Tensor V2

The aim of this issue is tu mutualize several Tensor related issues and propose a corresponding design that is, so far, partially implemented in !13 (closed).

A proposed design would be

Tensor
   // active area description
   first data coordinates in active area
   dimensions of the active area
   // stored area description
   first data coordinates in stored area
   dimensions of the stored area
   data type
   backend
   storage object // actual type depending of backend

Storage object represents raw storage that knows nothing about the Tensor.
In !13 (closed) TensorImpl partially correspond to the stored area description.

An open issue is how to have an efficient access to backend-independant data in the TensorImpl as both TensorImpl concrete objects and Tensor may need these informations.

Edited Oct 19, 2023 by laurent soulier