Template Numerical Library version\ main:0b2c40fb1
Loading...
Searching...
No Matches
Public Types | Public Member Functions | Protected Member Functions | Protected Attributes | List of all members
TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize > Class Template Reference

Data structure for Bisection Ellpack segments. More...

#include <TNL/Algorithms/Segments/BiEllpack.h>

Inheritance diagram for TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >:
[legend]
Collaboration diagram for TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >:
[legend]

Public Types

using ConstViewType = typename ViewType::ConstViewType
 Type of constant segments view.
using IndexAllocatorType = IndexAllocator
using OffsetsContainer = Containers::Vector< Index, Device, typename Base::IndexType, IndexAllocator >
template<typename Device_ = Device, typename Index_ = Index, typename IndexAllocator_ = typename Allocators::Default< Device_ >::template Allocator< Index_ >, ElementsOrganization Organization_ = Organization, int WarpSize_ = WarpSize>
using Self = BiEllpack< Device_, Index_, IndexAllocator_, Organization_, WarpSize_ >
 Templated type for creating BiEllpack segments with different template parameters.
template<typename Device_ = Device, typename Index_ = Index>
using ViewTemplate = BiEllpackView< Device_, Index_, Organization, WarpSize >
 Templated view type.
using ViewType = BiEllpackView< Device, Index, Organization, WarpSize >
 Type of segments view.
Public Types inherited from TNL::Algorithms::Segments::BiEllpackBase< Device, Index, DefaultElementsOrganization< Device >::getOrganization(), Backend::getWarpSize() >
using ConstOffsetsView
 The type for representing the constant vector view with offsets of the BiEllpack groups.
using DeviceType
 The device where the segments are operating.
using IndexType
 The type used for indexing of segments elements.
using OffsetsView
 The type for representing the vector view with offsets of the BiEllpack groups.
using SegmentViewType
 Accessor type for one particular segment.

Public Member Functions

 BiEllpack ()=default
 Constructor with no parameters to create empty segments.
 BiEllpack (BiEllpack &&segments) noexcept=default
 Move constructor.
 BiEllpack (const BiEllpack &segments)
 Copy constructor (makes deep copy).
template<typename SizesContainer, std::enable_if_t< IsArrayType< SizesContainer >::value, bool > = true>
 BiEllpack (const SizesContainer &segmentsSizes)
 Constructor that initializes segments based on their sizes.
template<typename ListIndex>
 BiEllpack (const std::initializer_list< ListIndex > &segmentsSizes)
 Constructor that initializes segments using an initializer list.
ConstViewType getConstView () const
 Returns a constant view for this instance of segments which can by used for example in lambda functions running in GPU kernels.
ViewType getView ()
 Returns a view for this instance of segments which can by used for example in lambda functions running in GPU kernels.
template<typename SizesHolder>
void initGroupPointers (const SizesHolder &segmentsSizes)
void load (File &file)
 Method for loading the segments from a file in a binary form.
BiEllpack & operator= (BiEllpack &&segments) noexcept(false)
 Move-assignment operator.
BiEllpack & operator= (const BiEllpack &segments)
 Copy-assignment operator (makes a deep copy).
template<typename Device_, typename Index_, typename IndexAllocator_, ElementsOrganization Organization_>
BiEllpack & operator= (const BiEllpack< Device_, Index_, IndexAllocator_, Organization_, WarpSize > &segments)
 Assignment operator for segments with different template parameters.
void reset ()
 Reset the segments to empty states (it means that there is no segment in the segments).
void save (File &file) const
 Method for saving the segments to a file in a binary form.
template<typename SizesHolder>
void setSegmentsSizes (const SizesHolder &sizes)
 Set sizes of particular segments.
Public Member Functions inherited from TNL::Algorithms::Segments::BiEllpackBase< Device, Index, DefaultElementsOrganization< Device >::getOrganization(), Backend::getWarpSize() >
__cuda_callable__ BiEllpackBase ()=default
 Default constructor with no parameters to create empty segments view.
void forAllElements (Function &&function) const
void forAllElementsIf (Condition condition, Function function) const
void forAllSegments (Function &&function) const
void forElements (IndexType begin, IndexType end, Function &&function) const
void forElementsIf (IndexType begin, IndexType end, Condition condition, Function function) const
void forSegments (IndexType begin, IndexType end, Function &&function) const
__cuda_callable__ IndexType getElementCount () const
 Returns the number of elements managed by all segments.
__cuda_callable__ IndexType getGlobalIndex (Index segmentIdx, Index localIdx) const
 Computes the global index of an element managed by the segments.
__cuda_callable__ OffsetsView getGroupPointersView ()
 Returns a view containing the pointers to the beginning of each segment.
__cuda_callable__ IndexType getSegmentCount () const
 Returns the number of segments.
__cuda_callable__ IndexType getSegmentsCount () const
 Returns the number of segments. Deprecated, use getSegmentCount().
__cuda_callable__ IndexType getSegmentSize (IndexType segmentIdx) const
 Returns the size of a particular segment denoted by segmentIdx.
__cuda_callable__ OffsetsView getSegmentsPermutationView ()
 Returns a view containing the permutation of segments.
__cuda_callable__ SegmentViewType getSegmentView (IndexType segmentIdx) const
 Returns a segment view (i.e., a segment accessor) for the specified segment index.
__cuda_callable__ IndexType getSize () const
 Returns the number of elements managed by all segments.
__cuda_callable__ IndexType getStorageSize () const
 Returns number of elements that needs to be allocated by a container connected to this segments.
__cuda_callable__ IndexType getVirtualSegments () const
BiEllpackBase & operator= (const BiEllpackBase &)=delete
 Copy-assignment operator.
void printStructure (std::ostream &str) const
 Prints the structure of the segments to the specified output stream.

Protected Member Functions

Index getStripLength (Index strip) const
template<typename SizesHolder>
void initSegmentsPermutation (const SizesHolder &segmentsSize)
template<typename SizesHolder>
void verifySegmentLengths (const SizesHolder &segmentsSizes)
template<typename SizesHolder>
void verifySegmentPerm (const SizesHolder &segmentsSizes)
Protected Member Functions inherited from TNL::Algorithms::Segments::BiEllpackBase< Device, Index, DefaultElementsOrganization< Device >::getOrganization(), Backend::getWarpSize() >
__cuda_callable__ void bind (IndexType size, IndexType storageSize, OffsetsView segmentsPermutation, OffsetsView groupPointers)
 Re-initializes the internal attributes of the base class.

Protected Attributes

OffsetsContainer groupPointers
OffsetsContainer segmentsPermutation
Protected Attributes inherited from TNL::Algorithms::Segments::BiEllpackBase< Device, Index, DefaultElementsOrganization< Device >::getOrganization(), Backend::getWarpSize() >
OffsetsView groupPointers
OffsetsView segmentsPermutation
IndexType size
IndexType storageSize

Additional Inherited Members

Static Public Member Functions inherited from TNL::Algorithms::Segments::BiEllpackBase< Device, Index, DefaultElementsOrganization< Device >::getOrganization(), Backend::getWarpSize() >
static constexpr int getLogWarpSize ()
 Returns the log2 of the warp size used for the segments.
static constexpr ElementsOrganization getOrganization ()
 Returns the data layout for the biellpack format (it is always row-major order).
static std::string getSegmentsType ()
 Returns string with the segments type.
static std::string getSerializationType ()
 Returns string with the serialization type.
static constexpr int getWarpSize ()
 Returns the warp size used for the segments.
static constexpr bool havePadding ()
 This function denotes that the biellpack format uses padding elements.

Detailed Description

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
class TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >

Data structure for Bisection Ellpack segments.

Bisection Ellpack segments are inspired by the following paper:

[C. Zheng, S. Gu, T.-X. Gu, B. Yang, X.-P. Liu, BiELL: A bisection ELLPACK-based storage format for optimizing SpMV on GPUs, Journal of Parallel and Distributed Computing, Volume 74, Issue 7, 2014, pp. 2639-2647](https://www.sciencedirect.com/science/article/pii/S0743731514000458).

This format is designed to improve load balancing for segments with unevenly distributed sizes. It uses more meta-information compared to, for example, TNL::Algorithms::Segments::CSR, which can introduce overhead. However, it can be beneficial for reduction operations, particularly for longer segments with significant size variations.

See TNL::Algorithms::Segments for more details about segments.

Template Parameters
DeviceThe type of device on which the segments will operate.
IndexThe type used for indexing elements managed by the segments.
IndexAllocatorThe allocator used for managing index containers.
OrganizationThe organization of the elements in the segments—either row-major or column-major order.
WarpSizeThe warp size used for the segments.

Member Typedef Documentation

◆ Self

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
template<typename Device_ = Device, typename Index_ = Index, typename IndexAllocator_ = typename Allocators::Default< Device_ >::template Allocator< Index_ >, ElementsOrganization Organization_ = Organization, int WarpSize_ = WarpSize>
using TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::Self = BiEllpack< Device_, Index_, IndexAllocator_, Organization_, WarpSize_ >

Templated type for creating BiEllpack segments with different template parameters.

Template Parameters
Device_is alternative device type.
Index_is alternative index type.
IndexAllocator_is alternative index allocator type.

◆ ViewTemplate

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
template<typename Device_ = Device, typename Index_ = Index>
using TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::ViewTemplate = BiEllpackView< Device_, Index_, Organization, WarpSize >

Templated view type.

Template Parameters
Device_is alternative device type for the view.
Index_is alternative index type for the view.

Constructor & Destructor Documentation

◆ BiEllpack() [1/2]

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
template<typename SizesContainer, std::enable_if_t< IsArrayType< SizesContainer >::value, bool > = true>
TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::BiEllpack ( const SizesContainer & segmentsSizes)

Constructor that initializes segments based on their sizes.

The number of segments is determined by the size of segmentsSizes. Each element in this container specifies the size of a corresponding segment.

Template Parameters
SizesContainerThe type of container used to store segment sizes. It can be, for example, TNL::Containers::Array or TNL::Containers::Vector.
Parameters
segmentsSizesAn instance of the container holding the sizes of the segments.

See the following example:

1#include <iostream>
2#include <TNL/Containers/Vector.h>
3#include <TNL/Algorithms/Segments/traverse.h>
4#include <TNL/Devices/Host.h>
5#include <TNL/Devices/Cuda.h>
6
7template< typename Segments >
8void
9SegmentsExample()
10{
11 using Device = typename Segments::DeviceType;
12
13 /***
14 * Create segments with given segments sizes.
15 */
16 TNL::Containers::Vector< int, Device > segmentsSizes{ 1, 2, 3, 4, 5 };
17 Segments segments( segmentsSizes );
18 std::cout << "Segments sizes are: " << segments << '\n';
19
20 /***
21 * Allocate array for the segments;
22 */
23 TNL::Containers::Array< double, Device > data( segments.getStorageSize(), 0.0 );
24
25 /***
26 * Insert data into particular segments.
27 */
28 auto data_view = data.getView();
29 TNL::Algorithms::Segments::forAllElements(
30 segments,
31 [ = ] __cuda_callable__( int segmentIdx, int localIdx, int globalIdx ) mutable
32 {
33 if( localIdx <= segmentIdx )
34 data_view[ globalIdx ] = segmentIdx;
35 } );
36
37 /***
38 * Print the data managed by the segments.
39 */
40 auto fetch = [ = ] __cuda_callable__( int globalIdx ) -> double
41 {
42 return data_view[ globalIdx ];
43 };
44 std::cout << TNL::Algorithms::Segments::print( segments, fetch ) << '\n';
45}
46
47int
48main( int argc, char* argv[] )
49{
50 std::cout << "Example of CSR segments on host:\n";
51 SegmentsExample< TNL::Algorithms::Segments::CSR< TNL::Devices::Host, int > >();
52
53 std::cout << "Example of Ellpack segments on host:\n";
54 SegmentsExample< TNL::Algorithms::Segments::Ellpack< TNL::Devices::Host, int > >();
55
56#ifdef __CUDACC__
57 std::cout << "Example of CSR segments on CUDA GPU:\n";
58 SegmentsExample< TNL::Algorithms::Segments::CSR< TNL::Devices::Cuda, int > >();
59
60 std::cout << "Example of Ellpack segments on CUDA GPU:\n";
61 SegmentsExample< TNL::Algorithms::Segments::Ellpack< TNL::Devices::Cuda, int > >();
62#endif
63 return EXIT_SUCCESS;
64}
__cuda_callable__
#define __cuda_callable__
Definition Macros.h:49
TNL::Containers::Array
Array is responsible for memory management, access to array elements, and general array operations.
Definition Array.h:65
TNL::Containers::Vector
Vector extends Array with algebraic operations.
Definition Vector.h:37
TNL::Algorithms::Segments
Namespace for the segments data structures.
Definition _NamespaceDoxy.h:7
TNL::Algorithms::Segments::print
SegmentsPrinter< typename Segments::ConstViewType, Fetch > print(const Segments &segments, Fetch fetch)
Print segments sizes, i.e. the segments setup.
TNL::Algorithms::Segments::forAllElements
void forAllElements(const Segments &segments, Function &&function, LaunchConfiguration launchConfig=Algorithms::Segments::LaunchConfiguration())
Iterates in parallel over all elements of all segments and applies the specified lambda function.

The expected output is:

Example of CSR segments on host:
Segments sizes are: [ 1, 2, 3, 4, 5 ]
Segment 0: [ 0 ]
Segment 1: [ 1, 1 ]
Segment 2: [ 2, 2, 2 ]
Segment 3: [ 3, 3, 3, 3 ]
Segment 4: [ 4, 4, 4, 4, 4 ]
Example of Ellpack segments on host:
Segments sizes are: [ 5, 5, 5, 5, 5 ]
Segment 0: [ 0, 0, 0, 0, 0 ]
Segment 1: [ 1, 1, 0, 0, 0 ]
Segment 2: [ 2, 2, 2, 0, 0 ]
Segment 3: [ 3, 3, 3, 3, 0 ]
Segment 4: [ 4, 4, 4, 4, 4 ]

◆ BiEllpack() [2/2]

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
template<typename ListIndex>
TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::BiEllpack ( const std::initializer_list< ListIndex > & segmentsSizes)

Constructor that initializes segments using an initializer list.

The number of segments is determined by the size of segmentsSizes. Each element in this initializer list specifies the size of a corresponding segment.

Template Parameters
ListIndexThe type used for indexing elements in the initializer list.
Parameters
segmentsSizesAn initializer list defining the sizes of the segments.

See the following example:

1#include <iostream>
2#include <TNL/Containers/Vector.h>
3#include <TNL/Algorithms/Segments/traverse.h>
4#include <TNL/Devices/Host.h>
5#include <TNL/Devices/Cuda.h>
6
7template< typename Segments >
8void
9SegmentsExample()
10{
11 using Device = typename Segments::DeviceType;
12
13 /***
14 * Create segments with given segments sizes.
15 */
16 Segments segments{ 1, 2, 3, 4, 5 };
17 std::cout << "Segments sizes are: " << segments << '\n';
18
19 /***
20 * Allocate array for the segments;
21 */
22 TNL::Containers::Array< double, Device > data( segments.getStorageSize(), 0.0 );
23
24 /***
25 * Insert data into particular segments.
26 */
27 auto data_view = data.getView();
28 TNL::Algorithms::Segments::forAllElements(
29 segments,
30 [ = ] __cuda_callable__( int segmentIdx, int localIdx, int globalIdx ) mutable
31 {
32 if( localIdx <= segmentIdx )
33 data_view[ globalIdx ] = segmentIdx;
34 } );
35
36 /***
37 * Print the data managed by the segments.
38 */
39 auto fetch = [ = ] __cuda_callable__( int globalIdx ) -> double
40 {
41 return data_view[ globalIdx ];
42 };
43 std::cout << TNL::Algorithms::Segments::print( segments, fetch ) << '\n';
44}
45
46int
47main( int argc, char* argv[] )
48{
49 std::cout << "Example of CSR segments on host:\n";
50 SegmentsExample< TNL::Algorithms::Segments::CSR< TNL::Devices::Host, int > >();
51
52 std::cout << "Example of Ellpack segments on host:\n";
53 SegmentsExample< TNL::Algorithms::Segments::Ellpack< TNL::Devices::Host, int > >();
54
55#ifdef __CUDACC__
56 std::cout << "Example of CSR segments on CUDA GPU:\n";
57 SegmentsExample< TNL::Algorithms::Segments::CSR< TNL::Devices::Cuda, int > >();
58
59 std::cout << "Example of Ellpack segments on CUDA GPU:\n";
60 SegmentsExample< TNL::Algorithms::Segments::Ellpack< TNL::Devices::Cuda, int > >();
61#endif
62 return EXIT_SUCCESS;
63}

The expected output is:

Example of CSR segments on host:
Segments sizes are: [ 1, 2, 3, 4, 5 ]
Segment 0: [ 0 ]
Segment 1: [ 1, 1 ]
Segment 2: [ 2, 2, 2 ]
Segment 3: [ 3, 3, 3, 3 ]
Segment 4: [ 4, 4, 4, 4, 4 ]
Example of Ellpack segments on host:
Segments sizes are: [ 5, 5, 5, 5, 5 ]
Segment 0: [ 0, 0, 0, 0, 0 ]
Segment 1: [ 1, 1, 0, 0, 0 ]
Segment 2: [ 2, 2, 2, 0, 0 ]
Segment 3: [ 3, 3, 3, 3, 0 ]
Segment 4: [ 4, 4, 4, 4, 4 ]

Member Function Documentation

◆ load()

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
void TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::load ( File & file)

Method for loading the segments from a file in a binary form.

Parameters
fileis the source file.

◆ operator=()

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
template<typename Device_, typename Index_, typename IndexAllocator_, ElementsOrganization Organization_>
BiEllpack & TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::operator= ( const BiEllpack< Device_, Index_, IndexAllocator_, Organization_, WarpSize > & segments)

Assignment operator for segments with different template parameters.

Performs a deep copy of the source segments.

Template Parameters
Device_The device type of the source segments.
Index_The index type of the source segments.
IndexAllocator_The index allocator type of the source segments.
Parameters
segmentsThe source segments object.
Returns
A reference to this instance.

◆ save()

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
void TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::save ( File & file) const

Method for saving the segments to a file in a binary form.

Parameters
fileis the target file.

◆ setSegmentsSizes()

template<typename Device, typename Index, typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >, ElementsOrganization Organization = DefaultElementsOrganization< Device >::getOrganization(), int WarpSize = Backend::getWarpSize()>
template<typename SizesHolder>
void TNL::Algorithms::Segments::BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::setSegmentsSizes ( const SizesHolder & sizes)

Set sizes of particular segments.

Template Parameters
SizesContaineris a container with segments sizes. It can be TNL::Containers::Array or TNL::Containers::Vector for example.
Parameters
sizesis an instance of the container with segments sizes.
The documentation for this class was generated from the following file:
Generated on for Template Numerical Library by doxygen 1.16.1