devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

bug #1562: optimize evaluation of small products of the form s*A*B by rewriting them as: s*(A.lazyProduct(B)) to save a costly temporary. Measured speedup from 2x to 5x...

Browse Source 
(grafted from d428a199ab
)
This commit is contained in:
Gael Guennebaud
2018-07-02 11:41:09 +02:00
parent 5c4b03c53e
commit 32cb4853c6
4 changed files with 43 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
Eigen/src/Core/ProductEvaluators.h
View File

@@ -396,7 +396,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
// but easier on the compiler side
call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::assign_op<typename Dst::Scalar,Scalar>());
}
template<typename Dst>
static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
{
@@ -410,6 +410,32 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode>
// dst.noalias() -= lhs.lazyProduct(rhs);
call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<typename Dst::Scalar,Scalar>());
}
// Catch "dst {,+,-}= (s*A)*B" and evaluate it lazily by moving out the scalar factor:
// dst {,+,-}= s * (A.lazyProduct(B))
// This is a huge benefit for heap-allocated matrix types as it save one costly allocation.
// For them, this strategy is also faster than simply by-passing the heap allocation through
// stack allocation.
// For fixed sizes matrices, this is less obvious, it is sometimes x2 faster, but sometimes x3 slower,
// and the behavior depends also a lot on the compiler... so let's be conservative and enable them for dynamic-size only,
// that is when coming from generic_product_impl<...,GemmProduct> in file GeneralMatrixMatrix.h
template<typename Dst, typename Scalar1, typename Scalar2, typename Plain1, typename Xpr2, typename Func>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
void eval_dynamic(Dst& dst, const CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>, Xpr2>& lhs, const Rhs& rhs, const Func &func)
{
call_assignment_no_alias(dst, lhs.lhs().functor().m_other * lhs.rhs().lazyProduct(rhs), func);
}
// Here, we we always have LhsT==Lhs, but we need to make it a template type to make the above
// overload more specialized.
template<typename Dst, typename LhsT, typename Func>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
void eval_dynamic(Dst& dst, const LhsT& lhs, const Rhs& rhs, const Func &func)
{
call_assignment_no_alias(dst, lhs.lazyProduct(rhs), func);
}
// template<typename Dst>
// static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)

6
Eigen/src/Core/products/GeneralMatrixMatrix.h
View File

@@ -428,7 +428,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
{
if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
lazyproduct::evalTo(dst, lhs, rhs);
lazyproduct::eval_dynamic(dst, lhs, rhs, internal::assign_op<typename Dst::Scalar,Scalar>());
else
{
dst.setZero();
@@ -440,7 +440,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
{
if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
lazyproduct::addTo(dst, lhs, rhs);
lazyproduct::eval_dynamic(dst, lhs, rhs, internal::add_assign_op<typename Dst::Scalar,Scalar>());
else
scaleAndAddTo(dst,lhs, rhs, Scalar(1));
}
@@ -449,7 +449,7 @@ struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
{
if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
lazyproduct::subTo(dst, lhs, rhs);
lazyproduct::eval_dynamic(dst, lhs, rhs, internal::sub_assign_op<typename Dst::Scalar,Scalar>());
else
scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
}