Optimizing data traversal on Tensor V2

As mentioned there https://gitlab.eclipse.org/eclipse/aidge/aidge_backend_cpu/-/blame/fix/TensorPIMPL/include/aidge/backend/cpu/data/TensorImpl.hpp#L150 data traversal can be optimised on Tensor V2 when some active area dimensions are the same as the full tensor area.

The algorithm consistes in computing the largest subtensor which is the same in active area and full area. It can be done by multiplying the dimensions from the last one to the first one for active area and full area, until the values differ.
The last dimension for which both subtensors size are the same determines which subtensor data can be traversed without jumps and, thus with a fast loop.
In the link above, it would determined the largest possible memcopy to do this extract.

This optimization can be used by all operators that supports restriction to the active area (by default, using an operator on tensor with restricted active area is UB).

illustration with a typical memory layout

changed milestone to %v0.1.0

added 1 deleted label

changed the description

mentioned in issue #58 (closed)

NB in order to make this easier for operator implementers, the "contiguity" dimension and and the length of continuous data can be precalculate for a TensorImpl, accessible through adequate getters.

mentioned in merge request !13 (closed)

Currently, this issue seems obsolete considering the updated design of the Tensor and TensorImpl classes, and therefore does not address a concrete problem.

closed