diff options
Diffstat (limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h')
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h | 5 |
1 files changed, 4 insertions, 1 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h index f594725a8..fb1f1f6ea 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h @@ -93,7 +93,10 @@ struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 1, 2> { template<int LoadMode, typename Index> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TgtPacket packet(Index index) const { const int SrcPacketSize = internal::unpacket_traits<SrcPacket>::size; - if (index + SrcPacketSize < m_maxIndex) { + // Only call m_impl.packet() when we have direct access to the underlying data. This + // ensures that we don't compute the subexpression twice. We may however load some + // coefficients twice, but in practice this doesn't negatively impact performance. + if (m_impl.data() && (index + SrcPacketSize < m_maxIndex)) { // Force unaligned memory loads since we can't ensure alignment anymore return internal::pcast<SrcPacket, TgtPacket>(m_impl.template packet<Unaligned>(index)); } else { |