Implementing CUDA support in matrix fromats. (bccd1586) · Commits · TNL / tnl-dev

src/implementation/core/arrays/tnlArray_impl.h

+4 −2

Original line number	Diff line number	Diff line
		@@ -418,10 +418,12 @@ ostream& operator << ( ostream& str, const tnlArray< Element, Device, Index >& v

		#ifdef TEMPLATE_EXPLICIT_INSTANTIATION

		extern template class tnlArray< float, tnlHost, int >;
		// TODO: this does not work with CUDA 5.5 - fix it later

		/*extern template class tnlArray< float, tnlHost, int >;
		extern template class tnlArray< double, tnlHost, int >;
		extern template class tnlArray< float, tnlHost, long int >;
		extern template class tnlArray< double, tnlHost, long int >;
		extern template class tnlArray< double, tnlHost, long int >;*/

		#ifdef HAVE_CUDA
		/*extern template class tnlArray< float, tnlCuda, int >;

src/implementation/matrices/tnlCSRMatrix_impl.h

+41 −21

Original line number	Diff line number	Diff line
		@@ -149,6 +149,16 @@ bool tnlCSRMatrix< Real, Device, Index >::operator != ( const tnlCSRMatrix< Real
		return ! ( ( *this ) == matrix );
		}

		template< typename Real,
		typename Device,
		typename Index >
		bool tnlCSRMatrix< Real, Device, Index >::setElementFast( const IndexType row,
		const IndexType column,
		const Real& value )
		{
		return this->addElementFast( row, column, value, 0.0 );
		}

		template< typename Real,
		typename Device,
		typename Index >
		@@ -156,13 +166,14 @@ bool tnlCSRMatrix< Real, Device, Index >::setElement( const IndexType row,
		const IndexType column,
		const Real& value )
		{
		return this->addElement( row, column, value, 0.0 );
		return this->addElementFast( row, column, value, 0.0 );
		}


		template< typename Real,
		typename Device,
		typename Index >
		bool tnlCSRMatrix< Real, Device, Index >::addElement( const IndexType row,
		bool tnlCSRMatrix< Real, Device, Index >::addElementFast( const IndexType row,
		const IndexType column,
		const RealType& value,
		const RealType& thisElementMultiplicator )
		@@ -210,7 +221,7 @@ bool tnlCSRMatrix< Real, Device, Index >::addElement( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		bool tnlCSRMatrix< Real, Device, Index > :: setRow( const IndexType row,
		bool tnlCSRMatrix< Real, Device, Index > :: setRowFast( const IndexType row,
		const IndexType* columnIndexes,
		const RealType* values,
		const IndexType elements )
		@@ -234,7 +245,7 @@ bool tnlCSRMatrix< Real, Device, Index > :: setRow( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		bool tnlCSRMatrix< Real, Device, Index > :: addRow( const IndexType row,
		bool tnlCSRMatrix< Real, Device, Index > :: addRowFast( const IndexType row,
		const IndexType* columns,
		const RealType* values,
		const IndexType numberOfElements,
		@@ -247,7 +258,7 @@ bool tnlCSRMatrix< Real, Device, Index > :: addRow( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		Real tnlCSRMatrix< Real, Device, Index >::getElement( const IndexType row,
		Real tnlCSRMatrix< Real, Device, Index >::getElementFast( const IndexType row,
		const IndexType column ) const
		{
		IndexType elementPtr = this->rowPointers[ row ];
		@@ -262,7 +273,16 @@ Real tnlCSRMatrix< Real, Device, Index >::getElement( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		void tnlCSRMatrix< Real, Device, Index >::getRow( const IndexType row,
		Real tnlCSRMatrix< Real, Device, Index >::getElement( const IndexType row,
		const IndexType column ) const
		{
		return this->getElementFast( row, column );
		}

		template< typename Real,
		typename Device,
		typename Index >
		void tnlCSRMatrix< Real, Device, Index >::getRowFast( const IndexType row,
		IndexType* columns,
		RealType* values ) const
		{

src/implementation/matrices/tnlChunkedEllpackMatrix_impl.h

+40 −20

Original line number	Diff line number	Diff line
		@@ -296,6 +296,16 @@ bool tnlChunkedEllpackMatrix< Real, Device, Index >::operator != ( const tnlChun
		return ! ( ( *this ) == matrix );
		}

		template< typename Real,
		typename Device,
		typename Index >
		bool tnlChunkedEllpackMatrix< Real, Device, Index >::setElementFast( const IndexType row,
		const IndexType column,
		const Real& value )
		{
		return this->addElement( row, column, value, 0.0 );
		}

		template< typename Real,
		typename Device,
		typename Index >
		@@ -306,10 +316,11 @@ bool tnlChunkedEllpackMatrix< Real, Device, Index >::setElement( const IndexType
		return this->addElement( row, column, value, 0.0 );
		}


		template< typename Real,
		typename Device,
		typename Index >
		bool tnlChunkedEllpackMatrix< Real, Device, Index >::addElement( const IndexType row,
		bool tnlChunkedEllpackMatrix< Real, Device, Index >::addElementFast( const IndexType row,
		const IndexType column,
		const RealType& value,
		const RealType& thisElementMultiplicator )
		@@ -366,7 +377,7 @@ bool tnlChunkedEllpackMatrix< Real, Device, Index >::addElement( const IndexType
		template< typename Real,
		typename Device,
		typename Index >
		bool tnlChunkedEllpackMatrix< Real, Device, Index > :: setRow( const IndexType row,
		bool tnlChunkedEllpackMatrix< Real, Device, Index > :: setRowFast( const IndexType row,
		const IndexType* columnIndexes,
		const RealType* values,
		const IndexType elements )
		@@ -394,7 +405,7 @@ bool tnlChunkedEllpackMatrix< Real, Device, Index > :: setRow( const IndexType r
		template< typename Real,
		typename Device,
		typename Index >
		bool tnlChunkedEllpackMatrix< Real, Device, Index > :: addRow( const IndexType row,
		bool tnlChunkedEllpackMatrix< Real, Device, Index > :: addRowFast( const IndexType row,
		const IndexType* columns,
		const RealType* values,
		const IndexType numberOfElements,
		@@ -407,7 +418,7 @@ bool tnlChunkedEllpackMatrix< Real, Device, Index > :: addRow( const IndexType r
		template< typename Real,
		typename Device,
		typename Index >
		Real tnlChunkedEllpackMatrix< Real, Device, Index >::getElement( const IndexType row,
		Real tnlChunkedEllpackMatrix< Real, Device, Index >::getElementFast( const IndexType row,
		const IndexType column ) const
		{
		const IndexType& sliceIndex = slicesToRowsMapping[ row ];
		@@ -427,7 +438,16 @@ Real tnlChunkedEllpackMatrix< Real, Device, Index >::getElement( const IndexType
		template< typename Real,
		typename Device,
		typename Index >
		void tnlChunkedEllpackMatrix< Real, Device, Index >::getRow( const IndexType row,
		Real tnlChunkedEllpackMatrix< Real, Device, Index >::getElement( const IndexType row,
		const IndexType column ) const
		{
		return this->getElementFast( row, column );
		}

		template< typename Real,
		typename Device,
		typename Index >
		void tnlChunkedEllpackMatrix< Real, Device, Index >::getRowFast( const IndexType row,
		IndexType* columns,
		RealType* values ) const
		{

src/implementation/matrices/tnlDenseMatrix_impl.h

+98 −7

Original line number	Diff line number	Diff line
		@@ -117,6 +117,17 @@ void tnlDenseMatrix< Real, Device, Index >::reset()
		this->values.reset();
		}

		template< typename Real,
		typename Device,
		typename Index >
		bool tnlDenseMatrix< Real, Device, Index >::setElementFast( const IndexType row,
		const IndexType column,
		const RealType& value )
		{
		this->values.operator[]( this->getElementIndex( row, column ) ) = value;
		return true;
		}

		template< typename Real,
		typename Device,
		typename Index >
		@@ -128,6 +139,23 @@ bool tnlDenseMatrix< Real, Device, Index >::setElement( const IndexType row,
		return true;
		}


		template< typename Real,
		typename Device,
		typename Index >
		bool tnlDenseMatrix< Real, Device, Index >::addElementFast( const IndexType row,
		const IndexType column,
		const RealType& value,
		const RealType& thisElementMultiplicator )
		{
		const IndexType elementIndex = this->getElementIndex( row, column );
		if( thisElementMultiplicator == 1.0 )
		this->values.operator[]( elementIndex ) += value;
		else
		this->values.operator[]( elementIndex ) =
		thisElementMultiplicator * this->values.operator[]( elementIndex ) + value;
		}

		template< typename Real,
		typename Device,
		typename Index >
		@@ -138,10 +166,29 @@ bool tnlDenseMatrix< Real, Device, Index >::addElement( const IndexType row,
		{
		const IndexType elementIndex = this->getElementIndex( row, column );
		if( thisElementMultiplicator == 1.0 )
		values->operator[]( elementIndex ) += value;
		this->values.setElement( elementIndex,
		this->values.getElement( elementIndex ) + value );
		else
		values->operator[]( elementIndex ) =
		thisElementMultiplicator * values->operator[]( elementIndex ) + value;
		this->values.setElement( elementIndex,
		thisElementMultiplicator * this->values.getElement( elementIndex ) + value );
		}


		template< typename Real,
		typename Device,
		typename Index >
		bool tnlDenseMatrix< Real, Device, Index >::setRowFast( const IndexType row,
		const IndexType* columns,
		const RealType* values,
		const IndexType elements )
		{
		tnlAssert( elements <= this->getColumns(),
		cerr << " elements = " << elements
		<< " this->columns = " << this->getColumns()
		<< " this->getName() = " << this->getName() );
		for( IndexType i = 0; i < elements; i++ )
		this->setElementFast( row, columns[ i ], values[ i ] );
		return true;
		}

		template< typename Real,
		@@ -152,12 +199,31 @@ bool tnlDenseMatrix< Real, Device, Index >::setRow( const IndexType row,
		const RealType* values,
		const IndexType elements )
		{
		tnlAssert( elements <= this->getDimensions().y(),
		tnlAssert( elements <= this->getColumns(),
		cerr << " elements = " << elements
		<< " this->columns = " << this->getDimensions().y()
		<< " this->columns = " << this->getColumns()
		<< " this->getName() = " << this->getName() );
		for( IndexType i = 0; i < elements; i++ )
		this->operator()( row, columns[ i ] ) = values[ i ];
		this->setElement( row, columns[ i ], values[ i ] );
		return true;
		}

		template< typename Real,
		typename Device,
		typename Index >
		bool tnlDenseMatrix< Real, Device, Index >::addRowFast( const IndexType row,
		const IndexType* columns,
		const RealType* values,
		const IndexType elements,
		const RealType& thisRowMultiplicator )
		{
		tnlAssert( elements <= this->columns,
		cerr << " elements = " << elements
		<< " this->columns = " << this->columns
		<< " this->getName() = " << this->getName() );
		for( IndexType i = 0; i < elements; i++ )
		this->setElementFast( row, columns[ i ],
		thisRowMultiplicator * this->getElementFast( row, columns[ i ] ) + values[ i ] );
		return true;
		}

		@@ -180,6 +246,16 @@ bool tnlDenseMatrix< Real, Device, Index >::addRow( const IndexType row,
		return true;
		}


		template< typename Real,
		typename Device,
		typename Index >
		Real tnlDenseMatrix< Real, Device, Index >::getElementFast( const IndexType row,
		const IndexType column ) const
		{
		return this->values.operator[]( this->getElementIndex( row, column ) );
		}

		template< typename Real,
		typename Device,
		typename Index >
		@@ -189,6 +265,20 @@ Real tnlDenseMatrix< Real, Device, Index >::getElement( const IndexType row,
		return this->values.getElement( this->getElementIndex( row, column ) );
		}

		template< typename Real,
		typename Device,
		typename Index >
		void tnlDenseMatrix< Real, Device, Index >::getRowFast( const IndexType row,
		IndexType* columns,
		RealType* values ) const
		{
		for( IndexType i = 0; i < this->getColumns(); i++ )
		{
		columns[ i ] = i;
		values[ i ] = this->getElementFast( row, i );
		}
		}

		template< typename Real,
		typename Device,
		typename Index >
		@@ -203,6 +293,7 @@ void tnlDenseMatrix< Real, Device, Index >::getRow( const IndexType row,
		}
		}


		template< typename Real,
		typename Device,
		typename Index >

src/implementation/matrices/tnlEllpackMatrix_impl.h

+39 −18

Original line number	Diff line number	Diff line
		@@ -155,6 +155,16 @@ bool tnlEllpackMatrix< Real, Device, Index >::operator != ( const tnlEllpackMatr
		return ! ( ( *this ) == matrix );
		}

		template< typename Real,
		typename Device,
		typename Index >
		bool tnlEllpackMatrix< Real, Device, Index > :: setElementFast( const IndexType row,
		const IndexType column,
		const Real& value )
		{
		return this->addElement( row, column, value, 0.0 );
		}

		template< typename Real,
		typename Device,
		typename Index >
		@@ -165,10 +175,11 @@ bool tnlEllpackMatrix< Real, Device, Index > :: setElement( const IndexType row,
		return this->addElement( row, column, value, 0.0 );
		}


		template< typename Real,
		typename Device,
		typename Index >
		bool tnlEllpackMatrix< Real, Device, Index > :: addElement( const IndexType row,
		bool tnlEllpackMatrix< Real, Device, Index > :: addElementFast( const IndexType row,
		const IndexType column,
		const RealType& value,
		const RealType& thisElementMultiplicator )
		@@ -215,7 +226,7 @@ bool tnlEllpackMatrix< Real, Device, Index > :: addElement( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		bool tnlEllpackMatrix< Real, Device, Index > :: setRow( const IndexType row,
		bool tnlEllpackMatrix< Real, Device, Index > :: setRowFast( const IndexType row,
		const IndexType* columnIndexes,
		const RealType* values,
		const IndexType elements )
		@@ -237,7 +248,7 @@ bool tnlEllpackMatrix< Real, Device, Index > :: setRow( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		bool tnlEllpackMatrix< Real, Device, Index > :: addRow( const IndexType row,
		bool tnlEllpackMatrix< Real, Device, Index > :: addRowFast( const IndexType row,
		const IndexType* columns,
		const RealType* values,
		const IndexType numberOfElements,
		@@ -250,7 +261,7 @@ bool tnlEllpackMatrix< Real, Device, Index > :: addRow( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		Real tnlEllpackMatrix< Real, Device, Index >::getElement( const IndexType row,
		Real tnlEllpackMatrix< Real, Device, Index >::getElementFast( const IndexType row,
		const IndexType column ) const
		{
		IndexType elementPtr( row * this->rowLengths );
		@@ -264,7 +275,17 @@ Real tnlEllpackMatrix< Real, Device, Index >::getElement( const IndexType row,
		template< typename Real,
		typename Device,
		typename Index >
		void tnlEllpackMatrix< Real, Device, Index >::getRow( const IndexType row,
		Real tnlEllpackMatrix< Real, Device, Index >::getElement( const IndexType row,
		const IndexType column ) const
		{
		return this->getElementFast( row, column );
		}


		template< typename Real,
		typename Device,
		typename Index >
		void tnlEllpackMatrix< Real, Device, Index >::getRowFast( const IndexType row,
		IndexType* columns,
		RealType* values ) const
		{