GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-12 11:49:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

In sparse matrix product, enable sorted insertion when doing two transposition is defenitely not optimal.

Browse Source
This commit is contained in:
Gael Guennebaud 2014-08-29 11:55:03 +02:00
parent c3e4080474
commit 1ed9e2d004
1 changed files with 45 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
View File

@ -15,7 +15,7 @@ namespace Eigen {
namespace internal { namespace internal {
template<typename Lhs, typename Rhs, typename ResultType> template<typename Lhs, typename Rhs, typename ResultType>
static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res) static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res, bool sortedInsertion = false)
{ {
typedef typename remove_all<Lhs>::type::Scalar Scalar; typedef typename remove_all<Lhs>::type::Scalar Scalar;
typedef typename remove_all<Lhs>::type::Index Index; typedef typename remove_all<Lhs>::type::Index Index;
@ -64,7 +64,8 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
values[i] += x * y; values[i] += x * y;
} }
} }
if(!sortedInsertion)
{
// unordered insertion // unordered insertion
for(Index k=0; k<nnz; ++k) for(Index k=0; k<nnz; ++k)
{ {
@ -72,22 +73,20 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
res.insertBackByOuterInnerUnordered(j,i) = values[i]; res.insertBackByOuterInnerUnordered(j,i) = values[i];
mask[i] = false; mask[i] = false;
} }
}
#if 0 else
{
// alternative ordered insertion code: // alternative ordered insertion code:
const Index t200 = rows/(log2(200)*1.39);
Index t200 = rows/(log2(200)*1.39); const Index t = (rows*100)/139;
Index t = (rows*100)/139;
// FIXME reserve nnz non zeros // FIXME reserve nnz non zeros
// FIXME implement fast sort algorithms for very small nnz // FIXME implement faster sorting algorithms for very small nnz
// if the result is sparse enough => use a quick sort // if the result is sparse enough => use a quick sort
// otherwise => loop through the entire vector // otherwise => loop through the entire vector
// In order to avoid to perform an expensive log2 when the // In order to avoid to perform an expensive log2 when the
// result is clearly very sparse we use a linear bound up to 200. // result is clearly very sparse we use a linear bound up to 200.
//if((nnz<200 && nnz<t200) || nnz * log2(nnz) < t) if((nnz<200 && nnz<t200) || nnz * log2(nnz) < t)
//res.startVec(j);
if(true)
{ {
if(nnz>1) std::sort(indices.data(),indices.data()+nnz); if(nnz>1) std::sort(indices.data(),indices.data()+nnz);
for(Index k=0; k<nnz; ++k) for(Index k=0; k<nnz; ++k)
@ -109,8 +108,7 @@ static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& r
} }
} }
} }
#endif }
} }
res.finalize(); res.finalize();
} }
@ -137,11 +135,21 @@ struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,C
typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::Index> RowMajorMatrix; typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::Index> RowMajorMatrix;
typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::Index> ColMajorMatrix; typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::Index> ColMajorMatrix;
ColMajorMatrix resCol(lhs.rows(),rhs.cols()); ColMajorMatrix resCol(lhs.rows(),rhs.cols());
internal::conservative_sparse_sparse_product_impl<Lhs,Rhs,ColMajorMatrix>(lhs, rhs, resCol); // FIXME, the following heuristic is probably not very good.
// sort the non zeros: if(lhs.rows()>=rhs.cols())
{
// perform sorted insertion
internal::conservative_sparse_sparse_product_impl<Lhs,Rhs,ColMajorMatrix>(lhs, rhs, resCol, true);
res = resCol;
}
else
{
// ressort to transpose to sort the entries
internal::conservative_sparse_sparse_product_impl<Lhs,Rhs,ColMajorMatrix>(lhs, rhs, resCol, false);
RowMajorMatrix resRow(resCol); RowMajorMatrix resRow(resCol);
res = resRow; res = resRow;
} }
}
}; };
template<typename Lhs, typename Rhs, typename ResultType> template<typename Lhs, typename Rhs, typename ResultType>