Base class for GPU-based nonlinear factors. More...

#include <nonlinear_factor_gpu.hpp>

Inheritance diagram for gtsam_points::NonlinearFactorGPU:

[legend]

Collaboration diagram for gtsam_points::NonlinearFactorGPU:

Public Types
using	shared_ptr = gtsam_points::shared_ptr< NonlinearFactorGPU >

Public Member Functions
template<typename CONTAINER >
	NonlinearFactorGPU (const CONTAINER &keys)

virtual size_t	linearization_input_size () const =0
	Size of data to be uploaded to the GPU before linearization.

virtual size_t	linearization_output_size () const =0
	Size of data to be downloaded from the GPU after linearization.

virtual size_t	evaluation_input_size () const =0
	Size of data to be uploaded to the GPU before cost evaluation.

virtual size_t	evaluation_output_size () const =0
	Size of data to be downloaded from the GPU after cost evaluation.

virtual void	set_linearization_point (const gtsam::Values &values, void *lin_input_cpu)=0
	Write linearization input data to the upload buffer.

virtual void	issue_linearize (const void lin_input_cpu, const void lin_input_gpu, void *lin_output_gpu)=0
	Issue linearization task.

virtual void	store_linearized (const void *lin_output_cpu)=0
	Read linearization output data from the download buffer.

virtual void	set_evaluation_point (const gtsam::Values &values, void *eval_input_cpu)=0
	Write cost evaluation input data to the upload buffer.

virtual void	issue_compute_error (const void lin_input_cpu, const void eval_input_cpu, const void lin_input_gpu, const void eval_input_gpu, void *eval_output_gpu)=0
	Issue cost evaluation task.

virtual void	store_computed_error (const void *eval_output_cpu)=0
	Read cost evaluation output data from the download buffer.

virtual void	sync ()=0
	Perform CPU-GPU synchronization and wait for the task.

Detailed Description

Base class for GPU-based nonlinear factors.

Note: To efficiently perform linearization (and cost evaluation) on a GPU, we issue all linearization tasks of GPU-based factors and copy all required data for linearization (e.g., current estimate) at once.

To allow gtsam_points::NonlinearFactorSetGPU to manage linearization, you need to implement the following methods:

linearization_(input|output)_size() : Define the size of input and result data for linearization
set_linearization_point() : Write the data to be uploaded to the GPU before linearization
issue_linearization() : Issue the linearization task
sync() : Perform CPU-GPU synchronization to wait for linearization completion
store_linearized() : Read back the data from the download buffer after linearization

For example, implementation for the linearization of the standard ICP factor would be as follows:

linearization_input_size() : sizeof(Eigen::Isometry3f) Current estimate of T_target_source
linearization_output_size() : sizeof(LinearizedSystem6) Linearized Hessian factor
set_linearization_point() : Write the current T_target_source to the lin_input_cpu
issue_linearization() : Issue the ICP computation task
sync() : Wait for the ICP computation task
store_linearized() : Read the linearized factor from lin_output_cpu

Optimizers in gtsam_points calls NonlinearFactorSetGPU's linearization routine before calling linearize() of each factor. You should thus store the linearized factor to a temporary member and just return it when linearize() is called.

Member Function Documentation

◆ evaluation_input_size()

virtual size_t gtsam_points::NonlinearFactorGPU::evaluation_input_size ( ) const

pure virtual

Size of data to be uploaded to the GPU before cost evaluation.

Returns: size_t

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ evaluation_output_size()

virtual size_t gtsam_points::NonlinearFactorGPU::evaluation_output_size ( ) const

pure virtual

Size of data to be downloaded from the GPU after cost evaluation.

Returns: size_t

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ issue_compute_error()

virtual void gtsam_points::NonlinearFactorGPU::issue_compute_error	(	const void *	lin_input_cpu,
		const void *	eval_input_cpu,
		const void *	lin_input_gpu,
		const void *	eval_input_gpu,
		void *	eval_output_gpu
	)

pure virtual

Issue cost evaluation task.

Parameters

lin_input_cpu	Linearization input data on the CPU memory (size == linearization_input_size)
eval_input_cpu	Cost evaluation input data on the CPU memory (size == evaluation_input_size)
lin_input_gpu	Linearization input data on the GPU memory (size == linearization_input_size)
eval_input_gpu	Cost evaluation input data on the GPU memory (size == evaluation_input_size)
eval_output_gpu	Output data destination on the GPU memory (size == evaluation_output_size)

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ issue_linearize()

virtual void gtsam_points::NonlinearFactorGPU::issue_linearize	(	const void *	lin_input_cpu,
		const void *	lin_input_gpu,
		void *	lin_output_gpu
	)

pure virtual

Issue linearization task.

Parameters

lin_input_cpu	Linearization input data on the CPU memory (size == linearization_input_size)
lin_input_gpu	Linearization input data on the GPU memory (size == linearization_input_size)
lin_output_gpu	Output data destination on the GPU memory (size == linearization_output_size)

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ linearization_input_size()

virtual size_t gtsam_points::NonlinearFactorGPU::linearization_input_size ( ) const

pure virtual

Size of data to be uploaded to the GPU before linearization.

Returns: size_t

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ linearization_output_size()

virtual size_t gtsam_points::NonlinearFactorGPU::linearization_output_size ( ) const

pure virtual

Size of data to be downloaded from the GPU after linearization.

Returns: size_t

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ set_evaluation_point()

virtual void gtsam_points::NonlinearFactorGPU::set_evaluation_point	(	const gtsam::Values &	values,
		void *	eval_input_cpu
	)

pure virtual

Write cost evaluation input data to the upload buffer.

Parameters

values	Current estimate
eval_input_cpu	Data upload buffer (size == evaluation_input_size)

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ set_linearization_point()

virtual void gtsam_points::NonlinearFactorGPU::set_linearization_point	(	const gtsam::Values &	values,
		void *	lin_input_cpu
	)

pure virtual

Write linearization input data to the upload buffer.

Parameters

values	Current estimate
lin_input_cpu	Data upload buffer (size == linearization_input_size)

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ store_computed_error()

virtual void gtsam_points::NonlinearFactorGPU::store_computed_error ( const void * eval_output_cpu )

pure virtual

Read cost evaluation output data from the download buffer.

Parameters

eval_output_cpu Data down load buffer (size == evaluation_output_size)

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ store_linearized()

virtual void gtsam_points::NonlinearFactorGPU::store_linearized ( const void * lin_output_cpu )

pure virtual

Read linearization output data from the download buffer.

Parameters

lin_output_cpu Data download buffer (size == linearization_output_size)

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

◆ sync()

virtual void gtsam_points::NonlinearFactorGPU::sync ( )

pure virtual

Perform CPU-GPU synchronization and wait for the task.

Implemented in gtsam_points::IntegratedVGICPFactorGPU.

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

/home/runner/work/gtsam_points/gtsam_points/include/gtsam_points/factors/nonlinear_factor_gpu.hpp

Public Types

Public Member Functions

Detailed Description

Member Function Documentation

◆ evaluation_input_size()

◆ evaluation_output_size()

◆ issue_compute_error()

◆ issue_linearize()

◆ linearization_input_size()

◆ linearization_output_size()

◆ set_evaluation_point()

◆ set_linearization_point()

◆ store_computed_error()

◆ store_linearized()

◆ sync()