GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-06 19:29:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix Alignment in coeff-based product, and enable unaligned vectorization

Browse Source
This commit is contained in:
Gael Guennebaud 2015-12-08 11:28:05 +01:00
parent 03ad4fc504
commit 543bd28a24
1 changed files with 11 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
Eigen/src/Core/ProductEvaluators.h
View File

@ -463,19 +463,16 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
LhsAlignment = LhsEtorType::Alignment,
RhsAlignment = RhsEtorType::Alignment,
LhsIsAligned = int(LhsAlignment) >= int(unpacket_traits<PacketScalar>::alignment),
RhsIsAligned = int(RhsAlignment) >= int(unpacket_traits<PacketScalar>::alignment),
LhsRowMajor = LhsFlags & RowMajorBit,
RhsRowMajor = RhsFlags & RowMajorBit,
SameType = is_same<typename LhsNestedCleaned::Scalar,typename RhsNestedCleaned::Scalar>::value,
CanVectorizeRhs = RhsRowMajor && (RhsFlags & PacketAccessBit)
&& (ColsAtCompileTime == Dynamic || ( (ColsAtCompileTime % PacketSize) == 0 && RhsIsAligned ) ),
&& (ColsAtCompileTime == Dynamic || ((ColsAtCompileTime % PacketSize) == 0) ),
CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit)
&& (RowsAtCompileTime == Dynamic || ( (RowsAtCompileTime % PacketSize) == 0 && LhsIsAligned ) ),
&& (RowsAtCompileTime == Dynamic || ((RowsAtCompileTime % PacketSize) == 0) ),
EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
: (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
@ -487,8 +484,11 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
| (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0)
| (XprType::IsVectorAtCompileTime ? LinearAccessBit : 0),
Alignment = CanVectorizeLhs ? LhsAlignment
: CanVectorizeRhs ? RhsAlignment
LhsOuterStrideBytes = LhsNestedCleaned::OuterStrideAtCompileTime * sizeof(typename LhsNestedCleaned::Scalar),
RhsOuterStrideBytes = RhsNestedCleaned::OuterStrideAtCompileTime * sizeof(typename RhsNestedCleaned::Scalar),
Alignment = CanVectorizeLhs ? (LhsOuterStrideBytes<0 || (int(LhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,LhsAlignment))!=0 ? 0 : LhsAlignment)
: CanVectorizeRhs ? (RhsOuterStrideBytes<0 || (int(RhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,RhsAlignment))!=0 ? 0 : RhsAlignment)
: 0,
/* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside
@ -500,7 +500,6 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
&& LhsRowMajor
&& (!RhsRowMajor)
&& (LhsFlags & RhsFlags & ActualPacketAccessBit)
&& (LhsIsAligned && RhsIsAligned)
&& (InnerSize % packet_traits<Scalar>::size == 0)
};
@ -524,10 +523,9 @@ struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape,
const PacketType packet(Index row, Index col) const
{
PacketType res;
typedef etor_product_packet_impl<Flags&RowMajorBit ? RowMajor : ColMajor,
Unroll ? InnerSize : Dynamic,
typedef etor_product_packet_impl<bool(int(Flags)&RowMajorBit) ? RowMajor : ColMajor,
Unroll ? int(InnerSize) : Dynamic,
LhsEtorType, RhsEtorType, PacketType, LoadMode> PacketImpl;
PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res);
return res;
}