onnxruntime::optimizer_utils Namespace Reference

Functions
InlinedVector< std::unique_ptr < RewriteRule > >	GenerateRewriteRules (TransformerLevel level, const InlinedHashSet< std::string > &rules_to_disable={})
std::string	GenerateRuleBasedTransformerName (TransformerLevel level)
std::unique_ptr < RuleBasedGraphTransformer >	GenerateRuleBasedGraphTransformer (TransformerLevel level, const InlinedHashSet< std::string > &rules_to_disable, const InlinedHashSet< std::string_view > &compatible_execution_providers)
InlinedVector< std::unique_ptr < GraphTransformer > >	GenerateTransformers (TransformerLevel level, const SessionOptions &session_options, const IExecutionProvider &execution_provider, const InlinedHashSet< std::string > &rules_and_transformers_to_disable={})
InlinedVector< std::unique_ptr < GraphTransformer > >	GenerateTransformersForMinimalBuild (TransformerLevel level, const SessionOptions &session_options, const SatApplyContextVariant &apply_context, const IExecutionProvider &cpu_execution_provider, const InlinedHashSet< std::string > &rules_and_transformers_to_disable={})

Function Documentation

InlinedVector<std::unique_ptr<RewriteRule> > onnxruntime::optimizer_utils::GenerateRewriteRules	(	TransformerLevel	level,
		const InlinedHashSet< std::string > &	rules_to_disable = {}
	)

Generates all predefined rules for this level. If rules_to_enable is not empty, it returns the intersection of predefined rules and rules_to_enable. TODO: This is visible for testing at the moment, but we should rather make it private.

std::unique_ptr<RuleBasedGraphTransformer> onnxruntime::optimizer_utils::GenerateRuleBasedGraphTransformer	(	TransformerLevel	level,
		const InlinedHashSet< std::string > &	rules_to_disable,
		const InlinedHashSet< std::string_view > &	compatible_execution_providers
	)

Generates all rule-based transformers for this level.

std::string onnxruntime::optimizer_utils::GenerateRuleBasedTransformerName

(

TransformerLevel

level

)

Given a TransformerLevel, this method generates a name for the rule-based graph transformer of that level.

InlinedVector<std::unique_ptr<GraphTransformer> > onnxruntime::optimizer_utils::GenerateTransformers	(	TransformerLevel	level,
		const SessionOptions &	session_options,
		const IExecutionProvider &	execution_provider,
		const InlinedHashSet< std::string > &	rules_and_transformers_to_disable = {}
	)

Generates all predefined (both rule-based and non-rule-based) transformers for this level. Any transformers or rewrite rules named in rules_and_transformers_to_disable will be excluded.

InlinedVector<std::unique_ptr<GraphTransformer> > onnxruntime::optimizer_utils::GenerateTransformersForMinimalBuild	(	TransformerLevel	level,
		const SessionOptions &	session_options,
		const SatApplyContextVariant &	apply_context,
		const IExecutionProvider &	cpu_execution_provider,
		const InlinedHashSet< std::string > &	rules_and_transformers_to_disable = {}
	)

Generates all predefined transformers which can be used to provide runtime optimizations for this level in a minimal build. Any transformers or rewrite rules named in rules_and_transformers_to_disable will be excluded.

This is a distinct function from GenerateTransformers() because:

• An ORT format model used in a minimal build will have been pre-optimized to at least level 1 when created, so level 1 transformers are not included.
• In a minimal build we have limited optimization/Graph capabilities
  • Graph::Resolve is not available so the transformer must keep the Graph in a valid state
  • Limited graph_utils capabilities are included
• Only a small subset of transformers support storing/replaying runtime optimizations with an ORT format model
  • this capability is provided by the SelectionActionTransformer infrastructure
  • the logic to determine the set of nodes a transformer should modify is captured during creation of the ORT format model
  • this information is saved in the ORT format model
  • only the logic to modify the set of nodes is included in the minimal build
• The QDQFinalCleanupTransformer and NhwcTransformer transformers are also supported in a minimal build