Deduction of optimal return type

[mmoeller]

Dear all,

is it possible to deduce the optimal return type from a vector or matrix expression?

Please consider the following demo code:

viennacl::vector<double> v1(20),v2(20);

// Create expression (not the vector)
auto tmp v1+v2;

// Create new vector from the expression
viennacl::vector<double> a(tmp);

In my application,I need to specialize some of my routines for ViennaCL, and therefore, I need to know
EITHER: a base class of all vector/matrix expressions so that I can test against std::is_base_of<???,Expr>::value
OR: a common type/attribute of all vector/matrix expressions, so that I can specialize using the presence of that type/attribute
AND: the optimal return type (here: vector and double)

Best regads,
Matthias

[Karl Rupp]

Hi,

the following is for ViennaCL 1.x.y:
For vector expressions (including operations where the operands may not be vectors, but the result is a vector) such as v1 + v2, the return type is viennacl::vector_expression<LHS, RHS, OP>
For matrix expressions the respective type is viennacl::matrix_expression<LHS, RHS, OP>.
In a few cases there are also scalar expressions used (e.g. dot products), where the return type is viennacl::scalar_expression<LHS, RHS, OP>.

The common base class for vectors and vector proxy objects is viennacl::vector_base<T>, for dense matrices it is viennacl::matrix_base<T>. Currently there is no common type for sparse matrices, but this may change.

ViennaCL 2.0.0 with see a complete redesign (second half of 2016), so don't expect these types to be still in use then.

Best regards,
Karli

[mmoeller]

Hi,

thank you very much for this information. What I would need is a mechanism to deduce the return type from an expression, e.g.

viennacl::vector<double> v1(20),v2(20);
auto tmp v1+v2;

typedef viennacl::return_type<typename decltype(tmp)> return_type;
return_type a(tmp);

It would also suffice if I can extract the type (matrix/vector) and the data type (float/double) separately.

Best regards,
Matthias

[Karl Rupp]

Hi,

the return type deduction (following your definition of a return type above) is as follows: For vector_expression<> it is always a vector<>, for matrix_expression it is always a matrix<>. Same for scalar_expression and scalar<>.

You can obtain the data type via

typedef viennacl::result_of::cpu_scalar_type<T>::type   data_type;

(I think this is what you were asking for. Please let me know if that does not answer your question.)

Best regards,
Karli