octave-lyh: liboctave/MArray.cc comparison

optimize accumdim with summation

author	Jaroslav Hajek <highegg@gmail.com>
date	Fri, 05 Mar 2010 12:31:30 +0100
parents	b47ab50a6aa8
children	fd0a3ac60b0e

comparison

equal deleted inserted replaced

-:aeb5b1e47978
+:a0b51ac0f88a
 octave_quit ();
 octave_idx_type len = std::min (idx.length (n), vals.length ());
 idx.loop (len, _idxbinop_helper<T, xmax> (this->fortran_vec (), vals.data ()));
+}
+#include <iostream>
+template <class T>
+void MArray<T>::idx_add_nd (const idx_vector& idx, const MArray<T>& vals, int dim)
+{
+int nd = std::max (this->ndims (), vals.ndims ());
+if (dim < 0)
+dim = vals.dims ().first_non_singleton ();
+else if (dim > nd)
+nd = dim;
+// Check dimensions.
+dim_vector ddv = Array<T>::dims ().redim (nd);
+dim_vector sdv = vals.dims ().redim (nd);
+octave_idx_type ext = idx.extent (ddv (dim));
+if (ext > ddv(dim))
+{
+ddv(dim) = ext;
+Array<T>::resize (ddv);
+ext = ddv(dim);
+}
+octave_idx_type l,n,u,ns;
+get_extent_triplet (ddv, dim, l, n, u);
+ns = sdv(dim);
+sdv(dim) = ddv(dim) = 0;
+if (ddv != sdv)
+(*current_liboctave_error_handler)
+("accumdim: dimension mismatch");
+T *dst = Array<T>::fortran_vec ();
+const T *src = vals.data ();
+octave_idx_type len = idx.length (ns);
+if (l == 1)
+{
+for (octave_idx_type j = 0; j < u; j++)
+{
+octave_quit ();
+idx.loop (len, _idxadda_helper<T> (dst + j*n, src + j*ns));
+}
+}
+else
+{
+for (octave_idx_type j = 0; j < u; j++)
+{
+octave_quit ();
+for (octave_idx_type i = 0; i < len; i++)
+{
+octave_idx_type k = idx(i);
+mx_inline_add2 (l, dst + l*k, src + l*i);
+}
+dst += l*n;
+src += l*ns;
+}
+}
 }
 // N-dimensional array with math ops.
 template <class T>
 void

Mercurial > hg > octave-lyh