*.

Definition at line 2260 of file vectorization.h.

◆ streaming_store()

void VectorizedArray< double >::streaming_store ( double * ptr ) const

inline

Write the content of the calling class into memory in form of n_array_elements to the given address using non-temporal stores that bypass the processor's caches, using _mm_stream_pd store intrinsics on supported CPUs. The destination of the store ptr must be aligned by the amount of bytes in the vectorized array.

This store operation can be faster than usual store operations in case the store is streaming because it avoids the read-for-ownership transfer typically invoked in standard stores. This approximately works as follows (see the literature on computer architecture for details): When an algorithm stores some results to a memory address, a processor typically wants to move it into some of its caches as it expects the data to be re-used again at some point. Since caches are organized in lines of sizes either 64 byte or 128 byte but writes are usually smaller, a processor must first load in the destination cache line upon a write because only part of the cache line is overwritten initially. If a series of stores write data in a chunk bigger than any of its caches could handle, the data finally has to be moved out from the caches to main memory. But since all addressed have first been read, this doubles the load on main memory, which can incur a performance penalty. Furthermore, the organization of caches in a multicore context also requires reading an address before something can be written to cache to that address, see e.g. the Wikipedia article on the MESI protocol for details. The instruction underlying this function call signals to the processor that these two prerequisites on a store are relaxed: Firstly, one expects the whole cache line to be overwritten (that the memory subsystem then handles appropriately), so no need to first read the "remainder" of the cache line. Secondly, the data behind that particular memory will not be subject to cache coherency protocol as it will be in main memory both when the same processor wants to access it again as well as any other processors in a multicore chip. Due to this particular setup, any subsequent access to the data written by this function will need to query main memory, which is slower than an access from a cache both latency-wise and throughput-wise. Thus, this command should only be used for large stores that will collectively not fit into caches, as performance will be degraded otherwise. For a typical use case, see also this blog article.

Note that streaming stores are only available in the specialized SSE/AVX classes of VectorizedArray of type double or float, not in the generic base class.

Note: Memory must be aligned by 16 bytes.

Definition at line 2269 of file vectorization.h.

◆ gather()

void VectorizedArray< double >::gather	(	const double *	base_ptr,
		const unsigned int *	offsets
	)

inline

Load n_array_elements from memory into the calling class, starting at the given address and with given offsets, each entry from the offset providing one element of the vectorized array.

This operation corresponds to the following code (but uses a more efficient implementation in case the hardware allows for that):

for (unsigned int v=0; v<VectorizedArray<Number>::n_array_elements; ++v)

this->operator[](v) = base_ptr[offsets[v]];

Definition at line 2288 of file vectorization.h.

◆ scatter()

void VectorizedArray< double >::scatter	(	const unsigned int *	offsets,
		double *	base_ptr
	)		const

inline

Write the content of the calling class into memory in form of n_array_elements to the given address and the given offsets, filling the elements of the vectorized array into each offset.

This operation corresponds to the following code (but uses a more efficient implementation in case the hardware allows for that):

for (unsigned int v=0; v<VectorizedArray<Number>::n_array_elements; ++v)

base_ptr[offsets[v]] = this->operator[](v);

Definition at line 2308 of file vectorization.h.

◆ get_sqrt()

VectorizedArray VectorizedArray< double >::get_sqrt ( ) const

inlineprivate

Return the square root of this field. Not for use in user code. Use sqrt(x) instead.

Definition at line 2328 of file vectorization.h.

◆ get_abs()

VectorizedArray VectorizedArray< double >::get_abs ( ) const

inlineprivate

Return the absolute value of this field. Not for use in user code. Use abs(x) instead.

Definition at line 2341 of file vectorization.h.

◆ get_max()

VectorizedArray VectorizedArray< double >::get_max ( const VectorizedArray< double > & other ) const

inlineprivate

Return the component-wise maximum of this field and another one. Not for use in user code. Use max(x,y) instead.

Definition at line 2359 of file vectorization.h.

◆ get_min()

VectorizedArray VectorizedArray< double >::get_min ( const VectorizedArray< double > & other ) const

inlineprivate

Return the component-wise minimum of this field and another one. Not for use in user code. Use min(x,y) instead.

Definition at line 2372 of file vectorization.h.

Friends And Related Function Documentation

◆ std::sqrt

template<typename Number2 >

VectorizedArray<Number2> std::sqrt ( const VectorizedArray< Number2 > & )

friend

Make a few functions friends.

Member Data Documentation

◆ n_array_elements

const unsigned int VectorizedArray< double >::n_array_elements = 2

static

This gives the number of vectors collected in this class.

Definition at line 2147 of file vectorization.h.

◆ data

__m128d VectorizedArray< double >::data

Actual data field. Since this class represents a POD data type, it remains public.

Definition at line 2319 of file vectorization.h.

The documentation for this class was generated from the following file:

deal.II/base/vectorization.h

Public Member Functions

Public Attributes

Static Public Attributes

Private Member Functions

Friends

Detailed Description

template<> class VectorizedArray< double >

Member Function Documentation

◆ operator=()

◆ operator[]() [1/2]

◆ operator[]() [2/2]

◆ operator+=()

◆ operator-=()

◆ operator*=()

◆ operator/=()

◆ load()

◆ store()

◆ streaming_store()

◆ gather()

◆ scatter()

◆ get_sqrt()

◆ get_abs()

◆ get_max()

◆ get_min()

Friends And Related Function Documentation

◆ std::sqrt

Member Data Documentation

◆ n_array_elements

◆ data

template<>
class VectorizedArray< double >