CUDA Residual operation implementing the BinaryOperation interface. More...

Inheritance diagram for Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >:

[legend]

Collaboration diagram for Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >:

[legend]

Public Types
using	ConfigType = ResidualConfig
using	CudaExecutionContext = ExecutionContext<DeviceType::Cuda>
using	MR = CudaDeviceMemoryResource
using	NativeType = typename Mila::Dnn::Compute::Cuda::TensorDataTypeMap<TPrecision>::device_type
Public Types inherited from Mila::Dnn::Compute::BinaryOperation< DeviceType::Cuda, TInputA, TInputA, TInputA >
using	MR
using	ParameterGradTensor
using	ParameterTensor
using	TensorLeftType
using	TensorOutputType
using	TensorRightType
Public Types inherited from Mila::Dnn::Compute::Operation< TDeviceType, TPrecision >
using	DataTypeTraits

Public Member Functions
	CudaResidualOp (IExecutionContext *context, const ResidualConfig &config)
	Construct with a non-null CUDA execution context.
void	backward (const ITensor &input_A, const ITensor &input_B, const ITensor &output_grad, ITensor &A_grad, ITensor &B_grad) const override
	Backward pass.
void	build (const BuildContext &config) override
	Prepare the operation for a concrete input shape.
void	forward (const ITensor &input_A, const ITensor &input_B, ITensor &output) const override
	Forward pass: element-wise addition Y = X1 + X2 using CUDA kernel.
std::string	getName () const override
	Human-readable operation name.
OperationType	getOperationType () const override
	Operation type identifier.
Public Member Functions inherited from Mila::Dnn::Compute::BinaryOperation< DeviceType::Cuda, TInputA, TInputA, TInputA >
virtual	~BinaryOperation ()=default
Public Member Functions inherited from Mila::Dnn::Compute::Operation< TDeviceType, TPrecision >
virtual	~Operation ()=default
virtual void	clearGradients () noexcept
	Clear any cached gradient pointers held by the operation.
virtual TensorDataType	getDataType () const
	Tensor data type for this operation.
virtual DeviceType	getDeviceType () const
	Device type for this operation.
virtual std::size_t	getStateMemorySize () const
	Returns the number of bytes of state memory allocated by this operation.
virtual bool	isBuilt () const
	Whether build() completed successfully for a concrete input shape.
virtual bool	isEvalMode () const
	Query whether operation is configured for training.
virtual void	setGradients (ITensor weight_grad, ITensor bias_grad)
	Bind module-owned gradient tensors to the operation.
virtual void	setParameters (ITensor weight, ITensor bias)
	Bind module-owned parameter tensors to the operation.
virtual void	setTrainingMode (TrainingMode training_mode)
	Configure operation training-mode behavior.

Private Attributes
ResidualConfig	config_
CudaExecutionContext *	context_ { nullptr }
int64_t	max_size_ { 0 }
float	scale_ { 1.0f }

Additional Inherited Members
Static Public Attributes inherited from Mila::Dnn::Compute::Operation< TDeviceType, TPrecision >
static constexpr TensorDataType	data_type
static constexpr DeviceType	device_type
Static Protected Member Functions inherited from Mila::Dnn::Compute::BinaryOperation< DeviceType::Cuda, TInputA, TInputA, TInputA >
static const TensorLeftType &	asLeftTensor (const ITensor &t)
static TensorOutputType &	asOutputTensor (ITensor &t)
static const TensorRightType &	asRightTensor (const ITensor &t)
Protected Attributes inherited from Mila::Dnn::Compute::Operation< TDeviceType, TPrecision >
bool	is_built_
TrainingMode	training_mode_

Detailed Description

template<TensorDataType TInputA, TensorDataType TInputB = TInputA, TensorDataType TPrecision = TInputA>
class Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >

CUDA Residual operation implementing the BinaryOperation interface.

Template Parameters

TInputA	Precision of first input tensor
TInputB	Precision of second input tensor
TPrecision	Computation precision

Member Typedef Documentation

◆ ConfigType

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

using Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::ConfigType = ResidualConfig

◆ CudaExecutionContext

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

using Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::CudaExecutionContext = ExecutionContext<DeviceType::Cuda>

◆ MR

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

using Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::MR = CudaDeviceMemoryResource

◆ NativeType

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

using Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::NativeType = typename Mila::Dnn::Compute::Cuda::TensorDataTypeMap<TPrecision>::device_type

Constructor & Destructor Documentation

◆ CudaResidualOp()

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

Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::CudaResidualOp	(	IExecutionContext *	context,
		const ResidualConfig &	config )

inline

Construct with a non-null CUDA execution context.

The context parameter is required and validated at runtime. Callers must pass a std::shared_ptr<ExecutionContext<DeviceType::Cuda>>.

Member Function Documentation

◆ backward()

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

void Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::backward	(	const ITensor &	input_A,
		const ITensor &	input_B,
		const ITensor &	output_grad,
		ITensor &	A_grad,
		ITensor &	B_grad ) const

inlineoverridevirtual

Backward pass.

Implements Mila::Dnn::Compute::BinaryOperation< DeviceType::Cuda, TInputA, TInputA, TInputA >.

◆ build()

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

void Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::build ( const BuildContext & build_context )

inlineoverridevirtual

Prepare the operation for a concrete input shape.

Default implementation is a no-op. Operations requiring shape-dependent setup should override this method.

Reimplemented from Mila::Dnn::Compute::Operation< TDeviceType, TPrecision >.

◆ forward()

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

void Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::forward	(	const ITensor &	input_A,
		const ITensor &	input_B,
		ITensor &	output ) const

inlineoverridevirtual

Forward pass: element-wise addition Y = X1 + X2 using CUDA kernel.

Input and output buffers are obtained from ITensor::rawData() and must be device pointers compatible with the CUDA kernels.

Implements Mila::Dnn::Compute::BinaryOperation< DeviceType::Cuda, TInputA, TInputA, TInputA >.

◆ getName()

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

std::string Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::getName ( ) const

inlineoverridevirtual

Human-readable operation name.

Implements Mila::Dnn::Compute::Operation< TDeviceType, TPrecision >.

◆ getOperationType()

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

OperationType Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::getOperationType ( ) const

inlineoverridevirtual

Operation type identifier.

Implements Mila::Dnn::Compute::Operation< TDeviceType, TPrecision >.

Member Data Documentation

◆ config_

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

ResidualConfig Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::config_

private

◆ context_

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

CudaExecutionContext* Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::context_ { nullptr }

private

◆ max_size_

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

int64_t Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::max_size_ { 0 }

private

◆ scale_

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

float Mila::Dnn::Compute::Cuda::Residual::CudaResidualOp< TInputA, TInputB, TPrecision >::scale_ { 1.0f }

private

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

/__w/Mila/Mila/Mila/Src/Dnn/Compute/Devices/Cuda/Operations/Residual/CudaResidualOp.ixx

Public Types

Public Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

Member Typedef Documentation

◆ ConfigType

◆ CudaExecutionContext

◆ MR

◆ NativeType

Constructor & Destructor Documentation

◆ CudaResidualOp()

Member Function Documentation

◆ backward()

◆ build()

◆ forward()

◆ getName()

◆ getOperationType()

Member Data Documentation

◆ config_

◆ context_

◆ max_size_

◆ scale_