Export refactor

Context

&12

Changes

• Mutualize meminfo between exports
  • Make work Aidge memory manager (currently a simple version is used ...)
  • Add meminfo in export_node
• Create ExportScheduler to mutualize function between exports
  • Better handle input/ouput of graph by avoiding double memory allocation
  • Regisration done with backends
  • Generate mem manager and scheduling
• Mutualize Meminfo in aidge_core
• Mutualize functions to generate files in export
• Refactor Export node
  • Add a dictionnary attributes to stack all the attributes desirable for the export
  • Get recursively attributes for MetaOperator (tracked by #180 (closed))
  • Add function to be agnostic from data form (get_input_channel etc ...)
  • Mutualize export and configuration function, with this change, dev of export only need to inform parameters when creating export_node and no longer to worry about copying kernels etc ...
  • Add function to automatically retrieve Chan, H or W regardless of the dataformat, remove logic from template and adding it in NodeExport instead
• Change key for selecting an ExportNode (backend + dataformat ...)
• Update Registration system to use backend and SpecImpl
• Make SpecImpl Hashable (tracked by: #179 (closed))
• remove producer set to constant trick to distinguish input producer from other producers in the graph
• Change structure of export to: &12

Other changes:

• Change type for templated operators (example: Conv -> Conv{NbDims}D) (fix: #168 (closed))
• Change constructor binding name for these operators (ConvOp{NbDims}D-> Conv{NbDims}DOp)
• Fix aidge#180 (closed)

Issue to solve

• When registering a MetaOperator every member should have an implementation Can be bypass by setting dummy implmentations:

@ExportLibAidgeARM.register("Pad2D", aidge_core.ImplSpec(aidge_core.IOSpec(aidge_core.dtype.any)))
class Pad_ARMCortexM(ExportNodeCpp):
    def __init__(self, node, mem_info, is_input, is_output):
        raise NotImplementedError("Pad2D nodes is not implemented")

Tracked by: #177 (closed)

• When creating a new backend, we need to add a "fake" tensor backend:

aidge_core.register_Tensor(["aidge_arm", aidge_core.dtype.float32],
                           aidge_core.get_key_value_Tensor(["cpu", aidge_core.dtype.float32]))

Tracked by: #178 (closed)

added Enhancement ⭐ Refactoring🎨 labels

assigned to @cmoineau

mentioned in task aidge#128 (closed)

added 1 commit

• bd723d05 - Add NodeExportCpp to specify NodeExport and how to generate forward and node definition.

Compare with previous version

added 1 commit

• 9885ceaf - Add back data format when chosing how to retrieve nb chan ...

Compare with previous version

added 3 commits

• 92b34948 - Add basic export registry system.
• 54121fb7 - Fix circular import issues.
• 80ca13c8 - Add exportable function to check if a node is exportable instead of having...

Compare with previous version

changed the description

marked the checklist item Add function to be agnostic from data form (get_input_channel etc ...) as completed

marked the checklist item Add a dictionnary attributes to stack all the attributes desirable for the export as completed

marked the checklist item Refactor Export node as completed

resolved all threads

added 1 commit

• 1c773b62 - Add meminfo to ExportNode.

Compare with previous version

marked this merge request as draft

added 1 commit

• cbf9c772 - Add In/Out Format in NodeExport.

Compare with previous version

added 4 commits

• 987cbfc8 - Introduce Scheduler export + move generic C++ file generation to core.
• a5e8976d - Fix multiple typo due to old naming.
• eff7dc2f - Raise error if no input set instead of just a warning.
• 91458013 - Main.cpp now include header with input name.

Compare with previous version

marked the checklist item Add meminfo in export_node ? as completed

marked the checklist item Add is_output and is_input attribute in ExportNode as completed

marked the checklist item Generate mem manager and scheduling as completed

marked the checklist item Better handle input/ouput as completed