From 0e83d3759580a23d17dd9caec2cb03e0f855cd1e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Jakub=20Klinkovsk=C3=BD?= <klinkjak@fjfi.cvut.cz>
Date: Sun, 25 Jun 2017 15:40:08 +0200
Subject: [PATCH] Added getRowLengthFast method to matrices
---
src/TNL/Matrices/CSR.h | 3 +++
src/TNL/Matrices/CSR_impl.h | 9 +++++++++
src/TNL/Matrices/ChunkedEllpack.h | 3 +++
src/TNL/Matrices/ChunkedEllpack_impl.h | 14 +++++++++++++-
src/TNL/Matrices/Dense.h | 3 +++
src/TNL/Matrices/Dense_impl.h | 9 +++++++++
src/TNL/Matrices/Ellpack.h | 3 +++
src/TNL/Matrices/Ellpack_impl.h | 9 +++++++++
src/TNL/Matrices/Multidiagonal.h | 3 +++
src/TNL/Matrices/Multidiagonal_impl.h | 16 ++++++++++++++++
src/TNL/Matrices/SlicedEllpack.h | 3 +++
src/TNL/Matrices/SlicedEllpack_impl.h | 11 +++++++++++
src/TNL/Matrices/Tridiagonal.h | 3 +++
src/TNL/Matrices/Tridiagonal_impl.h | 9 +++++++++
14 files changed, 97 insertions(+), 1 deletion(-)
diff --git a/src/TNL/Matrices/CSR.h b/src/TNL/Matrices/CSR.h
index f274200b22..97b046fba6 100644
--- a/src/TNL/Matrices/CSR.h
+++ b/src/TNL/Matrices/CSR.h
@@ -63,6 +63,9 @@ class CSR : public Sparse< Real, Device, Index >
IndexType getRowLength( const IndexType row ) const;
+ __cuda_callable__
+ IndexType getRowLengthFast( const IndexType row ) const;
+
template< typename Real2, typename Device2, typename Index2 >
void setLike( const CSR< Real2, Device2, Index2 >& matrix );
diff --git a/src/TNL/Matrices/CSR_impl.h b/src/TNL/Matrices/CSR_impl.h
index 3563fa789e..abc4e28231 100644
--- a/src/TNL/Matrices/CSR_impl.h
+++ b/src/TNL/Matrices/CSR_impl.h
@@ -117,6 +117,15 @@ template< typename Real,
typename Device,
typename Index >
Index CSR< Real, Device, Index >::getRowLength( const IndexType row ) const
+{
+ return this->rowPointers.getElement( row + 1 ) - this->rowPointers.getElement( row );
+}
+
+template< typename Real,
+ typename Device,
+ typename Index >
+__cuda_callable__
+Index CSR< Real, Device, Index >::getRowLengthFast( const IndexType row ) const
{
return this->rowPointers[ row + 1 ] - this->rowPointers[ row ];
}
diff --git a/src/TNL/Matrices/ChunkedEllpack.h b/src/TNL/Matrices/ChunkedEllpack.h
index dace4893cc..c86081216e 100644
--- a/src/TNL/Matrices/ChunkedEllpack.h
+++ b/src/TNL/Matrices/ChunkedEllpack.h
@@ -92,6 +92,9 @@ class ChunkedEllpack : public Sparse< Real, Device, Index >
IndexType getRowLength( const IndexType row ) const;
+ __cuda_callable__
+ IndexType getRowLengthFast( const IndexType row ) const;
+
template< typename Real2, typename Device2, typename Index2 >
void setLike( const ChunkedEllpack< Real2, Device2, Index2 >& matrix );
diff --git a/src/TNL/Matrices/ChunkedEllpack_impl.h b/src/TNL/Matrices/ChunkedEllpack_impl.h
index 3840384b90..bd94540b79 100644
--- a/src/TNL/Matrices/ChunkedEllpack_impl.h
+++ b/src/TNL/Matrices/ChunkedEllpack_impl.h
@@ -262,9 +262,21 @@ template< typename Real,
typename Index >
Index ChunkedEllpack< Real, Device, Index >::getRowLength( const IndexType row ) const
{
- const IndexType& sliceIndex = rowToSliceMapping[ row ];
+ const IndexType& sliceIndex = rowToSliceMapping.getElement( row );
TNL_ASSERT( sliceIndex < this->rows, );
const IndexType& chunkSize = slices.getElement( sliceIndex ).chunkSize;
+ return rowPointers.getElement( row + 1 ) - rowPointers.getElement( row );
+}
+
+template< typename Real,
+ typename Device,
+ typename Index >
+__cuda_callable__
+Index ChunkedEllpack< Real, Device, Index >::getRowLengthFast( const IndexType row ) const
+{
+ const IndexType& sliceIndex = rowToSliceMapping[ row ];
+ TNL_ASSERT( sliceIndex < this->rows, );
+ const IndexType& chunkSize = slices[ sliceIndex ].chunkSize;
return rowPointers[ row + 1 ] - rowPointers[ row ];
}
diff --git a/src/TNL/Matrices/Dense.h b/src/TNL/Matrices/Dense.h
index bd5daa284d..5d94b5bda6 100644
--- a/src/TNL/Matrices/Dense.h
+++ b/src/TNL/Matrices/Dense.h
@@ -66,6 +66,9 @@ class Dense : public Matrix< Real, Device, Index >
*/
IndexType getRowLength( const IndexType row ) const;
+ __cuda_callable__
+ IndexType getRowLengthFast( const IndexType row ) const;
+
IndexType getMaxRowLength() const;
IndexType getNumberOfMatrixElements() const;
diff --git a/src/TNL/Matrices/Dense_impl.h b/src/TNL/Matrices/Dense_impl.h
index 55282e8d62..08f66239c3 100644
--- a/src/TNL/Matrices/Dense_impl.h
+++ b/src/TNL/Matrices/Dense_impl.h
@@ -99,6 +99,15 @@ Index Dense< Real, Device, Index >::getRowLength( const IndexType row ) const
return this->getColumns();
}
+template< typename Real,
+ typename Device,
+ typename Index >
+__cuda_callable__
+Index Dense< Real, Device, Index >::getRowLengthFast( const IndexType row ) const
+{
+ return this->getColumns();
+}
+
template< typename Real,
typename Device,
typename Index >
diff --git a/src/TNL/Matrices/Ellpack.h b/src/TNL/Matrices/Ellpack.h
index 65f06a550e..4e1b92b949 100644
--- a/src/TNL/Matrices/Ellpack.h
+++ b/src/TNL/Matrices/Ellpack.h
@@ -56,6 +56,9 @@ class Ellpack : public Sparse< Real, Device, Index >
IndexType getRowLength( const IndexType row ) const;
+ __cuda_callable__
+ IndexType getRowLengthFast( const IndexType row ) const;
+
template< typename Real2, typename Device2, typename Index2 >
void setLike( const Ellpack< Real2, Device2, Index2 >& matrix );
diff --git a/src/TNL/Matrices/Ellpack_impl.h b/src/TNL/Matrices/Ellpack_impl.h
index 9f12aee4c9..ca23f4168d 100644
--- a/src/TNL/Matrices/Ellpack_impl.h
+++ b/src/TNL/Matrices/Ellpack_impl.h
@@ -113,6 +113,15 @@ Index Ellpack< Real, Device, Index >::getRowLength( const IndexType row ) const
return this->rowLengths;
}
+template< typename Real,
+ typename Device,
+ typename Index >
+__cuda_callable__
+Index Ellpack< Real, Device, Index >::getRowLengthFast( const IndexType row ) const
+{
+ return this->rowLengths;
+}
+
template< typename Real,
typename Device,
typename Index >
diff --git a/src/TNL/Matrices/Multidiagonal.h b/src/TNL/Matrices/Multidiagonal.h
index 058c05217e..884e804ec3 100644
--- a/src/TNL/Matrices/Multidiagonal.h
+++ b/src/TNL/Matrices/Multidiagonal.h
@@ -53,6 +53,9 @@ class Multidiagonal : public Matrix< Real, Device, Index >
IndexType getRowLength( const IndexType row ) const;
+ __cuda_callable__
+ IndexType getRowLengthFast( const IndexType row ) const;
+
IndexType getMaxRowLength() const;
template< typename Vector >
diff --git a/src/TNL/Matrices/Multidiagonal_impl.h b/src/TNL/Matrices/Multidiagonal_impl.h
index 91588a515e..ea8c4050db 100644
--- a/src/TNL/Matrices/Multidiagonal_impl.h
+++ b/src/TNL/Matrices/Multidiagonal_impl.h
@@ -105,6 +105,22 @@ Index Multidiagonal< Real, Device, Index >::getRowLength( const IndexType row )
return rowLength;
}
+template< typename Real,
+ typename Device,
+ typename Index >
+__cuda_callable__
+Index Multidiagonal< Real, Device, Index >::getRowLengthFast( const IndexType row ) const
+{
+ IndexType rowLength( 0 );
+ for( IndexType i = 0; i < diagonalsShift.getSize(); i++ )
+ {
+ const IndexType column = row + diagonalsShift[ i ];
+ if( column >= 0 && column < this->getColumns() )
+ rowLength++;
+ }
+ return rowLength;
+}
+
template< typename Real,
typename Device,
typename Index >
diff --git a/src/TNL/Matrices/SlicedEllpack.h b/src/TNL/Matrices/SlicedEllpack.h
index ef1d542bf3..2bd6e762d7 100644
--- a/src/TNL/Matrices/SlicedEllpack.h
+++ b/src/TNL/Matrices/SlicedEllpack.h
@@ -84,6 +84,9 @@ class SlicedEllpack : public Sparse< Real, Device, Index >
IndexType getRowLength( const IndexType row ) const;
+ __cuda_callable__
+ IndexType getRowLengthFast( const IndexType row ) const;
+
template< typename Real2, typename Device2, typename Index2 >
void setLike( const SlicedEllpack< Real2, Device2, Index2, SliceSize >& matrix );
diff --git a/src/TNL/Matrices/SlicedEllpack_impl.h b/src/TNL/Matrices/SlicedEllpack_impl.h
index 0f22e69b9c..bcbed7d986 100644
--- a/src/TNL/Matrices/SlicedEllpack_impl.h
+++ b/src/TNL/Matrices/SlicedEllpack_impl.h
@@ -108,6 +108,17 @@ Index SlicedEllpack< Real, Device, Index, SliceSize >::getRowLength( const Index
return this->sliceCompressedRowLengths.getElement( slice );
}
+template< typename Real,
+ typename Device,
+ typename Index,
+ int SliceSize >
+__cuda_callable__
+Index SlicedEllpack< Real, Device, Index, SliceSize >::getRowLengthFast( const IndexType row ) const
+{
+ const IndexType slice = row / SliceSize;
+ return this->sliceCompressedRowLengths[ slice ];
+}
+
template< typename Real,
typename Device,
typename Index,
diff --git a/src/TNL/Matrices/Tridiagonal.h b/src/TNL/Matrices/Tridiagonal.h
index 396ea25192..5ebd526231 100644
--- a/src/TNL/Matrices/Tridiagonal.h
+++ b/src/TNL/Matrices/Tridiagonal.h
@@ -54,6 +54,9 @@ class Tridiagonal : public Matrix< Real, Device, Index >
IndexType getRowLength( const IndexType row ) const;
+ __cuda_callable__
+ IndexType getRowLengthFast( const IndexType row ) const;
+
IndexType getMaxRowLength() const;
template< typename Real2, typename Device2, typename Index2 >
diff --git a/src/TNL/Matrices/Tridiagonal_impl.h b/src/TNL/Matrices/Tridiagonal_impl.h
index c1239627da..fb070c7f09 100644
--- a/src/TNL/Matrices/Tridiagonal_impl.h
+++ b/src/TNL/Matrices/Tridiagonal_impl.h
@@ -98,6 +98,15 @@ template< typename Real,
typename Device,
typename Index >
Index Tridiagonal< Real, Device, Index >::getRowLength( const IndexType row ) const
+{
+ return this->getRowLengthFast( row );
+}
+
+template< typename Real,
+ typename Device,
+ typename Index >
+__cuda_callable__
+Index Tridiagonal< Real, Device, Index >::getRowLengthFast( const IndexType row ) const
{
const IndexType diagonalLength = min( this->getRows(), this->getColumns() );
if( row == 0 )
--
GitLab