Implement the diag() operations about viennacl-dev

[ptillet]

In the process I introduce some "element_accessor" in addition of the index set... This allows for some memory-less patterns to be used, such as an "infinite" vector 1,2,3,..... , repmat, diag, etc...
Is there any additional cool feature that can be done using this mechanism, for cases where the elements of the vector don't come from an handle (or not only) ?

[karlrupp]

Please be more conservative with respect to features. Now as the generator is no longer super-experimental, we have a hard time removing features once they are introduced. ;-) Could you please elaborate more on what exactly 'element_accessor' is supposed to do (other than what is provided by range(), slide(), and finally index_set()?

[ptillet]

element_accessor, is just a way to specify where the elements of a given vector is located, because it may not come necessarily from a handle, for example, a constant vector would implement something like

std::string constant_element::access(index){ return val_name_; }

where val_name_ is the argument name of the corresponding CPU scalar. Or a vector 1,2,.....,N would be like

std::string linear_element::access(index){ return index; }

We can play a bit with matrices to introduce "memory free" r-value symbolic expressions :P For example, eye(N) would map to something like:

std::string identity_element::access(index1, index2){ return "(" + index1 + ")==(" + index2 + ")";

The idea is just that diag() would dispatch the access to a diag_accessor, that would be for example implemented like

std::string diag_mat_element::access(index1, index2){ return vec_handle_name_+"["+index1+"]"; }

At first I wanted to incorporate this information in index_set, but it seemed like an orthogonal issue :) This is used internally only so I wouldn't describe this as a "feature"

[ptillet]

I have reimplemented diag() and also repmat()... A "pitfall" of the "element_accessor" approach is that diag() and repmat accept only single elements and not expressions :P diag(x + y) is not legal (but can be replaced by diag(x) + diag(y) ).

[karlrupp]

Cool, thanks! What about still allowing the user to pass an expression, but internally immediately convert it to the expanded form.
Example: User writes diag(A + alpha * B), we transform this immediately to diag(A) + alpha * diag(B);