Commit 7fd8fa6b authored by Tomáš Oberhuber's avatar Tomáš Oberhuber
Browse files

Refactoring of ChunkedEllpack segments.

parent 3ce63c9f

src/TNL/Containers/Segments/ChunkedEllpack.h

+1 −6
Original line number Diff line number Diff line
@@ -36,7 +36,7 @@ class ChunkedEllpack 
      using ViewTemplate = ChunkedEllpackView< Device_, Index_, RowMajorOrder >;

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

      using SegmentViewType = SegmentView< IndexType, RowMajorOrder >;

      using ChunkedEllpackSliceInfoType = ChunkedEllpackSliceInfo< IndexType >;

      using ChunkedEllpackSliceInfoType = details::ChunkedEllpackSliceInfo< IndexType >;

      //TODO: using ChunkedEllpackSliceInfoAllocator = typename IndexAllocatorType::retype< ChunkedEllpackSliceInfoType >;

      using ChunkedEllpackSliceInfoAllocator = typename Allocators::Default< Device >::template Allocator< ChunkedEllpackSliceInfoType >;

      using ChunkedEllpackSliceInfoContainer = Containers::Array< ChunkedEllpackSliceInfoType, DeviceType, IndexType, ChunkedEllpackSliceInfoAllocator >;
@@ -66,7 +66,6 @@ class ChunkedEllpack 
      __cuda_callable__

      IndexType getSegmentsCount() const;



      __cuda_callable__

      IndexType getSegmentSize( const IndexType segmentIdx ) const;



      /**
@@ -75,16 +74,12 @@ class ChunkedEllpack 
      __cuda_callable__

      IndexType getSize() const;





      __cuda_callable__

      IndexType getStorageSize() const;



      __cuda_callable__

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



      __cuda_callable__

      void getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const;



      __cuda_callable__

      SegmentViewType getSegmentView( const IndexType segmentIdx ) const;

src/TNL/Containers/Segments/ChunkedEllpack.hpp

+14 −121
Original line number Diff line number Diff line
@@ -289,21 +289,15 @@ template< typename Device, 
          typename Index,

          typename IndexAllocator,

          bool RowMajorOrder >

__cuda_callable__

Index

ChunkedEllpack< Device, Index, IndexAllocator, RowMajorOrder >::

getSegmentSize( const IndexType segmentIdx ) const

{

   const IndexType& sliceIndex = rowToSliceMapping[ segmentIdx ];

   TNL_ASSERT_LE( sliceIndex, this->getSegmentsCount(), "" );

   IndexType firstChunkOfSegment( 0 );

   if( segmentIdx != slices[ sliceIndex ].firstRow )

      firstChunkOfSegment = rowToChunkMapping[ segmentIdx - 1 ];



   const IndexType lastChunkOfSegment = rowToChunkMapping[ segmentIdx ];

   const IndexType segmentChunksCount = lastChunkOfSegment - firstChunkOfSegment;

   const IndexType chunkSize = slices[ sliceIndex ].chunkSize;

   return chunkSize * segmentChunksCount;

   return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getSegmentSize(

      rowToSliceMapping.getView(),

      slices.getView(),

      rowToChunkMapping.getView(),

      segmentIdx );

}



template< typename Device,
@@ -339,37 +333,13 @@ Index 
ChunkedEllpack< Device, Index, IndexAllocator, RowMajorOrder >::

getGlobalIndex( const Index segmentIdx, const Index localIdx ) const

{

   const IndexType& sliceIndex = rowToSliceMapping[ segmentIdx ];

   TNL_ASSERT_LE( sliceIndex, this->rows, "" );

   IndexType firstChunkOfSegment( 0 );

   if( segmentIdx != slices[ sliceIndex ].firstRow )

      firstChunkOfSegment = rowToChunkMapping[ segmentIdx - 1 ];

   

   const IndexType lastChunkOfSegment = rowToChunkMapping[ segmentIdx ];

   const IndexType segmentChunksCount = lastChunkOfSegment - firstChunkOfSegment;

   const IndexType sliceOffset = slices[ sliceIndex ].pointer;

   const IndexType chunkSize = slices[ sliceIndex ].chunkSize;

   TNL_ASSERT_LE( localIdx, segmentChunksCount * chunkSize, "" );



   if( RowMajorOrder )

      return sliceOffset + firstChunkOfSegment * chunkSize + localIdx;

   else

   {

      const IndexType inChunkOffset = localIdx % chunkSize;

      const IndexType chunkIdx = localIdx / chunkSize;

      return sliceOffset + inChunkOffset * segmentChunksCount + chunkIdx;

   }

}



template< typename Device,

          typename Index,

          typename IndexAllocator,

          bool RowMajorOrder >

__cuda_callable__

void

ChunkedEllpack< Device, Index, IndexAllocator, RowMajorOrder >::

getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const

{

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getGlobalIndex(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         chunksInSlice,

         segmentIdx,

         localIdx );

}



template< typename Device,
@@ -381,16 +351,6 @@ auto 
ChunkedEllpack< Device, Index, IndexAllocator, RowMajorOrder >::

getSegmentView( const IndexType segmentIdx ) const -> SegmentViewType

{

   /*const IndexType sliceIdx = segmentIdx / SliceSize;

   const IndexType segmentInSliceIdx = segmentIdx % SliceSize;

   const IndexType& sliceOffset = this->sliceOffsets[ sliceIdx ];

   const IndexType& segmentSize = this->sliceSegmentSizes[ sliceIdx ];



   if( RowMajorOrder )

      return SegmentViewType( sliceOffset + segmentInSliceIdx * segmentSize, segmentSize, 1 );

   else

      return SegmentViewType( sliceOffset + segmentInSliceIdx, segmentSize, SliceSize );

      */

}



template< typename Device,
@@ -402,38 +362,7 @@ void 
ChunkedEllpack< Device, Index, IndexAllocator, RowMajorOrder >::

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

{

/*   const auto sliceSegmentSizes_view = this->sliceSegmentSizes.getConstView();

   const auto sliceOffsets_view = this->sliceOffsets.getConstView();

   if( RowMajorOrder )

   {

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

         const IndexType sliceIdx = segmentIdx / SliceSize;

         const IndexType segmentInSliceIdx = segmentIdx % SliceSize;

         const IndexType segmentSize = sliceSegmentSizes_view[ sliceIdx ];

         const IndexType begin = sliceOffsets_view[ sliceIdx ] + segmentInSliceIdx * segmentSize;

         const IndexType end = begin + segmentSize;

         IndexType localIdx( 0 );

         for( IndexType globalIdx = begin; globalIdx < end; globalIdx++  )

            if( ! f( segmentIdx, localIdx++, globalIdx, args... ) )

               break;

      };

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

   }

   else

   {

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

         const IndexType sliceIdx = segmentIdx / SliceSize;

         const IndexType segmentInSliceIdx = segmentIdx % SliceSize;

         const IndexType segmentSize = sliceSegmentSizes_view[ sliceIdx ];

         const IndexType begin = sliceOffsets_view[ sliceIdx ] + segmentInSliceIdx;

         const IndexType end = sliceOffsets_view[ sliceIdx + 1 ];

         IndexType localIdx( 0 );

         for( IndexType globalIdx = begin; globalIdx < end; globalIdx += SliceSize )

            if( ! f( segmentIdx, localIdx++, globalIdx, args... ) )

               break;

      };

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

   }*/

   this->getView().forSegments( first, last, f, args... );

}



template< typename Device,
@@ -457,43 +386,7 @@ void 
ChunkedEllpack< Device, Index, IndexAllocator, RowMajorOrder >::

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

{

/*   using RealType = decltype( fetch( IndexType(), IndexType(), IndexType(), std::declval< bool& >(), args... ) );

   const auto sliceSegmentSizes_view = this->sliceSegmentSizes.getConstView();

   const auto sliceOffsets_view = this->sliceOffsets.getConstView();

   if( RowMajorOrder )

   {

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

         const IndexType sliceIdx = segmentIdx / SliceSize;

         const IndexType segmentInSliceIdx = segmentIdx % SliceSize;

         const IndexType segmentSize = sliceSegmentSizes_view[ sliceIdx ];

         const IndexType begin = sliceOffsets_view[ sliceIdx ] + segmentInSliceIdx * segmentSize;

         const IndexType end = begin + segmentSize;

         RealType aux( zero );

         bool compute( true );

         IndexType localIdx( 0 );

         for( IndexType globalIdx = begin; globalIdx< end; globalIdx++  )

            reduction( aux, fetch( segmentIdx, localIdx++, globalIdx, compute, args... ) );

         keeper( segmentIdx, aux );

      };

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

   }

   else

   {

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

         const IndexType sliceIdx = segmentIdx / SliceSize;

         const IndexType segmentInSliceIdx = segmentIdx % SliceSize;

         const IndexType segmentSize = sliceSegmentSizes_view[ sliceIdx ];

         const IndexType begin = sliceOffsets_view[ sliceIdx ] + segmentInSliceIdx;

         const IndexType end = sliceOffsets_view[ sliceIdx + 1 ];

         RealType aux( zero );

         bool compute( true );

         IndexType localIdx( 0 );

         for( IndexType globalIdx = begin; globalIdx < end; globalIdx += SliceSize  )

            reduction( aux, fetch( segmentIdx, localIdx++, globalIdx, compute, args... ) );

         keeper( segmentIdx, aux );

      };

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

   }*/

   this->getView().segmentsReduction( first, last, fetch, reduction, keeper, zero, args... );

}



template< typename Device,

src/TNL/Containers/Segments/ChunkedEllpackView.h

+2 −35
Original line number Diff line number Diff line
@@ -14,42 +14,12 @@ 


#include <TNL/Containers/Vector.h>

#include <TNL/Containers/Segments/ChunkedEllpackSegmentView.h>

#include <TNL/Containers/Segments/details/ChunkedEllpack.h>



namespace TNL {

   namespace Containers {

      namespace Segments {



/***

 * In the ChunkedEllpack, the segments are split into slices. This is done

 * in ChunkedEllpack::resolveSliceSizes. All segments elements in each slice

 * are split into chunks. All chunks in one slice have the same size, but the size

 * of chunks can be different in each slice.

 */

template< typename Index >

struct ChunkedEllpackSliceInfo

{

   /**

    * The size of the slice, it means the number of the matrix rows covered by

    * the slice.

    */

   Index size;



   /**

    * The chunk size, i.e. maximal number of non-zero elements that can be stored

    * in the chunk.

    */

   Index chunkSize;



   /**

    * Index of the first segment covered be this slice.

    */

   Index firstSegment;



   /**

    * Position of the first element of this slice.

    */

   Index pointer;

};



template< typename Device,

          typename Index,
@@ -67,7 +37,7 @@ class ChunkedEllpackView 
      using ViewTemplate = ChunkedEllpackView< Device_, Index_ >;

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

      using SegmentViewType = ChunkedEllpackSegmentView< IndexType >;

      using ChunkedEllpackSliceInfoType = ChunkedEllpackSliceInfo< IndexType >;

      using ChunkedEllpackSliceInfoType = details::ChunkedEllpackSliceInfo< IndexType >;

      using ChunkedEllpackSliceInfoAllocator = typename Allocators::Default< Device >::template Allocator< ChunkedEllpackSliceInfoType >;

      using ChunkedEllpackSliceInfoContainer = Containers::Array< ChunkedEllpackSliceInfoType, DeviceType, IndexType, ChunkedEllpackSliceInfoAllocator >;

      using ChunkedEllpackSliceInfoContainerView = typename ChunkedEllpackSliceInfoContainer::ViewType;
@@ -140,9 +110,6 @@ class ChunkedEllpackView 
      __cuda_callable__

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



      __cuda_callable__

      void getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const;



      __cuda_callable__

      SegmentViewType getSegmentView( const IndexType segmentIdx ) const;

src/TNL/Containers/Segments/ChunkedEllpackView.hpp

+73 −64
Original line number Diff line number Diff line
@@ -179,16 +179,28 @@ Index 
ChunkedEllpackView< Device, Index, RowMajorOrder >::

getSegmentSize( const IndexType segmentIdx ) const

{

   const IndexType& sliceIndex = rowToSliceMapping[ segmentIdx ];

   TNL_ASSERT_LE( sliceIndex, this->getSegmentsCount(), "" );

   IndexType firstChunkOfSegment( 0 );

   if( segmentIdx != slices[ sliceIndex ].firstSegment )

      firstChunkOfSegment = rowToChunkMapping[ segmentIdx - 1 ];



   const IndexType lastChunkOfSegment = rowToChunkMapping[ segmentIdx ];

   const IndexType segmentChunksCount = lastChunkOfSegment - firstChunkOfSegment;

   const IndexType chunkSize = slices[ sliceIndex ].chunkSize;

   return chunkSize * segmentChunksCount;

   if( std::is_same< DeviceType, Devices::Host >::value )

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getSegmentSizeDirect(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         segmentIdx );

   if( std::is_same< DeviceType, Devices::Cuda >::value )

   {

#ifdef __CUDA_ARCH__

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getSegmentSizeDirect(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         segmentIdx );

#else

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getSegmentSize(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         segmentIdx );

#endif

   }

}



template< typename Device,
@@ -221,38 +233,36 @@ Index 
ChunkedEllpackView< Device, Index, RowMajorOrder >::

getGlobalIndex( const Index segmentIdx, const Index localIdx ) const

{

   const IndexType& sliceIndex = rowToSliceMapping[ segmentIdx ];

   TNL_ASSERT_LE( sliceIndex, this->size, "" );

   IndexType firstChunkOfSegment( 0 );

   if( segmentIdx != slices[ sliceIndex ].firstSegment )

      firstChunkOfSegment = rowToChunkMapping[ segmentIdx - 1 ];

   

   const IndexType lastChunkOfSegment = rowToChunkMapping[ segmentIdx ];

   const IndexType segmentChunksCount = lastChunkOfSegment - firstChunkOfSegment;

   const IndexType sliceOffset = slices[ sliceIndex ].pointer;

   const IndexType chunkSize = slices[ sliceIndex ].chunkSize;

   TNL_ASSERT_LE( localIdx, segmentChunksCount * chunkSize, "" );



   if( RowMajorOrder )

      return sliceOffset + firstChunkOfSegment * chunkSize + localIdx;

   else

   {

      const IndexType inChunkOffset = localIdx % chunkSize;

      const IndexType chunkIdx = localIdx / chunkSize;

      return sliceOffset + inChunkOffset * chunksInSlice + firstChunkOfSegment + chunkIdx;

   if( std::is_same< DeviceType, Devices::Host >::value )

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getGlobalIndexDirect(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         chunksInSlice,

         segmentIdx,

         localIdx );

   if( std::is_same< DeviceType, Devices::Cuda >::value )

   {

#ifdef __CUDA_ARCH__

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getGlobalIndexDirect(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         chunksInSlice,

         segmentIdx,

         localIdx );

#else

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getGlobalIndex(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         chunksInSlice,

         segmentIdx,

         localIdx );

#endif

   }

}



template< typename Device,

          typename Index,

          bool RowMajorOrder >

__cuda_callable__

void

ChunkedEllpackView< Device, Index, RowMajorOrder >::

getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const

{

}



template< typename Device,

          typename Index,

          bool RowMajorOrder >
@@ -261,32 +271,31 @@ auto 
ChunkedEllpackView< Device, Index, RowMajorOrder >::

getSegmentView( const IndexType segmentIdx ) const -> SegmentViewType

{

   const IndexType& sliceIndex = rowToSliceMapping[ segmentIdx ];

   TNL_ASSERT_LE( sliceIndex, this->size, "" );

   IndexType firstChunkOfSegment( 0 );

   if( segmentIdx != slices[ sliceIndex ].firstSegment )

      firstChunkOfSegment = rowToChunkMapping[ segmentIdx - 1 ];



   const IndexType lastChunkOfSegment = rowToChunkMapping[ segmentIdx ];

   const IndexType segmentChunksCount = lastChunkOfSegment - firstChunkOfSegment;

   const IndexType sliceOffset = slices[ sliceIndex ].pointer;

   const IndexType chunkSize = slices[ sliceIndex ].chunkSize;

   const IndexType segmentSize = segmentChunksCount * chunkSize;



   if( RowMajorOrder )

      return SegmentViewType( sliceOffset + firstChunkOfSegment * chunkSize,

                              segmentSize,

                              chunkSize,

                              chunksInSlice );

   else // TODO FIX !!!!!!!!!!!!!!

      return SegmentViewType( sliceOffset + firstChunkOfSegment,

                              segmentSize,

                              chunkSize,

                              chunksInSlice );

   

   

   

   

   if( std::is_same< DeviceType, Devices::Host >::value )

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getSegmentViewDirect(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         chunksInSlice,

         segmentIdx );

   if( std::is_same< DeviceType, Devices::Cuda >::value )

   {

#ifdef __CUDA_ARCH__

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getSegmentViewDirect(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         chunksInSlice,

         segmentIdx );

#else

      return details::ChunkedEllpack< IndexType, DeviceType, RowMajorOrder >::getSegmentView(

         rowToSliceMapping,

         slices,

         rowToChunkMapping,

         chunksInSlice,

         segmentIdx );

#endif

   }

}



template< typename Device,

src/TNL/Containers/Segments/details/ChunkedEllpack.h

0 → 100644
+230 −0

File added.

Preview size limit exceeded, changes collapsed.