inline
arma_warn_unused
typename
enable_if2
<(is_arma_sparse_type<T1>::value) && (is_arma_sparse_type<T2>::value) && (is_same_type<typename T1::elem_type, typename T2::elem_type>::value),
typename T1::elem_type
>::result
dot
(
const SpBase<typename T1::elem_type, T1>& x,
const SpBase<typename T2::elem_type, T2>& y
)
{
arma_extra_debug_sigprint();
const SpProxy<T1> pa(x.get_ref());
const SpProxy<T2> pb(y.get_ref());
arma_debug_assert_same_size(pa.get_n_rows(), pa.get_n_cols(), pb.get_n_rows(), pb.get_n_cols(), "dot()");
typedef typename T1::elem_type eT;
if((&(x.get_ref()) == &(y.get_ref())) && (SpProxy<T1>::must_use_iterator == false))
{
// We can do it directly!
return op_dot::direct_dot_arma(pa.get_n_nonzero(), pa.get_values(), pa.get_values());
}
else
{
// Iterate over both objects and see when they are the same
eT result = eT(0);
typename SpProxy<T1>::const_iterator_type a_it = pa.begin();
typename SpProxy<T2>::const_iterator_type b_it = pb.begin();
while((a_it.pos() < pa.get_n_nonzero()) && (b_it.pos() < pb.get_n_nonzero()))
{
if(a_it == b_it)
{
result += (*a_it) * (*b_it);
++a_it;
++b_it;
}
else if((a_it.col() < b_it.col()) || ((a_it.col() == b_it.col()) && (a_it.row() < b_it.row())))
{
// a_it is "behind"
++a_it;
}
else
{
// b_it is "behind"
++b_it;
}
}
return result;
}
}
inline
arma_warn_unused
typename
enable_if2
<(is_arma_type<T1>::value) && (is_arma_sparse_type<T2>::value) && (is_same_type<typename T1::elem_type, typename T2::elem_type>::value),
typename T1::elem_type
>::result
dot
(
const Base<typename T1::elem_type, T1>& x,
const SpBase<typename T2::elem_type, T2>& y
)
{
arma_extra_debug_sigprint();
const Proxy<T1> pa(x.get_ref());
const SpProxy<T2> pb(y.get_ref());
arma_debug_assert_same_size(pa.get_n_rows(), pa.get_n_cols(), pb.get_n_rows(), pb.get_n_cols(), "dot()");
typedef typename T1::elem_type eT;
eT result = eT(0);
typename SpProxy<T2>::const_iterator_type it = pb.begin();
// prefer_at_accessor won't save us operations
while(it.pos() < pb.get_n_nonzero())
{
result += (*it) * pa.at(it.row(), it.col());
++it;
}
return result;
}
inline
typename
enable_if2
<
(is_arma_type<T1>::value && is_arma_sparse_type<T2>::value && is_same_type<typename T1::elem_type, typename T2::elem_type>::value),
Mat<typename T1::elem_type>
>::result
operator/
(
const Base<typename T1::elem_type, T1>& x,
const SpBase<typename T2::elem_type, T2>& y
)
{
arma_extra_debug_sigprint();
const Proxy<T1> pa(x.get_ref());
const SpProxy<T2> pb(y.get_ref());
arma_debug_assert_same_size(pa.get_n_rows(), pa.get_n_cols(), pb.get_n_rows(), pb.get_n_cols(), "element-wise division");
Mat<typename T1::elem_type> result(pa.get_n_rows(), pa.get_n_cols());
result.fill(Datum<typename T1::elem_type>::inf);
// Now divide each element
typename SpProxy<T2>::const_iterator_type it = pb.begin();
while(it.pos() < pb.get_n_nonzero())
{
if(Proxy<T1>::prefer_at_accessor == false)
{
const uword index = (it.col() * result.n_rows) + it.row();
result[index] = pa[index] / (*it);
}
else
{
result.at(it.row(), it.col()) = pa.at(it.row(), it.col()) / (*it);
}
++it;
}
return result;
}
inline
void
spop_mean::apply_noalias_slow
(
SpMat<typename T1::elem_type>& out,
const SpProxy<T1>& p,
const uword dim
)
{
arma_extra_debug_sigprint();
typedef typename T1::elem_type eT;
const uword p_n_rows = p.get_n_rows();
const uword p_n_cols = p.get_n_cols();
if(dim == 0) // find the mean in each column
{
arma_extra_debug_print("spop_mean::apply_noalias(): dim = 0");
out.set_size((p_n_rows > 0) ? 1 : 0, p_n_cols);
if( (p_n_rows == 0) || (p.get_n_nonzero() == 0) ) { return; }
for(uword col = 0; col < p_n_cols; ++col)
{
// Do we have to use an iterator or can we use memory directly?
if(SpProxy<T1>::must_use_iterator)
{
typename SpProxy<T1>::const_iterator_type it = p.begin_col(col);
typename SpProxy<T1>::const_iterator_type end = p.begin_col(col + 1);
const uword n_zero = p_n_rows - (end.pos() - it.pos());
out.at(0,col) = spop_mean::iterator_mean(it, end, n_zero, eT(0));
}
else
{
out.at(0,col) = spop_mean::direct_mean
(
&p.get_values()[p.get_col_ptrs()[col]],
p.get_col_ptrs()[col + 1] - p.get_col_ptrs()[col],
p_n_rows
);
}
}
}
else
if(dim == 1) // find the mean in each row
{
arma_extra_debug_print("spop_mean::apply_noalias(): dim = 1");
out.set_size(p_n_rows, (p_n_cols > 0) ? 1 : 0);
if( (p_n_cols == 0) || (p.get_n_nonzero() == 0) ) { return; }
for(uword row = 0; row < p_n_rows; ++row)
{
// We must use an iterator regardless of how it is stored.
typename SpProxy<T1>::const_row_iterator_type it = p.begin_row(row);
typename SpProxy<T1>::const_row_iterator_type end = p.end_row(row);
const uword n_zero = p_n_cols - (end.pos() - it.pos());
out.at(row,0) = spop_mean::iterator_mean(it, end, n_zero, eT(0));
}
}
}
inline
void
spop_var::apply_noalias
(
SpMat<typename T1::pod_type>& out_ref,
const SpProxy<T1>& p,
const uword norm_type,
const uword dim
)
{
arma_extra_debug_sigprint();
typedef typename T1::elem_type in_eT;
//typedef typename T1::pod_type out_eT;
const uword p_n_rows = p.get_n_rows();
const uword p_n_cols = p.get_n_cols();
if(dim == 0)
{
arma_extra_debug_print("spop_var::apply(), dim = 0");
arma_debug_check((p_n_rows == 0), "var(): given object has zero rows");
out_ref.set_size(1, p_n_cols);
for(uword col = 0; col < p_n_cols; ++col)
{
if(SpProxy<T1>::must_use_iterator == true)
{
// We must use an iterator; we can't access memory directly.
typename SpProxy<T1>::const_iterator_type it = p.begin_col(col);
typename SpProxy<T1>::const_iterator_type end = p.begin_col(col + 1);
const uword n_zero = p.get_n_rows() - (end.pos() - it.pos());
// in_eT is used just to get the specialization right (complex / noncomplex)
out_ref.at(col) = spop_var::iterator_var(it, end, n_zero, norm_type, in_eT(0));
}
else
{
// We can use direct memory access to calculate the variance.
out_ref.at(col) = spop_var::direct_var
(
&p.get_values()[p.get_col_ptrs()[col]],
p.get_col_ptrs()[col + 1] - p.get_col_ptrs()[col],
p.get_n_rows(),
norm_type
);
}
}
}
else if(dim == 1)
{
arma_extra_debug_print("spop_var::apply_noalias(), dim = 1");
arma_debug_check((p_n_cols == 0), "var(): given object has zero columns");
out_ref.set_size(p_n_rows, 1);
for(uword row = 0; row < p_n_rows; ++row)
{
// We have to use an iterator here regardless of whether or not we can
// directly access memory.
typename SpProxy<T1>::const_row_iterator_type it = p.begin_row(row);
typename SpProxy<T1>::const_row_iterator_type end = p.end_row(row);
const uword n_zero = p.get_n_cols() - (end.pos() - it.pos());
out_ref.at(row) = spop_var::iterator_var(it, end, n_zero, norm_type, in_eT(0));
}
}
}
inline
typename
enable_if2
<
(is_arma_sparse_type<T1>::value && is_arma_sparse_type<T2>::value && is_same_type<typename T1::elem_type, typename T2::elem_type>::value),
SpMat<typename T1::elem_type>
>::result
operator%
(
const SpBase<typename T1::elem_type, T1>& x,
const SpBase<typename T2::elem_type, T2>& y
)
{
arma_extra_debug_sigprint();
const SpProxy<T1> pa(x.get_ref());
const SpProxy<T2> pb(y.get_ref());
arma_debug_assert_same_size(pa.get_n_rows(), pa.get_n_cols(), pb.get_n_rows(), pb.get_n_cols(), "element-wise multiplication");
SpMat<typename T1::elem_type> result(pa.get_n_rows(), pa.get_n_cols());
// Resize memory to correct size.
result.mem_resize(n_unique(x, y, op_n_unique_mul()));
// Now iterate across both matrices.
typename SpProxy<T1>::const_iterator_type x_it = pa.begin();
typename SpProxy<T2>::const_iterator_type y_it = pb.begin();
uword cur_val = 0;
while((x_it.pos() < pa.get_n_nonzero()) || (y_it.pos() < pb.get_n_nonzero()))
{
if(x_it == y_it)
{
const typename T1::elem_type val = (*x_it) * (*y_it);
if (val != 0)
{
access::rw(result.values[cur_val]) = val;
access::rw(result.row_indices[cur_val]) = x_it.row();
++access::rw(result.col_ptrs[x_it.col() + 1]);
++cur_val;
}
++x_it;
++y_it;
}
else
{
if((x_it.col() < y_it.col()) || ((x_it.col() == y_it.col()) && (x_it.row() < y_it.row()))) // if y is closer to the end
{
++x_it;
}
else
{
++y_it;
}
}
}
// Fix column pointers to be cumulative.
for(uword c = 1; c <= result.n_cols; ++c)
{
access::rw(result.col_ptrs[c]) += result.col_ptrs[c - 1];
}
return result;
}
inline
typename
enable_if2
<
(is_arma_type<T1>::value && is_arma_sparse_type<T2>::value && is_same_type<typename T1::elem_type, typename T2::elem_type>::value),
SpMat<typename T1::elem_type>
>::result
operator%
(
const Base<typename T1::elem_type, T1>& x,
const SpBase<typename T2::elem_type, T2>& y
)
{
arma_extra_debug_sigprint();
const Proxy<T1> pa(x.get_ref());
const SpProxy<T2> pb(y.get_ref());
arma_debug_assert_same_size(pa.get_n_rows(), pa.get_n_cols(), pb.get_n_rows(), pb.get_n_cols(), "element-wise multiplication");
SpMat<typename T1::elem_type> result(pa.get_n_rows(), pa.get_n_cols());
if(Proxy<T1>::prefer_at_accessor == false)
{
// use direct operator[] access
// count new size
uword new_n_nonzero = 0;
typename SpProxy<T2>::const_iterator_type it = pb.begin();
while(it.pos() < pb.get_n_nonzero())
{
if(((*it) * pa[(it.col() * pa.get_n_rows()) + it.row()]) != 0)
{
++new_n_nonzero;
}
++it;
}
// Resize memory accordingly.
result.mem_resize(new_n_nonzero);
uword cur_val = 0;
typename SpProxy<T2>::const_iterator_type it2 = pb.begin();
while(it2.pos() < pb.get_n_nonzero())
{
const typename T1::elem_type val = (*it2) * pa[(it2.col() * pa.get_n_rows()) + it2.row()];
if(val != 0)
{
access::rw(result.values[cur_val]) = val;
access::rw(result.row_indices[cur_val]) = it2.row();
++access::rw(result.col_ptrs[it2.col() + 1]);
++cur_val;
}
++it2;
}
}
else
{
// use at() access
// count new size
uword new_n_nonzero = 0;
typename SpProxy<T2>::const_iterator_type it = pb.begin();
while(it.pos() < pb.get_n_nonzero())
{
if(((*it) * pa.at(it.row(), it.col())) != 0)
{
++new_n_nonzero;
}
++it;
}
// Resize memory accordingly.
result.mem_resize(new_n_nonzero);
uword cur_val = 0;
typename SpProxy<T2>::const_iterator_type it2 = pb.begin();
while(it2.pos() < pb.get_n_nonzero())
{
const typename T1::elem_type val = (*it2) * pa.at(it2.row(), it2.col());
if(val != 0)
{
access::rw(result.values[cur_val]) = val;
access::rw(result.row_indices[cur_val]) = it2.row();
++access::rw(result.col_ptrs[it2.col() + 1]);
++cur_val;
}
++it2;
}
}
// Fix column pointers.
for(uword c = 1; c <= result.n_cols; ++c)
{
access::rw(result.col_ptrs[c]) += result.col_ptrs[c - 1];
}
return result;
}
inline
void
spop_var::apply_noalias
(
SpMat<typename T1::pod_type>& out,
const SpProxy<T1>& p,
const uword norm_type,
const uword dim
)
{
arma_extra_debug_sigprint();
typedef typename T1::elem_type in_eT;
//typedef typename T1::pod_type out_eT;
const uword p_n_rows = p.get_n_rows();
const uword p_n_cols = p.get_n_cols();
// TODO: this is slow; rewrite based on the approach used by sparse mean()
if(dim == 0) // find variance in each column
{
arma_extra_debug_print("spop_var::apply_noalias(): dim = 0");
out.set_size((p_n_rows > 0) ? 1 : 0, p_n_cols);
if( (p_n_rows == 0) || (p.get_n_nonzero() == 0) ) { return; }
for(uword col = 0; col < p_n_cols; ++col)
{
if(SpProxy<T1>::must_use_iterator)
{
// We must use an iterator; we can't access memory directly.
typename SpProxy<T1>::const_iterator_type it = p.begin_col(col);
typename SpProxy<T1>::const_iterator_type end = p.begin_col(col + 1);
const uword n_zero = p_n_rows - (end.pos() - it.pos());
// in_eT is used just to get the specialization right (complex / noncomplex)
out.at(0, col) = spop_var::iterator_var(it, end, n_zero, norm_type, in_eT(0));
}
else
{
// We can use direct memory access to calculate the variance.
out.at(0, col) = spop_var::direct_var
(
&p.get_values()[p.get_col_ptrs()[col]],
p.get_col_ptrs()[col + 1] - p.get_col_ptrs()[col],
p_n_rows,
norm_type
);
}
}
}
else
if(dim == 1) // find variance in each row
{
arma_extra_debug_print("spop_var::apply_noalias(): dim = 1");
out.set_size(p_n_rows, (p_n_cols > 0) ? 1 : 0);
if( (p_n_cols == 0) || (p.get_n_nonzero() == 0) ) { return; }
for(uword row = 0; row < p_n_rows; ++row)
{
// We have to use an iterator here regardless of whether or not we can
// directly access memory.
typename SpProxy<T1>::const_row_iterator_type it = p.begin_row(row);
typename SpProxy<T1>::const_row_iterator_type end = p.end_row(row);
const uword n_zero = p_n_cols - (end.pos() - it.pos());
out.at(row, 0) = spop_var::iterator_var(it, end, n_zero, norm_type, in_eT(0));
}
}
}
arma_hot
inline
void
spglue_minus::apply_noalias(SpMat<eT>& result, const SpProxy<T1>& pa, const SpProxy<T2>& pb)
{
arma_extra_debug_sigprint();
arma_debug_assert_same_size(pa.get_n_rows(), pa.get_n_cols(), pb.get_n_rows(), pb.get_n_cols(), "subtraction");
result.set_size(pa.get_n_rows(), pa.get_n_cols());
// Resize memory to correct size.
result.mem_resize(n_unique(pa, pb, op_n_unique_sub()));
// Now iterate across both matrices.
typename SpProxy<T1>::const_iterator_type x_it = pa.begin();
typename SpProxy<T2>::const_iterator_type y_it = pb.begin();
uword cur_val = 0;
while((x_it.pos() < pa.get_n_nonzero()) || (y_it.pos() < pb.get_n_nonzero()))
{
if(x_it == y_it)
{
const typename T1::elem_type val = (*x_it) - (*y_it);
if (val != 0)
{
access::rw(result.values[cur_val]) = val;
access::rw(result.row_indices[cur_val]) = x_it.row();
++access::rw(result.col_ptrs[x_it.col() + 1]);
++cur_val;
}
++x_it;
++y_it;
}
else
{
if((x_it.col() < y_it.col()) || ((x_it.col() == y_it.col()) && (x_it.row() < y_it.row()))) // if y is closer to the end
{
access::rw(result.values[cur_val]) = (*x_it);
access::rw(result.row_indices[cur_val]) = x_it.row();
++access::rw(result.col_ptrs[x_it.col() + 1]);
++cur_val;
++x_it;
}
else
{
access::rw(result.values[cur_val]) = -(*y_it);
access::rw(result.row_indices[cur_val]) = y_it.row();
++access::rw(result.col_ptrs[y_it.col() + 1]);
++cur_val;
++y_it;
}
}
}
// Fix column pointers to be cumulative.
for(uword c = 1; c <= result.n_cols; ++c)
{
access::rw(result.col_ptrs[c]) += result.col_ptrs[c - 1];
}
}
inline
typename
enable_if2
<
(is_arma_sparse_type<T1>::value && is_arma_type<T2>::value && is_same_type<typename T1::elem_type, typename T2::elem_type>::value),
SpMat<typename T1::elem_type>
>::result
operator/
(
const SpBase<typename T1::elem_type, T1>& x,
const Base<typename T2::elem_type, T2>& y
)
{
arma_extra_debug_sigprint();
const SpProxy<T1> pa(x.get_ref());
const Proxy<T2> pb(y.get_ref());
arma_debug_assert_same_size(pa.get_n_rows(), pa.get_n_cols(), pb.get_n_rows(), pb.get_n_cols(), "element-wise division");
SpMat<typename T1::elem_type> result(pa.get_n_rows(), pa.get_n_cols());
// The compiler should be smart enough to optimize out the inner if/else statement entirely
typename SpProxy<T1>::const_iterator_type it = pa.begin();
uword new_n_nonzero;
while(it.pos() < pa.get_n_nonzero())
{
if(Proxy<T2>::prefer_at_accessor == false)
{
const typename T1::elem_type val = (*it) / pb[(it.col() * pb.get_n_rows()) + it.row()];
if(val != 0)
{
++new_n_nonzero;
}
}
else
{
const typename T1::elem_type val = (*it) / pb.at(it.row(), it.col());
if(val != 0)
{
++new_n_nonzero;
}
}
++it;
}
result.mem_resize(new_n_nonzero);
typename SpProxy<T1>::const_iterator_type it2 = pa.begin();
uword cur_pos = 0;
while(it2.pos() < pa.get_n_nonzero())
{
if(Proxy<T2>::prefer_at_accessor == false)
{
const typename T1::elem_type val = (*it2) / pb[(it2.col() * pb.get_n_rows()) + it2.row()];
if(val != 0)
{
access::rw(result.values[cur_pos]) = val;
access::rw(result.row_indices[cur_pos]) = it2.row();
++access::rw(result.col_ptrs[it2.col() + 1]);
++cur_pos;
}
}
else
{
const typename T1::elem_type val = (*it2) / pb.at(it2.row(), it2.col());
if(val != 0)
{
access::rw(result.values[cur_pos]) = val;
access::rw(result.row_indices[cur_pos]) = it2.row();
++access::rw(result.col_ptrs[it2.col() + 1]);
++cur_pos;
}
}
++it2;
}
// Fix column pointers
for(uword col = 1; col <= result.n_cols; ++col)
{
access::rw(result.col_ptrs[col]) += result.col_ptrs[col - 1];
}
return result;
}
