From 4e1309ce0ab8b17502dc836da61a75a0d6d46ae0 Mon Sep 17 00:00:00 2001
From: Lukas Cejka <lukas.ostatek@gmail.com>
Date: Mon, 24 Jun 2019 23:28:19 +0200
Subject: [PATCH] Implemented fix for case when roundToMultiple would give a
number smaller than the number of rows, thus causing index to be out of
bounds on CUDA
---
src/TNL/Matrices/EllpackSymmetric_impl.h | 20 +++++++++++++++++---
1 file changed, 17 insertions(+), 3 deletions(-)
diff --git a/src/TNL/Matrices/EllpackSymmetric_impl.h b/src/TNL/Matrices/EllpackSymmetric_impl.h
index fa8952b4ff..f64cef4c5b 100644
--- a/src/TNL/Matrices/EllpackSymmetric_impl.h
+++ b/src/TNL/Matrices/EllpackSymmetric_impl.h
@@ -57,13 +57,27 @@ void EllpackSymmetric< Real, Device, Index >::setDimensions( const IndexType row
TNL_ASSERT( rows > 0 && columns > 0,
std::cerr << "rows = " << rows
<< " columns = " << columns <<std::endl );
+
this->rows = rows;
- this->columns = columns;
+ this->columns = columns;
+
if( std::is_same< DeviceType, Devices::Cuda >::value )
- this->alignedRows = roundToMultiple( columns, Cuda::getWarpSize() );
+ {
+ this->alignedRows = roundToMultiple( columns, Devices::Cuda::getWarpSize() );
+
+ if( this->rows - this->alignedRows > 0 )
+ {
+ IndexType missingRows = this->rows - this->alignedRows;
+ missingRows = roundToMultiple( missingRows, Devices::Cuda::getWarpSize() );
+ this->alignedRows += missingRows;
+
+// this->alignedRows += roundToMultiple( this->rows - this->alignedRows, Devices::Cuda::getWarpSize() );
+ }
+ }
else this->alignedRows = rows;
+
if( this->rowLengths != 0 )
- allocateElements();
+ allocateElements();
}
template< typename Real,
--
GitLab