[Add] Possibility to create a GenericOperator from any Operator

Context

One envisionned use-case is to be able to replace actual Operator obejcts with GenericOperator that would behave differently even regarding the size of inputs/outputs Tensors.

For example, a Conv that works on 4-bits packed data, which input/output size would be different from the standard Conv operator because of memory alignments. This would be useful for the export of sub 8-bits kernels in the Cortex-M export.

Major modifications

Add possibility to create a GenericOperator from any Operator.

• add: toGenericOp() recipe to convert operator into GenericOperator
• add: Operator::inputCategory() getter for input categories
• add: comparison operator= to DynamicAttributes
• add: GenericOperator constructor from Operator
• add: unit-tests for toGenericOperator() recipe
• upd: make std::array and std::pair recursively hashable in DynamicAttributes
• upd: make almost every StaticAttributes convertible to DynamicAttributes

requested review from @thibaultallenet

assigned to @olivierbichler

added Enhancement ⭐ label

What is the use case of this MR?

For the envisionned use case shouldn't we use a MetaOperator with operator Pack/Unpack?

This solution is good for creating a fast solution for the problem at hand I am just talking a more long term solution for the future.

changed the description

added 1 commit

• ea6ac3c2 - Add toGenericOp recipie + bind + test

Compare with previous version

added 1 commit

• a677e993 - Fixed issue

Compare with previous version

added 1 commit

• b5ed3f15 - Fixed unsupported std::void_t

Compare with previous version

When building ./setup.sh -m core -m backend_cpu -m backend_cuda -m quantization --debug --python -j 32

Commits :

• core : b5ed3f15 (latest commit on this MR)
• backend_cpu : 8b99d4f1bfec0cb1aa91e3dea655558eeadc7968
• backend_cuda : 75102058f9e75db69a9836b27c900507cbc723ca
• quantization : f50c8606001b0fe800362bf15c607a605bacd432

Several errors are raised on DynamicAttributes.hpp during the build of aidge_quantization.

See file "log_error" for complete log : log_error.txt

@olivierbichler can you explain your changes in DynamicAttributes.hpp?

Yes, I added the makeTypeConditionallyAvailable function that is called when StaticAttributes::getAttrs() is called. Its purpose is to register the types used in StaticAttributes to make them fully usable in DynamicAttributes, by providing an AnyUtils for these types. This allows seamless copy of static attributes to dynamic ones!

Could it open the path to having DynamicAttributes for every operators?

It makes (almost) any StaticAttributes convertible to DynamicAttributes seamlessly, yes. We could have from the beginning just use DynamicAttributes everywhere, but StaticAttributes's advantage is that it has 0 overhead vs a simple member variable!

added 1 commit

• 5fd06284 - Added missing hash specialization for std::pair

Compare with previous version

added 1 commit

• d7c2bda1 - Added more checks

Compare with previous version

added 1 commit

• 3c91814d - Fixed segfault

Compare with previous version

added 24 commits

• 3c91814d...0da4f0a8 - 23 commits from branch dev
• 49400010 - Merge branch 'dev' into genericop_from_op

Compare with previous version

enabled an automatic merge when all merge checks for 49400010 pass

Lets merge this so @thibaultallenet can move forward on his other MR.

mentioned in commit 8833a2bd

merged

changed milestone to %aidge v0.5.0