Commit af358d68 authored by Tomáš Oberhuber's avatar Tomáš Oberhuber
Browse files

Adding kernel type parameter to CSR segments.

parent 4160e723

src/TNL/Algorithms/Segments/CSR.h

+4 −2
Original line number Diff line number Diff line
@@ -22,6 +22,7 @@ namespace TNL { 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_ = CSRScalar,

          typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index > >

class CSR

{
@@ -36,6 +37,7 @@ class CSR 
      using ViewType = CSRView< Device, Index >;

      using ConstViewType = CSRView< Device, std::add_const_t< IndexType > >;

      using SegmentViewType = SegmentView< IndexType, RowMajorOrder >;

      CSRKernelTypes KernelType = KernelType_;



      CSR();
@@ -114,8 +116,8 @@ class CSR 


      CSR& operator=( const CSR& rhsSegments ) = default;



      template< typename Device_, typename Index_, typename IndexAllocator_ >

      CSR& operator=( const CSR< Device_, Index_, IndexAllocator_ >& source );

      template< typename Device_, typename Index_, CSRKernelTypes KernelType__, typename IndexAllocator_ >

      CSR& operator=( const CSR< Device_, Index_, KernelType__, IndexAllocator_ >& source );



      void save( File& file ) const;

src/TNL/Algorithms/Segments/CSR.hpp

+51 −28
Original line number Diff line number Diff line
@@ -22,16 +22,18 @@ namespace TNL { 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

CSR()

{

}



template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

CSR( const SegmentsSizes& segmentsSizes )

{

   this->setSegmentsSizes( segmentsSizes );
@@ -39,16 +41,18 @@ CSR( const SegmentsSizes& segmentsSizes ) 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

CSR( const CSR& csr ) : offsets( csr.offsets )

{

}



template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

CSR( const CSR&& csr ) : offsets( std::move( csr.offsets ) )

{
@@ -56,9 +60,10 @@ CSR( const CSR&& csr ) : offsets( std::move( csr.offsets ) ) 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

String

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

getSerializationType()

{

   return "CSR< [any_device], " + TNL::getSerializationType< IndexType >() + " >";
@@ -66,9 +71,10 @@ getSerializationType() 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

String

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

getSegmentsType()

{

   return ViewType::getSegmentsType();
@@ -76,10 +82,11 @@ getSegmentsType() 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

   template< typename SizesHolder >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

setSegmentsSizes( const SizesHolder& sizes )

{

   details::CSR< Device, Index >::setSegmentsSizes( sizes, this->offsets );
@@ -87,9 +94,10 @@ setSegmentsSizes( const SizesHolder& sizes ) 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

reset()

{

   this->offsets.setSize( 1 );
@@ -99,9 +107,10 @@ reset() 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

typename CSR< Device, Index, IndexAllocator >::ViewType

CSR< Device, Index, IndexAllocator >::

typename CSR< Device, Index, KernelType_, IndexAllocator >::ViewType

CSR< Device, Index, KernelType_, IndexAllocator >::

getView()

{

   return ViewType( this->offsets.getView() );
@@ -109,9 +118,10 @@ getView() 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

auto

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

getConstView() const -> const ConstViewType

{

   return ConstViewType( this->offsets.getConstView() );
@@ -119,8 +129,9 @@ getConstView() const -> const ConstViewType 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

__cuda_callable__ auto CSR< Device, Index, IndexAllocator >::

__cuda_callable__ auto CSR< Device, Index, KernelType_, IndexAllocator >::

getSegmentsCount() const -> IndexType

{

   return this->offsets.getSize() - 1;
@@ -128,8 +139,9 @@ getSegmentsCount() const -> IndexType 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

__cuda_callable__ auto CSR< Device, Index, IndexAllocator >::

__cuda_callable__ auto CSR< Device, Index, KernelType_, IndexAllocator >::

getSegmentSize( const IndexType segmentIdx ) const -> IndexType

{

   return details::CSR< Device, Index >::getSegmentSize( this->offsets, segmentIdx );
@@ -137,8 +149,9 @@ getSegmentSize( const IndexType segmentIdx ) const -> IndexType 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

__cuda_callable__ auto CSR< Device, Index, IndexAllocator >::

__cuda_callable__ auto CSR< Device, Index, KernelType_, IndexAllocator >::

getSize() const -> IndexType

{

   return this->getStorageSize();
@@ -146,8 +159,9 @@ getSize() const -> IndexType 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

__cuda_callable__ auto CSR< Device, Index, IndexAllocator >::

__cuda_callable__ auto CSR< Device, Index, KernelType_, IndexAllocator >::

getStorageSize() const -> IndexType

{

   return details::CSR< Device, Index >::getStorageSize( this->offsets );
@@ -155,8 +169,9 @@ getStorageSize() const -> IndexType 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

__cuda_callable__ auto CSR< Device, Index, IndexAllocator >::

__cuda_callable__ auto CSR< Device, Index, KernelType_, IndexAllocator >::

getGlobalIndex( const Index segmentIdx, const Index localIdx ) const -> IndexType

{

   if( ! std::is_same< DeviceType, Devices::Host >::value )
@@ -172,10 +187,11 @@ getGlobalIndex( const Index segmentIdx, const Index localIdx ) const -> IndexTyp 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

__cuda_callable__

auto

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

getSegmentView( const IndexType segmentIdx ) const -> SegmentViewType

{

   return SegmentViewType( offsets[ segmentIdx ], offsets[ segmentIdx + 1 ] - offsets[ segmentIdx ] );
@@ -183,10 +199,11 @@ getSegmentView( const IndexType segmentIdx ) const -> SegmentViewType 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

   template< typename Function, typename... Args >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

forSegments( IndexType first, IndexType last, Function& f, Args... args ) const

{

   this->getConstView().forSegments( first, last, f, args... );
@@ -194,10 +211,11 @@ forSegments( IndexType first, IndexType last, Function& f, Args... args ) const 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator>

   template< typename Function, typename... Args >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

forAll( Function& f, Args... args ) const

{

   this->forSegments( 0, this->getSegmentsCount(), f, args... );
@@ -205,10 +223,11 @@ forAll( Function& f, Args... args ) const 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

   template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

segmentsReduction( IndexType first, IndexType last, Fetch& fetch, const Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const

{

   this->getConstView().segmentsReduction( first, last, fetch, reduction, keeper, zero, args... );
@@ -216,10 +235,11 @@ segmentsReduction( IndexType first, IndexType last, Fetch& fetch, const Reductio 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

   template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

allReduction( Fetch& fetch, const Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const

{

   this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
@@ -227,11 +247,12 @@ allReduction( Fetch& fetch, const Reduction& reduction, ResultKeeper& keeper, co 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

   template< typename Device_, typename Index_, typename IndexAllocator_ >

CSR< Device, Index, IndexAllocator >&

CSR< Device, Index, IndexAllocator >::

operator=( const CSR< Device_, Index_, IndexAllocator_ >& source )

   template< typename Device_, typename Index_, CSRKernelTypes KernelType__, typename IndexAllocator_ >

CSR< Device, Index, KernelType_, IndexAllocator >&

CSR< Device, Index, KernelType_, IndexAllocator >::

operator=( const CSR< Device_, Index_, KernelType__, IndexAllocator_ >& source )

{

   this->offsets = source.offsets;

   return *this;
@@ -239,9 +260,10 @@ operator=( const CSR< Device_, Index_, IndexAllocator_ >& source ) 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

save( File& file ) const

{

   file << this->offsets;
@@ -249,9 +271,10 @@ save( File& file ) const 


template< typename Device,

          typename Index,

          CSRKernelTypes KernelType_,

          typename IndexAllocator >

void

CSR< Device, Index, IndexAllocator >::

CSR< Device, Index, KernelType_, IndexAllocator >::

load( File& file )

{

   file >> this->offsets;

src/TNL/Algorithms/Segments/CSRView.h

+6 −2
Original line number Diff line number Diff line
@@ -19,8 +19,11 @@ namespace TNL { 
   namespace Algorithms {

      namespace Segments {



enum CSRKernelTypes { CSRScalarKernel, CSRVectorKernel, CSRLightKernel };



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ = CSRScalar >

class CSRView

{

   public:
@@ -34,6 +37,7 @@ class CSRView 
      using ViewTemplate = CSRView< Device_, Index_ >;

      using ConstViewType = CSRView< Device, std::add_const_t< Index > >;

      using SegmentViewType = SegmentView< IndexType, RowMajorOrder >;

      CSRKernelTypes KernelType = KernelType_;



      __cuda_callable__

      CSRView();

src/TNL/Algorithms/Segments/CSRView.hpp

+81 −56
Original line number Diff line number Diff line
@@ -22,122 +22,136 @@ namespace TNL { 




template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

CSRView()

{

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

CSRView( const OffsetsView& offsets_view )

   : offsets( offsets_view )

{

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

CSRView( const OffsetsView&& offsets_view )

   : offsets( offsets_view )

{

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

CSRView( const CSRView& csr_view )

   : offsets( csr_view.offsets )

{

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

CSRView( const CSRView&& csr_view )

   : offsets( std::move( csr_view.offsets ) )

{

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

String

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

getSerializationType()

{

   return "CSR< [any_device], " + TNL::getSerializationType< IndexType >() + " >";

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

String

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

getSegmentsType()

{

   return "CSR";

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

typename CSRView< Device, Index >::ViewType

CSRView< Device, Index >::

typename CSRView< Device, Index, KernelType_ >::ViewType

CSRView< Device, Index, KernelType_ >::

getView()

{

   return ViewType( this->offsets );

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

auto

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

getConstView() const -> const ConstViewType

{

   return ConstViewType( this->offsets.getConstView() );

}



template< typename Device,

          typename Index >

__cuda_callable__ auto CSRView< Device, Index >::

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__ auto CSRView< Device, Index, KernelType_ >::

getSegmentsCount() const -> IndexType

{

   return this->offsets.getSize() - 1;

}



template< typename Device,

          typename Index >

__cuda_callable__ auto CSRView< Device, Index >::

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__ auto CSRView< Device, Index, KernelType_ >::

getSegmentSize( const IndexType segmentIdx ) const -> IndexType

{

   return details::CSR< Device, Index >::getSegmentSize( this->offsets, segmentIdx );

}



template< typename Device,

          typename Index >

__cuda_callable__ auto CSRView< Device, Index >::

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__ auto CSRView< Device, Index, KernelType_ >::

getSize() const -> IndexType

{

   return this->getStorageSize();

}



template< typename Device,

          typename Index >

__cuda_callable__ auto CSRView< Device, Index >::

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__ auto CSRView< Device, Index, KernelType_ >::

getStorageSize() const -> IndexType

{

   return details::CSR< Device, Index >::getStorageSize( this->offsets );

}



template< typename Device,

          typename Index >

__cuda_callable__ auto CSRView< Device, Index >::

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__ auto CSRView< Device, Index, KernelType_ >::

getGlobalIndex( const Index segmentIdx, const Index localIdx ) const -> IndexType

{

   if( ! std::is_same< DeviceType, Devices::Host >::value )
@@ -152,20 +166,22 @@ getGlobalIndex( const Index segmentIdx, const Index localIdx ) const -> IndexTyp 
}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

__cuda_callable__

auto

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

getSegmentView( const IndexType segmentIdx ) const -> SegmentViewType

{

   return SegmentViewType( offsets[ segmentIdx ], offsets[ segmentIdx + 1 ] - offsets[ segmentIdx ], 1 );

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

   template< typename Function, typename... Args >

void

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

forSegments( IndexType first, IndexType last, Function& f, Args... args ) const

{

   const auto offsetsView = this->offsets;
@@ -181,24 +197,28 @@ forSegments( IndexType first, IndexType last, Function& f, Args... args ) const 
}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

   template< typename Function, typename... Args >

void

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

forAll( Function& f, Args... args ) const

{

   this->forSegments( 0, this->getSegmentsCount(), f, args... );

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

   template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >

void

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

segmentsReduction( IndexType first, IndexType last, Fetch& fetch, const Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const

{

   using RealType = typename details::FetchLambdaAdapter< Index, Fetch >::ReturnType;

   const auto offsetsView = this->offsets.getConstView();

   if( KernelType == CSRScalar )

   {

      auto l = [=] __cuda_callable__ ( const IndexType segmentIdx, Args... args ) mutable {

         const IndexType begin = offsetsView[ segmentIdx ];

         const IndexType end = offsetsView[ segmentIdx + 1 ];
@@ -211,21 +231,24 @@ segmentsReduction( IndexType first, IndexType last, Fetch& fetch, const Reductio 
      };

   Algorithms::ParallelFor< Device >::exec( first, last, l, args... );

   }

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

   template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >

void

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

allReduction( Fetch& fetch, const Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const

{

   this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );

}



template< typename Device,

          typename Index >

CSRView< Device, Index >&

CSRView< Device, Index >::

          typename Index,

          CSRKernelTypes KernelType_ >

CSRView< Device, Index, KernelType_ >&

CSRView< Device, Index, KernelType_ >::

operator=( const CSRView& view )

{

   this->offsets.bind( view.offsets );
@@ -233,18 +256,20 @@ operator=( const CSRView& view ) 
}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

void

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

save( File& file ) const

{

   file << this->offsets;

}



template< typename Device,

          typename Index >

          typename Index,

          CSRKernelTypes KernelType_ >

void

CSRView< Device, Index >::

CSRView< Device, Index, KernelType_ >::

load( File& file )

{

   file >> this->offsets;