Added forSegments to all segments and forRows (just in SparseMatrix for now). (b32d83f0) · Commits · TNL / tnl-dev

src/TNL/Algorithms/Segments/BiEllpack.h

+109 −98

Original line number	Diff line number	Diff line
		@@ -15,9 +15,12 @@
		#include <TNL/Algorithms/Segments/BiEllpackView.h>
		#include <TNL/Algorithms/Segments/SegmentView.h>

		namespace TNL {
		namespace Algorithms {
		namespace Segments {
		namespace TNL
		{
		namespace Algorithms
		{
		namespace Segments
		{

		template <typename Device,
		typename Index,
		@@ -27,16 +30,15 @@ template< typename Device,
		class BiEllpack
		{
		public:

		using DeviceType = Device;
		using IndexType = std::remove_const_t<Index>;
		using OffsetsHolder = Containers::Vector< Index, DeviceType, IndexType, IndexAllocator >;
		using OffsetsHolder = Containers::Vector< IndexType, DeviceType, IndexType, IndexAllocator>;
		static constexpr ElementsOrganization getOrganization() { return Organization; }
		using ViewType = BiEllpackView< Device, Index, Organization >;
		using ViewType = BiEllpackView< Device, Index, Organization, WarpSize >;
		template <typename Device_, typename Index_>
		using ViewTemplate = BiEllpackView< Device_, Index_, Organization >;
		using ConstViewType = BiEllpackView< Device, std::add_const_t< IndexType >, Organization >;
		using SegmentViewType = BiEllpackSegmentView< IndexType, Organization >;
		using ViewTemplate = BiEllpackView<Device_, Index_, Organization, WarpSize >;
		using ConstViewType = typename ViewType::ConstViewType;
		using SegmentViewType = typename ViewType::SegmentViewType;

		static constexpr bool havePadding() { return true; };

		@@ -60,7 +62,8 @@ class BiEllpack
		* \brief Number of segments.
		*/
		__cuda_callable__
		IndexType getSegmentsCount() const;
		IndexType
		getSegmentsCount() const;

		/**
		* \brief Set sizes of particular segments.
		@@ -76,16 +79,20 @@ class BiEllpack
		* \brief Number segments.
		*/
		__cuda_callable__
		IndexType getSize() const;
		IndexType
		getSize() const;

		__cuda_callable__
		IndexType getStorageSize() const;
		IndexType
		getStorageSize() const;

		__cuda_callable__
		IndexType getGlobalIndex( const IndexType segmentIdx, const IndexType localIdx ) const;
		IndexType
		getGlobalIndex(const IndexType segmentIdx, const IndexType localIdx) const;

		__cuda_callable__
		SegmentViewType getSegmentView( const IndexType segmentIdx ) const;
		SegmentViewType
		getSegmentView(const IndexType segmentIdx) const;

		/***
		* \brief Go over all segments and for each segment element call
		@@ -93,12 +100,17 @@ class BiEllpack
		* When its true, the for-loop continues. Once 'f' returns false, the for-loop
		* is terminated.
		*/
		template< typename Function, typename... Args >
		void forElements( IndexType first, IndexType last, Function& f, Args... args ) const;
		template <typename Function>
		void forElements(IndexType first, IndexType last, Function &&f) const;

		template <typename Function>
		void forEachElement(Function &&f) const;

		template< typename Function, typename... Args >
		void forEachElement( Function& f, Args... args ) const;
		template <typename Function>
		void forSegments(IndexType begin, IndexType end, Function &&f) const;

		template <typename Function>
		void forEachSegment(Function &&f) const;

		/***
		* \brief Go over all segments and perform a reduction in each of them.
		@@ -129,7 +141,6 @@ class BiEllpack
		void computeColumnSizes(const SizesHolder &segmentsSizes);

		protected:

		static constexpr int getWarpSize() { return WarpSize; };

		static constexpr int getLogWarpSize() { return std::log2(WarpSize); };

src/TNL/Algorithms/Segments/BiEllpack.hpp

+33 −6

Original line number	Diff line number	Diff line
		@@ -443,12 +443,12 @@ template< typename Device,
		typename IndexAllocator,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function, typename... Args >
		template< typename Function >
		void
		BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::
		forElements( IndexType first, IndexType last, Function& f, Args... args ) const
		forElements( IndexType first, IndexType last, Function&& f ) const
		{
		this->getConstView().forElements( first, last, f, args... );
		this->getConstView().forElements( first, last, f );
		}

		template< typename Device,
		@@ -456,14 +456,41 @@ template< typename Device,
		typename IndexAllocator,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function, typename... Args >
		template< typename Function >
		void
		BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::
		forEachElement( Function& f, Args... args ) const
		forEachElement( Function&& f ) const
		{
		this->forElements( 0, this->getSegmentsCount(), f, args... );
		this->forElements( 0, this->getSegmentsCount(), f );
		}

		template< typename Device,
		typename Index,
		typename IndexAllocator,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function >
		void
		BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::
		forSegments( IndexType begin, IndexType end, Function&& f ) const
		{
		this->getConstView().forSegments( begin, end, f );
		}

		template< typename Device,
		typename Index,
		typename IndexAllocator,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function >
		void
		BiEllpack< Device, Index, IndexAllocator, Organization, WarpSize >::
		forEachSegment( Function&& f ) const
		{
		this->getConstView().forEachSegment( f );
		}


		template< typename Device,
		typename Index,
		typename IndexAllocator,

src/TNL/Algorithms/Segments/BiEllpackSegmentView.h

+12 −5

Original line number	Diff line number	Diff line
		@@ -44,10 +44,11 @@ class BiEllpackSegmentView
		* \param groupsWidth is a static vector containing widths of the strip groups
		*/
		__cuda_callable__
		BiEllpackSegmentView( const IndexType offset,
		BiEllpackSegmentView( const IndexType segmentIdx,
		const IndexType offset,
		const IndexType inStripIdx,
		const GroupsWidthType& groupsWidth )
		: groupOffset( offset ), inStripIdx( inStripIdx ), segmentSize( TNL::sum( groupsWidth ) ), groupsWidth( groupsWidth ){};
		: segmentIdx( segmentIdx ), groupOffset( offset ), inStripIdx( inStripIdx ), segmentSize( TNL::sum( groupsWidth ) ), groupsWidth( groupsWidth ){};

		__cuda_callable__
		IndexType getSize() const
		@@ -79,9 +80,15 @@ class BiEllpackSegmentView
		return offset + inStripIdx + localIdx * groupHeight;
		};

		__cuda_callable__
		const IndexType& getSegmentIndex() const
		{
		return this->segmentIdx;
		};

		protected:

		IndexType groupOffset, inStripIdx, segmentSize;
		IndexType segmentIdx, groupOffset, inStripIdx, segmentSize;

		GroupsWidthType groupsWidth;
		};

src/TNL/Algorithms/Segments/BiEllpackView.h

+13 −8

Original line number	Diff line number	Diff line
		@@ -32,13 +32,13 @@ class BiEllpackView

		using DeviceType = Device;
		using IndexType = std::remove_const_t< Index >;
		using OffsetsView = typename Containers::VectorView< Index, DeviceType, IndexType >;
		using OffsetsView = typename Containers::VectorView< IndexType, DeviceType, IndexType >;
		using ConstOffsetsView = typename OffsetsView::ConstViewType;
		using ViewType = BiEllpackView;
		template< typename Device_, typename Index_ >
		using ViewTemplate = BiEllpackView< Device_, Index_ >;
		using ConstViewType = BiEllpackView< Device, std::add_const_t< Index > >;
		using SegmentViewType = BiEllpackSegmentView< IndexType, Organization >;
		using ViewTemplate = BiEllpackView< Device_, Index_, Organization, WarpSize >;
		using ConstViewType = BiEllpackView< Device, std::add_const_t< Index >, Organization, WarpSize >;
		using SegmentViewType = BiEllpackSegmentView< IndexType, Organization, WarpSize >;

		static constexpr bool havePadding() { return true; };

		@@ -111,12 +111,17 @@ class BiEllpackView
		* When its true, the for-loop continues. Once 'f' returns false, the for-loop
		* is terminated.
		*/
		template< typename Function, typename... Args >
		void forElements( IndexType first, IndexType last, Function& f, Args... args ) const;
		template< typename Function >
		void forElements( IndexType first, IndexType last, Function&& f ) const;

		template< typename Function, typename... Args >
		void forEachElement( Function& f, Args... args ) const;
		template< typename Function >
		void forEachElement( Function&& f ) const;

		template< typename Function >
		void forSegments( IndexType begin, IndexType end, Function&& f ) const;

		template< typename Function >
		void forEachSegment( Function&& f ) const;

		/***
		* \brief Go over all segments and perform a reduction in each of them.

src/TNL/Algorithms/Segments/BiEllpackView.hpp

+42 −8

Original line number	Diff line number	Diff line
		@@ -129,7 +129,12 @@ template< typename Device,
		__cuda_callable__ auto BiEllpackView< Device, Index, Organization, WarpSize >::
		getConstView() const -> const ConstViewType
		{
		return ConstViewType( size, storageSize, virtualRows, rowPermArray.getConstView(), groupPointers.getConstView() );
		BiEllpackView* this_ptr = const_cast< BiEllpackView* >( this );
		return ConstViewType( size,
		storageSize,
		virtualRows,
		this_ptr->rowPermArray.getView(),
		this_ptr->groupPointers.getView() );
		}

		template< typename Device,
		@@ -255,14 +260,14 @@ template< typename Device,
		typename Index,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function, typename... Args >
		template< typename Function >
		void
		BiEllpackView< Device, Index, Organization, WarpSize >::
		forElements( IndexType first, IndexType last, Function& f, Args... args ) const
		forElements( IndexType first, IndexType last, Function&& f ) const
		{
		const auto segmentsPermutationView = this->rowPermArray.getConstView();
		const auto groupPointersView = this->groupPointers.getConstView();
		auto work = [=] __cuda_callable__ ( IndexType segmentIdx, Args... args ) mutable {
		auto work = [=] __cuda_callable__ ( IndexType segmentIdx ) mutable {
		const IndexType strip = segmentIdx / getWarpSize();
		const IndexType firstGroupInStrip = strip * ( getLogWarpSize() + 1 );
		const IndexType rowStripPerm = segmentsPermutationView[ segmentIdx ] - strip * getWarpSize();
		@@ -298,19 +303,48 @@ forElements( IndexType first, IndexType last, Function& f, Args... args ) const
		groupHeight /= 2;
		}
		};
		Algorithms::ParallelFor< DeviceType >::exec( first, last , work, args... );
		Algorithms::ParallelFor< DeviceType >::exec( first, last , work );
		}

		template< typename Device,
		typename Index,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function, typename... Args >
		template< typename Function >
		void
		BiEllpackView< Device, Index, Organization, WarpSize >::
		forEachElement( Function& f, Args... args ) const
		forEachElement( Function&& f ) const
		{
		this->forElements( 0, this->getSegmentsCount(), f, args... );
		this->forElements( 0, this->getSegmentsCount(), f );
		}

		template< typename Device,
		typename Index,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function >
		void
		BiEllpackView< Device, Index, Organization, WarpSize >::
		forSegments( IndexType begin, IndexType end, Function&& function ) const
		{
		auto view = this->getConstView();
		auto f = [=] __cuda_callable__ ( IndexType segmentIdx ) mutable {
		auto segment = view.getSegmentView( segmentIdx );
		function( segment );
		};
		TNL::Algorithms::ParallelFor< DeviceType >::exec( begin, end, f );
		}

		template< typename Device,
		typename Index,
		ElementsOrganization Organization,
		int WarpSize >
		template< typename Function >
		void
		BiEllpackView< Device, Index, Organization, WarpSize >::
		forEachSegment( Function&& f ) const
		{
		this->forSegments( 0, this->getSegmentsCount(), f );
		}

		template< typename Device,