From fcb368b75abc9fbce5a0e4cf8e31c18935fc1fdb Mon Sep 17 00:00:00 2001
From: Tomas Oberhuber <tomas.oberhuber@fjfi.cvut.cz>
Date: Wed, 15 Jan 2020 18:05:53 +0100
Subject: [PATCH] Forwarding dense matrix methods calls to DenseMatrixView.
---
src/TNL/Containers/Segments/EllpackView.h | 2 +
src/TNL/Containers/Segments/EllpackView.hpp | 13 ++
src/TNL/Matrices/Dense.h | 4 +-
src/TNL/Matrices/Dense.hpp | 136 +++++---------------
src/TNL/Matrices/DenseMatrixView.h | 10 --
src/TNL/Matrices/DenseMatrixView.hpp | 77 ++++-------
src/UnitTests/Matrices/DenseMatrixTest.h | 4 +-
7 files changed, 71 insertions(+), 175 deletions(-)
diff --git a/src/TNL/Containers/Segments/EllpackView.h b/src/TNL/Containers/Segments/EllpackView.h
index 737810498a..f64b04068e 100644
--- a/src/TNL/Containers/Segments/EllpackView.h
+++ b/src/TNL/Containers/Segments/EllpackView.h
@@ -104,6 +104,8 @@ class EllpackView
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;
+ EllpackView& operator=( const EllpackView& view );
+
void save( File& file ) const;
void load( File& file );
diff --git a/src/TNL/Containers/Segments/EllpackView.hpp b/src/TNL/Containers/Segments/EllpackView.hpp
index 21be886541..719a349a98 100644
--- a/src/TNL/Containers/Segments/EllpackView.hpp
+++ b/src/TNL/Containers/Segments/EllpackView.hpp
@@ -291,6 +291,19 @@ allReduction( Fetch& fetch, Reduction& reduction, ResultKeeper& keeper, const Re
this->segmentsReduction( 0, this->getSegmentsCount(), fetch, reduction, keeper, zero, args... );
}
+template< typename Device,
+ typename Index,
+ bool RowMajorOrder,
+ int Alignment >
+EllpackView< Device, Index, RowMajorOrder, Alignment >&
+EllpackView< Device, Index, RowMajorOrder, Alignment >::
+operator=( const EllpackView< Device, Index, RowMajorOrder, Alignment >& view )
+{
+ this->segmentSize = view.segmentSize;
+ this->size = view.size;
+ this->alignedSize = view.alignedSize;
+}
+
template< typename Device,
typename Index,
bool RowMajorOrder,
diff --git a/src/TNL/Matrices/Dense.h b/src/TNL/Matrices/Dense.h
index 485c947010..2e71316e94 100644
--- a/src/TNL/Matrices/Dense.h
+++ b/src/TNL/Matrices/Dense.h
@@ -168,7 +168,7 @@ class Dense : public Matrix< Real, Device, Index >
const RealType& omega = 1.0 ) const;
// copy assignment
- //Dense& operator=( const Dense& matrix );
+ Dense& operator=( const Dense& matrix );
// cross-device copy assignment
template< typename Real_, typename Device_, typename Index_, bool RowMajorOrder_, typename RealAlocator_ >
@@ -200,6 +200,8 @@ class Dense : public Matrix< Real, Device, Index >
friend class DenseDeviceDependentCode< DeviceType >;
SegmentsType segments;
+
+ ViewType view;
};
} // namespace Matrices
diff --git a/src/TNL/Matrices/Dense.hpp b/src/TNL/Matrices/Dense.hpp
index 7e6f379486..b8fa969074 100644
--- a/src/TNL/Matrices/Dense.hpp
+++ b/src/TNL/Matrices/Dense.hpp
@@ -76,10 +76,7 @@ String
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
getSerializationType()
{
- return String( "Matrices::Dense< " ) +
- TNL::getSerializationType< RealType >() + ", [any_device], " +
- TNL::getSerializationType< IndexType >() + ", " +
- ( RowMajorOrder ? "true" : "false" ) + ", [any_allocator] >";
+ return ViewType::getSerializationType();
}
template< typename Real,
@@ -108,6 +105,7 @@ setDimensions( const IndexType rows,
this->segments.setSegmentsSizes( rows, columns );
this->values.setSize( rows * columns );
this->values = 0.0;
+ this->view = this->getView();
}
template< typename Real,
@@ -145,19 +143,7 @@ void
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
getCompressedRowLengths( Vector& rowLengths ) const
{
- rowLengths.setSize( this->getRows() );
- rowLengths = 0;
- auto rowLengths_view = rowLengths.getView();
- auto fetch = [] __cuda_callable__ ( IndexType row, IndexType column, const RealType& value ) -> IndexType {
- return ( value != 0.0 );
- };
- auto reduce = [] __cuda_callable__ ( IndexType& aux, const IndexType a ) {
- aux += a;
- };
- auto keep = [=] __cuda_callable__ ( const IndexType rowIdx, const IndexType value ) mutable {
- rowLengths_view[ rowIdx ] = value;
- };
- this->allRowsReduction( fetch, reduce, keep, 0 );
+ this->view.getCompressedRowLengths( rowLengths );
}
template< typename Real,
@@ -197,11 +183,7 @@ template< typename Real,
typename RealAllocator >
Index Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::getNumberOfNonzeroMatrixElements() const
{
- const auto values_view = this->values.getConstView();
- auto fetch = [=] __cuda_callable__ ( const IndexType i ) -> IndexType {
- return ( values_view[ i ] != 0.0 );
- };
- return Algorithms::Reduction< DeviceType >::reduce( this->values.getSize(), std::plus<>{}, fetch, 0 );
+ return this->view.getNumberOfNonzeroMatrixElements();
}
template< typename Real,
@@ -221,7 +203,7 @@ template< typename Real,
typename RealAllocator >
void Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::setValue( const Real& value )
{
- this->values = value;
+ this->view.setValue( value );
}
template< typename Real,
@@ -233,8 +215,7 @@ __cuda_callable__ auto
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
getRow( const IndexType& rowIdx ) const -> const RowView
{
- TNL_ASSERT_LT( rowIdx, this->getRows(), "Row index is larger than number of matrix rows." );
- return RowView( this->segments.getSegmentView( rowIdx ), this->values.getView() );
+ return this->view.getRow( rowIdx );
}
template< typename Real,
@@ -246,8 +227,7 @@ __cuda_callable__ auto
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
getRow( const IndexType& rowIdx ) -> RowView
{
- TNL_ASSERT_LT( rowIdx, this->getRows(), "Row index is larger than number of matrix rows." );
- return RowView( this->segments.getSegmentView( rowIdx ), this->values.getView() );
+ return this->view.getRow( rowIdx );
}
template< typename Real,
@@ -259,12 +239,7 @@ __cuda_callable__
Real& Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::operator()( const IndexType row,
const IndexType column )
{
- TNL_ASSERT_GE( row, 0, "Row index must be non-negative." );
- TNL_ASSERT_LT( row, this->getRows(), "Row index is out of bounds." );
- TNL_ASSERT_GE( column, 0, "Column index must be non-negative." );
- TNL_ASSERT_LT( column, this->getColumns(), "Column index is out of bounds." );
-
- return this->values.operator[]( this->getElementIndex( row, column ) );
+ return this->view.operator()( row, column );
}
template< typename Real,
@@ -276,12 +251,7 @@ __cuda_callable__
const Real& Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::operator()( const IndexType row,
const IndexType column ) const
{
- TNL_ASSERT_GE( row, 0, "Row index must be non-negative." );
- TNL_ASSERT_LT( row, this->getRows(), "Row index is out of bounds." );
- TNL_ASSERT_GE( column, 0, "Column index must be non-negative." );
- TNL_ASSERT_LT( column, this->getColumns(), "Column index is out of bounds." );
-
- return this->values.operator[]( this->getElementIndex( row, column ) );
+ return this->view.operator()( row, column );
}
template< typename Real,
@@ -295,7 +265,7 @@ setElement( const IndexType row,
const IndexType column,
const RealType& value )
{
- this->values.setElement( this->getElementIndex( row, column ), value );
+ this->view.setElement( row, column, value );
}
template< typename Real,
@@ -310,13 +280,7 @@ addElement( const IndexType row,
const RealType& value,
const RealType& thisElementMultiplicator )
{
- const IndexType elementIndex = this->getElementIndex( row, column );
- if( thisElementMultiplicator == 1.0 )
- this->values.setElement( elementIndex,
- this->values.getElement( elementIndex ) + value );
- else
- this->values.setElement( elementIndex,
- thisElementMultiplicator * this->values.getElement( elementIndex ) + value );
+ this->view.addElement( row, column, value, thisElementMultiplicator );
}
template< typename Real,
@@ -329,7 +293,7 @@ Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
getElement( const IndexType row,
const IndexType column ) const
{
- return this->values.getElement( this->getElementIndex( row, column ) );
+ return this->view.getElement( row, column );
}
template< typename Real,
@@ -342,12 +306,7 @@ void
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
rowsReduction( IndexType first, IndexType last, Fetch& fetch, Reduce& reduce, Keep& keep, const FetchValue& zero ) const
{
- const auto values_view = this->values.getConstView();
- auto fetch_ = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx, bool& compute ) mutable -> decltype( fetch( IndexType(), IndexType(), RealType() ) ) {
- return fetch( rowIdx, columnIdx, values_view[ globalIdx ] );
- return zero;
- };
- this->segments.segmentsReduction( first, last, fetch_, reduce, keep, zero );
+ this->view.rowsReduction( first, last, fetch, reduce, keep, zero );
}
template< typename Real,
@@ -373,12 +332,7 @@ void
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
forRows( IndexType first, IndexType last, Function& function ) const
{
- const auto values_view = this->values.getConstView();
- auto f = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx ) mutable -> bool {
- function( rowIdx, columnIdx, globalIdx, values_view[ globalIdx ] );
- return true;
- };
- this->segments.forSegments( first, last, f );
+ this->view.forRows( first, last, function );
}
template< typename Real,
@@ -391,12 +345,7 @@ void
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
forRows( IndexType first, IndexType last, Function& function )
{
- auto values_view = this->values.getView();
- auto f = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx ) mutable -> bool {
- function( rowIdx, columnIdx, globalIdx, values_view[ globalIdx ] );
- return true;
- };
- this->segments.forSegments( first, last, f );
+ this->view.forRows( first, last, function );
}
template< typename Real,
@@ -435,11 +384,7 @@ __cuda_callable__
typename Vector::RealType Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::rowVectorProduct( const IndexType row,
const Vector& vector ) const
{
- RealType sum( 0.0 );
- // TODO: Fix this
- //for( IndexType column = 0; column < this->getColumns(); column++ )
- // sum += this->getElementFast( row, column ) * vector[ column ];
- return sum;
+ return this->view.rowVectorProduct( row, vector );
}
template< typename Real,
@@ -453,27 +398,7 @@ void
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
vectorProduct( const InVector& inVector, OutVector& outVector ) const
{
- TNL_ASSERT( this->getColumns() == inVector.getSize(),
- std::cerr << "Matrix columns: " << this->getColumns() << std::endl
- << "Vector size: " << inVector.getSize() << std::endl );
- TNL_ASSERT( this->getRows() == outVector.getSize(),
- std::cerr << "Matrix rows: " << this->getRows() << std::endl
- << "Vector size: " << outVector.getSize() << std::endl );
-
- //DeviceDependentCode::vectorProduct( *this, inVector, outVector );
- const auto inVectorView = inVector.getConstView();
- auto outVectorView = outVector.getView();
- const auto valuesView = this->values.getConstView();
- auto fetch = [=] __cuda_callable__ ( IndexType row, IndexType column, IndexType offset, bool& compute ) -> RealType {
- return valuesView[ offset ] * inVectorView[ column ];
- };
- auto reduction = [] __cuda_callable__ ( RealType& sum, const RealType& value ) {
- sum += value;
- };
- auto keeper = [=] __cuda_callable__ ( IndexType row, const RealType& value ) mutable {
- outVectorView[ row ] = value;
- };
- this->segments.segmentsReduction( 0, this->getRows(), fetch, reduction, keeper, ( RealType ) 0.0 );
+ this->view.vectorProduct( inVector, outVector );
}
template< typename Real,
@@ -959,7 +884,7 @@ void Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::performSORItera
x[ row ] = ( 1.0 - omega ) * x[ row ] + omega / diagonalValue * ( b[ row ] - sum );
}
-/*template< typename Real,
+template< typename Real,
typename Device,
typename Index,
bool RowMajorOrder,
@@ -968,7 +893,9 @@ Dense< Real, Device, Index, RowMajorOrder, RealAllocator >&
Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
operator=( const Dense< Real, Device, Index, RowMajorOrder, RealAllocator >& matrix )
{
- const IndexType bufferRowsCount( 128 );
+ setLike( matrix );
+ this->values = matrix.values;
+ /*const IndexType bufferRowsCount( 128 );
const IndexType columns = this->getColumns();
const size_t bufferSize = bufferRowsCount * columns;
Containers::Vector< RealType, Device, IndexType, RealAllocatorType > sourceValuesBuffer( bufferSize );
@@ -1000,8 +927,8 @@ operator=( const Dense< Real, Device, Index, RowMajorOrder, RealAllocator >& mat
this->forRows( baseRow, lastRow, f2 );
baseRow += bufferRowsCount;
}
- return *this;
-}*/
+ return *this;*/
+}
template< typename Real,
typename Device,
@@ -1101,7 +1028,7 @@ template< typename Real,
typename RealAllocator >
void Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::save( const String& fileName ) const
{
- Object::save( fileName );
+ this->view.save( fileName );
}
template< typename Real,
@@ -1121,7 +1048,7 @@ template< typename Real,
typename RealAllocator >
void Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::save( File& file ) const
{
- Matrix< Real, Device, Index >::save( file );
+ this->view.save( file );
}
template< typename Real,
@@ -1141,13 +1068,7 @@ template< typename Real,
typename RealAllocator >
void Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::print( std::ostream& str ) const
{
- for( IndexType row = 0; row < this->getRows(); row++ )
- {
- str <<"Row: " << row << " -> ";
- for( IndexType column = 0; column < this->getColumns(); column++ )
- str << " Col:" << column << "->" << this->getElement( row, column ) << "\t";
- str << std::endl;
- }
+ this->view.print( str );
}
template< typename Real,
@@ -1156,8 +1077,9 @@ template< typename Real,
bool RowMajorOrder,
typename RealAllocator >
__cuda_callable__
-Index Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::getElementIndex( const IndexType row,
- const IndexType column ) const
+Index
+Dense< Real, Device, Index, RowMajorOrder, RealAllocator >::
+getElementIndex( const IndexType row, const IndexType column ) const
{
return this->segments.getGlobalIndex( row, column );
}
diff --git a/src/TNL/Matrices/DenseMatrixView.h b/src/TNL/Matrices/DenseMatrixView.h
index 9bad424b2a..d963dd7c3e 100644
--- a/src/TNL/Matrices/DenseMatrixView.h
+++ b/src/TNL/Matrices/DenseMatrixView.h
@@ -171,22 +171,12 @@ class DenseMatrixView : public MatrixView< Real, Device, Index >
Vector2& x,
const RealType& omega = 1.0 ) const;
- // copy assignment
DenseMatrixView& operator=( const DenseMatrixView& matrix );
- // cross-device copy assignment
- template< typename Real2, typename Device2, typename Index2,
- typename = typename Enabler< Device2 >::type >
- DenseMatrixView& operator=( const DenseMatrixView< Real2, Device2, Index2 >& matrix );
-
void save( const String& fileName ) const;
- void load( const String& fileName );
-
void save( File& file ) const;
- void load( File& file );
-
void print( std::ostream& str ) const;
protected:
diff --git a/src/TNL/Matrices/DenseMatrixView.hpp b/src/TNL/Matrices/DenseMatrixView.hpp
index 21f6d79ef5..527915d55f 100644
--- a/src/TNL/Matrices/DenseMatrixView.hpp
+++ b/src/TNL/Matrices/DenseMatrixView.hpp
@@ -83,7 +83,7 @@ getSerializationType()
{
return String( "Matrices::Dense< " ) +
TNL::getSerializationType< RealType >() + ", [any_device], " +
- TNL::getSerializationType< IndexType >() +
+ TNL::getSerializationType< IndexType >() + ", " +
( RowMajorOrder ? "true" : "false" ) + ", [any_allocator] >";
}
@@ -321,7 +321,7 @@ forRows( IndexType first, IndexType last, Function& function ) const
{
const auto values_view = this->values.getConstView();
auto f = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx ) mutable -> bool {
- function( rowIdx, columnIdx, values_view[ globalIdx ] );
+ function( rowIdx, columnIdx, columnIdx, values_view[ globalIdx ] );
return true;
};
this->segments.forSegments( first, last, f );
@@ -339,7 +339,7 @@ forRows( IndexType first, IndexType last, Function& function )
{
auto values_view = this->values.getView();
auto f = [=] __cuda_callable__ ( IndexType rowIdx, IndexType columnIdx, IndexType globalIdx ) mutable -> bool {
- function( rowIdx, columnIdx, values_view[ globalIdx ] );
+ function( rowIdx, columnIdx, globalIdx, values_view[ globalIdx ] );
return true;
};
this->segments.forSegments( first, last, f );
@@ -395,14 +395,22 @@ template< typename Real,
void DenseMatrixView< Real, Device, Index, RowMajorOrder >::vectorProduct( const InVector& inVector,
OutVector& outVector ) const
{
- TNL_ASSERT( this->getColumns() == inVector.getSize(),
- std::cerr << "Matrix columns: " << this->getColumns() << std::endl
- << "Vector size: " << inVector.getSize() << std::endl );
- TNL_ASSERT( this->getRows() == outVector.getSize(),
- std::cerr << "Matrix rows: " << this->getRows() << std::endl
- << "Vector size: " << outVector.getSize() << std::endl );
-
- //DeviceDependentCode::vectorProduct( *this, inVector, outVector );
+ TNL_ASSERT_EQ( this->getColumns(), inVector.getSize(), "Matrix columns count differs with input vector size." );
+ TNL_ASSERT_EQ( this->getRows(), outVector.getSize(), "Matrix rows count differs with output vector size." );
+
+ const auto inVectorView = inVector.getConstView();
+ auto outVectorView = outVector.getView();
+ const auto valuesView = this->values.getConstView();
+ auto fetch = [=] __cuda_callable__ ( IndexType row, IndexType column, IndexType offset, bool& compute ) -> RealType {
+ return valuesView[ offset ] * inVectorView[ column ];
+ };
+ auto reduction = [] __cuda_callable__ ( RealType& sum, const RealType& value ) {
+ sum += value;
+ };
+ auto keeper = [=] __cuda_callable__ ( IndexType row, const RealType& value ) mutable {
+ outVectorView[ row ] = value;
+ };
+ this->segments.segmentsReduction( 0, this->getRows(), fetch, reduction, keeper, ( RealType ) 0.0 );
}
template< typename Real,
@@ -885,39 +893,18 @@ void DenseMatrixView< Real, Device, Index, RowMajorOrder >::performSORIteration(
}
-// copy assignment
template< typename Real,
typename Device,
typename Index,
bool RowMajorOrder >
DenseMatrixView< Real, Device, Index, RowMajorOrder >&
-DenseMatrixView< Real, Device, Index, RowMajorOrder >::operator=( const DenseMatrixView& matrix )
-{
- this->setLike( matrix );
- this->values = matrix.values;
- return *this;
-}
-
-// cross-device copy assignment
-template< typename Real,
- typename Device,
- typename Index,
- bool RowMajorOrder >
- template< typename Real2, typename Device2, typename Index2, typename >
-DenseMatrixView< Real, Device, Index, RowMajorOrder >&
-DenseMatrixView< Real, Device, Index, RowMajorOrder >::operator=( const DenseMatrixView< Real2, Device2, Index2 >& matrix )
+DenseMatrixView< Real, Device, Index, RowMajorOrder >::
+operator=( const DenseMatrixView& matrix )
{
- static_assert( std::is_same< Device, Devices::Host >::value || std::is_same< Device, Devices::Cuda >::value,
- "unknown device" );
- static_assert( std::is_same< Device2, Devices::Host >::value || std::is_same< Device2, Devices::Cuda >::value,
- "unknown device" );
-
- this->setLike( matrix );
-
- throw Exceptions::NotImplementedError("Cross-device assignment for the Dense format is not implemented yet.");
+ MatrixView< Real, Device, Index >::operator=( matrix );
+ this->segments = matrix.segments;
}
-
template< typename Real,
typename Device,
typename Index,
@@ -927,15 +914,6 @@ void DenseMatrixView< Real, Device, Index, RowMajorOrder >::save( const String&
Object::save( fileName );
}
-template< typename Real,
- typename Device,
- typename Index,
- bool RowMajorOrder >
-void DenseMatrixView< Real, Device, Index, RowMajorOrder >::load( const String& fileName )
-{
- Object::load( fileName );
-}
-
template< typename Real,
typename Device,
typename Index,
@@ -945,15 +923,6 @@ void DenseMatrixView< Real, Device, Index, RowMajorOrder >::save( File& file ) c
MatrixView< Real, Device, Index >::save( file );
}
-template< typename Real,
- typename Device,
- typename Index,
- bool RowMajorOrder >
-void DenseMatrixView< Real, Device, Index, RowMajorOrder >::load( File& file )
-{
- MatrixView< Real, Device, Index >::load( file );
-}
-
template< typename Real,
typename Device,
typename Index,
diff --git a/src/UnitTests/Matrices/DenseMatrixTest.h b/src/UnitTests/Matrices/DenseMatrixTest.h
index 0f71580100..a3e7e8f61e 100644
--- a/src/UnitTests/Matrices/DenseMatrixTest.h
+++ b/src/UnitTests/Matrices/DenseMatrixTest.h
@@ -1234,9 +1234,7 @@ void test_SaveAndLoad()
ASSERT_NO_THROW( savedMatrix.save( TEST_FILE_NAME ) );
- Matrix loadedMatrix;
- loadedMatrix.reset();
- loadedMatrix.setDimensions( rows, cols );
+ Matrix loadedMatrix( rows, cols );
ASSERT_NO_THROW( loadedMatrix.load( TEST_FILE_NAME ) );
--
GitLab