Page MenuHomec4science
Files F73427890

crs_matrix_base.hpp
No OneTemporary
Actions

File Metadata

Created
Sun, Jul 21, 06:26

crs_matrix_base.hpp
View Options

#pragma once
#ifndef __CRS_MATRIX_BASE_HPP__
#define __CRS_MATRIX_BASE_HPP__
/// \file crs_matrix_base.hpp
/// \brief CRS matrix base object interfaces to user provided input matrices.
/// \author Kyungjoo Kim (kyukim@sandia.gov)
#include "util.hpp"
#include "coo.hpp"
namespace Tacho {
using namespace std;
template< typename , typename > class TaskView ;
template < typename CrsMatrixType >
struct GetCrsMatrixRowViewType {
typedef int type ;
};
template < typename CrsMatrixViewType , typename TaskFactoryType >
struct GetCrsMatrixRowViewType
< TaskView<CrsMatrixViewType,TaskFactoryType> >
{
typedef typename CrsMatrixViewType::row_view_type type ;
};
/// \class CrsMatrixBase
/// \breif CRS matrix base object using Kokkos view and subview
template<typename ValueType,
typename OrdinalType,
typename SizeType = OrdinalType,
typename SpaceType = void,
typename MemoryTraits = void>
class CrsMatrixBase {
public:
typedef ValueType value_type;
typedef OrdinalType ordinal_type;
typedef SpaceType space_type;
typedef SizeType size_type;
typedef MemoryTraits memory_traits;
// 1D view, layout does not matter; no template parameters for that
typedef Kokkos::View<size_type*, space_type,memory_traits> size_type_array;
typedef Kokkos::View<ordinal_type*,space_type,memory_traits> ordinal_type_array;
typedef Kokkos::View<value_type*, space_type,memory_traits> value_type_array;
typedef typename size_type_array::value_type* size_type_array_ptr;
typedef typename ordinal_type_array::value_type* ordinal_type_array_ptr;
typedef typename value_type_array::value_type* value_type_array_ptr;
// range type
template<typename T> using range_type = pair<T,T>;
// external interface
typedef Coo<CrsMatrixBase> ijv_type;
friend class CrsMatrixHelper;
private:
ordinal_type _m; //!< # of rows
ordinal_type _n; //!< # of cols
size_type _nnz; //!< # of nonzeros
size_type_array _ap; //!< pointers to column index and values
ordinal_type_array _aj; //!< column index compressed format
value_type_array _ax; //!< values
public:
typedef typename GetCrsMatrixRowViewType< ValueType >::type row_view_type ;
typedef Kokkos::View<row_view_type*,space_type> row_view_type_array;
row_view_type_array _all_row_views ;
protected:
void createInternalArrays(const ordinal_type m,
const ordinal_type n,
const size_type nnz) {
_m = m;
_n = n;
_nnz = nnz;
if (static_cast<ordinal_type>(_ap.dimension_0()) < m+1)
_ap = size_type_array("CrsMatrixBase::RowPtrArray", m+1);
if (static_cast<size_type>(_aj.dimension_0()) < nnz)
_aj = ordinal_type_array("CrsMatrixBase::ColsArray", nnz);
if (static_cast<size_type>(_ax.dimension_0()) < nnz)
_ax = value_type_array("CrsMatrixBase::ValuesArray", nnz);
}
// Copy sparse matrix structure from coordinate format in 'mm'
// to CRS format in Views _ap, _aj, a_x.
void ijv2crs(const vector<ijv_type> &mm) {
ordinal_type ii = 0;
size_type jj = 0;
ijv_type prev = mm[0];
_ap[ii++] = 0;
_aj[jj] = prev.Col();
_ax[jj] = prev.Val();
++jj;
for (typename vector<ijv_type>::const_iterator it=(mm.begin()+1);it<mm.end();++it) {
ijv_type aij = (*it);
// row index
if (aij.Row() != prev.Row()) {
_ap[ii++] = jj;
}
if (aij == prev) {
--jj;
_aj[jj] = aij.Col();
_ax[jj] += aij.Val();
} else {
_aj[jj] = aij.Col();
_ax[jj] = aij.Val();
}
++jj;
prev = aij;
}
// add the last index to terminate the storage
_ap[ii++] = jj;
_nnz = jj;
}
public:
KOKKOS_INLINE_FUNCTION
void setNumNonZeros() {
if (_m)
_nnz = _ap[_m];
}
KOKKOS_INLINE_FUNCTION
ordinal_type NumRows() const { return _m; }
KOKKOS_INLINE_FUNCTION
ordinal_type NumCols() const { return _n; }
KOKKOS_INLINE_FUNCTION
size_type NumNonZeros() const { return _nnz; }
KOKKOS_INLINE_FUNCTION
size_type_array_ptr RowPtr() const { return &_ap[0]; }
KOKKOS_INLINE_FUNCTION
ordinal_type_array_ptr ColPtr() const { return &_aj[0]; }
KOKKOS_INLINE_FUNCTION
value_type_array_ptr ValuePtr() const { return &_ax[0];}
KOKKOS_INLINE_FUNCTION
size_type RowPtr(const ordinal_type i) const { return _ap[i]; }
KOKKOS_INLINE_FUNCTION
ordinal_type_array_ptr ColsInRow(const ordinal_type i) const { return _aj.data() + _ap[i] ; }
KOKKOS_INLINE_FUNCTION
value_type_array_ptr ValuesInRow(const ordinal_type i) const { return _ax.data() + _ap[i] ; }
KOKKOS_INLINE_FUNCTION
ordinal_type NumNonZerosInRow(const ordinal_type i) const { return (_ap[i+1] - _ap[i]); }
KOKKOS_INLINE_FUNCTION
value_type& Value(const ordinal_type k) { return _ax[k]; }
KOKKOS_INLINE_FUNCTION
value_type Value(const ordinal_type k) const { return _ax[k]; }
/// \brief Default constructor.
KOKKOS_INLINE_FUNCTION
CrsMatrixBase()
: _m(0),
_n(0),
_nnz(0),
_ap(),
_aj(),
_ax()
{ }
/// \brief Constructor with label
CrsMatrixBase(const string & )
: _m(0),
_n(0),
_nnz(0),
_ap(),
_aj(),
_ax()
{ }
/// \brief Copy constructor (shallow copy), for deep-copy use a method copy
template<typename VT,
typename OT,
typename ST,
typename SpT,
typename MT>
CrsMatrixBase(const CrsMatrixBase<VT,OT,ST,SpT,MT> &b)
: _m(b._m),
_n(b._n),
_nnz(b._nnz),
_ap(b._ap),
_aj(b._aj),
_ax(b._ax)
{ }
/// \brief Constructor to allocate internal data structures.
CrsMatrixBase(const string & ,
const ordinal_type m,
const ordinal_type n,
const ordinal_type nnz)
: _m(m),
_n(n),
_nnz(nnz),
_ap("CrsMatrixBase::RowPtrArray", m+1),
_aj("CrsMatrixBase::ColsArray", nnz),
_ax("CrsMatrixBase::ValuesArray", nnz)
{ }
/// \brief Constructor to attach external arrays to the matrix.
CrsMatrixBase(const string &,
const ordinal_type m,
const ordinal_type n,
const ordinal_type nnz,
const size_type_array &ap,
const ordinal_type_array &aj,
const value_type_array &ax)
: _m(m),
_n(n),
_nnz(nnz),
_ap(ap),
_aj(aj),
_ax(ax)
{ }
// Allow the copy function access to the input CrsMatrixBase
// private data.
template<typename, typename, typename, typename, typename>
friend class CrsMatrixBase ;
public:
/// \brief deep copy of matrix b, potentially different spaces
template< typename SpT >
int
copy(const CrsMatrixBase<ValueType,OrdinalType,SizeType,SpT,MemoryTraits> &b) {
space_type::execution_space::fence();
createInternalArrays(b._m, b._n, b._nnz);
space_type::execution_space::fence();
const auto ap_range = range_type<ordinal_type>(0, min(_ap.dimension_0(), b._ap.dimension_0()));
const auto aj_range = range_type<size_type> (0, min(_aj.dimension_0(), b._aj.dimension_0()));
const auto ax_range = range_type<size_type> (0, min(_ax.dimension_0(), b._ax.dimension_0()));
Kokkos::deep_copy(Kokkos::subview( _ap, ap_range),
Kokkos::subview(b._ap, ap_range));
Kokkos::deep_copy(Kokkos::subview( _aj, aj_range),
Kokkos::subview(b._aj, aj_range));
Kokkos::deep_copy(Kokkos::subview( _ax, ax_range),
Kokkos::subview(b._ax, ax_range));
space_type::execution_space::fence();
return 0;
}
/// \brief deep copy of lower/upper triangular of matrix b
int
copy(const int uplo,
const CrsMatrixBase &b) {
createInternalArrays(b._m, b._n, b._nnz);
// assume that matrix b is sorted.
switch (uplo) {
case Uplo::Lower: {
_nnz = 0;
for (ordinal_type i=0;i<_m;++i) {
size_type jbegin = b._ap[i];
size_type jend = b._ap[i+1];
_ap[i] = _nnz;
for (size_type j=jbegin;j<jend && (i >= b._aj[j]);++j,++_nnz) {
_aj[_nnz] = b._aj[j];
_ax[_nnz] = b._ax[j];
}
}
_ap[_m] = _nnz;
break;
}
case Uplo::Upper: {
_nnz = 0;
for (ordinal_type i=0;i<_m;++i) {
size_type j = b._ap[i];
size_type jend = b._ap[i+1];
_ap[i] = _nnz;
for ( ;j<jend && (i > b._aj[j]);++j) ;
for ( ;j<jend;++j,++_nnz) {
_aj[_nnz] = b._aj[j];
_ax[_nnz] = b._ax[j];
}
}
_ap[_m] = _nnz;
break;
}
}
return 0;
}
/// \brief deep copy of matrix b with given permutation vectors
template<typename VT,
typename OT,
typename ST,
typename SpT,
typename MT>
int
copy(const typename CrsMatrixBase<VT,OT,ST,SpT,MT>::ordinal_type_array &p,
const typename CrsMatrixBase<VT,OT,ST,SpT,MT>::ordinal_type_array &ip,
const CrsMatrixBase<VT,OT,ST,SpT,MT> &b) {
createInternalArrays(b._m, b._n, b._nnz);
// Question:: do I need to use Kokkos::vector ?
// in other words, where do we permute matrix in factoriztion ?
// permuting a matrix is a kernel ?
vector<ijv_type> tmp;
// any chance to use parallel_for ?
_nnz = 0;
for (ordinal_type i=0;i<_m;++i) {
ordinal_type ii = ip[i];
size_type jbegin = b._ap[ii];
size_type jend = b._ap[ii+1];
_ap[i] = _nnz;
for (size_type j=jbegin;j<jend;++j) {
ordinal_type jj = p[b._aj[j]];
ijv_type aij(i, jj, b._ax[j]);
tmp.push_back(aij);
}
sort(tmp.begin(), tmp.end(), less<ijv_type>());
for (auto it=tmp.begin();it<tmp.end();++it) {
ijv_type aij = (*it);
_aj[_nnz] = aij.Col();
_ax[_nnz] = aij.Val();
++_nnz;
}
tmp.clear();
}
_ap[_m] = _nnz;
return 0;
}
/// \brief add the matrix b into this non-zero entires
template<typename VT,
typename OT,
typename ST,
typename SpT,
typename MT>
int
add(const CrsMatrixBase<VT,OT,ST,SpT,MT> &b) {
const ordinal_type m = min(b._m, _m);
for (ordinal_type i=0;i<m;++i) {
const size_type jaend = _ap[i+1];
const size_type jbend = b._ap[i+1];
size_type ja = _ap[i];
size_type jb = b._ap[i];
for ( ;jb<jbend;++jb) {
for ( ;(_aj[ja]<b._aj[jb] && ja<jaend);++ja);
_ax[ja] += (_aj[ja] == b._aj[jb])*b._ax[jb];
}
}
return 0;
}
int symmetrize(const int uplo,
const bool conjugate = false) {
vector<ijv_type> mm;
mm.reserve(_nnz*2);
for (ordinal_type i=0;i<_m;++i) {
const size_type jbegin = _ap[i];
const size_type jend = _ap[i+1];
for (size_type jj=jbegin;jj<jend;++jj) {
const ordinal_type j = _aj[jj];
const value_type val = (conjugate ? conj(_ax[j]) : _ax[j]);
if (uplo == Uplo::Lower && i > j) {
mm.push_back(ijv_type(i, j, val));
mm.push_back(ijv_type(j, i, val));
} else if (uplo == Uplo::Upper && i < j) {
mm.push_back(ijv_type(i, j, val));
mm.push_back(ijv_type(j, i, val));
} else if (i == j) {
mm.push_back(ijv_type(i, i, val));
}
}
}
sort(mm.begin(), mm.end(), less<ijv_type>());
createInternalArrays(_m, _n, mm.size());
ijv2crs(mm);
return 0;
}
int hermitianize(int uplo) {
return symmetrize(uplo, true);
}
ostream& showMe(ostream &os) const {
streamsize prec = os.precision();
os.precision(8);
os << scientific;
os << " -- CrsMatrixBase -- " << endl
<< " # of Rows = " << _m << endl
<< " # of Cols = " << _n << endl
<< " # of NonZeros = " << _nnz << endl
<< endl
<< " RowPtrArray length = " << _ap.dimension_0() << endl
<< " ColArray length = " << _aj.dimension_0() << endl
<< " ValueArray length = " << _ax.dimension_0() << endl
<< endl;
const int w = 10;
if (_ap.size() && _aj.size() && _ax.size()) {
os << setw(w) << "Row" << " "
<< setw(w) << "Col" << " "
<< setw(w) << "Val" << endl;
for (ordinal_type i=0;i<_m;++i) {
size_type jbegin = _ap[i], jend = _ap[i+1];
for (size_type j=jbegin;j<jend;++j) {
value_type val = _ax[j];
os << setw(w) << i << " "
<< setw(w) << _aj[j] << " "
<< setw(w) << val << endl;
}
}
}
os.unsetf(ios::scientific);
os.precision(prec);
return os;
}
int importMatrixMarket(ifstream &file) {
vector<ijv_type> mm;
const ordinal_type mm_base = 1;
{
string header;
if (file.is_open()) {
getline(file, header);
while (file.good()) {
char c = file.peek();
if (c == '%' || c == '\n') {
file.ignore(256, '\n');
continue;
}
break;
}
} else {
ERROR(MSG_INVALID_INPUT(file));
}
// check the header
bool symmetry = (header.find("symmetric") != string::npos);
// read matrix specification
ordinal_type m, n;
size_type nnz;
file >> m >> n >> nnz;
mm.reserve(nnz*(symmetry ? 2 : 1));
for (size_type i=0;i<nnz;++i) {
ordinal_type row, col;
value_type val;
file >> row >> col >> val;
row -= mm_base;
col -= mm_base;
mm.push_back(ijv_type(row, col, val));
if (symmetry && row != col)
mm.push_back(ijv_type(col, row, val));
}
sort(mm.begin(), mm.end(), less<ijv_type>());
// construct workspace and set variables
createInternalArrays(m, n, mm.size());
}
// change mm to crs
ijv2crs(mm);
return 0;
}
int exportMatrixMarket(ofstream &file,
const string comment,
const int uplo = 0) {
streamsize prec = file.precision();
file.precision(8);
file << scientific;
file << "%%MatrixMarket matrix coordinate "
<< (is_fundamental<value_type>::value ? "real " : "complex ")
<< ((uplo == Uplo::Upper || uplo == Uplo::Lower) ? "symmetric " : "general ")
<< endl;
file << comment << endl;
// cnt nnz
size_type nnz = 0;
for (ordinal_type i=0;i<_m;++i) {
const size_type jbegin = _ap[i], jend = _ap[i+1];
for (size_type j=jbegin;j<jend;++j) {
if (uplo == Uplo::Upper && i <= _aj[j]) ++nnz;
if (uplo == Uplo::Lower && i >= _aj[j]) ++nnz;
if (!uplo) ++nnz;
}
}
file << _m << " " << _n << " " << nnz << endl;
const int w = 10;
for (ordinal_type i=0;i<_m;++i) {
const size_type jbegin = _ap[i], jend = _ap[i+1];
for (size_type j=jbegin;j<jend;++j) {
bool flag = false;
if (uplo == Uplo::Upper && i <= _aj[j]) flag = true;
if (uplo == Uplo::Lower && i >= _aj[j]) flag = true;
if (!uplo) flag = true;
if (flag) {
value_type val = _ax[j];
file << setw(w) << ( i+1) << " "
<< setw(w) << (_aj[j]+1) << " "
<< setw(w) << val << endl;
}
}
}
file.unsetf(ios::scientific);
file.precision(prec);
return 0;
}
//----------------------------------------------------------------------
int convertGraph(size_type_array rptr,
ordinal_type_array cidx) const {
ordinal_type ii = 0;
size_type jj = 0;
for (ordinal_type i=0;i<_m;++i) {
size_type jbegin = _ap[i], jend = _ap[i+1];
rptr[ii++] = jj;
for (size_type j=jbegin;j<jend;++j)
if (i != _aj[j])
cidx[jj++] = _aj[j];
}
rptr[ii] = jj;
return 0;
}
//----------------------------------------------------------------------
};
}
#endif

Event Timeline

c4science · Help