Added methods forRows and rowsReduction to dense matrix. (873b9a9c) · Commits · TNL / tnl-dev

src/TNL/Matrices/Dense.h

+123 −107

Original line number	Diff line number	Diff line
		@@ -84,6 +84,10 @@ public:

		IndexType getNumberOfNonzeroMatrixElements() const;

		template< typename Vector >
		void getCompressedRowLengths( Vector& rowLengths ) const;


		void reset();

		__cuda_callable__
		@@ -115,11 +119,23 @@ public:
		Real getElement( const IndexType row,
		const IndexType column ) const;

		/*__cuda_callable__
		MatrixRow getRow( const IndexType rowIndex );
		template< typename Fetch, typename Reduce, typename Keep, typename FetchReal >
		void rowsReduction( IndexType first, IndexType last, Fetch& fetch, Reduce& reduce, Keep& keep, const FetchReal& zero ) const;

		__cuda_callable__
		const MatrixRow getRow( const IndexType rowIndex ) const;*/
		template< typename Fetch, typename Reduce, typename Keep, typename FetchReal >
		void allRowsReduction( Fetch& fetch, Reduce& reduce, Keep& keep, const FetchReal& zero ) const;

		template< typename Function >
		void forRows( IndexType first, IndexType last, Function& function ) const;

		template< typename Function >
		void forRows( IndexType first, IndexType last, Function& function );

		template< typename Function >
		void forAllRows( Function& function ) const;

		template< typename Function >
		void forAllRows( Function& function );

		template< typename Vector >
		__cuda_callable__

src/TNL/Matrices/Dense.hpp

+124 −2

Original line number	Diff line number	Diff line
		@@ -94,6 +94,31 @@ setCompressedRowLengths( ConstCompressedRowLengthsVectorView rowLengths )
		this->setDimensions( rowLengths.getSize(), max( rowLengths ) );
		}

		template< typename Real,
		typename Device,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		template< typename Vector >
		void
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		getCompressedRowLengths( Vector& rowLengths ) const
		{
		rowLengths.setSize( this->getRows() );
		rowLengths = 0;
		auto rowLengths_view = rowLengths.getView();
		auto fetch = [] __cuda_callable__ ( IndexType row, IndexType column, const RealType& value ) -> IndexType {
		return ( value != 0.0 );
		};
		auto reduce = [] __cuda_callable__ ( IndexType& aux, const IndexType a ) {
		aux += a;
		};
		auto keep = [=] __cuda_callable__ ( const IndexType rowIdx, const IndexType value ) mutable {
		rowLengths_view[ rowIdx ] = value;
		};
		this->allRowsReduction( fetch, reduce, keep, 0 );
		}

		template< typename Real,
		typename Device,
		typename Index,
		@@ -256,12 +281,109 @@ template< typename Real,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		Real Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::getElement( const IndexType row,
		Real
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		getElement( const IndexType row,
		const IndexType column ) const
		{
		return this->values.getElement( this->getElementIndex( row, column ) );
		}

		template< typename Real,
		typename Device,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		template< typename Fetch, typename Reduce, typename Keep, typename FetchValue >
		void
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		rowsReduction( IndexType first, IndexType last, Fetch& fetch, Reduce& reduce, Keep& keep, const FetchValue& zero ) const
		{
		const auto values_view = this->values.getConstView();
		auto fetch_ = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx, bool& compute ) mutable -> decltype( fetch( IndexType(), IndexType(), RealType() ) ) {
		return fetch( rowIdx, columnIdx, values_view[ globalIdx ] );
		return zero;
		};
		this->segments.segmentsReduction( first, last, fetch_, reduce, keep, zero );
		}

		template< typename Real,
		typename Device,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		template< typename Fetch, typename Reduce, typename Keep, typename FetchReal >
		void
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		allRowsReduction( Fetch& fetch, Reduce& reduce, Keep& keep, const FetchReal& zero ) const
		{
		this->rowsReduction( 0, this->getRows(), fetch, reduce, keep, zero );
		}

		template< typename Real,
		typename Device,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		template< typename Function >
		void
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		forRows( IndexType first, IndexType last, Function& function ) const
		{
		const auto values_view = this->values.getConstView();
		auto f = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx ) mutable -> bool {
		function( rowIdx, columnIdx, values_view[ globalIdx ] );
		return true;
		};
		this->segments.forSegments( first, last, f );

		}

		template< typename Real,
		typename Device,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		template< typename Function >
		void
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		forRows( IndexType first, IndexType last, Function& function )
		{
		auto values_view = this->values.getView();
		auto f = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx ) mutable -> bool {
		function( rowIdx, columnIdx, values_view[ globalIdx ] );
		return true;
		};
		this->segments.forSegments( first, last, f );

		}

		template< typename Real,
		typename Device,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		template< typename Function >
		void
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		forAllRows( Function& function ) const
		{
		this->forRows( 0, this->getRows(), function );
		}

		template< typename Real,
		typename Device,
		typename Index,
		bool RowMajorOrder,
		typename RealAllocator >
		template< typename Function >
		void
		Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
		forAllRows( Function& function )
		{
		this->forRows( 0, this->getRows(), function );
		}

		template< typename Real,
		typename Device,
		typename Index,