mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-13 18:37:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

Skip null numerators in triangular-vector-solve (as in BLAS TRSV).

Browse Source
This commit is contained in:
Gael Guennebaud 2018-07-09 11:13:19 +02:00
parent 359dd77ec3
commit ec323b7e66
1 changed files with 12 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
Eigen/src/Core/products/TriangularSolverVector.h
View File

@ -58,7 +58,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
{
// let's directly call the low level product function because:
// 1 - it is faster to compile
// 2 - it is slighlty faster at runtime
// 2 - it is slightly faster at runtime
Index startRow = IsLower ? pi : pi-actualPanelWidth;
Index startCol = IsLower ? 0 : pi;
@ -77,7 +77,7 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
if (k>0)
rhs[i] -= (cjLhs.row(i).segment(s,k).transpose().cwiseProduct(Map<const Matrix<RhsScalar,Dynamic,1> >(rhs+s,k))).sum();
if(!(Mode & UnitDiag))
if((!(Mode & UnitDiag)) && numext::not_equal_strict(rhs[i],RhsScalar(0)))
rhs[i] /= cjLhs(i,i);
}
}
@ -114,20 +114,23 @@ struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Con
for(Index k=0; k<actualPanelWidth; ++k)
{
Index i = IsLower ? pi+k : pi-k-1;
if(!(Mode & UnitDiag))
rhs[i] /= cjLhs.coeff(i,i);
if(numext::not_equal_strict(rhs[i],RhsScalar(0)))
{
if(!(Mode & UnitDiag))
rhs[i] /= cjLhs.coeff(i,i);
Index r = actualPanelWidth - k - 1; // remaining size
Index s = IsLower ? i+1 : i-r;
if (r>0)
Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
Index r = actualPanelWidth - k - 1; // remaining size
Index s = IsLower ? i+1 : i-r;
if (r>0)
Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
}
}
Index r = IsLower ? size - endBlock : startBlock; // remaining size
if (r > 0)
{
// let's directly call the low level product function because:
// 1 - it is faster to compile
// 2 - it is slighlty faster at runtime
// 2 - it is slightly faster at runtime
general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
r, actualPanelWidth,
LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride),