[Good First Issue]: Possible Conditional Compilation Optimization for type_to_fuse_map

[riverlijunjie]

Context

Conditional Compilation is very helpful to optimize openvino binaries size for specified platforms and models, seems it can be used for type_to_fuse_map to do some binaries optimization:

openvino/src/common/transformations/src/transformations/convert_precision.cpp

Lines 430 to 477 in f9605cd

	type_to_fuse_map type_to_fuse{
	{opset4::Convert::get_type_info_static(), fuse_type_to_convert},
	{opset4::ShapeOf::get_type_info_static(), fuse_type_to_shapeof},
	{opset6::Assign::get_type_info_static(),
	m_store_original_precision_as_rt_attribute ? store_original_type_as_attribute : wrap_into_original_type},
	{opset6::ReadValue::get_type_info_static(),
	m_store_original_precision_as_rt_attribute ? store_original_type_as_attribute : wrap_into_original_type},
	{opset3::NonMaxSuppression::get_type_info_static(), fuse_type_to_nms3},
	{opset4::NonMaxSuppression::get_type_info_static(), fuse_type_to_nms4},
	{opset5::NonMaxSuppression::get_type_info_static(), fuse_type_to_nms5},
	{opset9::NonMaxSuppression::get_type_info_static(), fuse_type_to_nms9},
	{op::v13::NMSRotated::get_type_info_static(), fuse_type_to_nms_rotated},
	{opset8::MatrixNms::get_type_info_static(), fuse_type_to_matrix_nms},
	{opset8::MulticlassNms::get_type_info_static(), fuse_type_to_multiclass_nms},
	{opset9::MulticlassNms::get_type_info_static(), fuse_type_to_multiclass_nms},
	{opset9::GenerateProposals::get_type_info_static(), fuse_type_to_generate_proposals},
	{opset6::CTCGreedyDecoderSeqLen::get_type_info_static(), fuse_type_to_ctc_greedy_decoder_seq_len},
	{opset1::TopK::get_type_info_static(), fuse_type_to_topk},
	{opset4::TopK::get_type_info_static(), fuse_type_to_topk},
	{opset11::TopK::get_type_info_static(), fuse_type_to_topk},
	{opset8::MaxPool::get_type_info_static(), fuse_type_to_maxpool},
	{opset4::NonZero::get_type_info_static(), fuse_type_to_nonzero},
	{opset4::Bucketize::get_type_info_static(), fuse_type_to_bucketize},
	{opset4::Equal::get_type_info_static(), fuse_type_to_binary_comparision<opset4::Equal>},
	{opset4::NotEqual::get_type_info_static(), fuse_type_to_binary_comparision<opset4::NotEqual>},
	{opset4::Greater::get_type_info_static(), fuse_type_to_binary_comparision<opset4::Greater>},
	{opset4::GreaterEqual::get_type_info_static(), fuse_type_to_binary_comparision<opset4::GreaterEqual>},
	{opset4::Less::get_type_info_static(), fuse_type_to_binary_comparision<opset4::Less>},
	{opset4::LessEqual::get_type_info_static(), fuse_type_to_binary_comparision<opset4::LessEqual>},
	{opset10::IsFinite::get_type_info_static(), fuse_type_to_binary_comparision<opset10::IsFinite>},
	{opset10::IsNaN::get_type_info_static(), fuse_type_to_binary_comparision<opset10::IsNaN>},
	{opset10::IsInf::get_type_info_static(), fuse_type_to_binary_comparision<opset10::IsInf>},
	{opset4::LogicalAnd::get_type_info_static(), fuse_type_to_logical<opset4::LogicalAnd>},
	{opset4::LogicalOr::get_type_info_static(), fuse_type_to_logical<opset4::LogicalOr>},
	{opset4::LogicalXor::get_type_info_static(), fuse_type_to_logical<opset4::LogicalXor>},
	{opset4::LogicalNot::get_type_info_static(), fuse_type_to_logical<opset4::LogicalNot>},
	{opset1::Xor::get_type_info_static(), fuse_type_to_logical<opset1::Xor>},
	{opset4::ReduceLogicalAnd::get_type_info_static(), fuse_type_to_reduce_logical<opset4::ReduceLogicalAnd>},
	{opset4::ReduceLogicalOr::get_type_info_static(), fuse_type_to_reduce_logical<opset4::ReduceLogicalOr>},
	{opset1::ShapeOf::get_type_info_static(), fuse_type_to_shapeof_v0},
	{opset4::Range::get_type_info_static(), fuse_type_to_range_v4},
	{opset9::Eye::get_type_info_static(), fuse_type_to_eye_v9},
	{opset10::Unique::get_type_info_static(), fuse_type_to_unique_v10},
	{opset8::RandomUniform::get_type_info_static(), fuse_type_to_random_uniform_v8},
	{opset13::Multinomial::get_type_info_static(), fuse_type_to_multinomial_v13},
	{opset1::PriorBox::get_type_info_static(), fuse_type_to_prior_box<opset1::PriorBox>},
	{opset8::PriorBox::get_type_info_static(), fuse_type_to_prior_box<opset8::PriorBox>},
	{opset1::PriorBoxClustered::get_type_info_static(), fuse_type_to_prior_box<opset1::PriorBoxClustered>}};

Need to consider usage of CC factory here, because in case of conditional compilation, this map will hold all the operations, which are not even used in the model.
See https://github.com/openvinotoolkit/openvino/blob/master/src/common/conditional_compilation/include/openvino/cc/factory.h

What needs to be done?

1. Apply Conditional Compilation for type_to_fuse_map to optimize binaries size.
2. Measure how many binaries size benefit can be achieved from this optimization.

Example Pull Requests

No response

Resources

• Contribution guide - start here!

• Intel DevHub Discord channel - engage in discussions, ask questions and talk to OpenVINO developers

• https://docs.openvino.ai/2023.3/opevino_conditional_compilation_deploying_guide.html

Contact points

@riverlijunjie

Ticket

No response

[RamyaKappagantu]

Hi @riverlijunjie , I would like to work on this task as a part of GSoC pre-requisite. Could you please assign this to me?

[p-wysocki]

Hello @RamyaKappagantu, are you still working on this? Is there anything we could help you with?

[MirImaadAli1]

.take

[github-actions]

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

[p-wysocki]

Hello @MirImaadAli1, are you still working on that issue? Do you need any help?

[11happy]

.take

[github-actions]

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

[11happy]

Hello @riverlijunjie , @p-wysocki is this the correct way to approach this problem,
first I define a CC domain , then I register a factory pattern via the factory class & then we use the createNodeifRequired method
I took reference from the tests implemented in this file src/common/conditional_compilation/tests/selective_build_analyzer.cpp

[11happy]

Currently I have it implemented this way on my end , would see if this corectly compiles & if I can see the binary size improvements . I did this selective build mentioned in this doc
https://github.com/openvinotoolkit/openvino/blob/master/src/common/conditional_compilation/docs/develop_cc_for_new_component.md
edit : this way is having some compilation issues that I am fixing.