Commit 156d2fa1 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Sliced-ellpack format for multimaps

parent 05ec2789

src/TNL/Experimental/Multimaps/EllpackIndexMultimap.h

+28 −11
Original line number Diff line number Diff line
@@ -15,9 +15,22 @@ 


namespace TNL {



template< typename Device >

struct EllpackIndexMultimapSliceSizeGetter

{

   static constexpr int SliceSize = 1;

};



template<>

struct EllpackIndexMultimapSliceSizeGetter< Devices::Cuda >

{

   static constexpr int SliceSize = 32;

};



template< typename Index = int,

          typename Device = Devices::Host,

          typename LocalIndex = Index >

          typename LocalIndex = Index,

          int SliceSize = EllpackIndexMultimapSliceSizeGetter< Device >::SliceSize >

class EllpackIndexMultimap

   : public virtual Object

{
@@ -25,17 +38,17 @@ class EllpackIndexMultimap 
      using DeviceType                 = Device;

      using IndexType                  = Index;

      using LocalIndexType             = LocalIndex;

      using ValuesAccessorType         = EllpackIndexMultimapValues< IndexType, DeviceType, LocalIndexType >;

      using ConstValuesAccessorType    = EllpackIndexMultimapValues< const IndexType, DeviceType, LocalIndexType >;

      using ValuesAccessorType         = EllpackIndexMultimapValues< IndexType, DeviceType, LocalIndexType, SliceSize >;

      using ConstValuesAccessorType    = EllpackIndexMultimapValues< const IndexType, DeviceType, LocalIndexType, SliceSize >;

      using ValuesAllocationVectorType = Containers::Vector< LocalIndexType, DeviceType, IndexType >;



      EllpackIndexMultimap() = default;



      template< typename Device_ >

      EllpackIndexMultimap( const EllpackIndexMultimap< Index, Device_, LocalIndex >& other );

      EllpackIndexMultimap( const EllpackIndexMultimap< Index, Device_, LocalIndex, SliceSize >& other );



      template< typename Device_ >

      EllpackIndexMultimap& operator=( const EllpackIndexMultimap< Index, Device_, LocalIndex >& other );

      EllpackIndexMultimap& operator=( const EllpackIndexMultimap< Index, Device_, LocalIndex, SliceSize >& other );



      static String getType();
@@ -50,8 +63,8 @@ class EllpackIndexMultimap 


      bool allocate( const ValuesAllocationVectorType& valuesCounts );



      template< typename Device_ >

      bool setLike( const EllpackIndexMultimap< Index, Device_, LocalIndex >& other );

      template< typename Device_, int SliceSize_ >

      bool setLike( const EllpackIndexMultimap< Index, Device_, LocalIndex, SliceSize_ >& other );



      __cuda_callable__

      ValuesAccessorType getValues( const IndexType& inputIndex );
@@ -59,7 +72,7 @@ class EllpackIndexMultimap 
      __cuda_callable__

      ConstValuesAccessorType getValues( const IndexType& inputIndex ) const;



      bool operator==( const EllpackIndexMultimap< Index, Device, LocalIndex >& other ) const;

      bool operator==( const EllpackIndexMultimap& other ) const;



      bool save( File& file ) const;
@@ -78,15 +91,19 @@ class EllpackIndexMultimap 
      IndexType keysRange = 0;

      LocalIndexType maxValuesCount = 0;



      __cuda_callable__

      IndexType getAllocationKeysRange( IndexType keysRange ) const;



      // friend class is needed for templated assignment operators

      template< typename Index_, typename Device_, typename LocalIndex_ >

      template< typename Index_, typename Device_, typename LocalIndex_, int SliceSize_ >

      friend class EllpackIndexMultimap;

};



template< typename Index,

          typename Device,

          typename LocalIndex >

std::ostream& operator << ( std::ostream& str, const EllpackIndexMultimap< Index, Device, LocalIndex >& multimap );

          typename LocalIndex,

          int SliceSize >

std::ostream& operator << ( std::ostream& str, const EllpackIndexMultimap< Index, Device, LocalIndex, SliceSize >& multimap );



} // namespace TNL

src/TNL/Experimental/Multimaps/EllpackIndexMultimapValues.h

+11 −9
Original line number Diff line number Diff line
@@ -19,19 +19,21 @@ namespace TNL { 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int SliceSize >

class EllpackIndexMultimap;



template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step = 1 >

class EllpackIndexMultimapValues

{

   public:

      using DeviceType     = Device;

      using IndexType      = Index;

      using LocalIndexType = LocalIndex;

      using NetworkType    = EllpackIndexMultimap< IndexType, DeviceType, LocalIndexType >;

      using NetworkType    = EllpackIndexMultimap< IndexType, DeviceType, LocalIndexType, step >;



      __cuda_callable__

      EllpackIndexMultimapValues();
@@ -85,8 +87,7 @@ class EllpackIndexMultimapValues 


      __cuda_callable__

      EllpackIndexMultimapValues( IndexType* values,

                                  ValuesCountType* valuesCounts,

                                  const IndexType& input,

                                  ValuesCountType* valuesCount,

                                  const LocalIndexType& allocatedSize );



      IndexType* values;
@@ -96,14 +97,15 @@ class EllpackIndexMultimapValues 
      // TODO: this is useless for a const-accessor (without setSize etc.)

      LocalIndexType allocatedSize;



      friend EllpackIndexMultimap< IndexType, DeviceType, LocalIndexType >;

      friend EllpackIndexMultimap< typename std::remove_const< IndexType >::type, DeviceType, LocalIndexType >;

      friend EllpackIndexMultimap< IndexType, DeviceType, LocalIndexType, step >;

      friend EllpackIndexMultimap< typename std::remove_const< IndexType >::type, DeviceType, LocalIndexType, step >;

};



template< typename Index,

          typename Device,

          typename LocalIndex >

std::ostream& operator << ( std::ostream& str, const EllpackIndexMultimapValues< Index, Device, LocalIndex >& ports );

          typename LocalIndex,

          int step >

std::ostream& operator << ( std::ostream& str, const EllpackIndexMultimapValues< Index, Device, LocalIndex, step >& ports );



} // namespace TNL

src/TNL/Experimental/Multimaps/EllpackIndexMultimapValues_impl.h

+68 −54
Original line number Diff line number Diff line
@@ -18,9 +18,10 @@ namespace TNL { 
 

template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

EllpackIndexMultimapValues()

: values( nullptr ), valuesCount( nullptr ), allocatedSize( 0 )

{
@@ -28,9 +29,10 @@ EllpackIndexMultimapValues() 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

EllpackIndexMultimapValues( EllpackIndexMultimapValues&& other )

: values( other.values ), valuesCount( other.valuesCount ), allocatedSize( other.allocatedSize )

{
@@ -41,10 +43,25 @@ EllpackIndexMultimapValues( EllpackIndexMultimapValues&& other ) 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

EllpackIndexMultimapValues< Index, Device, LocalIndex >&

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

EllpackIndexMultimapValues( IndexType* values,

                            ValuesCountType* valuesCount,

                            const LocalIndexType& allocatedSize )

: values( values ), valuesCount( valuesCount ), allocatedSize( allocatedSize )

{

   TNL_ASSERT( *(this->valuesCount) <= allocatedSize, );

}



template< typename Index,

          typename Device,

          typename LocalIndex,

          int step >

__cuda_callable__

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >&

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

operator=( const EllpackIndexMultimapValues& other )

{

   TNL_ASSERT( this->getSize() == other.getSize(), );
@@ -57,10 +74,11 @@ operator=( const EllpackIndexMultimapValues& other ) 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

EllpackIndexMultimapValues< Index, Device, LocalIndex >&

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >&

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

operator=( EllpackIndexMultimapValues&& other )

{

   this->values = other.values;
@@ -74,10 +92,11 @@ operator=( EllpackIndexMultimapValues&& other ) 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

void

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

bind( const EllpackIndexMultimapValues& other )

{

   this->values = other.values;
@@ -87,26 +106,11 @@ bind( const EllpackIndexMultimapValues& other ) 


template< typename Index,

          typename Device,

          typename LocalIndex >

__cuda_callable__

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues( IndexType* values,

                            ValuesCountType* valuesCounts,

                            const IndexType& input,

                            const LocalIndexType& allocatedSize )

{

   this->values = &values[ input * allocatedSize ];

   this->valuesCount = &valuesCounts[ input ];

   this->allocatedSize = allocatedSize;

   TNL_ASSERT( *(this->valuesCount) <= allocatedSize, );

}



template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

bool

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

setSize( const LocalIndexType& size )

{

   if( ! this->valuesCount || size > this->allocatedSize )
@@ -117,10 +121,11 @@ setSize( const LocalIndexType& size ) 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

LocalIndex

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

getSize() const

{

   if( ! valuesCount )
@@ -130,10 +135,11 @@ getSize() const 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

LocalIndex

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

getAllocatedSize() const

{

   return this->allocatedSize;
@@ -141,10 +147,11 @@ getAllocatedSize() const 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

void

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

setValue( const LocalIndexType& portIndex,

          const IndexType& value )

{
@@ -152,60 +159,64 @@ setValue( const LocalIndexType& portIndex, 
               std::cerr << " portIndex = " << portIndex

                         << " getSize() = " << this->getSize()

                         << std::endl );

   this->values[ portIndex ] = value;

   this->values[ portIndex * step ] = value;

}



template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

Index

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

getValue( const LocalIndexType& portIndex ) const

{

   TNL_ASSERT( portIndex < this->getSize(),

               std::cerr << " portIndex = " << portIndex

                         << " getSize() = " << this->getSize()

                         << std::endl );

   return this->values[ portIndex ];

   return this->values[ portIndex * step ];

}



template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

Index&

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

operator[]( const LocalIndexType& portIndex )

{

   TNL_ASSERT( portIndex < this->getSize(),

               std::cerr << " portIndex = " << portIndex

                         << " getSize() = " << this->getSize()

                         << std::endl );

   return this->values[ portIndex ];

   return this->values[ portIndex * step ];

}



template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

const Index&

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

operator[]( const LocalIndexType& portIndex ) const

{

   TNL_ASSERT( portIndex < this->getSize(),

               std::cerr << " portIndex = " << portIndex

                         << " getSize() = " << this->getSize()

                         << std::endl );

   return this->values[ portIndex ];

   return this->values[ portIndex * step ];

}



template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

bool

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

operator==( const EllpackIndexMultimapValues& other ) const

{

   if( this->getSize() != other.getSize() )
@@ -218,10 +229,11 @@ operator==( const EllpackIndexMultimapValues& other ) const 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

__cuda_callable__

bool

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

operator!=( const EllpackIndexMultimapValues& other ) const

{

   return ! ( *this == other );
@@ -229,9 +241,10 @@ operator!=( const EllpackIndexMultimapValues& other ) const 


template< typename Index,

          typename Device,

          typename LocalIndex >

          typename LocalIndex,

          int step >

void

EllpackIndexMultimapValues< Index, Device, LocalIndex >::

EllpackIndexMultimapValues< Index, Device, LocalIndex, step >::

print( std::ostream& str ) const

{

   str << "[ ";
@@ -246,8 +259,9 @@ print( std::ostream& str ) const 


template< typename Index,

          typename Device,

          typename LocalIndex >

std::ostream& operator << ( std::ostream& str, const EllpackIndexMultimapValues< Index, Device, LocalIndex >& ports )

          typename LocalIndex,

          int step >

std::ostream& operator << ( std::ostream& str, const EllpackIndexMultimapValues< Index, Device, LocalIndex, step >& ports )

{

   ports.print( str );

   return str;

src/TNL/Experimental/Multimaps/EllpackIndexMultimap_impl.h

+88 −50

File changed.

Preview size limit exceeded, changes collapsed.

src/TNL/Experimental/Multimaps/StaticEllpackIndexMultimap.h

+26 −9
Original line number Diff line number Diff line
@@ -15,10 +15,23 @@ 


namespace TNL {



template< typename Device >

struct StaticEllpackIndexMultimapSliceSizeGetter

{

   static constexpr int SliceSize = 1;

};



template<>

struct StaticEllpackIndexMultimapSliceSizeGetter< Devices::Cuda >

{

   static constexpr int SliceSize = 32;

};



template< int ValuesCount,

          typename Index = int,

          typename Device = Devices::Host,

          typename LocalIndex = Index >

          typename LocalIndex = Index,

          int SliceSize = StaticEllpackIndexMultimapSliceSizeGetter< Device >::SliceSize >

class StaticEllpackIndexMultimap

   : public virtual Object

{
@@ -26,16 +39,16 @@ class StaticEllpackIndexMultimap 
      using DeviceType                 = Device;

      using IndexType                  = Index;

      using LocalIndexType             = LocalIndex;

      using ValuesAccessorType         = StaticEllpackIndexMultimapValues< ValuesCount, IndexType, DeviceType, LocalIndexType >;

      using ConstValuesAccessorType    = StaticEllpackIndexMultimapValues< ValuesCount, const IndexType, DeviceType, LocalIndexType >;

      using ValuesAccessorType         = StaticEllpackIndexMultimapValues< ValuesCount, IndexType, DeviceType, LocalIndexType, SliceSize >;

      using ConstValuesAccessorType    = StaticEllpackIndexMultimapValues< ValuesCount, const IndexType, DeviceType, LocalIndexType, SliceSize >;



      StaticEllpackIndexMultimap() = default;



      template< typename Device_ >

      StaticEllpackIndexMultimap( const StaticEllpackIndexMultimap< ValuesCount, Index, Device_, LocalIndex >& other );

      StaticEllpackIndexMultimap( const StaticEllpackIndexMultimap< ValuesCount, Index, Device_, LocalIndex, SliceSize >& other );



      template< typename Device_ >

      StaticEllpackIndexMultimap& operator=( const StaticEllpackIndexMultimap< ValuesCount, Index, Device_, LocalIndex >& other );

      StaticEllpackIndexMultimap& operator=( const StaticEllpackIndexMultimap< ValuesCount, Index, Device_, LocalIndex, SliceSize >& other );



      static String getType();
@@ -49,7 +62,7 @@ class StaticEllpackIndexMultimap 
      bool allocate();



      template< typename Device_ >

      bool setLike( const StaticEllpackIndexMultimap< ValuesCount, Index, Device_, LocalIndex >& other );

      bool setLike( const StaticEllpackIndexMultimap< ValuesCount, Index, Device_, LocalIndex, SliceSize >& other );



      __cuda_callable__

      ValuesAccessorType getValues( const IndexType& inputIndex );
@@ -74,16 +87,20 @@ class StaticEllpackIndexMultimap 


      IndexType keysRange = 0;



      __cuda_callable__

      IndexType getAllocationKeysRange( IndexType keysRange ) const;



      // friend class is needed for templated assignment operators

      template< int ValuesCount_, typename Index_, typename Device_, typename LocalIndex_ >

      template< int ValuesCount_, typename Index_, typename Device_, typename LocalIndex_, int SliceSize_ >

      friend class StaticEllpackIndexMultimap;

};



template< int ValuesCount,

          typename Index,

          typename Device,

          typename LocalIndex >

std::ostream& operator << ( std::ostream& str, const StaticEllpackIndexMultimap< ValuesCount, Index, Device, LocalIndex >& multimap );

          typename LocalIndex,

          int SliceSize >

std::ostream& operator << ( std::ostream& str, const StaticEllpackIndexMultimap< ValuesCount, Index, Device, LocalIndex, SliceSize >& multimap );



} // namespace TNL