Reduce the cells' maximum eigenvalues and derive the next global time step size from the CFL condition. More...

Inheritance diagram for exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition:

[legend]

Collaboration diagram for exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition:

[legend]

Public Member Functions
	__init__ (self, solver, cfl_constant, patch_size, guard, max_eigenvalue=None)

	get_body_of_prepareTraversal (self)
	prepareTraversal() is static - called once per rank, before any instance of this action set is created - so it can only ever touch static members, never a per-instance one.

	get_body_of_unprepareTraversal (self)
	Called once per rank, after every instance of this action set for the current grid sweep has already been destroyed (see get_body_of_prepareTraversal()'s note on the static/per-instance split) - the one safe place both to read _maxEigenvalue (nothing can still be writing it) and to issue an MPI reduction: tarch::mpi::Rank::allReduce()'s own documentation requires "only one thread per rank" to ever call it, which a once-per-rank static method already guarantees.

	get_attributes (self)
	Return attributes as copied and pasted into the generated class.

	get_static_initialisations (self, full_qualified_classname)

	get_includes (self)
	Return include statements that you need.

	get_body_of_operation (self, operation_name)
	Return actual C++ code snippets to be inserted into C++ code.

	get_guard_of_operation (self, operation_name)
	Return the guard of an operation.

	get_action_set_name (self)
	You should replicate this function in each subclass, so you get meaningful action set names (otherwise, it will be AbstractFVActionSet0,1,2,...).

Public Member Functions inherited from exahype2.api.fv.actionsets.AbstractFVActionSet.AbstractFVActionSet
	user_should_modify_template (self)
	Is the user allowed to modify the output.

Public Member Functions inherited from peano4.api.solversteps.ActionSet.ActionSet
	get_constructor_body (self)
	Define a tailored constructor body.

	get_destructor_body (self)

	get_body_of_createGridControlEventsForNextTraversal (self)
	Return the grid control events that this action set wants to trigger.

	get_implementation_of_createGridControlEventsForNextTraversal (self)
	Return the C++ implementation of createGridControlEventsForNextTraversal()

	get_implementation_of_prepareTraversal (self)
	Return the C++ implementation of prepareTraversal()

	get_implementation_of_unprepareTraversal (self)
	Return the C++ implementation of unprepareTraversal()

	get_implementation_of_operation (self, operation_name)
	Return the C++ implementation of an operation.

bool	may_run_in_background_task (self)
	May a task graph compiler move touchCellFirstTime() into a task.

Data Fields
	guard

	cfl_constant

	patch_size

Data Fields inherited from peano4.api.solversteps.ActionSet.ActionSet
	descend_invocation_order

	parallel

Protected Attributes
	_max_eigenvalue

Protected Attributes inherited from exahype2.api.fv.actionsets.AbstractFVActionSet.AbstractFVActionSet
	_solver

Additional Inherited Members
Static Public Attributes inherited from peano4.api.solversteps.ActionSet.ActionSet
int	descend_invocation_order = 0

bool	parallel = False

str	OPERATION_BEGIN_TRAVERSAL = "beginTraversal"

str	OPERATION_END_TRAVERSAL = "endTraversal"

str	OPERATION_CREATE_PERSISTENT_VERTEX = "createPersistentVertex"

str	OPERATION_DESTROY_PERSISTENT_VERTEX = "destroyPersistentVertex"

str	OPERATION_CREATE_HANGING_VERTEX = "createHangingVertex"

str	OPERATION_DESTROY_HANGING_VERTEX = "destroyHangingVertex"

str	OPERATION_CREATE_PERSISTENT_FACE = "createPersistentFace"

str	OPERATION_DESTROY_PERSISTENT_FACE = "destroyPersistentFace"

str	OPERATION_CREATE_HANGING_FACE = "createHangingFace"

str	OPERATION_DESTROY_HANGING_FACE = "destroyHangingFace"

str	OPERATION_CREATE_CELL = "createCell"

str	OPERATION_DESTROY_CELL = "destroyCell"

str	OPERATION_TOUCH_VERTEX_FIRST_TIME = "touchVertexFirstTime"

str	OPERATION_TOUCH_VERTEX_LAST_TIME = "touchVertexLastTime"

str	OPERATION_TOUCH_FACE_FIRST_TIME = "touchFaceFirstTime"

str	OPERATION_TOUCH_FACE_LAST_TIME = "touchFaceLastTime"

str	OPERATION_TOUCH_CELL_FIRST_TIME = "touchCellFirstTime"

str	OPERATION_TOUCH_CELL_LAST_TIME = "touchCellLastTime"

Detailed Description

Reduce the cells' maximum eigenvalues and derive the next global time step size from the CFL condition.

exahype2.api.fv.actionsets.UpdateCell stores the maximum eigenvalue of every patch it has updated in the cell data solver._max_eigenvalue. This action set runs in a subsequent mesh traversal: touchCellFirstTime() folds each cell's value into the static _maxEigenvalue, and unprepareTraversal() turns the global maximum into the next time step size and feeds it into exahype2::Solver's own time step bookkeeping (updateCell()/finishTimeStep()/startTimeStep()). There is deliberately no separate, solver-level time stamp/time step size state on top of exahype2::Solver.

Global time stepping: all cells share one time stamp and time step size, which are taken from the solver (getMinTimeStamp() and getMinTimeStepSize()) rather than from per-cell bookkeeping.

The kernel updates all dimensions in one explicit Euler step, so the CFL condition reads dt * sum_d lambda_d/h_d <= 1, i.e. the admissible time step size is cfl_constant * h / (Dimensions * lambda_max), with h the minimal volume size and lambda_max the maximum eigenvalue.

As this is global time stepping, getMinTimeStamp()==getMaxTimeStamp() and getMinTimeStepSize()==getMaxTimeStepSize() always hold. Hence a single updateCell() call per time step, not one per cell, is enough to keep both the min and the max in sync.

Definition at line 8 of file ComputeGlobalCFLCondition.py.

Constructor & Destructor Documentation

◆ init()

exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition.__init__	(	self,
		solver,
		cfl_constant,
		patch_size,
		guard,
		max_eigenvalue = `None`
	)

Parameters

solver exahype2.api.Solver Has to have solver._max_eigenvalue set -

See also: exahype2.api.fv.RusanovWithExplicitEuler.

Parameters

cfl_constant float Safety factor the admissible time step size is scaled down by, relative to the theoretical CFL=1 stability limit (which already accounts for the number of dimensions, see class docstring) -

See also: exahype2.api.fv.RusanovWithExplicitEuler's own cfl_constant parameter, which this is meant to be passed straight through from.

Parameters

patch_size	int Number of Finite Volumes per patch and dimension. The CFL condition is evaluated over the volume size, i.e. the cell size divided by patch_size.
guard	str C++ boolean expression, evaluated per cell in touchCellFirstTime, guarding whether that cell contributes to the CFL condition. Has to match the guard of exahype2.api.fv.actionsets.UpdateCell, as only the cells updated there hold a valid maximum eigenvalue.
max_eigenvalue	peano4.api.datamodel.DaStGen2 The cell data holding the maximum eigenvalue (first entry). Defaults to solver._max_eigenvalue.

Reimplemented from exahype2.api.fv.actionsets.AbstractFVActionSet.AbstractFVActionSet.

Definition at line 40 of file ComputeGlobalCFLCondition.py.

Member Function Documentation

◆ get_action_set_name()

exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition.get_action_set_name ( self )

You should replicate this function in each subclass, so you get meaningful action set names (otherwise, it will be AbstractFVActionSet0,1,2,...).

Reimplemented from exahype2.api.fv.actionsets.AbstractFVActionSet.AbstractFVActionSet.

Definition at line 222 of file ComputeGlobalCFLCondition.py.

◆ get_attributes()

exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition.get_attributes ( self )

Return attributes as copied and pasted into the generated class.

Please note that action sets are not persistent, i.e. there is one object creation per grid sweep per tree.

Reimplemented from peano4.api.solversteps.ActionSet.ActionSet.

Definition at line 178 of file ComputeGlobalCFLCondition.py.

◆ get_body_of_operation()

exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition.get_body_of_operation	(	self,
		operation_name
	)

Return actual C++ code snippets to be inserted into C++ code.

See class' string constants starting with OPERATION_ for possible values of operation_name. The snippet is executed if and only if the guard returned by get_guard_of_operation() holds, i.e. the body should not evaluate the guard itself. See the section "Guard and body" in the class documentation.

Reimplemented from peano4.api.solversteps.ActionSet.ActionSet.

Definition at line 208 of file ComputeGlobalCFLCondition.py.

Referenced by peano4.api.solversteps.ActionSet.ActionSet.get_implementation_of_operation().

Here is the caller graph for this function:

◆ get_body_of_prepareTraversal()

exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition.get_body_of_prepareTraversal ( self )

prepareTraversal() is static - called once per rank, before any instance of this action set is created - so it can only ever touch static members, never a per-instance one.

_maxEigenvalue is static for exactly that reason: it is reset here, once, before every grid sweep, and then accumulated into by potentially many instances of this action set (one per tree, possibly running on different threads concurrently) as touchCellFirstTime() visits the cells - see _semaphore, which guards that concurrent access.

Reimplemented from peano4.api.solversteps.ActionSet.ActionSet.

Definition at line 77 of file ComputeGlobalCFLCondition.py.

Referenced by peano4.api.solversteps.ActionSet.ActionSet.get_implementation_of_prepareTraversal().

Here is the caller graph for this function:

◆ get_body_of_unprepareTraversal()

exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition.get_body_of_unprepareTraversal ( self )

Called once per rank, after every instance of this action set for the current grid sweep has already been destroyed (see get_body_of_prepareTraversal()'s note on the static/per-instance split) - the one safe place both to read _maxEigenvalue (nothing can still be writing it) and to issue an MPI reduction: tarch::mpi::Rank::allReduce()'s own documentation requires "only one thread per rank" to ever call it, which a once-per-rank static method already guarantees.

This combines two different kinds of "global", reduced two different ways:

The mesh size comes out of peano4::grid::GridStatistics, which peano4::parallel::SpacetreeSet reduces over all ranks at the end of every traversal (Grid statistics and other global mesh information). Every rank hence reads the same, global value. GridStatistics holds the cell size; the CFL condition needs the volume size, i.e. the cell size divided by the patch size.
_maxEigenvalue is a plain double, already only a local, per-rank accumulation (see _semaphore) - that one is reduced directly via tarch::mpi::Rank::allReduce(..., MPI_MAX, ...).

Once newTimeStepSize is known, it folds it into the solver's own time step bookkeeping (see this class' docstring for why a single updateCell() call, not one per cell, is correct for this solver). newTimeStepSize is not truncated by MaxTerminalTime: the last time step may overshoot the end time, and the simulation stops afterwards, i.e. MaxTerminalTime is a lower bound on the final time. The time stamp is advanced by the time step size the update used (getMaxTimeStepSize() before updateCell()), not by newTimeStepSize, which only the next update uses. The very first time step therefore runs with time step size 0 and leaves the time stamp at 0: it only yields the first admissible time step size.

updateCell(newTimeStepSize, newTimeStamp) folds the new values into exahype2::Solver's ...UpdatedThisTimeStep accumulators.
finishTimeStep() then MPI-all-reduces that accumulator across ranks - a no-op given newTimeStepSize/newTimeStamp are already identical on every rank, but calling it anyway keeps the full exahype2::Solver lifecycle honest for any future override that relies on it actually running.
startTimeStep(...) immediately promotes that same accumulator into the stable _minTimeStamp/_maxTimeStamp/_minTimeStepSize/ _maxTimeStepSize fields getMinTimeStamp()/.../getMaxTimeStepSize() expose, so there is no one-phase lag: the very first TimeStep phase must already leave a genuinely positive getMaxTimeStepSize() behind for the generated main loop's tarch::la::greater(maxTimeStepSize,0.0) guard to see, or the simulation could never advance past its first iteration.

Reimplemented from peano4.api.solversteps.ActionSet.ActionSet.

Definition at line 92 of file ComputeGlobalCFLCondition.py.

Referenced by peano4.api.solversteps.ActionSet.ActionSet.get_implementation_of_unprepareTraversal().

Here is the caller graph for this function:

◆ get_guard_of_operation()

exahype2.api.fv.actionsets.ComputeGlobalCFLCondition.ComputeGlobalCFLCondition.get_guard_of_operation	(	self,
		operation_name
	)

Return the guard of an operation.

The guard is a C++ boolean expression. The body of the operation (see get_body_of_operation()) is executed if and only if the guard holds. The guard can use the same arguments as the body, i.e. the marker and the fine and coarse grid data. By default, the guard is true, i.e. the body is always executed.

See class' string constants starting with OPERATION_ for possible values of operation_name.

Reimplemented from peano4.api.solversteps.ActionSet.ActionSet.

Definition at line 217 of file ComputeGlobalCFLCondition.py.