guwi17_at_[hidden] wrote:<br><br>&gt;BTW. I would not expect a big difference between a banded matrix with 
<br>
&gt;compile time known bandwidth compared to the current implementation. I 
<br>
&gt;still wait for someone who proves the opposite ;-)<br><br>I&#39;ve implemented banded_c_matrix.hpp, with compile time UPPER and LOWER, and your expectations were right: banded_c_matrix for LOWER = UPPER = 0 (extreme case) is as fast as (or marginally better than) a diagonal_matrix (from 
banded.hpp).<br><br>I suppose I cannot expect that the matrix_assign_* implementations for a packed matrix can take advantage of compile-time knowledge of the bandwidth size.<br><br>ublas users: anyone has useful advice on achieving optimal performance when using diagonal matrices?
<br>Any hint will be appreciated.<br><br>ublas developers: any plan on having a packed_c_tag representing compile-time knowledge of bandwidth size, and accompanying specialized algorithms?<br><br>Thanks in advance for your answers.
<br><span style="font-family: courier new,monospace;"></span>