From 01f367ca5e64c90d8c1db876e16bad37b3ef6e2d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Tom=C3=A1=C5=A1=20Oberhuber?= <oberhuber.tomas@gmail.com>
Date: Mon, 9 Dec 2019 20:07:29 +0100
Subject: [PATCH] Implementing SlicedEllpack segments.
---
src/TNL/Containers/Segments/CSR.hpp | 2 +-
src/TNL/Containers/Segments/Ellpack.hpp | 5 +-
src/TNL/Containers/Segments/SlicedEllpack.h | 14 +-
src/TNL/Containers/Segments/SlicedEllpack.hpp | 191 ++++++++++++------
4 files changed, 141 insertions(+), 71 deletions(-)
diff --git a/src/TNL/Containers/Segments/CSR.hpp b/src/TNL/Containers/Segments/CSR.hpp
index 486149e042..ef7431038c 100644
--- a/src/TNL/Containers/Segments/CSR.hpp
+++ b/src/TNL/Containers/Segments/CSR.hpp
@@ -153,7 +153,7 @@ void
CSR< Device, Index >::
forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
{
- const auto offsetsView = this->offsets.getView();
+ const auto offsetsView = this->offsets.getConstView();
auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) mutable {
const IndexType begin = offsetsView[ i ];
const IndexType end = offsetsView[ i + 1 ];
diff --git a/src/TNL/Containers/Segments/Ellpack.hpp b/src/TNL/Containers/Segments/Ellpack.hpp
index 034b0820eb..d3d90be5e0 100644
--- a/src/TNL/Containers/Segments/Ellpack.hpp
+++ b/src/TNL/Containers/Segments/Ellpack.hpp
@@ -189,11 +189,10 @@ void
Ellpack< Device, Index, RowMajorOrder, Alignment >::
forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
{
- const auto offsetsView = this->offsets.getView();
if( RowMajorOrder )
{
const IndexType segmentSize = this->segmentSize;
- auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) {
+ auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) mutable {
const IndexType begin = i * segmentSize;
const IndexType end = begin + segmentSize;
for( IndexType j = begin; j < end; j++ )
@@ -206,7 +205,7 @@ forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
{
const IndexType storageSize = this->getStorageSize();
const IndexType alignedSize = this->alignedSize;
- auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) {
+ auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) mutable {
const IndexType begin = i;
const IndexType end = storageSize;
for( IndexType j = begin; j < end; j += alignedSize )
diff --git a/src/TNL/Containers/Segments/SlicedEllpack.h b/src/TNL/Containers/Segments/SlicedEllpack.h
index a5ef9d1211..ecc2c8c7ef 100644
--- a/src/TNL/Containers/Segments/SlicedEllpack.h
+++ b/src/TNL/Containers/Segments/SlicedEllpack.h
@@ -42,7 +42,13 @@ class SlicedEllpack
* \brief Set sizes of particular segments.
*/
template< typename SizesHolder = OffsetsHolder >
- void setSizes( const SizesHolder& sizes );
+ void setSegmentsSizes( const SizesHolder& sizes );
+
+ __cuda_callable__
+ IndexType getSegmentsCount() const;
+
+ __cuda_callable__
+ IndexType getSegmentSize( const IndexType segmentIdx ) const;
/**
* \brief Number segments.
@@ -50,8 +56,6 @@ class SlicedEllpack
__cuda_callable__
IndexType getSize() const;
- __cuda_callable__
- IndexType getSegmentSize( const IndexType segmentIdx ) const;
__cuda_callable__
IndexType getStorageSize() const;
@@ -90,9 +94,9 @@ class SlicedEllpack
protected:
- IndexType size;
+ IndexType size, alignedSize, segmentsCount;
- OffsetHolder sliceOffsets;
+ OffsetsHolder sliceOffsets, sliceSegmentSizes;
};
} // namespace Segements
diff --git a/src/TNL/Containers/Segments/SlicedEllpack.hpp b/src/TNL/Containers/Segments/SlicedEllpack.hpp
index c91a13473f..e23ee5f15b 100644
--- a/src/TNL/Containers/Segments/SlicedEllpack.hpp
+++ b/src/TNL/Containers/Segments/SlicedEllpack.hpp
@@ -26,17 +26,30 @@ template< typename Device,
int SliceSize >
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
SlicedEllpack()
- : size( 0 )
+ : size( 0 ), alignedSize( 0 ), segmentsCount( 0 )
{
}
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack( const Vector< IndexType, DeviceType, IndexType >& sizes )
+ : size( 0 ), alignedSize( 0 ), segmentsCount( 0 )
+{
+ this->setSegmentsSizes( sizes );
+}
+
template< typename Device,
typename Index,
bool RowMajorOrder,
int SliceSize >
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
SlicedEllpack( const SlicedEllpack& slicedEllpack )
- : size( slicedEllpack.size ), sliceOffsets( slicedEllpack.sliceOffsets )
+ : size( slicedEllpack.size ), alignedSize( slicedEllpack.alignedSize ),
+ segmentsCount( slicedEllpack.segmentsCount ), sliceOffsets( slicedEllpack.sliceOffsets ),
+ sliceSegmentSizes( slicedEllpack.sliceSegmentSizes )
{
}
@@ -46,7 +59,9 @@ template< typename Device,
int SliceSize >
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
SlicedEllpack( const SlicedEllpack&& slicedEllpack )
- : size( slicedEllpack.size ), sliceOffsets( slicedEllpack.sliceOffsets )
+ : size( slicedEllpack.size ), alignedSize( slicedEllpack.alignedSize ),
+ segmentsCount( slicedEllpack.segmentsCount ), sliceOffsets( slicedEllpack.sliceOffsets ),
+ sliceSegmentSizes( slicedEllpack.sliceSegmentSizes )
{
}
@@ -57,36 +72,36 @@ template< typename Device,
template< typename SizesHolder >
void
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
-setSizes( const SizesHolder& sizes )
+setSegmentsSizes( const SizesHolder& sizes )
{
- this->size = sizes.getSize();
- const IndexType segmentsCount = roundUpDivision( this->size, getSliceSize() );
- this->segmentOffsets.setSize( segmentsCount + 1 );
+ this->segmentsCount = sizes.getSize();
+ const IndexType slicesCount = roundUpDivision( this->segmentsCount, getSliceSize() );
+ this->sliceOffsets.setSize( slicesCount + 1 );
+ this->sliceOffsets = 0;
+ this->sliceSegmentSizes.setSize( slicesCount );
Ellpack< DeviceType, IndexType, true > ellpack;
- ellpack.setSizes( segmentsCount, SliceSize );
+ ellpack.setSegmentsSizes( slicesCount, SliceSize );
- const IndexType _size = this->getSize();
+ const IndexType _size = sizes.getSize();
const auto sizes_view = sizes.getConstView();
- auto offsets_view = this->segmentOffsets().getView();
+ auto slices_view = this->sliceOffsets.getView();
+ auto slice_segment_size_view = this->sliceSegmentSizes.getView();
auto fetch = [=] __cuda_callable__ ( IndexType segmentIdx, IndexType globalIdx ) -> IndexType {
- if( globalIdx < size )
+ if( globalIdx < _size )
return sizes_view[ globalIdx ];
+ return 0;
};
auto reduce = [] __cuda_callable__ ( IndexType& aux, const IndexType i ) {
aux = TNL::max( aux, i );
};
- auto keep = [=] __cuda_callable__ ( IndexType i, IndexType res ) {
- offsets_view[ i ] = res;
- }
-
- std::cerr << offsets_view << std::endl;
-
-
-
- if( RowMajorOrder )
- this->alignedSize = this->size;
- else
- this->alignedSize = roundUpDivision( size, this->getSliceSize() ) * this->getSliceSize();
+ auto keep = [=] __cuda_callable__ ( IndexType i, IndexType res ) mutable {
+ slices_view[ i ] = res * SliceSize;
+ slice_segment_size_view[ i ] = res;
+ };
+ ellpack.allReduction( fetch, reduce, keep, std::numeric_limits< IndexType >::min() );
+ this->sliceOffsets.template scan< Algorithms::ScanType::Exclusive >();
+ this->size = sum( sizes );
+ this->alignedSize = this->sliceOffsets.getElement( slicesCount );
}
template< typename Device,
@@ -96,9 +111,9 @@ template< typename Device,
__cuda_callable__
Index
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
-getSize() const
+getSegmentsCount() const
{
- return this->size;
+ return this->segmentsCount;
}
template< typename Device,
@@ -110,7 +125,29 @@ Index
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
getSegmentSize( const IndexType segmentIdx ) const
{
- return this->segmentSize;
+ const Index sliceIdx = segmentIdx / SliceSize;
+ if( std::is_same< DeviceType, Devices::Host >::value )
+ return this->sliceSegmentSizes[ sliceIdx ];
+ else
+ {
+#ifdef __CUDA_ARCH__
+ return this->sliceSegmentSizes[ sliceIdx ];
+#else
+ return this->sliceSegmentSizes.getElement( sliceIdx );
+#endif
+ }
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+Index
+SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+getSize() const
+{
+ return this->size;
}
template< typename Device,
@@ -122,7 +159,7 @@ Index
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
getStorageSize() const
{
- return this->alignedSize * this->segmentSize;
+ return this->alignedSize;
}
template< typename Device,
@@ -134,10 +171,28 @@ Index
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
{
+ const IndexType sliceIdx = segmentIdx / SliceSize;
+ const IndexType segmentInSliceIdx = segmentIdx % SliceSize;
+ IndexType sliceOffset, segmentSize;
+ if( std::is_same< DeviceType, Devices::Host >::value )
+ {
+ sliceOffset = this->sliceOffsets[ sliceIdx ];
+ segmentSize = this->sliceSegmentSizes[ sliceIdx ];
+ }
+ else
+ {
+#ifdef __CUDA__ARCH__
+ sliceOffset = this->sliceOffsets[ sliceIdx ];
+ segmentSize = this->sliceSegmentSizes[ sliceIdx ];
+#else
+ sliceOffset = this->sliceOffsets.getElement( sliceIdx );
+ segmentSize = this->sliceSegmentSizes.getElement( sliceIdx );
+#endif
+ }
if( RowMajorOrder )
- return segmentIdx * this->segmentSize + localIdx;
+ return sliceOffset + segmentInSliceIdx * segmentSize + localIdx;
else
- return segmentIdx + this->alignedSize * localIdx;
+ return sliceOffset + segmentInSliceIdx + SliceSize * localIdx;
}
template< typename Device,
@@ -160,28 +215,32 @@ void
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
{
- const auto offsetsView = this->offsets.getView();
+ const auto sliceSegmentSizes_view = this->sliceSegmentSizes.getConstView();
+ const auto sliceOffsets_view = this->sliceOffsets.getConstView();
if( RowMajorOrder )
{
- const IndexType segmentSize = this->segmentSize;
- auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) {
- const IndexType begin = i * segmentSize;
+ auto l = [=] __cuda_callable__ ( const IndexType segmentIdx, Args... args ) {
+ 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;
- for( IndexType j = begin; j < end; j++ )
- if( ! f( i, j, args... ) )
+ for( IndexType globalIdx = begin; globalIdx < end; globalIdx++ )
+ if( ! f( segmentIdx, globalIdx, args... ) )
break;
};
Algorithms::ParallelFor< Device >::exec( first, last, l, args... );
}
else
{
- const IndexType storageSize = this->getStorageSize();
- const IndexType alignedSize = this->alignedSize;
- auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) {
- const IndexType begin = i;
- const IndexType end = storageSize;
- for( IndexType j = begin; j < end; j += alignedSize )
- if( ! f( i, j, args... ) )
+ auto l = [=] __cuda_callable__ ( const IndexType segmentIdx, Args... args ) {
+ 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 ];
+ for( IndexType globalIdx = begin; globalIdx < end; globalIdx += SliceSize )
+ if( ! f( segmentIdx, globalIdx, args... ) )
break;
};
Algorithms::ParallelFor< Device >::exec( first, last, l, args... );
@@ -197,7 +256,7 @@ void
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
forAll( Function& f, Args... args ) const
{
- this->forSegments( 0, this->getSize(), f, args... );
+ this->forSegments( 0, this->getSegmentsCount(), f, args... );
}
template< typename Device,
@@ -209,32 +268,36 @@ void
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
{
+ using RealType = decltype( fetch( IndexType(), IndexType() ) );
+ const auto sliceSegmentSizes_view = this->sliceSegmentSizes.getConstView();
+ const auto sliceOffsets_view = this->sliceOffsets.getConstView();
if( RowMajorOrder )
{
- using RealType = decltype( fetch( IndexType(), IndexType() ) );
- const IndexType segmentSize = this->segmentSize;
- auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) mutable {
- const IndexType begin = i * segmentSize;
+ 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 );
- for( IndexType j = begin; j < end; j++ )
- reduction( aux, fetch( i, j, args... ) );
- keeper( i, aux );
+ for( IndexType globalIdx = begin; globalIdx< end; globalIdx++ )
+ reduction( aux, fetch( segmentIdx, globalIdx, args... ) );
+ keeper( segmentIdx, aux );
};
Algorithms::ParallelFor< Device >::exec( first, last, l, args... );
}
else
{
- using RealType = decltype( fetch( IndexType(), IndexType() ) );
- const IndexType storageSize = this->getStorageSize();
- const IndexType alignedSize = this->alignedSize;
- auto l = [=] __cuda_callable__ ( const IndexType i, Args... args ) mutable {
- const IndexType begin = i;
- const IndexType end = storageSize;
+ 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 );
- for( IndexType j = begin; j < end; j += alignedSize )
- reduction( aux, fetch( i, j, args... ) );
- keeper( i, aux );
+ for( IndexType globalIdx = begin; globalIdx < end; globalIdx += SliceSize )
+ reduction( aux, fetch( segmentIdx, globalIdx, args... ) );
+ keeper( segmentIdx, aux );
};
Algorithms::ParallelFor< Device >::exec( first, last, l, args... );
}
@@ -249,7 +312,7 @@ void
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
{
- this->segmentsReduction( 0, this->getSize(), fetch, reduction, keeper, zero, args... );
+ this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
}
template< typename Device,
@@ -260,9 +323,11 @@ void
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
save( File& file ) const
{
- file.save( &segmentSize );
file.save( &size );
file.save( &alignedSize );
+ file.save( &segmentsCount );
+ this->sliceOffsets.save( file );
+ this->sliceSegmentSizes.save( file );
}
template< typename Device,
@@ -273,9 +338,11 @@ void
SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
load( File& file )
{
- file.load( &segmentSize );
file.load( &size );
file.load( &alignedSize );
+ file.load( &segmentsCount );
+ this->sliceOffsets.load( file );
+ this->sliceSegmentSizes.load( file );
}
} // namespace Segments
--
GitLab