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

- cleaner use of OpenMP (no code duplication anymore)

Browse Source 
using a macro and _Pragma.
- use OpenMP also in cacheOptimalProduct and in the
  vectorized paths as well
- kill the vector assignment unroller. implement in
  operator= the logic for assigning a row-vector in
  a col-vector.
- CMakeLists support for building tests/examples
  with -fopenmp and/or -msse2
- updates in bench/, especially replace identity()
  by ones() which prevents underflows from perturbing
  bench results.
This commit is contained in:
Benoit Jacob
2008-04-11 14:28:42 +00:00
parent 7bee90a62a
commit dcebc46cdc
8 changed files with 179 additions and 231 deletions
Download Patch File Download Diff File Expand all files Collapse all files

112
Eigen/src/Core/Product.h
View File

@@ -283,64 +283,70 @@ void Product<Lhs,Rhs,EvalMode>::_cacheOptimalEval(DestDerived& res) const
#ifdef EIGEN_VECTORIZE
if( (Flags & VectorizableBit) && (!(Lhs::Flags & RowMajorBit)) )
{
for(int k=0; k<m_rhs.cols(); k++)
{
int j=0;
for(; j<cols4; j+=4)
{
const typename ei_packet_traits<Scalar>::type tmp0 = ei_pset1(m_rhs.coeff(j+0,k));
const typename ei_packet_traits<Scalar>::type tmp1 = ei_pset1(m_rhs.coeff(j+1,k));
const typename ei_packet_traits<Scalar>::type tmp2 = ei_pset1(m_rhs.coeff(j+2,k));
const typename ei_packet_traits<Scalar>::type tmp3 = ei_pset1(m_rhs.coeff(j+3,k));
for (int i=0; i<m_lhs.rows(); i+=ei_packet_traits<Scalar>::size)
{
res.writePacketCoeff(i,k,
ei_padd(
res.packetCoeff(i,k),
ei_padd(
ei_padd(
ei_pmul(tmp0, m_lhs.packetCoeff(i,j)),
ei_pmul(tmp1, m_lhs.packetCoeff(i,j+1))),
ei_padd(
ei_pmul(tmp2, m_lhs.packetCoeff(i,j+2)),
ei_pmul(tmp3, m_lhs.packetCoeff(i,j+3))
)
)
)
);
}
#define EIGEN_THE_PARALLELIZABLE_LOOP \
for(int k=0; k<this->cols(); k++) \
{ \
int j=0; \
for(; j<cols4; j+=4) \
{ \
const typename ei_packet_traits<Scalar>::type tmp0 = ei_pset1(m_rhs.coeff(j+0,k)); \
const typename ei_packet_traits<Scalar>::type tmp1 = ei_pset1(m_rhs.coeff(j+1,k)); \
const typename ei_packet_traits<Scalar>::type tmp2 = ei_pset1(m_rhs.coeff(j+2,k)); \
const typename ei_packet_traits<Scalar>::type tmp3 = ei_pset1(m_rhs.coeff(j+3,k)); \
for (int i=0; i<this->rows(); i+=ei_packet_traits<Scalar>::size) \
{ \
res.writePacketCoeff(i,k,\
ei_padd( \
res.packetCoeff(i,k), \
ei_padd( \
ei_padd( \
ei_pmul(tmp0, m_lhs.packetCoeff(i,j)), \
ei_pmul(tmp1, m_lhs.packetCoeff(i,j+1))), \
ei_padd( \
ei_pmul(tmp2, m_lhs.packetCoeff(i,j+2)), \
ei_pmul(tmp3, m_lhs.packetCoeff(i,j+3)) \
) \
) \
) \
); \
} \
} \
for(; j<m_lhs.cols(); ++j) \
{ \
const typename ei_packet_traits<Scalar>::type tmp = ei_pset1(m_rhs.coeff(j,k)); \
for (int i=0; i<this->rows(); ++i) \
res.writePacketCoeff(i,k,ei_pmul(tmp, m_lhs.packetCoeff(i,j))); \
} \
}
for(; j<m_lhs.cols(); ++j)
{
const typename ei_packet_traits<Scalar>::type tmp = ei_pset1(m_rhs.coeff(j,k));
for (int i=0; i<m_lhs.rows(); ++i)
res.writePacketCoeff(i,k,ei_pmul(tmp, m_lhs.packetCoeff(i,j)));
}
}
EIGEN_RUN_PARALLELIZABLE_LOOP(Flags & DestDerived::Flags & LargeBit)
#undef EIGEN_THE_PARALLELIZABLE_LOOP
}
else
#endif
#endif // EIGEN_VECTORIZE
{
for(int k=0; k<m_rhs.cols(); ++k)
{
int j=0;
for(; j<cols4; j+=4)
{
const Scalar tmp0 = m_rhs.coeff(j ,k);
const Scalar tmp1 = m_rhs.coeff(j+1,k);
const Scalar tmp2 = m_rhs.coeff(j+2,k);
const Scalar tmp3 = m_rhs.coeff(j+3,k);
for (int i=0; i<m_lhs.rows(); ++i)
res.coeffRef(i,k) += tmp0 * m_lhs.coeff(i,j) + tmp1 * m_lhs.coeff(i,j+1)
+ tmp2 * m_lhs.coeff(i,j+2) + tmp3 * m_lhs.coeff(i,j+3);
#define EIGEN_THE_PARALLELIZABLE_LOOP \
for(int k=0; k<this->cols(); ++k) \
{ \
int j=0; \
for(; j<cols4; j+=4) \
{ \
const Scalar tmp0 = m_rhs.coeff(j ,k); \
const Scalar tmp1 = m_rhs.coeff(j+1,k); \
const Scalar tmp2 = m_rhs.coeff(j+2,k); \
const Scalar tmp3 = m_rhs.coeff(j+3,k); \
for (int i=0; i<this->rows(); ++i) \
res.coeffRef(i,k) += tmp0 * m_lhs.coeff(i,j) + tmp1 * m_lhs.coeff(i,j+1) \
+ tmp2 * m_lhs.coeff(i,j+2) + tmp3 * m_lhs.coeff(i,j+3); \
} \
for(; j<m_lhs.cols(); ++j) \
{ \
const Scalar tmp = m_rhs.coeff(j,k); \
for (int i=0; i<this->rows(); ++i) \
res.coeffRef(i,k) += tmp * m_lhs.coeff(i,j); \
} \
}
for(; j<m_lhs.cols(); ++j)
{
const Scalar tmp = m_rhs.coeff(j,k);
for (int i=0; i<m_lhs.rows(); ++i)
res.coeffRef(i,k) += tmp * m_lhs.coeff(i,j);
}
}
EIGEN_RUN_PARALLELIZABLE_LOOP(Flags & DestDerived::Flags & LargeBit)
#undef EIGEN_THE_PARALLELIZABLE_LOOP
}
}