mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-07 18:27:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enable vectorization of gemv for PacketSize>4 through unaligned loads (still better than no vectorization)

Browse Source
This commit is contained in:
Gael Guennebaud 2014-03-28 09:11:06 +01:00
parent 51e85c936d
commit c94fde118a
1 changed files with 13 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
Eigen/src/Core/products/GeneralMatrixVector.h
View File

@ -136,12 +136,17 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,ColMajor,Co
// find how many columns do we have to skip to be aligned with the result (if possible)
Index skipColumns = 0;
// if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
// TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4.
if( (size_t(lhs)%sizeof(LhsScalar)) || (size_t(res)%sizeof(ResScalar)) || LhsPacketSize > 4)
if( (size_t(lhs)%sizeof(LhsScalar)) || (size_t(res)%sizeof(ResScalar)) )
{
alignedSize = 0;
alignedStart = 0;
}
else if(LhsPacketSize > 4)
{
// TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4.
// Currently, it seems to be better to perform unaligned loads anyway
alignmentPattern = NoneAligned;
}
else if (LhsPacketSize>1)
{
eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || size<LhsPacketSize);
@ -401,13 +406,16 @@ EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,RowMajor,Co
// find how many rows do we have to skip to be aligned with rhs (if possible)
Index skipRows = 0;
// if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
// TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4.
if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (size_t(lhs)%sizeof(LhsScalar)) || (size_t(rhs)%sizeof(RhsScalar)) ||
(LhsPacketSize > 4))
if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) || (size_t(lhs)%sizeof(LhsScalar)) || (size_t(rhs)%sizeof(RhsScalar)) )
{
alignedSize = 0;
alignedStart = 0;
}
else if(LhsPacketSize > 4)
{
// TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4.
alignmentPattern = NoneAligned;
}
else if (LhsPacketSize>1)
{
eigen_internal_assert(size_t(lhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || depth<LhsPacketSize);