inline
void
op_sp_plus::apply(SpMat<typename T1::elem_type>& out, const SpToDOp<T1,op_sp_plus>& in)
{
arma_extra_debug_sigprint();
typedef typename T1::elem_type eT;
// Note that T1 will be a sparse type, so we use SpProxy.
const SpProxy<T1> proxy(in.m);
const uword n_rows = proxy.get_n_rows();
const uword n_cols = proxy.get_n_cols();
out.set_size(n_rows, n_cols);
const eT k = in.aux;
// We have to loop over all the elements.
for(uword c = 0; c < n_cols; ++c)
for(uword r = 0; r < n_rows; ++r)
{
out.at(r, c) = proxy.at(r, c) + k;
}
}
inline
void
spop_repmat::apply(SpMat<typename T1::elem_type>& out, const SpOp<T1, spop_repmat>& in)
{
arma_extra_debug_sigprint();
typedef typename T1::elem_type eT;
const unwrap_spmat<T1> U(in.m);
const SpMat<eT>& X = U.M;
const uword X_n_rows = X.n_rows;
const uword X_n_cols = X.n_cols;
const uword copies_per_row = in.aux_uword_a;
const uword copies_per_col = in.aux_uword_b;
// out.set_size(X_n_rows * copies_per_row, X_n_cols * copies_per_col);
//
// const uword out_n_rows = out.n_rows;
// const uword out_n_cols = out.n_cols;
//
// if( (out_n_rows > 0) && (out_n_cols > 0) )
// {
// for(uword col = 0; col < out_n_cols; col += X_n_cols)
// for(uword row = 0; row < out_n_rows; row += X_n_rows)
// {
// out.submat(row, col, row+X_n_rows-1, col+X_n_cols-1) = X;
// }
// }
SpMat<eT> tmp(X_n_rows * copies_per_row, X_n_cols);
if(tmp.n_elem > 0)
{
for(uword row = 0; row < tmp.n_rows; row += X_n_rows)
{
tmp.submat(row, 0, row+X_n_rows-1, X_n_cols-1) = X;
}
}
// tmp contains copies of the input matrix, so no need to check for aliasing
out.set_size(X_n_rows * copies_per_row, X_n_cols * copies_per_col);
const uword out_n_rows = out.n_rows;
const uword out_n_cols = out.n_cols;
if( (out_n_rows > 0) && (out_n_cols > 0) )
{
for(uword col = 0; col < out_n_cols; col += X_n_cols)
{
out.submat(0, col, out_n_rows-1, col+X_n_cols-1) = tmp;
}
}
}
arma_hot
inline
void
spop_htrans::apply(SpMat<typename T1::elem_type>& out, const SpOp<T1,spop_htrans>& in, const typename arma_cx_only<typename T1::elem_type>::result* junk)
{
arma_extra_debug_sigprint();
arma_ignore(junk);
typedef typename T1::elem_type eT;
typedef typename umat::elem_type ueT;
const SpProxy<T1> p(in.m);
const uword N = p.get_n_nonzero();
if(N == uword(0))
{
out.set_size(p.get_n_cols(), p.get_n_rows());
return;
}
umat locs(2, N);
Col<eT> vals(N);
eT* vals_ptr = vals.memptr();
typename SpProxy<T1>::const_iterator_type it = p.begin();
for(uword count = 0; count < N; ++count)
{
ueT* locs_ptr = locs.colptr(count);
locs_ptr[0] = it.col();
locs_ptr[1] = it.row();
vals_ptr[count] = std::conj(*it);
++it;
}
SpMat<eT> tmp(locs, vals, p.get_n_cols(), p.get_n_rows());
out.steal_mem(tmp);
}
arma_hot
inline
void
spglue_plus::apply_noalias(SpMat<eT>& out, 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(), "addition");
if( (pa.get_n_nonzero() != 0) && (pb.get_n_nonzero() != 0) )
{
out.set_size(pa.get_n_rows(), pa.get_n_cols());
// Resize memory to correct size.
out.mem_resize(n_unique(pa, pb, op_n_unique_add()));
// 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();
typename SpProxy<T1>::const_iterator_type x_end = pa.end();
typename SpProxy<T2>::const_iterator_type y_end = pb.end();
uword cur_val = 0;
while( (x_it != x_end) || (y_it != y_end) )
{
if(x_it == y_it)
{
const eT val = (*x_it) + (*y_it);
if (val != eT(0))
{
access::rw(out.values[cur_val]) = val;
access::rw(out.row_indices[cur_val]) = x_it.row();
++access::rw(out.col_ptrs[x_it.col() + 1]);
++cur_val;
}
++x_it;
++y_it;
}
else
{
const uword x_it_row = x_it.row();
const uword x_it_col = x_it.col();
const uword y_it_row = y_it.row();
const uword y_it_col = y_it.col();
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(out.values[cur_val]) = (*x_it);
access::rw(out.row_indices[cur_val]) = x_it_row;
++access::rw(out.col_ptrs[x_it_col + 1]);
++cur_val;
++x_it;
}
else
{
access::rw(out.values[cur_val]) = (*y_it);
access::rw(out.row_indices[cur_val]) = y_it_row;
++access::rw(out.col_ptrs[y_it_col + 1]);
++cur_val;
++y_it;
}
}
}
const uword out_n_cols = out.n_cols;
uword* col_ptrs = access::rwp(out.col_ptrs);
// Fix column pointers to be cumulative.
for(uword c = 1; c <= out_n_cols; ++c)
{
col_ptrs[c] += col_ptrs[c - 1];
}
}
else
{
if(pa.get_n_nonzero() == 0)
{
out = pb.Q;
return;
}
if(pb.get_n_nonzero() == 0)
{
out = pa.Q;
return;
}
}
}
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
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];
}
}