mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-12-21 07:19:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

Updated the tensor type casting code as follow: in the case where TgtRatio < SrcRatio, disable the vectorization of the source expression unless is has direct-access.

Browse Source
This commit is contained in:
Benoit Steiner 2015-03-02 10:11:40 -08:00
parent d9cb604a5d
commit b0f2b6f297
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
View File

@ -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 {