- disable PR automerge on compiler update (too risky)
- change init file to be a namespace package
- remove version.py (infos can still be retrieved through package meta)
this avoids potential conflicts with other packages we release
- various fixes for compiler compatibility
closes#1272
- poetry lock required to update package name
- related Makefile and licenses update to keep consistency (will be removed
once the PyPi package is available)
- modified compile_torch_model to only accept datasets with one input
mypy error with better numpy typing hints appeared
closes#1252
- corrected docstring that was mistaken on what was returned
- updated pyproject.toml to ignore warnings that happened naturally in
networkx and that was blocking proper test execution (no way around that
this is code from networkx that triggered the warning)
- add a test case for the newly supported fusing topology
closes#499
- happens when a constant is bigger than the variable input of the
subgraph, would require special handling, hard to write and requires a
concatenation/stacking operator see #587closes#645
- add an attribute fusable to False for the operations that should not be
explicitely fused from the original addition of GenericFunction
- add op_kind instance attribute to differentiate between TLU and memory
operations for GenericFunction
refs #600
- MLIRConverter becomes an abstract base class
- pass the needed informations in an extra dict to MLIR converters
- NPMLIRConverter handles the specifics of numpy tables generation
- if only some values are problematic, flood fill the resulting table with
other valid values
- if no valid value was generated an AssertionError will be thrown
- removed underlying_type_constructor from BaseDataType as it was scalar
specific and put it in values
- update inputset_eval to keep a sample of intermediate node values
- allows to get the proper value constructor to be used in
UnivariateFunction get_table and have tensors as inputs
- renamed output_index to output_idx in BaseTracer
- update tracing and fusing code to manage output_idx correctly
- update OPGraph evaluate and update_values_with_bounds to manage
output_idx
- update tests checking graph validity to have output_idx set properly
- the support of actual multi-output nodes is in #81
- add op_attributes on ArbitraryFunction
- this works if all constants are smaller than the input tensor
- otherwise it requires more advanced code and a concatenate operator which
currently does not exist
