The main class for calculating the atom group properties on GPU. More...

#include <colvaratoms_gpu.h>

Collaboration diagram for colvars_gpu::colvaratoms_gpu:

[legend]

Public Member Functions
	colvaratoms_gpu ()
	Constructor. More...

	~colvaratoms_gpu ()
	Destructor.

int	init_gpu ()
	Initialize the object. More...

int	destroy_gpu ()
	Destructor. More...

int	sync_to_gpu_buffers (const cvm::atom_group *cpu_atoms)
	Synchronize atom-wise data fields from the CPU buffers. More...

int	clear_gpu_buffers (const cvm::atom_group *cpu_atoms)
	Clear GPU atom-wise data fields. More...

int	add_reset_atoms_data_nodes (const cvm::atom_group *cpu_atoms, cudaGraph_t &graph, std::unordered_map< std::string, cudaGraphNode_t > &nodes_map)
	Add a node for resetting some atom-wise data fields to the CUDA graph. More...

int	add_read_positions_nodes (const cvm::atom_group *cpu_atoms, cudaGraph_t &graph, std::unordered_map< std::string, cudaGraphNode_t > &nodes_map)
	Add a node for reading positions and velocities to the CUDA graph. More...

int	add_calc_required_properties_nodes (const cvm::atom_group *cpu_atoms, cudaGraph_t &graph, std::unordered_map< std::string, cudaGraphNode_t > &nodes_map, const std::vector< cudaGraphNode_t > &extra_initial_dependencies={})
	Add nodes for calculating the required properties to the CUDA graph. More...

int	add_update_cpu_buffers_nodes (cvm::atom_group *cpu_atoms, cudaGraph_t &graph, std::unordered_map< std::string, cudaGraphNode_t > &nodes_map)
	Add nodes for updating the CPU buffers after the GPU calculation. More...

int	after_read_data_sync (cvm::atom_group *cpu_atoms, bool copy_to_cpu, cudaStream_t stream)
	Update the CPU COM, COG and rotation object after GPU synchronization. More...

int	begin_apply_force_gpu ()
	Clear the CPU force buffer for scalar components before applying forces on GPU.

int	add_apply_force_nodes (const cvm::atom_group *cpu_atoms, cudaGraph_t &graph, std::unordered_map< std::string, cudaGraphNode_t > &nodes_map, const std::vector< cudaGraphNode_t > &extra_initial_dependencies={})
	Add nodes for applying forces to the CUDA graph. More...

int	add_calc_fit_gradients_nodes (cvm::atom_group *cpu_atoms, cudaGraph_t &graph, std::unordered_map< std::string, cudaGraphNode_t > &nodes_map, bool use_cpu_buffers=false)
	Add nodes for calculating the fit gradients to the CUDA graph. More...

int	read_positions_gpu_debug (cvm::atom_group *cpu_atoms, bool change_fitting_group, size_t change_atom_i, int xyz, bool to_cpu, double sign, cudaStream_t stream)
	Function for reading atom positions used for debug gradients. More...

int	calc_required_properties_gpu_debug (cvm::atom_group *cpu_atoms, bool to_cpu, cudaStream_t stream)
	Function for calculating the required properties used for debug gradients. More...

void	do_feature_side_effects_gpu (cvm::atom_group *cpu_atoms, int id)
	Function to be called when a colvardeps feature is enabled. More...

int	setup_rotation (const cvm::atom_group *cpu_atoms)
	Setup the rotation object and copy the reference positions to GPU. More...

int	setup_rotation_derivative (const cvm::atom_group *cpu_atoms)
	Setup the rotation derivative object on GPU. More...

colvaratoms_gpu_buffer_t &	get_gpu_buffers ()
	Get the GPU buffers. More...

int	read_total_forces (cvm::atom_group *cpu_atoms)
	Read the total forces from the proxy buffer to the GPU buffer. More...

void	apply_colvar_force_from_cpu (cvm::real const &cpu_force)
	Function to intercept the forces applied from the CPU interface. More...

void	set_use_cpu_group_force (bool yesno)
	Set whether to use the CPU atom group force. More...

Private Attributes
colvaratoms_gpu_buffer_t	gpu_buffers

colvaratoms_gpu_calc_fit_info_t	calc_fit_gradients_gpu_info
	Temporary variables for calc_fit_gradients GPU kernel.

colvaratoms_gpu_calc_fit_info_t	calc_fit_forces_gpu_info
	Temporary variables for calc_fit_forces (or "calc_fit_gradients" for vector CVCs) GPU kernel.

colvaratoms_gpu_debug_graph_t	debug_graphs
	Separate CUDA graphs for supporting debug gradients.

cvm::real *	h_sum_applied_colvar_force
	For intercepting the forces applied from the CPU interface.

bool	use_apply_colvar_force
	If the CPU code path use apply_colvar_force(), this will be set to true, and then reset to false in begin_apply_force_gpu()

bool	use_group_force
	If the CPU code path use group_force_object, this will be set to true, and then reset to false in begin_apply_force_gpu()

colvars_gpu::rotation_gpu	rot_gpu
	GPU rotation object.

colvars_gpu::rotation_derivative_gpu *	rot_deriv_gpu
	GPU Rotation derivative;.

Detailed Description

The main class for calculating the atom group properties on GPU.

Constructor & Destructor Documentation

◆ colvaratoms_gpu()

colvars_gpu::colvaratoms_gpu::colvaratoms_gpu ( )

Constructor.

All variables are expected to initialize to 0.

Member Function Documentation

◆ add_apply_force_nodes()

int colvars_gpu::colvaratoms_gpu::add_apply_force_nodes	(	const cvm::atom_group *	cpu_atoms,
		cudaGraph_t &	graph,
		std::unordered_map< std::string, cudaGraphNode_t > &	nodes_map,
		const std::vector< cudaGraphNode_t > &	extra_initial_dependencies = `{}`
	)

Add nodes for applying forces to the CUDA graph.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	graph	CUDA graph object
[in,out]	nodes_map	A map to store the added nodes with operation name
[in]	extra_initial_dependencies	Additional dependencies to be added to the initial nodes

Returns: COLVARS_OK if succeeded

◆ add_calc_fit_gradients_nodes()

int colvars_gpu::colvaratoms_gpu::add_calc_fit_gradients_nodes	(	cvm::atom_group *	cpu_atoms,
		cudaGraph_t &	graph,
		std::unordered_map< std::string, cudaGraphNode_t > &	nodes_map,
		bool	use_cpu_buffers = `false`
	)

Add nodes for calculating the fit gradients to the CUDA graph.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	graph	CUDA graph object
[in,out]	nodes_map	A map to store the added nodes with operation name
[in]	use_cpu_buffers	If true, add a node for copying fit gradients from GPU to CPU

Returns: COLVARS_OK if succeeded

◆ add_calc_required_properties_nodes()

int colvars_gpu::colvaratoms_gpu::add_calc_required_properties_nodes	(	const cvm::atom_group *	cpu_atoms,
		cudaGraph_t &	graph,
		std::unordered_map< std::string, cudaGraphNode_t > &	nodes_map,
		const std::vector< cudaGraphNode_t > &	extra_initial_dependencies = `{}`
	)

Add nodes for calculating the required properties to the CUDA graph.

This function adds nodes for calculating COM, COG, rotation matrix, rotated positions, and reference COG to the CUDA graph.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	graph	CUDA graph object
[in,out]	nodes_map	A map to store the added nodes with operation name
[in]	extra_initial_dependencies	Additional dependencies to be added to the initial nodes

Returns: COLVARS_OK if succeeded

◆ add_read_positions_nodes()

int colvars_gpu::colvaratoms_gpu::add_read_positions_nodes	(	const cvm::atom_group *	cpu_atoms,
		cudaGraph_t &	graph,
		std::unordered_map< std::string, cudaGraphNode_t > &	nodes_map
	)

Add a node for reading positions and velocities to the CUDA graph.

This function adds cudaMemcpy nodes to read positions and velocities from the CPU buffers to the GPU buffers to the CUDA graph.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	graph	CUDA graph object
[in,out]	nodes_map	A map to store the added nodes with operation name

Returns: COLVARS_OK if succeeded

◆ add_reset_atoms_data_nodes()

int colvars_gpu::colvaratoms_gpu::add_reset_atoms_data_nodes	(	const cvm::atom_group *	cpu_atoms,
		cudaGraph_t &	graph,
		std::unordered_map< std::string, cudaGraphNode_t > &	nodes_map
	)

Add a node for resetting some atom-wise data fields to the CUDA graph.

This function adds cudaMemset nodes to clear gradients, total forces and applied forces to the CUDA graph. The positions and velocities are not cleared since they are expected to be overwritten.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	graph	CUDA graph object
[in,out]	nodes_map	A map to store the added nodes with operation name

Returns: COLVARS_OK if succeeded

◆ add_update_cpu_buffers_nodes()

int colvars_gpu::colvaratoms_gpu::add_update_cpu_buffers_nodes	(	cvm::atom_group *	cpu_atoms,
		cudaGraph_t &	graph,
		std::unordered_map< std::string, cudaGraphNode_t > &	nodes_map
	)

Add nodes for updating the CPU buffers after the GPU calculation.

This function adds cudaMemcpy nodes to update COM, COG, original COG, rotation matrix, and fit gradients to the CPU buffers.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	graph	CUDA graph object
[in,out]	nodes_map	A map to store the added nodes with operation name

Returns: COLVARS_OK if succeeded

◆ after_read_data_sync()

int colvars_gpu::colvaratoms_gpu::after_read_data_sync	(	cvm::atom_group *	cpu_atoms,
		bool	copy_to_cpu,
		cudaStream_t	stream
	)

Update the CPU COM, COG and rotation object after GPU synchronization.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	copy_to_cpu	If true, copy the COM, COG and rotation object to CPU buffers
[in]	stream	CUDA stream to be synchronized

Returns: COLVARS_OK if succeeded

◆ apply_colvar_force_from_cpu()

void colvars_gpu::colvaratoms_gpu::apply_colvar_force_from_cpu ( cvm::real const & cpu_force )

Function to intercept the forces applied from the CPU interface.

This function is called when apply_colvar_force() is called on CPU. It stores the applied force in a temporary variable, which will be added to the GPU buffer.

Parameters

[in] cpu_force Force applied from the CPU interface

◆ calc_required_properties_gpu_debug()

int colvars_gpu::colvaratoms_gpu::calc_required_properties_gpu_debug	(	cvm::atom_group *	cpu_atoms,
		bool	to_cpu,
		cudaStream_t	stream
	)

Function for calculating the required properties used for debug gradients.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	to_cpu	If true, copy the calculated properties to CPU buffers
[in]	stream	CUDA stream to be used

Returns: COLVARS_OK if succeeded

◆ clear_gpu_buffers()

int colvars_gpu::colvaratoms_gpu::clear_gpu_buffers ( const cvm::atom_group * cpu_atoms )

Clear GPU atom-wise data fields.

This function clears the GPU buffers for proxy indices, charges, masses, weights, positions, gradients, total forces, velocities, and applied forces.

Parameters

[in] cpu_atoms CPU atom group class

Returns: COLVARS_OK if succeeded

◆ destroy_gpu()

int colvars_gpu::colvaratoms_gpu::destroy_gpu ( )

Destructor.

Returns: COLVARS_OK if succeeded

◆ do_feature_side_effects_gpu()

void colvars_gpu::colvaratoms_gpu::do_feature_side_effects_gpu	(	cvm::atom_group *	cpu_atoms,
		int	id
	)

Function to be called when a colvardeps feature is enabled.

This function performs any side effects needed when a feature is enabled. Specifically, it allocates the GPU buffers for unrotated positions if f_ag_fit_gradients is enabled, and sets up the rotation object if f_ag_rotate is enabled.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	id	Feature ID that has been enabled

◆ get_gpu_buffers()

colvaratoms_gpu_buffer_t & colvars_gpu::colvaratoms_gpu::get_gpu_buffers ( )

inline

Get the GPU buffers.

Returns: Reference to the GPU buffers struct

◆ init_gpu()

int colvars_gpu::colvaratoms_gpu::init_gpu ( )

Initialize the object.

This function re-allocates the GPU buffers for COM, COG, atomic counters, reference COG, and the temporary variables used for fit gradients. However, it does not (re)allocate the atom-wise data fields like positions. The atom-wise data fields are expected to be updated when sync_to_gpu_buffers is called.

Returns: COLVARS_OK if succeeded

◆ read_positions_gpu_debug()

int colvars_gpu::colvaratoms_gpu::read_positions_gpu_debug	(	cvm::atom_group *	cpu_atoms,
		bool	change_fitting_group,
		size_t	change_atom_i,
		int	xyz,
		bool	to_cpu,
		double	sign,
		cudaStream_t	stream
	)

Function for reading atom positions used for debug gradients.

This function reads the atom positions from the proxy buffer to the atom group GPU buffer, and optionally modifies one of the coordinates for testing the numerical gradients.

Parameters

[in]	cpu_atoms	CPU atom group class
[in]	change_fitting_group	If true, change the fitting group atom position
[in]	change_atom_i	Index of the atom to be changed
[in]	xyz	Coordinate index (0, 1, or 2) to be changed
[in]	to_cpu	If true, copy the changed positions to CPU buffers
[in]	sign	Sign of the change (+1 or -1)
[in]	stream	CUDA stream to be used

Returns: COLVARS_OK if succeeded

◆ read_total_forces()

int colvars_gpu::colvaratoms_gpu::read_total_forces ( cvm::atom_group * cpu_atoms )

Read the total forces from the proxy buffer to the GPU buffer.

Parameters

[in] cpu_atoms CPU atom group class

Returns: COLVARS_OK if succeeded

◆ set_use_cpu_group_force()

void colvars_gpu::colvaratoms_gpu::set_use_cpu_group_force ( bool yesno )

inline

Set whether to use the CPU atom group force.

This function is called when group_force_object is used on CPU. It sets a flag to indicate that the group force should be added to the GPU buffer.

Parameters

[in] yesno True to use the CPU group force, false otherwise

◆ setup_rotation()

int colvars_gpu::colvaratoms_gpu::setup_rotation ( const cvm::atom_group * cpu_atoms )

Setup the rotation object and copy the reference positions to GPU.

Parameters

[in] cpu_atoms CPU atom group class

Returns: COLVARS_OK if succeeded

◆ setup_rotation_derivative()

int colvars_gpu::colvaratoms_gpu::setup_rotation_derivative ( const cvm::atom_group * cpu_atoms )

Setup the rotation derivative object on GPU.

Parameters

[in] cpu_atoms CPU atom group class

Returns: COLVARS_OK if succeeded

◆ sync_to_gpu_buffers()

int colvars_gpu::colvaratoms_gpu::sync_to_gpu_buffers ( const cvm::atom_group * cpu_atoms )

Synchronize atom-wise data fields from the CPU buffers.

This function (re)allocates the GPU buffers for proxy indices, charges, masses, weights, positions, velocities, total forces, gradients, and forces that will be applied. Then the proxy indices, charges, masses, positions, velocities, gradients and total forces will be updated from the corresponding CPU buffers.

Parameters

[in] cpu_atoms CPU atom group class

Returns: COLVARS_OK if succeeded

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

colvaratoms_gpu.h
colvaratoms_gpu.cpp

Public Member Functions

Private Attributes

Detailed Description

Constructor & Destructor Documentation

◆ colvaratoms_gpu()

Member Function Documentation

◆ add_apply_force_nodes()

◆ add_calc_fit_gradients_nodes()

◆ add_calc_required_properties_nodes()

◆ add_read_positions_nodes()

◆ add_reset_atoms_data_nodes()

◆ add_update_cpu_buffers_nodes()

◆ after_read_data_sync()

◆ apply_colvar_force_from_cpu()

◆ calc_required_properties_gpu_debug()

◆ clear_gpu_buffers()

◆ destroy_gpu()

◆ do_feature_side_effects_gpu()

◆ get_gpu_buffers()

◆ init_gpu()

◆ read_positions_gpu_debug()

◆ read_total_forces()

◆ set_use_cpu_group_force()

◆ setup_rotation()

◆ setup_rotation_derivative()

◆ sync_to_gpu_buffers()