phynexis-deps/armadillo-code
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

simplification + speedup

Browse Source
This commit is contained in:
conrad
2026-07-01 14:45:17 +10:00
parent c6f4af01d8
commit 23f652e542
2 changed files with 18 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/armadillo_bits/op_vectorise_bones.hpp
-2
View File
@@ -34,8 +34,6 @@ struct op_vectorise_col
template<typename T1> inline static void apply_direct(Mat_noalias<typename T1::elem_type>& out, const T1& expr);
template<typename eT> inline static void apply_subview(Mat<eT>& out, const subview<eT>& sv);
template<typename T1> inline static void apply_proxy(Mat<typename T1::elem_type>& out, const Proxy<T1>& P);
};
include/armadillo_bits/op_vectorise_meat.hpp
+18 -106
View File
@@ -80,39 +80,12 @@ op_vectorise_col::apply_direct(Mat<typename T1::elem_type>& out, const T1& expr)
}
}
else
if( (quasi_unwrap<T1>::has_orig_mem) || (is_Mat<typename Proxy<T1>::stored_type>::value) || (arma_config::openmp && Proxy<T1>::use_mp) )
{
const quasi_unwrap<T1> U(expr);
Mat<eT> tmp = expr;
if(U.is_alias(out))
{
Mat<eT> tmp(U.M.memptr(), U.M.n_elem, 1);
out.steal_mem(tmp);
}
else
{
out.set_size(U.M.n_elem, 1);
arrayops::copy(out.memptr(), U.M.memptr(), U.M.n_elem);
}
}
else
{
const Proxy<T1> P(expr);
tmp.set_size(tmp.n_elem, 1);
if(P.is_alias(out))
{
Mat<eT> tmp;
op_vectorise_col::apply_proxy(tmp, P);
out.steal_mem(tmp);
}
else
{
op_vectorise_col::apply_proxy(out, P);
}
out.steal_mem(tmp);
}
}
@@ -133,12 +106,23 @@ op_vectorise_col::apply(Mat_noalias<typename T1::elem_type>& out, const Op<T1,op
template<typename T1>
inline
void
op_vectorise_col::apply_direct(Mat_noalias<typename T1::elem_type>& out, const T1& expr)
op_vectorise_col::apply_direct(Mat_noalias<typename T1::elem_type>& actual_out, const T1& expr)
{
arma_debug_sigprint();
typedef typename T1::elem_type eT;
Mat<eT>& out = actual_out;
if(is_Mat<T1>::value)
{
const plain_unwrap<T1> U(expr);
out.set_size(U.M.n_elem, 1);
arrayops::copy(out.memptr(), U.M.memptr(), U.M.n_elem);
}
else
if(is_subview<T1>::value)
{
const subview<eT>& sv = reinterpret_cast< const subview<eT>& >(expr);
@@ -146,19 +130,12 @@ op_vectorise_col::apply_direct(Mat_noalias<typename T1::elem_type>& out, const T
op_vectorise_col::apply_subview(out, sv);
}
else
if( (quasi_unwrap<T1>::has_orig_mem) || (is_Mat<typename Proxy<T1>::stored_type>::value) || (arma_config::openmp && Proxy<T1>::use_mp) )
{
const quasi_unwrap<T1> U(expr);
Mat<eT> tmp = expr;
out.set_size(U.M.n_elem, 1);
tmp.set_size(tmp.n_elem, 1);
arrayops::copy(out.memptr(), U.M.memptr(), U.M.n_elem);
}
else
{
const Proxy<T1> P(expr);
op_vectorise_col::apply_proxy(out, P);
out.steal_mem(tmp);
}
}
@@ -190,71 +167,6 @@ op_vectorise_col::apply_subview(Mat<eT>& out, const subview<eT>& sv)
template<typename T1>
inline
void
op_vectorise_col::apply_proxy(Mat<typename T1::elem_type>& out, const Proxy<T1>& P)
{
arma_debug_sigprint();
typedef typename T1::elem_type eT;
const uword N = P.get_n_elem();
out.set_size(N, 1);
if(N == 0) { return; }
eT* outmem = out.memptr();
if(Proxy<T1>::use_at == false)
{
// TODO: add handling of aligned access ?
typename Proxy<T1>::ea_type A = P.get_ea();
uword i,j;
for(i=0, j=1; j < N; i+=2, j+=2)
{
const eT tmp_i = A[i];
const eT tmp_j = A[j];
outmem[i] = tmp_i;
outmem[j] = tmp_j;
}
if(i < N)
{
outmem[i] = A[i];
}
}
else
{
const uword n_rows = P.get_n_rows();
const uword n_cols = P.get_n_cols();
if(n_rows == 1)
{
for(uword i=0; i < n_cols; ++i)
{
outmem[i] = P.at(0,i);
}
}
else
{
for(uword col=0; col < n_cols; ++col)
for(uword row=0; row < n_rows; ++row)
{
*outmem = P.at(row,col);
outmem++;
}
}
}
}
template<typename T1>
inline
void