diff --git a/src/Benchmarks/SpMV/spmv.h b/src/Benchmarks/SpMV/spmv.h
index a6acb52fd0d6ff86985df379205b89de5874d263..66f4fb236a6d10fe5e777278d37a749cfe10d16a 100644
--- a/src/Benchmarks/SpMV/spmv.h
+++ b/src/Benchmarks/SpMV/spmv.h
@@ -45,14 +45,14 @@ using SlicedEllpackAlias = Matrices::SlicedEllpack< Real, Device, Index >;
template< typename Real, typename Device, typename Index >
using SparseMatrix_CSR = Matrices::SparseMatrix< Real, Containers::Segments::CSR, Device, Index >;
-template< typename Device, typename Index >
-using EllpackSegments = Containers::Segments::Ellpack< Device, Index >;
+template< typename Device, typename Index, typename IndexAllocator >
+using EllpackSegments = Containers::Segments::Ellpack< Device, Index, IndexAllocator >;
template< typename Real, typename Device, typename Index >
using SparseMatrix_Ellpack = Matrices::SparseMatrix< Real, EllpackSegments, Device, Index >;
-template< typename Device, typename Index >
-using SlicedEllpackSegments = Containers::Segments::SlicedEllpack< Device, Index >;
+template< typename Device, typename Index, typename IndexAllocator >
+using SlicedEllpackSegments = Containers::Segments::SlicedEllpack< Device, Index, IndexAllocator >;
template< typename Real, typename Device, typename Index >
using SparseMatrix_SlicedEllpack = Matrices::SparseMatrix< Real, SlicedEllpackSegments, Device, Index >;
diff --git a/src/TNL/Containers/Segments/CSR.h b/src/TNL/Containers/Segments/CSR.h
index ecd1de983ce309be9f2ea20575e45a9267b6b261..b83e43f1d146091219e3948afdbb94bfa0ae0b4e 100644
--- a/src/TNL/Containers/Segments/CSR.h
+++ b/src/TNL/Containers/Segments/CSR.h
@@ -10,22 +10,28 @@
#pragma once
+#include <type_traits>
+
#include <TNL/Containers/Vector.h>
+#include <TNL/Containers/Segments/CSRView.h>
namespace TNL {
namespace Containers {
namespace Segments {
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index > >
class CSR
{
public:
using DeviceType = Device;
using IndexType = Index;
- using OffsetsHolder = Containers::Vector< IndexType, DeviceType, IndexType >;
+ using OffsetsHolder = Containers::Vector< IndexType, DeviceType, typename std::remove_const< IndexType >::type, IndexAllocator >;
using SegmentsSizes = OffsetsHolder;
+ using ViewType = CSRView< Device, Index >;
+ using ConstViewType = CSRView< Device, std::add_const_t< Index > >;
CSR();
@@ -41,6 +47,10 @@ class CSR
template< typename SizesHolder = OffsetsHolder >
void setSegmentsSizes( const SizesHolder& sizes );
+ ViewType getView();
+
+ ConstViewType getConstView() const;
+
/**
* \brief Number segments.
*/
diff --git a/src/TNL/Containers/Segments/CSR.hpp b/src/TNL/Containers/Segments/CSR.hpp
index ccb483125d423edde3237e8a69b3e24b83a3334e..a8f12e7dc3e3db1504a54a045e92b8cb5f0cbbfc 100644
--- a/src/TNL/Containers/Segments/CSR.hpp
+++ b/src/TNL/Containers/Segments/CSR.hpp
@@ -13,6 +13,7 @@
#include <TNL/Containers/Vector.h>
#include <TNL/Algorithms/ParallelFor.h>
#include <TNL/Containers/Segments/CSR.h>
+#include <TNL/Containers/Segments/details/CSR.h>
namespace TNL {
namespace Containers {
@@ -20,64 +21,92 @@ namespace TNL {
template< typename Device,
- typename Index >
-CSR< Device, Index >::
+ typename Index,
+ typename IndexAllocator >
+CSR< Device, Index, IndexAllocator >::
CSR()
{
}
template< typename Device,
- typename Index >
-CSR< Device, Index >::
+ typename Index,
+ typename IndexAllocator >
+CSR< Device, Index, IndexAllocator >::
CSR( const SegmentsSizes& segmentsSizes )
{
this->setSegmentsSizes( segmentsSizes );
}
template< typename Device,
- typename Index >
-CSR< Device, Index >::
+ typename Index,
+ typename IndexAllocator >
+CSR< Device, Index, IndexAllocator >::
CSR( const CSR& csr ) : offsets( csr.offsets )
{
}
template< typename Device,
- typename Index >
-CSR< Device, Index >::
+ typename Index,
+ typename IndexAllocator >
+CSR< Device, Index, IndexAllocator >::
CSR( const CSR&& csr ) : offsets( std::move( csr.offsets ) )
{
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
template< typename SizesHolder >
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
setSegmentsSizes( const SizesHolder& sizes )
{
- this->offsets.setSize( sizes.getSize() + 1 );
+ details::CSR< Device, Index >::setSegmentsSizes( sizes, this->offsets );
+ /*this->offsets.setSize( sizes.getSize() + 1 );
auto view = this->offsets.getView( 0, sizes.getSize() );
view = sizes;
this->offsets.setElement( sizes.getSize(), 0 );
- this->offsets.template scan< Algorithms::ScanType::Exclusive >();
+ this->offsets.template scan< Algorithms::ScanType::Exclusive >();*/
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
+typename CSR< Device, Index, IndexAllocator >::ViewType
+CSR< Device, Index, IndexAllocator >::
+getView()
+{
+ return ViewType( this->offsets.getView() );
+}
+
+template< typename Device,
+ typename Index,
+ typename IndexAllocator >
+typename CSR< Device, Index, IndexAllocator >::ConstViewType
+CSR< Device, Index, IndexAllocator >::
+getConstView() const
+{
+ return ConstViewType( this->offsets.getConstView() );
+}
+
+template< typename Device,
+ typename Index,
+ typename IndexAllocator >
__cuda_callable__
Index
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
getSegmentsCount() const
{
return this->offsets.getSize() - 1;
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
__cuda_callable__
Index
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
getSegmentSize( const IndexType segmentIdx ) const
{
if( ! std::is_same< DeviceType, Devices::Host >::value )
@@ -92,20 +121,22 @@ getSegmentSize( const IndexType segmentIdx ) const
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
__cuda_callable__
Index
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
getSize() const
{
return this->getStorageSize();
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
__cuda_callable__
Index
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
getStorageSize() const
{
if( ! std::is_same< DeviceType, Devices::Host >::value )
@@ -120,10 +151,11 @@ getStorageSize() const
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
__cuda_callable__
Index
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
{
if( ! std::is_same< DeviceType, Devices::Host >::value )
@@ -138,19 +170,21 @@ getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
__cuda_callable__
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const
{
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
template< typename Function, typename... Args >
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
{
const auto offsetsView = this->offsets.getConstView();
@@ -166,20 +200,22 @@ forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator>
template< typename Function, typename... Args >
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
forAll( Function& f, Args... args ) const
{
this->forSegments( 0, this->getSize(), f, args... );
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
{
using RealType = decltype( fetch( IndexType(), IndexType() ) );
@@ -196,28 +232,31 @@ segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& red
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
{
this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
save( File& file ) const
{
file << this->offsets;
}
template< typename Device,
- typename Index >
+ typename Index,
+ typename IndexAllocator >
void
-CSR< Device, Index >::
+CSR< Device, Index, IndexAllocator >::
load( File& file )
{
file >> this->offsets;
diff --git a/src/TNL/Containers/Segments/CSRView.h b/src/TNL/Containers/Segments/CSRView.h
index 5eeb7ecb3efafb5258f9100505874df4f8322538..2f89579702b543cf76f47d32e20bebf2c497828c 100644
--- a/src/TNL/Containers/Segments/CSRView.h
+++ b/src/TNL/Containers/Segments/CSRView.h
@@ -10,6 +10,8 @@
#pragma once
+#include <type_traits>
+
#include <TNL/Containers/Vector.h>
namespace TNL {
@@ -24,13 +26,19 @@ class CSRView
using DeviceType = Device;
using IndexType = Index;
- using OffsetsHolderView = typedef Containers::Vector< IndexType, DeviceType, IndexType >::ViewType;
+ using OffsetsView = typename Containers::VectorView< IndexType, DeviceType, IndexType >;
+ using ConstOffsetsView = typename Containers::Vector< IndexType, DeviceType, IndexType >::ConstViewType;
+ using ViewType = CSRView;
+ using ConstViewType = CSRView< Device, std::add_const_t< Index > >;
__cuda_callable__
CSRView();
__cuda_callable__
- CSRView( const OffsetsHolderView& offsets );
+ CSRView( const OffsetsView&& offsets );
+
+ __cuda_callable__
+ CSRView( const ConstOffsetsView&& offsets );
__cuda_callable__
CSRView( const CSRView& csr_view );
@@ -38,6 +46,10 @@ class CSRView
__cuda_callable__
CSRView( const CSRView&& csr_view );
+ ViewType getView();
+
+ ConstViewType getConstView() const;
+
/**
* \brief Number segments.
*/
@@ -96,7 +108,7 @@ class CSRView
protected:
- OffsetsHolderView offsets;
+ OffsetsView offsets;
};
} // namespace Segements
} // namespace Conatiners
diff --git a/src/TNL/Containers/Segments/CSRView.hpp b/src/TNL/Containers/Segments/CSRView.hpp
index 30ed24071e7b42b8cd27cf959e890366682319ed..f50a74985f050f18b5913008636552e1bbf4f760 100644
--- a/src/TNL/Containers/Segments/CSRView.hpp
+++ b/src/TNL/Containers/Segments/CSRView.hpp
@@ -31,7 +31,7 @@ template< typename Device,
typename Index >
__cuda_callable__
CSRView< Device, Index >::
-CSRView( const OffsetsHolderView& offsets_view )
+CSRView( const OffsetsView&& offsets_view )
: offsets( offsets_view )
{
}
@@ -40,10 +40,18 @@ template< typename Device,
typename Index >
__cuda_callable__
CSRView< Device, Index >::
-CSRView( const CSRView& csr_view )
- : offsets( csr_view.offsest )
+CSRView( const ConstOffsetsView&& offsets_view )
+ : offsets( offsets_view )
{
+}
+template< typename Device,
+ typename Index >
+__cuda_callable__
+CSRView< Device, Index >::
+CSRView( const CSRView& csr_view )
+ : offsets( csr_view.offsets )
+{
}
template< typename Device,
@@ -51,9 +59,26 @@ template< typename Device,
__cuda_callable__
CSRView< Device, Index >::
CSRView( const CSRView&& csr_view )
- : offsets( std::move( csr_view.offsest ) )
+ : offsets( std::move( csr_view.offsets ) )
+{
+}
+
+template< typename Device,
+ typename Index >
+typename CSRView< Device, Index >::ViewType
+CSRView< Device, Index >::
+getView()
{
+ return ViewType( this->offsets );
+}
+template< typename Device,
+ typename Index >
+typename CSRView< Device, Index >::ConstViewType
+CSRView< Device, Index >::
+getConstView() const
+{
+ return ConstViewType( this->offsets.getConstView() );
}
template< typename Device,
diff --git a/src/TNL/Containers/Segments/Ellpack.h b/src/TNL/Containers/Segments/Ellpack.h
index b08ad0f04f9d316f6e2ce62aea5d8990c1204978..9c81a84281925b2ec971bceba5a161aa464c83e6 100644
--- a/src/TNL/Containers/Segments/Ellpack.h
+++ b/src/TNL/Containers/Segments/Ellpack.h
@@ -11,6 +11,7 @@
#pragma once
#include <TNL/Containers/Vector.h>
+#include <TNL/Containers/Segments/EllpackView.h>
namespace TNL {
namespace Containers {
@@ -18,6 +19,7 @@ namespace TNL {
template< typename Device,
typename Index,
+ typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >,
bool RowMajorOrder = std::is_same< Device, Devices::Host >::value,
int Alignment = 32 >
class Ellpack
@@ -30,6 +32,9 @@ class Ellpack
static constexpr bool getRowMajorOrder() { return RowMajorOrder; }
using OffsetsHolder = Containers::Vector< IndexType, DeviceType, IndexType >;
using SegmentsSizes = OffsetsHolder;
+ using ViewType = EllpackView< Device, Index, RowMajorOrder, Alignment >;
+ //using ConstViewType = EllpackView< Device, std::add_const_t< Index >, RowMajorOrder, Alignment >;
+
Ellpack();
@@ -41,6 +46,10 @@ class Ellpack
Ellpack( const Ellpack&& segments );
+ ViewType getView();
+
+ //ConstViewType getConstView() const;
+
/**
* \brief Set sizes of particular segments.
*/
diff --git a/src/TNL/Containers/Segments/Ellpack.hpp b/src/TNL/Containers/Segments/Ellpack.hpp
index 337009e99a7b0a3493b5015190c5d915ea6714ce..482c87d4f36274edf8aef622f576ce07773df3b9 100644
--- a/src/TNL/Containers/Segments/Ellpack.hpp
+++ b/src/TNL/Containers/Segments/Ellpack.hpp
@@ -21,9 +21,10 @@ namespace TNL {
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
Ellpack()
: segmentSize( 0 ), size( 0 ), alignedSize( 0 )
{
@@ -31,9 +32,10 @@ Ellpack()
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
Ellpack( const SegmentsSizes& segmentsSizes )
: segmentSize( 0 ), size( 0 ), alignedSize( 0 )
{
@@ -42,9 +44,10 @@ Ellpack( const SegmentsSizes& segmentsSizes )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
Ellpack( const IndexType segmentsCount, const IndexType segmentSize )
: segmentSize( 0 ), size( 0 ), alignedSize( 0 )
{
@@ -53,9 +56,10 @@ Ellpack( const IndexType segmentsCount, const IndexType segmentSize )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
Ellpack( const Ellpack& ellpack )
: segmentSize( ellpack.segmentSize ), size( ellpack.size ), alignedSize( ellpack.alignedSize )
{
@@ -63,9 +67,10 @@ Ellpack( const Ellpack& ellpack )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
Ellpack( const Ellpack&& ellpack )
: segmentSize( ellpack.segmentSize ), size( ellpack.size ), alignedSize( ellpack.alignedSize )
{
@@ -73,11 +78,35 @@ Ellpack( const Ellpack&& ellpack )
template< typename Device,
typename Index,
+ typename IndexAllocator,
+ bool RowMajorOrder,
+ int Alignment >
+typename Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::ViewType
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
+getView()
+{
+ return ViewType( segmentSize, size, alignedSize );
+}
+
+/*template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+typename Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::ConstViewType
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
+getConstView() const
+{
+ return ConstViewType( segmentSize, size, alignedSize );
+}*/
+
+template< typename Device,
+ typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
template< typename SizesHolder >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
setSegmentsSizes( const SizesHolder& sizes )
{
this->segmentSize = max( sizes );
@@ -90,10 +119,11 @@ setSegmentsSizes( const SizesHolder& sizes )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
setSegmentsSizes( const IndexType segmentsCount, const IndexType segmentSize )
{
this->segmentSize = segmentSize;
@@ -107,11 +137,12 @@ setSegmentsSizes( const IndexType segmentsCount, const IndexType segmentSize )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
__cuda_callable__
Index
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
getSegmentsCount() const
{
return this->size;
@@ -119,11 +150,12 @@ getSegmentsCount() const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
__cuda_callable__
Index
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
getSegmentSize( const IndexType segmentIdx ) const
{
return this->segmentSize;
@@ -131,11 +163,12 @@ getSegmentSize( const IndexType segmentIdx ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
__cuda_callable__
Index
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
getSize() const
{
return this->size * this->segmentSize;
@@ -144,11 +177,12 @@ getSize() const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
__cuda_callable__
Index
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
getStorageSize() const
{
return this->alignedSize * this->segmentSize;
@@ -156,11 +190,12 @@ getStorageSize() const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
__cuda_callable__
Index
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
{
if( RowMajorOrder )
@@ -171,22 +206,24 @@ getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
__cuda_callable__
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const
{
}
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
template< typename Function, typename... Args >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
{
if( RowMajorOrder )
@@ -220,11 +257,12 @@ forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
template< typename Function, typename... Args >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
forAll( Function& f, Args... args ) const
{
this->forSegments( 0, this->getSize(), f, args... );
@@ -232,11 +270,12 @@ forAll( Function& f, Args... args ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
{
if( RowMajorOrder )
@@ -272,11 +311,12 @@ segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& red
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
{
this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
@@ -284,10 +324,11 @@ allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Re
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
save( File& file ) const
{
file.save( &segmentSize );
@@ -297,10 +338,11 @@ save( File& file ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int Alignment >
void
-Ellpack< Device, Index, RowMajorOrder, Alignment >::
+Ellpack< Device, Index, IndexAllocator, RowMajorOrder, Alignment >::
load( File& file )
{
file.load( &segmentSize );
diff --git a/src/TNL/Containers/Segments/EllpackView.h b/src/TNL/Containers/Segments/EllpackView.h
new file mode 100644
index 0000000000000000000000000000000000000000..adbfee629c03d9ff49c572781cabc4a95c0ee0ba
--- /dev/null
+++ b/src/TNL/Containers/Segments/EllpackView.h
@@ -0,0 +1,111 @@
+/***************************************************************************
+ EllpackView.h - description
+ -------------------
+ begin : Dec 12, 2019
+ copyright : (C) 2019 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+#include <type_traits>
+
+#include <TNL/Containers/Vector.h>
+
+
+namespace TNL {
+ namespace Containers {
+ namespace Segments {
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder = std::is_same< Device, Devices::Host >::value,
+ int Alignment = 32 >
+class EllpackView
+{
+ public:
+
+ using DeviceType = Device;
+ using IndexType = Index;
+ static constexpr int getAlignment() { return Alignment; }
+ static constexpr bool getRowMajorOrder() { return RowMajorOrder; }
+ using OffsetsHolder = Containers::Vector< IndexType, DeviceType, IndexType >;
+ using SegmentsSizes = OffsetsHolder;
+ using ViewType = EllpackView;
+ //using ConstViewType = EllpackView< Device, std::add_const_t< Index > >;
+
+ __cuda_callable__
+ EllpackView();
+
+ __cuda_callable__
+ EllpackView( IndexType segmentSize, IndexType size, IndexType alignedSize );
+
+ __cuda_callable__
+ EllpackView( const EllpackView& ellpackView );
+
+ __cuda_callable__
+ EllpackView( const EllpackView&& ellpackView );
+
+ ViewType getView();
+
+ //ConstViewType getConstView() const;
+
+ /**
+ * \brief Number segments.
+ */
+ __cuda_callable__
+ IndexType getSegmentsCount() const;
+
+ __cuda_callable__
+ IndexType getSegmentSize( const IndexType segmentIdx ) const;
+
+ __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;
+
+ /***
+ * \brief Go over all segments and for each segment element call
+ * function 'f' with arguments 'args'. The return type of 'f' is bool.
+ * When its true, the for-loop continues. Once 'f' returns false, the for-loop
+ * is terminated.
+ */
+ template< typename Function, typename... Args >
+ void forSegments( IndexType first, IndexType last, Function& f, Args... args ) const;
+
+ template< typename Function, typename... Args >
+ void forAll( Function& f, Args... args ) const;
+
+
+ /***
+ * \brief Go over all segments and perform a reduction in each of them.
+ */
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ void save( File& file ) const;
+
+ void load( File& file );
+
+ protected:
+
+ IndexType segmentSize, size, alignedSize;
+};
+
+ } // namespace Segements
+ } // namespace Conatiners
+} // namespace TNL
+
+#include <TNL/Containers/Segments/EllpackView.hpp>
diff --git a/src/TNL/Containers/Segments/EllpackView.hpp b/src/TNL/Containers/Segments/EllpackView.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..d124633ff68f1d64230d754a6d81dd87a6fc0117
--- /dev/null
+++ b/src/TNL/Containers/Segments/EllpackView.hpp
@@ -0,0 +1,293 @@
+/***************************************************************************
+ EllpackView.hpp - description
+ -------------------
+ begin : Dec 12, 2019
+ copyright : (C) 2019 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+#include <TNL/Containers/Vector.h>
+#include <TNL/Algorithms/ParallelFor.h>
+#include <TNL/Containers/Segments/EllpackView.h>
+
+namespace TNL {
+ namespace Containers {
+ namespace Segments {
+
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+EllpackView()
+ : segmentSize( 0 ), size( 0 ), alignedSize( 0 )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+EllpackView( IndexType segmentSize, IndexType size, IndexType alignedSize )
+ : segmentSize( segmentSize ), size( size ), alignedSize( alignedSize )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+EllpackView( const EllpackView& ellpack )
+ : segmentSize( ellpack.segmentSize ), size( ellpack.size ), alignedSize( ellpack.alignedSize )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+EllpackView( const EllpackView&& ellpack )
+ : segmentSize( ellpack.segmentSize ), size( ellpack.size ), alignedSize( ellpack.alignedSize )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+typename EllpackView< Device, Index, RowMajorOrder, Alignment >::ViewType
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getView()
+{
+ return ViewType( segmentSize, size, alignedSize );
+}
+
+/*template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+typename EllpackView< Device, Index, RowMajorOrder, Alignment >::ConstViewType
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getConstView() const
+{
+ return ConstViewType( segmentSize, size, alignedSize );
+}*/
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+Index
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getSegmentsCount() const
+{
+ return this->size;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+Index
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getSegmentSize( const IndexType segmentIdx ) const
+{
+ return this->segmentSize;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+Index
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getSize() const
+{
+ return this->size * this->segmentSize;
+}
+
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+Index
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getStorageSize() const
+{
+ return this->alignedSize * this->segmentSize;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+Index
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
+{
+ if( RowMajorOrder )
+ return segmentIdx * this->segmentSize + localIdx;
+ else
+ return segmentIdx + this->alignedSize * localIdx;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+__cuda_callable__
+void
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+ template< typename Function, typename... Args >
+void
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
+{
+ if( RowMajorOrder )
+ {
+ const IndexType segmentSize = this->segmentSize;
+ auto l = [=] __cuda_callable__ ( const IndexType segmentIdx, Args... args ) mutable {
+ const IndexType begin = segmentIdx * 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
+ {
+ const IndexType storageSize = this->getStorageSize();
+ const IndexType alignedSize = this->alignedSize;
+ auto l = [=] __cuda_callable__ ( const IndexType segmentIdx, Args... args ) mutable {
+ const IndexType begin = segmentIdx;
+ const IndexType end = storageSize;
+ IndexType localIdx( 0 );
+ for( IndexType globalIdx = begin; globalIdx < end; globalIdx += alignedSize )
+ if( ! f( segmentIdx, localIdx++, globalIdx, args... ) )
+ break;
+ };
+ Algorithms::ParallelFor< Device >::exec( first, last, l, args... );
+ }
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+ template< typename Function, typename... Args >
+void
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+forAll( Function& f, Args... args ) const
+{
+ this->forSegments( 0, this->getSize(), f, args... );
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+void
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
+{
+ 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;
+ const IndexType end = begin + segmentSize;
+ RealType aux( zero );
+ for( IndexType j = begin; j < end; j++ )
+ reduction( aux, fetch( i, j, args... ) );
+ keeper( i, 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;
+ RealType aux( zero );
+ for( IndexType j = begin; j < end; j += alignedSize )
+ reduction( aux, fetch( i, j, args... ) );
+ keeper( i, aux );
+ };
+ Algorithms::ParallelFor< Device >::exec( first, last, l, args... );
+ }
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+void
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
+{
+ this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+void
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+save( File& file ) const
+{
+ file.save( &segmentSize );
+ file.save( &size );
+ file.save( &alignedSize );
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+void
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+load( File& file )
+{
+ file.load( &segmentSize );
+ file.load( &size );
+ file.load( &alignedSize );
+}
+
+ } // namespace Segments
+ } // namespace Conatiners
+} // namespace TNL
diff --git a/src/TNL/Containers/Segments/SlicedEllpack.h b/src/TNL/Containers/Segments/SlicedEllpack.h
index ecc2c8c7ef1d8fa24d418372c07c2f769ab75cc9..fc514c51f3edcff69682e35378515c9d90dc8ffc 100644
--- a/src/TNL/Containers/Segments/SlicedEllpack.h
+++ b/src/TNL/Containers/Segments/SlicedEllpack.h
@@ -11,6 +11,7 @@
#pragma once
#include <TNL/Containers/Vector.h>
+#include <TNL/Containers/Segments/SlicedEllpackView.h>
namespace TNL {
namespace Containers {
@@ -18,6 +19,7 @@ namespace TNL {
template< typename Device,
typename Index,
+ typename IndexAllocator = typename Allocators::Default< Device >::template Allocator< Index >,
bool RowMajorOrder = std::is_same< Device, Devices::Host >::value,
int SliceSize = 32 >
class SlicedEllpack
@@ -26,9 +28,11 @@ class SlicedEllpack
using DeviceType = Device;
using IndexType = Index;
- using OffsetsHolder = Containers::Vector< IndexType, DeviceType, IndexType >;
+ using OffsetsHolder = Containers::Vector< IndexType, DeviceType, typename std::remove_const< IndexType >::type, IndexAllocator >;
static constexpr int getSliceSize() { return SliceSize; }
static constexpr bool getRowMajorOrder() { return RowMajorOrder; }
+ using ViewType = SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >;
+ using ConstViewType = SlicedEllpackView< Device, std::add_const_t< Index >, RowMajorOrder, SliceSize >;
SlicedEllpack();
@@ -38,6 +42,10 @@ class SlicedEllpack
SlicedEllpack( const SlicedEllpack&& segments );
+ ViewType getView();
+
+ ConstViewType getConstView() const;
+
/**
* \brief Set sizes of particular segments.
*/
diff --git a/src/TNL/Containers/Segments/SlicedEllpack.hpp b/src/TNL/Containers/Segments/SlicedEllpack.hpp
index d721edb00b2119bce445a8ec6c3929414f5cd262..bdf28ff73b46de6e9423eebcd03849d7ef5cca2b 100644
--- a/src/TNL/Containers/Segments/SlicedEllpack.hpp
+++ b/src/TNL/Containers/Segments/SlicedEllpack.hpp
@@ -22,9 +22,10 @@ namespace TNL {
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
SlicedEllpack()
: size( 0 ), alignedSize( 0 ), segmentsCount( 0 )
{
@@ -32,9 +33,10 @@ SlicedEllpack()
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
SlicedEllpack( const Vector< IndexType, DeviceType, IndexType >& sizes )
: size( 0 ), alignedSize( 0 ), segmentsCount( 0 )
{
@@ -43,9 +45,10 @@ SlicedEllpack( const Vector< IndexType, DeviceType, IndexType >& sizes )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
SlicedEllpack( const SlicedEllpack& slicedEllpack )
: size( slicedEllpack.size ), alignedSize( slicedEllpack.alignedSize ),
segmentsCount( slicedEllpack.segmentsCount ), sliceOffsets( slicedEllpack.sliceOffsets ),
@@ -55,9 +58,10 @@ SlicedEllpack( const SlicedEllpack& slicedEllpack )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
SlicedEllpack( const SlicedEllpack&& slicedEllpack )
: size( slicedEllpack.size ), alignedSize( slicedEllpack.alignedSize ),
segmentsCount( slicedEllpack.segmentsCount ), sliceOffsets( slicedEllpack.sliceOffsets ),
@@ -67,11 +71,36 @@ SlicedEllpack( const SlicedEllpack&& slicedEllpack )
template< typename Device,
typename Index,
+ typename IndexAllocator,
+ bool RowMajorOrder,
+ int SliceSize >
+typename SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::ViewType
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
+getView()
+{
+ return ViewType( size, alignedSize, segmentsCount, sliceOffsets.getView(), sliceSegmentSizes.getView() );
+}
+
+template< typename Device,
+ typename Index,
+ typename IndexAllocator,
+ bool RowMajorOrder,
+ int SliceSize >
+typename SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::ConstViewType
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
+getConstView() const
+{
+ return ConstViewType( size, alignedSize, segmentsCount, sliceOffsets.getConstView(), sliceSegmentSizes.getConstView() );
+}
+
+template< typename Device,
+ typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
template< typename SizesHolder >
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
setSegmentsSizes( const SizesHolder& sizes )
{
this->segmentsCount = sizes.getSize();
@@ -79,7 +108,7 @@ setSegmentsSizes( const SizesHolder& sizes )
this->sliceOffsets.setSize( slicesCount + 1 );
this->sliceOffsets = 0;
this->sliceSegmentSizes.setSize( slicesCount );
- Ellpack< DeviceType, IndexType, true > ellpack;
+ Ellpack< DeviceType, IndexType, IndexAllocator, true > ellpack;
ellpack.setSegmentsSizes( slicesCount, SliceSize );
const IndexType _size = sizes.getSize();
@@ -106,11 +135,12 @@ setSegmentsSizes( const SizesHolder& sizes )
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
__cuda_callable__
Index
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
getSegmentsCount() const
{
return this->segmentsCount;
@@ -118,11 +148,12 @@ getSegmentsCount() const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
__cuda_callable__
Index
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
getSegmentSize( const IndexType segmentIdx ) const
{
const Index sliceIdx = segmentIdx / SliceSize;
@@ -140,11 +171,12 @@ getSegmentSize( const IndexType segmentIdx ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
__cuda_callable__
Index
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
getSize() const
{
return this->size;
@@ -152,11 +184,12 @@ getSize() const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
__cuda_callable__
Index
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
getStorageSize() const
{
return this->alignedSize;
@@ -164,11 +197,12 @@ getStorageSize() const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
__cuda_callable__
Index
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
{
const IndexType sliceIdx = segmentIdx / SliceSize;
@@ -197,22 +231,24 @@ getGlobalIndex( const Index segmentIdx, const Index localIdx ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
__cuda_callable__
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const
{
}
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
template< typename Function, typename... Args >
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
{
const auto sliceSegmentSizes_view = this->sliceSegmentSizes.getConstView();
@@ -251,11 +287,12 @@ forSegments( IndexType first, IndexType last, Function& f, Args... args ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
template< typename Function, typename... Args >
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
forAll( Function& f, Args... args ) const
{
this->forSegments( 0, this->getSegmentsCount(), f, args... );
@@ -263,11 +300,12 @@ forAll( Function& f, Args... args ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, 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() ) );
@@ -307,11 +345,12 @@ segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& red
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
{
this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
@@ -319,10 +358,11 @@ allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Re
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
save( File& file ) const
{
file.save( &size );
@@ -334,10 +374,11 @@ save( File& file ) const
template< typename Device,
typename Index,
+ typename IndexAllocator,
bool RowMajorOrder,
int SliceSize >
void
-SlicedEllpack< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpack< Device, Index, IndexAllocator, RowMajorOrder, SliceSize >::
load( File& file )
{
file.load( &size );
diff --git a/src/TNL/Containers/Segments/SlicedEllpackView.h b/src/TNL/Containers/Segments/SlicedEllpackView.h
new file mode 100644
index 0000000000000000000000000000000000000000..275baacf5f5cdc2f367bee8ece316a4106f47ef8
--- /dev/null
+++ b/src/TNL/Containers/Segments/SlicedEllpackView.h
@@ -0,0 +1,116 @@
+/***************************************************************************
+ SlicedEllpackView.h - description
+ -------------------
+ begin : Dec 12, 2019
+ copyright : (C) 2019 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+#include <type_traits>
+
+#include <TNL/Containers/Vector.h>
+
+namespace TNL {
+ namespace Containers {
+ namespace Segments {
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder = std::is_same< Device, Devices::Host >::value,
+ int SliceSize = 32 >
+class SlicedEllpackView
+{
+ public:
+
+ using DeviceType = Device;
+ using IndexType = Index;
+ using OffsetsView = typename Containers::VectorView< IndexType, DeviceType, typename std::remove_const < IndexType >::type >;
+ static constexpr int getSliceSize() { return SliceSize; }
+ static constexpr bool getRowMajorOrder() { return RowMajorOrder; }
+ using ViewType = SlicedEllpackView;
+ using ConstViewType = SlicedEllpackView< Device, std::add_const_t< Index > >;
+
+ __cuda_callable__
+ SlicedEllpackView();
+
+ __cuda_callable__
+ SlicedEllpackView( IndexType size,
+ IndexType alignedSize,
+ IndexType segmentsCount,
+ OffsetsView&& sliceOffsets,
+ OffsetsView&& sliceSegmentSizes );
+
+ __cuda_callable__
+ SlicedEllpackView( const SlicedEllpackView& slicedEllpackView );
+
+ __cuda_callable__
+ SlicedEllpackView( const SlicedEllpackView&& slicedEllpackView );
+
+ ViewType getView();
+
+ ConstViewType getConstView() const;
+
+ __cuda_callable__
+ IndexType getSegmentsCount() const;
+
+ __cuda_callable__
+ IndexType getSegmentSize( const IndexType segmentIdx ) const;
+
+ /**
+ * \brief Number segments.
+ */
+ __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;
+
+ /***
+ * \brief Go over all segments and for each segment element call
+ * function 'f' with arguments 'args'. The return type of 'f' is bool.
+ * When its true, the for-loop continues. Once 'f' returns false, the for-loop
+ * is terminated.
+ */
+ template< typename Function, typename... Args >
+ void forSegments( IndexType first, IndexType last, Function& f, Args... args ) const;
+
+ template< typename Function, typename... Args >
+ void forAll( Function& f, Args... args ) const;
+
+
+ /***
+ * \brief Go over all segments and perform a reduction in each of them.
+ */
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ void save( File& file ) const;
+
+ void load( File& file );
+
+ protected:
+
+ IndexType size, alignedSize, segmentsCount;
+
+ OffsetsView sliceOffsets, sliceSegmentSizes;
+};
+
+ } // namespace Segements
+ } // namespace Conatiners
+} // namespace TNL
+
+#include <TNL/Containers/Segments/SlicedEllpackView.hpp>
diff --git a/src/TNL/Containers/Segments/SlicedEllpackView.hpp b/src/TNL/Containers/Segments/SlicedEllpackView.hpp
new file mode 100644
index 0000000000000000000000000000000000000000..f2e03bd38aab3de36e32feb929202b2ede2b8c34
--- /dev/null
+++ b/src/TNL/Containers/Segments/SlicedEllpackView.hpp
@@ -0,0 +1,342 @@
+/***************************************************************************
+ SlicedEllpackView.hpp - description
+ -------------------
+ begin : Dec 4, 2019
+ copyright : (C) 2019 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+#include <TNL/Containers/Vector.h>
+#include <TNL/Algorithms/ParallelFor.h>
+#include <TNL/Containers/Segments/SlicedEllpackView.h>
+
+namespace TNL {
+ namespace Containers {
+ namespace Segments {
+
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpackView()
+ : size( 0 ), alignedSize( 0 ), segmentsCount( 0 )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpackView( IndexType size,
+ IndexType alignedSize,
+ IndexType segmentsCount,
+ OffsetsView&& sliceOffsets,
+ OffsetsView&& sliceSegmentSizes )
+ : size( size ), alignedSize( alignedSize ), segmentsCount( segmentsCount ),
+ sliceOffsets( std::forward< OffsetsView >( sliceOffsets ) ), sliceSegmentSizes( std::forward< OffsetsView >( sliceSegmentSizes ) )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpackView( const SlicedEllpackView& slicedEllpackView )
+ : size( slicedEllpackView.size ), alignedSize( slicedEllpackView.alignedSize ),
+ segmentsCount( slicedEllpackView.segmentsCount ), sliceOffsets( slicedEllpackView.sliceOffsets ),
+ sliceSegmentSizes( slicedEllpackView.sliceSegmentSizes )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+SlicedEllpackView( const SlicedEllpackView&& slicedEllpackView )
+ : size( slicedEllpackView.size ), alignedSize( slicedEllpackView.alignedSize ),
+ segmentsCount( slicedEllpackView.segmentsCount ), sliceOffsets( slicedEllpackView.sliceOffsets ),
+ sliceSegmentSizes( slicedEllpackView.sliceSegmentSizes )
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+typename SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::ViewType
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+getView()
+{
+ return ViewType( size, alignedSize, segmentsCount, sliceOffsets, sliceSegmentSizes );
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+typename SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::ConstViewType
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+getConstView() const
+{
+ return ConstViewType( size, alignedSize, segmentsCount, sliceOffsets.getConstView(), sliceSegmentSizes.getConstView() );
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+Index
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+getSegmentsCount() const
+{
+ return this->segmentsCount;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+Index
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+getSegmentSize( const IndexType segmentIdx ) const
+{
+ 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
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+getSize() const
+{
+ return this->size;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+Index
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+getStorageSize() const
+{
+ return this->alignedSize;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+Index
+SlicedEllpackView< 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 sliceOffset + segmentInSliceIdx * segmentSize + localIdx;
+ else
+ return sliceOffset + segmentInSliceIdx + SliceSize * localIdx;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+__cuda_callable__
+void
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+getSegmentAndLocalIndex( const Index globalIdx, Index& segmentIdx, Index& localIdx ) const
+{
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+ template< typename Function, typename... Args >
+void
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+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 ) {
+ 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 ) {
+ 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... );
+ }
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+ template< typename Function, typename... Args >
+void
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+forAll( Function& f, Args... args ) const
+{
+ this->forSegments( 0, this->getSegmentsCount(), f, args... );
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+void
+SlicedEllpackView< 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 )
+ {
+ 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 globalIdx = begin; globalIdx< end; globalIdx++ )
+ reduction( aux, fetch( segmentIdx, globalIdx, 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 );
+ 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... );
+ }
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+void
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const
+{
+ this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+void
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+save( File& file ) const
+{
+ file.save( &size );
+ file.save( &alignedSize );
+ file.save( &segmentsCount );
+ file << this->sliceOffsets;
+ file << this->sliceSegmentSizes;
+}
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int SliceSize >
+void
+SlicedEllpackView< Device, Index, RowMajorOrder, SliceSize >::
+load( File& file )
+{
+ file.load( &size );
+ file.load( &alignedSize );
+ file.load( &segmentsCount );
+ file >> this->sliceOffsets;
+ file >> this->sliceSegmentSizes;
+}
+
+ } // namespace Segments
+ } // namespace Conatiners
+} // namespace TNL
diff --git a/src/TNL/Containers/Segments/details/CSR.h b/src/TNL/Containers/Segments/details/CSR.h
new file mode 100644
index 0000000000000000000000000000000000000000..47e768d289cb307957c117402a4a45ce8cd54c7e
--- /dev/null
+++ b/src/TNL/Containers/Segments/details/CSR.h
@@ -0,0 +1,89 @@
+/***************************************************************************
+ CSR.h - description
+ -------------------
+ begin : Dec 12, 2019
+ copyright : (C) 2019 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+
+namespace TNL {
+ namespace Containers {
+ namespace Segments {
+ namespace details {
+
+template< typename Device,
+ typename Index >
+class CSR
+{
+ public:
+
+ using DeviceType = Device;
+ using IndexType = Index;
+
+ template< typename SizesHolder, typename CSROffsets >
+ static void setSegmentsSizes( const SizesHolder& sizes, CSROffsets& offsets )
+ {
+ offsets.setSize( sizes.getSize() + 1 );
+ auto view = offsets.getView( 0, sizes.getSize() );
+ view = sizes;
+ offsets.setElement( sizes.getSize(), 0 );
+ offsets.template scan< Algorithms::ScanType::Exclusive >();
+ }
+
+ /***
+ * \brief Returns size of the segment number \r segmentIdx
+ */
+ __cuda_callable__
+ IndexType getSegmentSize( const IndexType segmentIdx ) const;
+
+ /***
+ * \brief Returns number of elements managed by all segments.
+ */
+ __cuda_callable__
+ IndexType getSize() const;
+
+ /***
+ * \brief Returns number of elements that needs to be allocated.
+ */
+ __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;
+
+ /***
+ * \brief Go over all segments and for each segment element call
+ * function 'f' with arguments 'args'. The return type of 'f' is bool.
+ * When its true, the for-loop continues. Once 'f' returns false, the for-loop
+ * is terminated.
+ */
+ template< typename Function, typename... Args >
+ void forSegments( IndexType first, IndexType last, Function& f, Args... args ) const;
+
+ template< typename Function, typename... Args >
+ void forAll( Function& f, Args... args ) const;
+
+
+ /***
+ * \brief Go over all segments and perform a reduction in each of them.
+ */
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+};
+ } // namespace details
+ } // namespace Segements
+ } // namespace Conatiners
+} // namespace TNL
+
+#include <TNL/Containers/Segments/CSR.hpp>
diff --git a/src/TNL/Containers/Segments/details/Ellpack.h b/src/TNL/Containers/Segments/details/Ellpack.h
new file mode 100644
index 0000000000000000000000000000000000000000..b08ad0f04f9d316f6e2ce62aea5d8990c1204978
--- /dev/null
+++ b/src/TNL/Containers/Segments/details/Ellpack.h
@@ -0,0 +1,107 @@
+/***************************************************************************
+ Ellpack.h - description
+ -------------------
+ begin : Dec 3, 2019
+ copyright : (C) 2019 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+#include <TNL/Containers/Vector.h>
+
+namespace TNL {
+ namespace Containers {
+ namespace Segments {
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder = std::is_same< Device, Devices::Host >::value,
+ int Alignment = 32 >
+class Ellpack
+{
+ public:
+
+ using DeviceType = Device;
+ using IndexType = Index;
+ static constexpr int getAlignment() { return Alignment; }
+ static constexpr bool getRowMajorOrder() { return RowMajorOrder; }
+ using OffsetsHolder = Containers::Vector< IndexType, DeviceType, IndexType >;
+ using SegmentsSizes = OffsetsHolder;
+
+ Ellpack();
+
+ Ellpack( const SegmentsSizes& sizes );
+
+ Ellpack( const IndexType segmentsCount, const IndexType segmentSize );
+
+ Ellpack( const Ellpack& segments );
+
+ Ellpack( const Ellpack&& segments );
+
+ /**
+ * \brief Set sizes of particular segments.
+ */
+ template< typename SizesHolder = OffsetsHolder >
+ void setSegmentsSizes( const SizesHolder& sizes );
+
+ void setSegmentsSizes( const IndexType segmentsCount, const IndexType segmentSize );
+ /**
+ * \brief Number segments.
+ */
+ __cuda_callable__
+ IndexType getSegmentsCount() const;
+
+ __cuda_callable__
+ IndexType getSegmentSize( const IndexType segmentIdx ) const;
+
+ __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;
+
+ /***
+ * \brief Go over all segments and for each segment element call
+ * function 'f' with arguments 'args'. The return type of 'f' is bool.
+ * When its true, the for-loop continues. Once 'f' returns false, the for-loop
+ * is terminated.
+ */
+ template< typename Function, typename... Args >
+ void forSegments( IndexType first, IndexType last, Function& f, Args... args ) const;
+
+ template< typename Function, typename... Args >
+ void forAll( Function& f, Args... args ) const;
+
+
+ /***
+ * \brief Go over all segments and perform a reduction in each of them.
+ */
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ void save( File& file ) const;
+
+ void load( File& file );
+
+ protected:
+
+ IndexType segmentSize, size, alignedSize;
+};
+
+ } // namespace Segements
+ } // namespace Conatiners
+} // namespace TNL
+
+#include <TNL/Containers/Segments/Ellpack.hpp>
diff --git a/src/TNL/Containers/Segments/details/SlicedEllpack.h b/src/TNL/Containers/Segments/details/SlicedEllpack.h
new file mode 100644
index 0000000000000000000000000000000000000000..ecc2c8c7ef1d8fa24d418372c07c2f769ab75cc9
--- /dev/null
+++ b/src/TNL/Containers/Segments/details/SlicedEllpack.h
@@ -0,0 +1,106 @@
+/***************************************************************************
+ SlicedEllpack.h - description
+ -------------------
+ begin : Dec 4, 2019
+ copyright : (C) 2019 by Tomas Oberhuber
+ email : tomas.oberhuber@fjfi.cvut.cz
+ ***************************************************************************/
+
+/* See Copyright Notice in tnl/Copyright */
+
+#pragma once
+
+#include <TNL/Containers/Vector.h>
+
+namespace TNL {
+ namespace Containers {
+ namespace Segments {
+
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder = std::is_same< Device, Devices::Host >::value,
+ int SliceSize = 32 >
+class SlicedEllpack
+{
+ public:
+
+ using DeviceType = Device;
+ using IndexType = Index;
+ using OffsetsHolder = Containers::Vector< IndexType, DeviceType, IndexType >;
+ static constexpr int getSliceSize() { return SliceSize; }
+ static constexpr bool getRowMajorOrder() { return RowMajorOrder; }
+
+ SlicedEllpack();
+
+ SlicedEllpack( const Vector< IndexType, DeviceType, IndexType >& sizes );
+
+ SlicedEllpack( const SlicedEllpack& segments );
+
+ SlicedEllpack( const SlicedEllpack&& segments );
+
+ /**
+ * \brief Set sizes of particular segments.
+ */
+ template< typename SizesHolder = OffsetsHolder >
+ void setSegmentsSizes( const SizesHolder& sizes );
+
+ __cuda_callable__
+ IndexType getSegmentsCount() const;
+
+ __cuda_callable__
+ IndexType getSegmentSize( const IndexType segmentIdx ) const;
+
+ /**
+ * \brief Number segments.
+ */
+ __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;
+
+ /***
+ * \brief Go over all segments and for each segment element call
+ * function 'f' with arguments 'args'. The return type of 'f' is bool.
+ * When its true, the for-loop continues. Once 'f' returns false, the for-loop
+ * is terminated.
+ */
+ template< typename Function, typename... Args >
+ void forSegments( IndexType first, IndexType last, Function& f, Args... args ) const;
+
+ template< typename Function, typename... Args >
+ void forAll( Function& f, Args... args ) const;
+
+
+ /***
+ * \brief Go over all segments and perform a reduction in each of them.
+ */
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void segmentsReduction( IndexType first, IndexType last, Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ template< typename Fetch, typename Reduction, typename ResultKeeper, typename Real, typename... Args >
+ void allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Real& zero, Args... args ) const;
+
+ void save( File& file ) const;
+
+ void load( File& file );
+
+ protected:
+
+ IndexType size, alignedSize, segmentsCount;
+
+ OffsetsHolder sliceOffsets, sliceSegmentSizes;
+};
+
+ } // namespace Segements
+ } // namespace Conatiners
+} // namespace TNL
+
+#include <TNL/Containers/Segments/SlicedEllpack.hpp>
diff --git a/src/TNL/Matrices/SparseMatrix.h b/src/TNL/Matrices/SparseMatrix.h
index b6a618e105f30ce43b276eea26ff095fabd1a30d..1512f8574d1e4a29534b61a63f1ebed40c25af82 100644
--- a/src/TNL/Matrices/SparseMatrix.h
+++ b/src/TNL/Matrices/SparseMatrix.h
@@ -17,7 +17,7 @@ namespace TNL {
namespace Matrices {
template< typename Real,
- template< typename Device_, typename Index_ > class Segments,
+ template< typename Device_, typename Index_, typename IndexAllocator_ > class Segments,
typename Device = Devices::Host,
typename Index = int,
typename RealAllocator = typename Allocators::Default< Device >::template Allocator< Real >,
@@ -27,9 +27,9 @@ class SparseMatrix : public Matrix< Real, Device, Index, RealAllocator >
public:
using RealType = Real;
- template< typename Device_, typename Index_ >
- using SegmentsTemplate = Segments< Device_, Index_ >;
- using SegmentsType = Segments< Device, Index >;
+ template< typename Device_, typename Index_, typename IndexAllocator_ >
+ using SegmentsTemplate = Segments< Device_, Index_, IndexAllocator_ >;
+ using SegmentsType = Segments< Device, Index, IndexAllocator >;
using DeviceType = Device;
using IndexType = Index;
using RealAllocatorType = RealAllocator;
@@ -77,7 +77,7 @@ class SparseMatrix : public Matrix< Real, Device, Index, RealAllocator >
__cuda_callable__
IndexType getNonZeroRowLengthFast( const IndexType row ) const;
- template< typename Real2, template< typename, typename > class Segments2, typename Device2, typename Index2, typename RealAllocator2, typename IndexAllocator2 >
+ template< typename Real2, template< typename, typename, typename > class Segments2, typename Device2, typename Index2, typename RealAllocator2, typename IndexAllocator2 >
void setLike( const SparseMatrix< Real2, Segments2, Device2, Index2, RealAllocator2, IndexAllocator2 >& matrix );
IndexType getNumberOfNonzeroMatrixElements() const;
@@ -191,7 +191,7 @@ class SparseMatrix : public Matrix< Real, Device, Index, RealAllocator >
// cross-device copy assignment
template< typename Real2,
- template< typename, typename > class Segments2,
+ template< typename, typename, typename > class Segments2,
typename Device2,
typename Index2,
typename RealAllocator2,
diff --git a/src/TNL/Matrices/SparseMatrix.hpp b/src/TNL/Matrices/SparseMatrix.hpp
index 0d9ee0b0607e8579b3e89a03dd1c6616bf9b65f6..e24ed2f448b8b9dc8fe7a78f7f0b8af0431fbf77 100644
--- a/src/TNL/Matrices/SparseMatrix.hpp
+++ b/src/TNL/Matrices/SparseMatrix.hpp
@@ -18,7 +18,7 @@ namespace TNL {
namespace Matrices {
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -31,7 +31,7 @@ SparseMatrix( const RealAllocatorType& realAllocator,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -43,7 +43,7 @@ SparseMatrix( const SparseMatrix& m )
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -55,7 +55,7 @@ SparseMatrix( const SparseMatrix&& m )
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -70,7 +70,7 @@ SparseMatrix( const IndexType rows,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -86,7 +86,7 @@ getSerializationType()
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -99,7 +99,7 @@ getSerializationTypeVirtual() const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -117,7 +117,7 @@ setCompressedRowLengths( ConstCompressedRowLengthsVectorView rowLengths )
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -143,7 +143,7 @@ getCompressedRowLengths( Vector& rowLengths ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -156,7 +156,7 @@ getRowLength( const IndexType row ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -170,7 +170,7 @@ getRowLengthFast( const IndexType row ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -183,7 +183,7 @@ getNonZeroRowLength( const IndexType row ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -197,12 +197,12 @@ getNonZeroRowLengthFast( const IndexType row ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
typename IndexAllocator >
- template< typename Real2, template< typename, typename > class Segments2, typename Device2, typename Index2, typename RealAllocator2, typename IndexAllocator2 >
+ template< typename Real2, template< typename, typename, typename > class Segments2, typename Device2, typename Index2, typename RealAllocator2, typename IndexAllocator2 >
void
SparseMatrix< Real, Segments, Device, Index, RealAllocator, IndexAllocator >::
setLike( const SparseMatrix< Real2, Segments2, Device2, Index2, RealAllocator2, IndexAllocator2 >& matrix )
@@ -211,7 +211,7 @@ setLike( const SparseMatrix< Real2, Segments2, Device2, Index2, RealAllocator2,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -229,7 +229,7 @@ getNumberOfNonzeroMatrixElements() const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -244,7 +244,7 @@ reset()
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -260,7 +260,7 @@ setElementFast( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -275,7 +275,7 @@ setElement( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -292,7 +292,7 @@ addElementFast( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -358,7 +358,7 @@ addElement( const IndexType row,
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -374,7 +374,7 @@ setRowFast( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -403,7 +403,7 @@ setRow( const IndexType row,
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -421,7 +421,7 @@ addRowFast( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -439,7 +439,7 @@ addRow( const IndexType row,
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -454,7 +454,7 @@ getElementFast( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -477,7 +477,7 @@ getElement( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -493,7 +493,7 @@ getRowFast( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -509,7 +509,7 @@ rowVectorProduct( const IndexType row,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -544,7 +544,7 @@ vectorProduct( const InVector& inVector,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -567,7 +567,7 @@ rowsReduction( IndexType first, IndexType last, Fetch& fetch, Reduce& reduce, Ke
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -581,7 +581,7 @@ allRowsReduction( Fetch& fetch, Reduce& reduce, Keep& keep, const FetchReal& zer
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -605,7 +605,7 @@ forRows( IndexType first, IndexType last, Function& function ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -619,7 +619,7 @@ forAllRows( Function& function ) const
}
/*template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -635,7 +635,7 @@ addMatrix( const SparseMatrix< Real2, Segments2, Device, Index2, RealAllocator2,
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -650,7 +650,7 @@ getTransposition( const SparseMatrix< Real2, Device, Index2 >& matrix,
}*/
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -668,7 +668,7 @@ performSORIteration( const Vector1& b,
// copy assignment
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -686,13 +686,13 @@ operator=( const SparseMatrix& matrix )
// cross-device copy assignment
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
typename IndexAllocator >
template< typename Real2,
- template< typename, typename > class Segments2,
+ template< typename, typename, typename > class Segments2,
typename Device2,
typename Index2,
typename RealAllocator2,
@@ -729,7 +729,7 @@ operator=( const SparseMatrix< Real2, Segments2, Device2, Index2, RealAllocator2
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -744,7 +744,7 @@ save( File& file ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -759,7 +759,7 @@ load( File& file )
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -772,7 +772,7 @@ save( const String& fileName ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -785,7 +785,7 @@ load( const String& fileName )
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
@@ -811,7 +811,7 @@ print( std::ostream& str ) const
}
template< typename Real,
- template< typename, typename > class Segments,
+ template< typename, typename, typename > class Segments,
typename Device,
typename Index,
typename RealAllocator,
diff --git a/src/UnitTests/Containers/Segments/SegmentsTest.hpp b/src/UnitTests/Containers/Segments/SegmentsTest.hpp
index acc75655f7003f80d920f64ceb710844ccdc376b..5e74f96b039b23102908c772f576b9592c1bb079 100644
--- a/src/UnitTests/Containers/Segments/SegmentsTest.hpp
+++ b/src/UnitTests/Containers/Segments/SegmentsTest.hpp
@@ -52,6 +52,16 @@ void test_SetSegmentsSizes_EqualSizes()
for( IndexType i = 0; i < segmentsCount; i++ )
EXPECT_EQ( segments3.getSegmentSize( i ), segmentSize );
+
+ using SegmentsView = typename Segments::ViewType;
+
+ SegmentsView segmentsView = segments.getView();
+ EXPECT_EQ( segmentsView.getSegmentsCount(), segmentsCount );
+ EXPECT_EQ( segmentsView.getSize(), segmentsCount * segmentSize );
+ EXPECT_LE( segmentsView.getSize(), segments.getStorageSize() );
+
+ for( IndexType i = 0; i < segmentsCount; i++ )
+ EXPECT_EQ( segmentsView.getSegmentSize( i ), segmentSize );
}
template< typename Segments >
@@ -89,6 +99,16 @@ void test_SetSegmentsSizes_EqualSizes_EllpackOnly()
for( IndexType i = 0; i < segmentsCount; i++ )
EXPECT_EQ( segments3.getSegmentSize( i ), segmentSize );
+
+ using SegmentsView = typename Segments::ViewType;
+
+ SegmentsView segmentsView = segments.getView();
+ EXPECT_EQ( segmentsView.getSegmentsCount(), segmentsCount );
+ EXPECT_EQ( segmentsView.getSize(), segmentsCount * segmentSize );
+ EXPECT_LE( segmentsView.getSize(), segments.getStorageSize() );
+
+ for( IndexType i = 0; i < segmentsCount; i++ )
+ EXPECT_EQ( segmentsView.getSegmentSize( i ), segmentSize );
}
template< typename Segments >
@@ -136,6 +156,11 @@ void test_AllReduction_MaximumInSegments()
for( IndexType i = 0; i < segmentsCount; i++ )
EXPECT_EQ( result.getElement( i ), ( i + 1 ) * segmentSize );
+
+ result_view = 0;
+ segments.getView().allReduction( fetch, reduce, keep, std::numeric_limits< IndexType >::min() );
+ for( IndexType i = 0; i < segmentsCount; i++ )
+ EXPECT_EQ( result.getElement( i ), ( i + 1 ) * segmentSize );
}
#endif
diff --git a/src/UnitTests/Matrices/SparseMatrixTest_Ellpack_segments.h b/src/UnitTests/Matrices/SparseMatrixTest_Ellpack_segments.h
index 2c0514c0aba69e5e26abdf1434b7a2d333c25b63..16c22d9cad957dd62cfaec4fe35303f41a5c002b 100644
--- a/src/UnitTests/Matrices/SparseMatrixTest_Ellpack_segments.h
+++ b/src/UnitTests/Matrices/SparseMatrixTest_Ellpack_segments.h
@@ -28,14 +28,14 @@ protected:
////
// Row-major format is used for the host system
-template< typename Device, typename Index >
-using RowMajorEllpack = TNL::Containers::Segments::Ellpack< Device, Index, true, 32 >;
+template< typename Device, typename Index, typename IndexAlocator >
+using RowMajorEllpack = TNL::Containers::Segments::Ellpack< Device, Index, IndexAlocator, true, 32 >;
////
// Column-major format is used for GPUs
-template< typename Device, typename Index >
-using ColumnMajorEllpack = TNL::Containers::Segments::Ellpack< Device, Index, false, 32 >;
+template< typename Device, typename Index, typename IndexAllocator >
+using ColumnMajorEllpack = TNL::Containers::Segments::Ellpack< Device, Index, IndexAllocator, false, 32 >;
// types for which MatrixTest is instantiated
using EllpackMatrixTypes = ::testing::Types
diff --git a/src/UnitTests/Matrices/SparseMatrixTest_SlicedEllpack_segments.h b/src/UnitTests/Matrices/SparseMatrixTest_SlicedEllpack_segments.h
index 5efcb1eae0a01e9d4267c8e07cdcd396ecb1015e..8597121e463a1d8349c1d8f93a9dd71a88f576fd 100644
--- a/src/UnitTests/Matrices/SparseMatrixTest_SlicedEllpack_segments.h
+++ b/src/UnitTests/Matrices/SparseMatrixTest_SlicedEllpack_segments.h
@@ -28,14 +28,14 @@ protected:
////
// Row-major format is used for the host system
-template< typename Device, typename Index >
-using RowMajorSlicedEllpack = TNL::Containers::Segments::SlicedEllpack< Device, Index, true, 32 >;
+template< typename Device, typename Index, typename IndexAllocator >
+using RowMajorSlicedEllpack = TNL::Containers::Segments::SlicedEllpack< Device, Index, IndexAllocator, true, 32 >;
////
// Column-major format is used for GPUs
-template< typename Device, typename Index >
-using ColumnMajorSlicedEllpack = TNL::Containers::Segments::SlicedEllpack< Device, Index, false, 32 >;
+template< typename Device, typename Index, typename IndexAllocator >
+using ColumnMajorSlicedEllpack = TNL::Containers::Segments::SlicedEllpack< Device, Index, IndexAllocator, false, 32 >;
// types for which MatrixTest is instantiated
using SlicedEllpackMatrixTypes = ::testing::Types