GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-07-31 01:03:38 +08:00
Code Issues Packages Projects Releases Wiki Activity

implement __gnuc_forget_about_setZero_its_over_now

Browse Source
This commit is contained in:
Gael Guennebaud 2009-09-18 15:36:05 +02:00
parent 6b5f96cb03
commit 0b60027f3c
2 changed files with 37 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

58
Eigen/src/Sparse/SparseProduct.h
View File

@ -305,23 +305,9 @@ inline Derived& SparseMatrixBase<Derived>::operator=(const SparseProduct<Lhs,Rhs
}
// dense = sparse * dense
// template<typename Derived>
// template<typename Lhs, typename Rhs>
// Derived& MatrixBase<Derived>::lazyAssign(const SparseProduct<Lhs,Rhs,SparseTimeDenseProduct>& product)
// {
// typedef typename ei_cleantype<Lhs>::type _Lhs;
// typedef typename _Lhs::InnerIterator LhsInnerIterator;
// enum { LhsIsRowMajor = (_Lhs::Flags&RowMajorBit)==RowMajorBit };
// derived().setZero();
// for (int j=0; j<product.lhs().outerSize(); ++j)
// for (LhsInnerIterator i(product.lhs(),j); i; ++i)
// derived().row(LhsIsRowMajor ? j : i.index()) += i.value() * product.rhs().row(LhsIsRowMajor ? i.index() : j);
// return derived();
// }
template<typename Derived>
template<typename Lhs, typename Rhs>
Derived& MatrixBase<Derived>::lazyAssign(const SparseProduct<Lhs,Rhs,SparseTimeDenseProduct>& product)
// Note that here we force no inlining and separate the setZero() because GCC messes up otherwise
template<typename Lhs, typename Rhs, typename Dest>
EIGEN_DONT_INLINE void ei_sparse_time_dense_product(const Lhs& lhs, const Rhs& rhs, Dest& dst)
{
typedef typename ei_cleantype<Lhs>::type _Lhs;
typedef typename ei_cleantype<Rhs>::type _Rhs;
@ -335,34 +321,44 @@ Derived& MatrixBase<Derived>::lazyAssign(const SparseProduct<Lhs,Rhs,SparseTimeD
|| ( (_Lhs::Flags&LowerTriangularBit) && LhsIsRowMajor) ),
ProcessSecondHalf = LhsIsSelfAdjoint && (!ProcessFirstHalf)
};
derived().setZero();
for (int j=0; j<product.lhs().outerSize(); ++j)
for (int j=0; j<lhs.outerSize(); ++j)
{
LhsInnerIterator i(product.lhs(),j);
LhsInnerIterator i(lhs,j);
if (ProcessSecondHalf && i && (i.index()==j))
{
derived().row(j) += i.value() * product.rhs().row(j);
dst.row(j) += i.value() * rhs.row(j);
++i;
}
Block<Derived,1,Derived::ColsAtCompileTime> res(derived().row(LhsIsRowMajor ? j : 0));
for (; (ProcessFirstHalf ? i && i.index() < j : i) ; ++i)
typename Rhs::Scalar rhs_j = rhs.coeff(j,0);
Block<Dest,1,Dest::ColsAtCompileTime> dst_j(dst.row(LhsIsRowMajor ? j : 0));
for(; (ProcessFirstHalf ? i && i.index() < j : i) ; ++i)
{
if (LhsIsSelfAdjoint)
if(LhsIsSelfAdjoint)
{
int a = LhsIsRowMajor ? j : i.index();
int b = LhsIsRowMajor ? i.index() : j;
Scalar v = i.value();
derived().row(a) += (v) * product.rhs().row(b);
derived().row(b) += ei_conj(v) * product.rhs().row(a);
typename Lhs::Scalar v = i.value();
dst.row(a) += (v) * rhs.row(b);
dst.row(b) += ei_conj(v) * rhs.row(a);
}
else if (LhsIsRowMajor)
res += i.value() * product.rhs().row(i.index());
else if(LhsIsRowMajor)
dst_j += i.value() * rhs.row(i.index());
else if(Rhs::ColsAtCompileTime==1)
dst.coeffRef(i.index()) += i.value() * rhs_j;
else
derived().row(i.index()) += i.value() * product.rhs().row(j);
dst.row(i.index()) += i.value() * rhs.row(j);
}
if (ProcessFirstHalf && i && (i.index()==j))
derived().row(j) += i.value() * product.rhs().row(j);
dst.row(j) += i.value() * rhs.row(j);
}
}
template<typename Derived>
template<typename Lhs, typename Rhs>
Derived& MatrixBase<Derived>::lazyAssign(const SparseProduct<Lhs,Rhs,SparseTimeDenseProduct>& product)
{
derived().setZero();
ei_sparse_time_dense_product(product.lhs(), product.rhs(), derived());
return derived();
}

16
bench/sparse_dense_product.cpp
View File

@ -91,11 +91,11 @@ int main(int argc, char *argv[])
{
std::cout << "Eigen sparse\t" << sm1.nonZeros()/float(sm1.rows()*sm1.cols())*100 << "%\n";
BENCH(for (int k=0; k<REPEAT; ++k) v2 = sm1 * v1;)
BENCH(asm("#myc"); v2 = sm1 * v1; asm("#myd");)
std::cout << " a * v:\t" << timer.value() << endl;
BENCH(for (int k=0; k<REPEAT; ++k) { asm("#mya"); v2 = sm1.transpose() * v1; asm("#myb"); })
BENCH( { asm("#mya"); v2 = sm1.transpose() * v1; asm("#myb"); })
std::cout << " a' * v:\t" << timer.value() << endl;
}
@ -123,18 +123,10 @@ int main(int argc, char *argv[])
Map<Matrix<Scalar,Dynamic,1> >(&gmmV1[0], cols) = v1;
Map<Matrix<Scalar,Dynamic,1> >(&gmmV2[0], cols) = v2;
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
gmm::mult(m1, gmmV1, gmmV2);
timer.stop();
BENCH( asm("#myx"); gmm::mult(m1, gmmV1, gmmV2); asm("#myy"); )
std::cout << " a * v:\t" << timer.value() << endl;
timer.reset();
timer.start();
for (int k=0; k<REPEAT; ++k)
gmm::mult(gmm::transposed(m1), gmmV1, gmmV2);
timer.stop();
BENCH( gmm::mult(gmm::transposed(m1), gmmV1, gmmV2); )
std::cout << " a' * v:\t" << timer.value() << endl;
}
#endif