Maxence Naud authored 3 months ago and Maxence Naud committed 3 months ago

- move some inline functions from header to source
- include what is used in header and source files to avoid relying on recursive includes
- remove 'constexpr' from 'Tensor(Vector)' constructor that applied forced inlining
- use C++ types and not C type (std::size_t instead of size_t)
- move some default constructor/operators to src
- explicitly delete some default class constructors for 'SingleGraphMatching', 'StaticAnalysis', 'MetaOpStats' and 'OperatorStats'
- add default parameters to 'ArgMax' constructor
- make Node::mAttrs available to the associated Operator instead of accessing it each time getOperator() is called
- make some for loop access elements by reference in 'FuseBatchNorm'

Jerome Hue authored 7 months ago and Maxence Naud committed 4 months ago

Add a .codespellrc and fix spelling mistakes in code.

feat(export): Add a way to give a custom forward template & add test argument in scheduler_export() function

Christophe Guillon authored 7 months ago and Olivier BICHLER committed 7 months ago

Add getOrderedNodes(reversed=false) method which produces
a topological order of the graph view (or of the reversed
graph view when reversed is true).
This order is also deterministic given the nodes set and
the ordered output nodes.
Cyclic graphs are handled by breaking back edges as declared
in the new setBackEdges() method of the Operator class.
The order is suitable for resolving constant propagation
of tensor types in one pass including for cyclic graphs
if back edges are declared correctly for operators.
It is also suitable for exporting to onnx graphs
which require a topological order of exported acyclic graphs.

Christophe Guillon authored 9 months ago and Maxence Naud committed 9 months ago

Define a Tensor with undefined dims by explicitly setting mSize to 0.
Provide undefined() method which must be used instead of empty()
to test whether a Tensor as associated dimensions and can be used
for forwarding dimensions downward.
This disambiguates from defined scalar Tensor which actually have
empty dimensions, but size of 1.
Note that as soon as a Tensor with undefined dimensions is resized,
it's dimensions are permanently defined (though may still change).

This change solves issues in operators forwardDims() when
input values are scalar, i.e. !undefined() and empty().

The previous behaviour was to keep the Tensor of the removed layer. Now it is also removed

[Fix] (tmp) GraphView::forward_dims() if dimsension provided and Tensor already associated