Mila::Dnn::Compute::OperationRegistry Class Reference

Central registry for typed, device-aware compute operations. More...

Classes
struct	TypeID
	Composite key for registry lookup. More...
struct	TypeIDHash

Public Member Functions
template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>
std::shared_ptr< BinaryOperation< TDeviceType, TInputA, TInputB, TComputePrecision > >	createBinaryOperation (const std::string &operation_name, IExecutionContext *context, const ComponentConfig &config) const
	Create a binary operation instance.
template<DeviceType TDeviceType, TensorDataType TPrecision, TensorDataType TInputA = TPrecision, TensorDataType TInputB = TInputA>
std::shared_ptr< PairedOperation< TDeviceType, TPrecision, TInputA, TInputB > >	createPairedOperation (const std::string &operation_name, IExecutionContext *context, const ComponentConfig &config) const
	Create a paired operation instance.
template<DeviceType TDeviceType, TensorDataType TInputType, TensorDataType TComputePrecision = TInputType>
std::shared_ptr< UnaryOperation< TDeviceType, TInputType, TComputePrecision > >	createUnaryOperation (const std::string &operation_name, IExecutionContext *context, const ComponentConfig &config) const
	Create a unary operation instance.
template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>
std::vector< std::string >	getRegisteredOperations () const
	Return all registered operation names across all arities for a given type configuration.
template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>
bool	isOperationRegistered (const std::string &operation_name) const
	Return true if an operation name is registered under any arity for a given type configuration.
template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>
void	registerBinaryOperation (const std::string &operation_name, std::function< std::shared_ptr< BinaryOperation< TDeviceType, TInputA, TInputB, TComputePrecision > >(IExecutionContext *, const ComponentConfig &)> creator)
	Register a binary operation factory.
template<DeviceType TDeviceType, TensorDataType TPrecision, TensorDataType TInputA = TPrecision, TensorDataType TInputB = TInputA>
void	registerPairedOperation (const std::string &operation_name, std::function< std::shared_ptr< PairedOperation< TDeviceType, TPrecision, TInputA, TInputB > >(IExecutionContext *, const ComponentConfig &)> creator)
	Register a paired operation factory.
template<DeviceType TDeviceType, TensorDataType TInputType, TensorDataType TComputePrecision = TInputType>
void	registerUnaryOperation (std::string_view operation_name, std::function< std::shared_ptr< UnaryOperation< TDeviceType, TInputType, TComputePrecision > >(IExecutionContext *, const ComponentConfig &)> creator)
	Register a unary operation factory.

Static Public Member Functions
static OperationRegistry &	instance ()

Private Types
using	GenericCreator
using	RegistryMap

Private Member Functions
	OperationRegistry ()=default
	OperationRegistry (const OperationRegistry &)=delete
OperationRegistry &	operator= (const OperationRegistry &)=delete

Private Attributes
RegistryMap	binary_registry_
RegistryMap	paired_registry_
RegistryMap	unary_registry_

Detailed Description

Central registry for typed, device-aware compute operations.

Maintains three independent stores keyed by (DeviceType, input types, compute precision) and operation name — one per operation arity. Separation prevents cross-arity static_pointer_cast UB when an op is retrieved under the wrong base type.

Member Typedef Documentation

GenericCreator

using Mila::Dnn::Compute::OperationRegistry::GenericCreator

private

Initial value:

 std::function<std::shared_ptr<void>(
            IExecutionContext*, const ComponentConfig& )>

RegistryMap

using Mila::Dnn::Compute::OperationRegistry::RegistryMap

private

Initial value:

 std::unordered_map<TypeID,
            std::unordered_map<std::string, GenericCreator>,
            TypeIDHash>

Constructor & Destructor Documentation

OperationRegistry() [1/2]

Mila::Dnn::Compute::OperationRegistry::OperationRegistry ( )

privatedefault

Here is the caller graph for this function:

OperationRegistry() [2/2]

Mila::Dnn::Compute::OperationRegistry::OperationRegistry ( const OperationRegistry & )

privatedelete

Here is the call graph for this function:

Member Function Documentation

createBinaryOperation()

template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>

std::shared_ptr< BinaryOperation< TDeviceType, TInputA, TInputB, TComputePrecision > > Mila::Dnn::Compute::OperationRegistry::createBinaryOperation ( const std::string & operation_name, IExecutionContext * context, const ComponentConfig & config ) const

inline

Create a binary operation instance.

Exceptions

std::runtime_error	if no matching registration exists.
std::invalid_argument	if context is null.

Here is the caller graph for this function:

createPairedOperation()

template<DeviceType TDeviceType, TensorDataType TPrecision, TensorDataType TInputA = TPrecision, TensorDataType TInputB = TInputA>

std::shared_ptr< PairedOperation< TDeviceType, TPrecision, TInputA, TInputB > > Mila::Dnn::Compute::OperationRegistry::createPairedOperation ( const std::string & operation_name, IExecutionContext * context, const ComponentConfig & config ) const

inline

Create a paired operation instance.

Exceptions

std::runtime_error	if no matching registration exists.
std::invalid_argument	if context is null.

createUnaryOperation()

template<DeviceType TDeviceType, TensorDataType TInputType, TensorDataType TComputePrecision = TInputType>

std::shared_ptr< UnaryOperation< TDeviceType, TInputType, TComputePrecision > > Mila::Dnn::Compute::OperationRegistry::createUnaryOperation ( const std::string & operation_name, IExecutionContext * context, const ComponentConfig & config ) const

inline

Create a unary operation instance.

Exceptions

std::runtime_error	if no matching registration exists.
std::invalid_argument	if context is null.

Here is the caller graph for this function:

getRegisteredOperations()

template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>

std::vector< std::string > Mila::Dnn::Compute::OperationRegistry::getRegisteredOperations ( ) const

inline

Return all registered operation names across all arities for a given type configuration.

instance()

OperationRegistry & Mila::Dnn::Compute::OperationRegistry::instance ( )

inlinestatic

Here is the call graph for this function:

Here is the caller graph for this function:

isOperationRegistered()

template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>

bool Mila::Dnn::Compute::OperationRegistry::isOperationRegistered ( const std::string & operation_name ) const

inline

Return true if an operation name is registered under any arity for a given type configuration.

operator=()

OperationRegistry & Mila::Dnn::Compute::OperationRegistry::operator= ( const OperationRegistry & )

privatedelete

Here is the call graph for this function:

registerBinaryOperation()

template<DeviceType TDeviceType, TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TComputePrecision = TInputA>

void Mila::Dnn::Compute::OperationRegistry::registerBinaryOperation ( const std::string & operation_name, std::function< std::shared_ptr< BinaryOperation< TDeviceType, TInputA, TInputB, TComputePrecision > >(IExecutionContext *, const ComponentConfig &)> creator )

inline

Register a binary operation factory.

Creator must return a std::shared_ptr<BinaryOperation<TDeviceType, TInputA, TInputB, TComputePrecision>>.

Here is the caller graph for this function:

registerPairedOperation()

template<DeviceType TDeviceType, TensorDataType TPrecision, TensorDataType TInputA = TPrecision, TensorDataType TInputB = TInputA>

void Mila::Dnn::Compute::OperationRegistry::registerPairedOperation ( const std::string & operation_name, std::function< std::shared_ptr< PairedOperation< TDeviceType, TPrecision, TInputA, TInputB > >(IExecutionContext *, const ComponentConfig &)> creator )

inline

Register a paired operation factory.

Creator must return a std::shared_ptr<PairedOperation<TDeviceType, TPrecision, TInputA, TInputB>>.

Here is the caller graph for this function:

registerUnaryOperation()

template<DeviceType TDeviceType, TensorDataType TInputType, TensorDataType TComputePrecision = TInputType>

void Mila::Dnn::Compute::OperationRegistry::registerUnaryOperation ( std::string_view operation_name, std::function< std::shared_ptr< UnaryOperation< TDeviceType, TInputType, TComputePrecision > >(IExecutionContext *, const ComponentConfig &)> creator )

inline

Register a unary operation factory.

Creator must return a std::shared_ptr<UnaryOperation<TDeviceType, TInputType, TComputePrecision>>.

Here is the caller graph for this function:

Member Data Documentation

binary_registry_

RegistryMap Mila::Dnn::Compute::OperationRegistry::binary_registry_

private

paired_registry_

RegistryMap Mila::Dnn::Compute::OperationRegistry::paired_registry_

private

unary_registry_

RegistryMap Mila::Dnn::Compute::OperationRegistry::unary_registry_

private

---

The documentation for this class was generated from the following file:

• /__w/Mila/Mila/Mila/Src/Dnn/Compute/Operations/OperationRegistry.ixx