CPU implementation of the Fully Connected operation for neural networks. More...

Inheritance diagram for Mila::Dnn::Compute::CpuLinearOp:

[legend]

Collaboration diagram for Mila::Dnn::Compute::CpuLinearOp:

[legend]

Public Types
using	MR = typename CpuDevice::MR

using	OperationBase = UnaryOperation< DeviceType::Cpu, float >

Public Types inherited from Mila::Dnn::Compute::UnaryOperation< DeviceType::Cpu, float >
using	MR = std::conditional_t< TDeviceType==DeviceType::Cuda, CudaMemoryResource, HostMemoryResource >
	Memory resource type based on device type.

Public Member Functions
	CpuLinearOp (const LinearConfig &config)
	Constructs a new CPU Fully Connected operation with the default device context.

	CpuLinearOp (std::shared_ptr< DeviceContext > context, const LinearConfig &config)
	Constructs a new CPU Fully Connected operation with a specific device context.

void	backward (Tensor< float, MR > &input_grad, const std::vector< std::shared_ptr< Tensor< float, MR > > > &parameter_grads, const Tensor< float, MR > &output_grad, const Tensor< float, MR > input, const Tensor< float, MR > weight, int B, int T, int C, int OC)
	Performs the backward pass of the Fully Connected operation.

void	forward (const Tensor< float, MR > &input, const std::vector< std::shared_ptr< Tensor< float, MR > > > &parameters, const OperationAttributes &properties, Tensor< float, MR > &output, std::vector< std::shared_ptr< Tensor< float, MR > > > &output_state) const override
	Performs the forward pass of the Linear operation.

std::string	getName () const override
	Gets the name of this operation.

Public Member Functions inherited from Mila::Dnn::Compute::UnaryOperation< DeviceType::Cpu, float >
	UnaryOperation (OperationType operation_type)
	Constructs a UnaryOperation with the specified operation type.

	UnaryOperation (OperationType operation_type, std::shared_ptr< DeviceContext > context)
	Constructs a UnaryOperation with the specified operation type and device context.

virtual	~UnaryOperation ()=default
	Virtual destructor for proper cleanup of derived classes.

virtual void	backward (const Tensor< float, MR > &grad, const std::vector< std::shared_ptr< Tensor< float, MR > > > &parameters, std::vector< std::shared_ptr< Tensor< float, MR > > > &output_grads) const
	Executes the backward pass of a unary operation.

virtual void	backward (const Tensor< float, MR > &input, const Tensor< float, MR > &output_grad, const std::vector< std::shared_ptr< Tensor< float, MR > > > &parameters, std::vector< std::shared_ptr< Tensor< float, MR > > > &parameter_grads, Tensor< float, MR > &input_grad, const OperationAttributes &properties, const std::vector< std::shared_ptr< Tensor< float, MR > > > &output_state) const
	Executes the comprehensive backward pass of a unary operation.

virtual void	forward (const Tensor< float, MR > &input, const std::vector< std::shared_ptr< Tensor< float, MR > > > &parameters, const OperationAttributes &properties, Tensor< float, MR > &output, std::vector< std::shared_ptr< Tensor< float, MR > > > &output_state) const=0
	Executes the forward pass of a unary operation.

Public Member Functions inherited from Mila::Dnn::Compute::OperationBase< TDeviceType, TInput1, TInput2, TOutput >
	OperationBase (OperationType operation_type, std::shared_ptr< DeviceContext > context)
	Constructs an OperationBase object with a specific device context and compute precision.

virtual	~OperationBase ()=default
	Virtual destructor for the OperationBase class.

std::shared_ptr< DeviceContext >	getDeviceContext () const
	Gets the device context associated with this operation.

DeviceType	getDeviceType () const
	Gets the device type for this operation.

OperationType	getOperationType () const
	Gets the operation type enumeration value.

Private Member Functions
void	forward_naive (const Tensor< float, MR > &input, const std::shared_ptr< Tensor< float, MR > > &weight, const std::shared_ptr< Tensor< float, MR > > &bias, Tensor< float, MR > &output, int outer_size, int C, int OC) const
	Naive implementation of the forward pass for the Fully Connected operation.

Private Attributes
LinearConfig	config_
	Configuration for the linear operation.

Detailed Description

CPU implementation of the Fully Connected operation for neural networks.

This class provides a CPU-based implementation of the Fully Connected operation, which performs a matrix multiplication between the input and a weight matrix, optionally adds a bias, and produces an output. This operation implements the standard linear layer commonly used in neural networks.

The implementation includes both a performance-optimized version with loop unrolling and a naive fallback implementation for special cases.

Template Parameters

float	The data type of the input tensor elements.
TDataType	The data type used for computation and output (defaults to the input type).

Member Typedef Documentation

◆ MR

using Mila::Dnn::Compute::CpuLinearOp::MR = typename CpuDevice::MR

◆ OperationBase

using Mila::Dnn::Compute::CpuLinearOp::OperationBase = UnaryOperation<DeviceType::Cpu, float>

Constructor & Destructor Documentation

◆ CpuLinearOp() [1/2]

Mila::Dnn::Compute::CpuLinearOp::CpuLinearOp ( const LinearConfig & config )

inline

Constructs a new CPU Fully Connected operation with the default device context.

CPU operations always use full precision regardless of policy settings.

Parameters

precision_policy Ignored for CPU operations, as they always use full precision.

◆ CpuLinearOp() [2/2]

Mila::Dnn::Compute::CpuLinearOp::CpuLinearOp	(	std::shared_ptr< DeviceContext >	context,
		const LinearConfig &	config
	)

inline

Constructs a new CPU Fully Connected operation with a specific device context.

CPU operations always use full precision regardless of policy settings.

Parameters

context	The device context to use for this operation.
precision_policy	Ignored for CPU operations, as they always use full precision.

Exceptions

std::runtime_error If the context is not for a CPU device.

Member Function Documentation

◆ backward()

void Mila::Dnn::Compute::CpuLinearOp::backward	(	Tensor< float, MR > &	input_grad,
		const std::vector< std::shared_ptr< Tensor< float, MR > > > &	parameter_grads,
		const Tensor< float, MR > &	output_grad,
		const Tensor< float, MR >	input,
		const Tensor< float, MR >	weight,
		int	B,
		int	T,
		int	C,
		int	OC
	)

inline

Performs the backward pass of the Fully Connected operation.

Computes gradients with respect to inputs, weights, and biases based on the output gradient.

Parameters

dinp	Pointer to the gradient buffer for input.
dweight	Pointer to the gradient buffer for weight parameters.
dbias	Pointer to the gradient buffer for bias parameters (can be NULL if no bias is used).
dout	Pointer to the gradient buffer from the output.
inp	Pointer to the original input values.
weight	Pointer to the weight parameters.
B	Batch size.
TDataType	Sequence length.
C	Input feature dimension.
OC	Output feature dimension.

Here is the call graph for this function:

◆ forward()

void Mila::Dnn::Compute::CpuLinearOp::forward	(	const Tensor< float, MR > &	input,
		const std::vector< std::shared_ptr< Tensor< float, MR > > > &	parameters,
		const OperationAttributes &	properties,
		Tensor< float, MR > &	output,
		std::vector< std::shared_ptr< Tensor< float, MR > > > &	output_state
	)		const

inlineoverride

Performs the forward pass of the Linear operation.

Computes the matrix multiplication between input and weights, adds bias if provided, and stores the result in the output tensor. Uses loop unrolling for performance optimization when possible, otherwise falls back to a naive implementation.

Parameters

input	Input tensor of shape [B, TDataType, C] where B is batch size, TDataType is sequence length, and C is input feature dimension.
parameters	Vector of parameter tensors [weight, bias] where weight is of shape [OC, C] and bias (optional) is of shape [OC].
properties	Additional attributes for the operation.
output	Output tensor of shape [B, TDataType, OC] where OC is output feature dimension.
output_state	Cache for intermediate results (not used in this operation).

Here is the call graph for this function:

◆ forward_naive()

void Mila::Dnn::Compute::CpuLinearOp::forward_naive	(	const Tensor< float, MR > &	input,
		const std::shared_ptr< Tensor< float, MR > > &	weight,
		const std::shared_ptr< Tensor< float, MR > > &	bias,
		Tensor< float, MR > &	output,
		int	outer_size,
		int	C,
		int	OC
	)		const

inlineprivate

Naive implementation of the forward pass for the Fully Connected operation.

This is a simple implementation without optimizations that serves as a fallback for cases where the optimized implementation cannot be used.

Parameters

input	Input tensor.
weight	Weight tensor.
bias	Bias tensor (optional).
output	Output tensor.
B	Batch size.
TDataType	Sequence length.
C	Input feature dimension.
OC	Output feature dimension.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ getName()

std::string Mila::Dnn::Compute::CpuLinearOp::getName ( ) const

inlineoverridevirtual

Gets the name of this operation.

Returns: std::string The name of the operation ("Cpu::LinearOp").

Implements Mila::Dnn::Compute::OperationBase< TDeviceType, TInput1, TInput2, TOutput >.

Member Data Documentation

◆ config_

LinearConfig Mila::Dnn::Compute::CpuLinearOp::config_

private

Configuration for the linear operation.

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

/home/runner/work/Mila/Mila/Mila/Src/Dnn/Compute/Operations/Cpu/CpuLinearOp.ixx

Public Types

Public Member Functions

Private Member Functions

Private Attributes

Detailed Description

Member Typedef Documentation

◆ MR

◆ OperationBase

Constructor & Destructor Documentation

◆ CpuLinearOp() [1/2]

◆ CpuLinearOp() [2/2]

Member Function Documentation

◆ backward()

◆ forward()

◆ forward_naive()

◆ getName()

Member Data Documentation

◆ config_