""" An implementation of a general solver base class """
from __future__ import print_function
# System library imports.
import os
import numpy
from compyle.profile import profile, profile_ctx
# PySPH imports
from pysph.base.kernels import CubicSpline
from pysph.sph.acceleration_eval import make_acceleration_evals
from pysph.sph.sph_compiler import SPHCompiler
from pysph.solver.utils import ProgressBar, load, dump
import logging
logger = logging.getLogger(__name__)
EPSILON = numpy.finfo(float).eps*2
[docs]class Solver(object):
"""Base class for all PySPH Solvers
"""
def __init__(self, dim=2, integrator=None, kernel=None,
n_damp=0, tf=1.0, dt=1e-3,
adaptive_timestep=False, cfl=0.3,
output_at_times=(),
fixed_h=False, **kwargs):
"""**Constructor**
Any additional keyword args are used to set the values of any
of the attributes.
Parameters
----------
dim : int
Dimension of the problem
integrator : pysph.sph.integrator.Integrator
Integrator to use
kernel : pysph.base.kernels.Kernel
SPH kernel to use
n_damp : int
Number of timesteps for which the initial damping is required.
This is used to improve stability for problems with strong
discontinuity in initial condition.
Setting it to zero will disable damping of the timesteps.
dt : double
Suggested initial time step for integration
tf : double
Final time for integration
adaptive_timestep : bint
Flag to use adaptive time steps
cfl : double
CFL number for adaptive time stepping
pfreq : int
Output files dumping frequency.
output_at_times : list/array
Optional list of output times to force dump the output file
fixed_h : bint
Flag for constant smoothing lengths `h`
reorder_freq : int
The number of iterations after which particles should
be re-ordered. If zero, do not do this.
Example
-------
>>> integrator = PECIntegrator(fluid=WCSPHStep())
>>> kernel = CubicSpline(dim=2)
>>> solver = Solver(dim=2, integrator=integrator, kernel=kernel,
... n_damp=50, tf=1.0, dt=1e-3, adaptive_timestep=True,
... pfreq=100, cfl=0.5, output_at_times=[1e-1, 1.0])
"""
self.integrator = integrator
self.dim = dim
if kernel is not None:
self.kernel = kernel
else:
self.kernel = CubicSpline(dim)
# set the particles to None
self.particles = None
self.acceleration_evals = None
self.nnps = None
# solver time and iteration count
self.t = 0
self.count = 0
self.execute_commands = None
# list of functions to be called before and after an integration step
self.pre_step_callbacks = []
self.post_step_callbacks = []
# List of functions to be called after each stage of the integrator.
self.post_stage_callbacks = []
# default output printing frequency
self.pfreq = 100
# Compress generated files.
self.compress_output = False
self.disable_output = False
# the process id for parallel runs
self.pid = None
# set the default rank to 0
self.rank = 0
# set the default comm to None.
self.comm = None
# set the default mode to serial
self.in_parallel = False
# arrays to print output
self.arrays_to_print = []
# the default parallel output mode
self.parallel_output_mode = "collected"
# flag to print all arrays
self.detailed_output = False
# flag to save Remote arrays
self.output_only_real = True
# output filename
self.fname = self.__class__.__name__
# output drectory
self.output_directory = self.fname+'_output'
# solution damping to avoid impulsive starts
self.n_damp = n_damp
# Use adaptive time steps and cfl number
self.adaptive_timestep = adaptive_timestep
self.cfl = cfl
# list of output times
self.output_at_times = numpy.asarray(output_at_times)
self.force_output = False
# default time step constants
self.tf = tf
self.dt = dt
self.max_steps = 1 << 31
self._prev_dt = None
self._damping_factor = 1.0
self._epsilon = EPSILON*tf
# flag for constant smoothing lengths
self.fixed_h = fixed_h
self.reorder_freq = 0
# Set all extra keyword arguments
for attr, value in kwargs.items():
if hasattr(self, attr):
setattr(self, attr, value)
else:
msg = 'Unknown keyword arg "%s" passed to constructor' % attr
raise TypeError(msg)
##########################################################################
# Public interface.
##########################################################################
[docs] def setup(self, particles, equations, nnps, kernel=None, fixed_h=False):
""" Setup the solver.
The solver's processor id is set if the in_parallel flag is set
to true.
The order of the integrating calcs is determined by the solver's
order attribute.
This is usually called at the start of a PySPH simulation.
"""
self.particles = particles
if kernel is not None:
self.kernel = kernel
mode = 'mpi' if self.in_parallel else 'serial'
self.acceleration_evals = make_acceleration_evals(
particles, equations, self.kernel, mode
)
sph_compiler = SPHCompiler(
self.acceleration_evals, self.integrator
)
sph_compiler.compile()
# Set the nnps for all concerned objects.
self.nnps = nnps
for ae in self.acceleration_evals:
ae.set_nnps(nnps)
self.integrator.set_nnps(nnps)
# set the parallel manager for the integrator
self.integrator.set_parallel_manager(self.pm)
# Set the post_stage_callback.
self.integrator.set_post_stage_callback(self._post_stage_callback)
# set integrator option for constant smoothing length
self.fixed_h = fixed_h
self.integrator.set_fixed_h(fixed_h)
logger.debug("Solver setup complete.")
[docs] def add_post_stage_callback(self, callback):
"""These callbacks are called *after* each integrator stage.
The callbacks are passed (current_time, dt, stage). See the the
`Integrator.one_timestep` methods for examples of how this is called.
Example
-------
>>> def post_stage_callback_function(t, dt, stage):
>>> # This function is called after every stage of integrator.
>>> print(t, dt, stage)
>>> # Do something
>>> solver.add_post_stage_callback(post_stage_callback_function)
"""
self.post_stage_callbacks.append(callback)
[docs] def add_post_step_callback(self, callback):
"""These callbacks are called *after* each timestep is performed.
The callbacks are passed the solver instance (i.e. self).
Example
-------
>>> def post_step_callback_function(solver):
>>> # This function is called after every time step.
>>> print(solver.t, solver.dt)
>>> # Do something
>>> solver.add_post_step_callback(post_step_callback_function)
"""
self.post_step_callbacks.append(callback)
[docs] def add_pre_step_callback(self, callback):
"""These callbacks are called *before* each timestep is performed.
The callbacks are passed the solver instance (i.e. self).
Example
-------
>>> def pre_step_callback_function(solver):
>>> # This function is called before every time step.
>>> print(solver.t, solver.dt)
>>> # Do something
>>> solver.add_pre_step_callback(pre_step_callback_function)
"""
self.pre_step_callbacks.append(callback)
[docs] def append_particle_arrrays(self, arrays):
""" Append the particle arrays to the existing particle arrays
"""
if not self.particles:
print('Warning! Particles not defined.')
return
for array in self.particles:
array_name = array.name
for arr in arrays:
if array_name == arr.name:
array.append_parray(arr)
self.setup(self.particles)
@profile
def reorder_particles(self):
"""Re-order particles so as to coalesce memory access.
"""
for i in range(len(self.particles)):
self.nnps.spatially_order_particles(i)
# We must update after the reorder.
self.nnps.update()
[docs] def set_adaptive_timestep(self, value):
"""Set it to True to use adaptive timestepping based on
cfl, viscous and force factor.
Look at pysph.sph.integrator.compute_time_step for more details.
"""
self.adaptive_timestep = value
[docs] def set_cfl(self, value):
'Set the CFL number for adaptive time stepping'
self.cfl = value
[docs] def set_final_time(self, tf):
""" Set the final time for the simulation """
self.tf = tf
self._epsilon = EPSILON*tf
[docs] def set_n_damp(self, ndamp):
"""Set the number of timesteps for which the timestep should be
initially damped.
"""
self.n_damp = ndamp
[docs] def set_time_step(self, dt):
""" Set the time step to use """
self.dt = dt
[docs] def set_print_freq(self, n):
""" Set the output print frequency """
self.pfreq = n
[docs] def set_disable_output(self, value):
"""Disable file output.
"""
self.disable_output = value
[docs] def set_arrays_to_print(self, array_names=None):
"""Only print the arrays with the given names.
"""
available_arrays = [array.name for array in self.particles]
if array_names:
for name in array_names:
if name not in available_arrays:
raise RuntimeError("Array %s not availabe" % (name))
for arr in self.particles:
if arr.name == name:
array = arr
break
self.arrays_to_print.append(array)
else:
self.arrays_to_print = self.particles
[docs] def set_output_fname(self, fname):
""" Set the output file name """
self.fname = fname
[docs] def set_output_printing_level(self, detailed_output):
""" Set the output printing level """
self.detailed_output = detailed_output
[docs] def set_output_only_real(self, output_only_real):
""" Set the flag to save out only real particles """
self.output_only_real = output_only_real
[docs] def set_output_directory(self, path):
""" Set the output directory """
self.output_directory = path
[docs] def set_output_at_times(self, output_at_times):
""" Set a list of output times """
self.output_at_times = numpy.asarray(output_at_times)
[docs] def set_max_steps(self, max_steps):
"""Set the maximum number of iterations to perform.
"""
self.max_steps = max_steps
[docs] def set_compress_output(self, compress):
"""Compress the dumped output files.
"""
self.compress_output = compress
[docs] def set_parallel_output_mode(self, mode="collected"):
"""Set the default solver dump mode in parallel.
The available modes are:
collected : Collect array data from all processors on root and
dump a single file.
distributed : Each processor dumps a file locally.
"""
assert mode in ("collected", "distributed")
self.parallel_output_mode = mode
[docs] def set_command_handler(self, callable, command_interval=1):
""" set the `callable` to be called at every `command_interval`
iteration
the `callable` is called with the solver instance as an argument
"""
self.execute_commands = callable
self.command_interval = command_interval
def set_parallel_manager(self, pm):
self.pm = pm
[docs] def set_reorder_freq(self, freq):
"""Set the reorder frequency in number of iterations.
"""
self.reorder_freq = freq
def barrier(self):
if self.comm:
self.comm.barrier()
[docs] def solve(self, show_progress=True):
""" Solve the system
Notes
-----
Pre-stepping functions are those that need to be called before
the integrator is called.
Similarly, post step functions are those that are called after
the stepping within the integrator.
"""
if self.in_parallel:
show = False
else:
show = show_progress
bar = ProgressBar(self.t, self.tf, show=show)
self._epsilon = EPSILON*self.tf
# Initial solution
self.dump_output()
self.barrier() # everybody waits for this to complete
reorder_freq = self.reorder_freq
if reorder_freq > 0:
self.reorder_particles()
# Compute the accelerations once for the predictor corrector
# integrator to work correctly at the first time step.
self.integrator.initial_acceleration(self.t, self.dt)
# Now get a suitable adaptive (if requested) and damped timestep to
# integrate with.
self.dt = self._get_timestep()
while (self.tf - self.t) > self._epsilon and \
(self.count < self.max_steps):
# perform any pre step functions
if self.pre_step_callbacks:
with profile_ctx('Solver.pre_step_callback'):
for callback in self.pre_step_callbacks:
callback(self)
if self.rank == 0:
logger.debug(
"Iteration=%d, time=%f, timestep=%f" %
(self.count, self.t, self.dt)
)
# perform the integration and update the time.
# print('Solver Iteration', self.count, self.dt, self.t)
self.integrator.step(self.t, self.dt)
# perform any post step functions
if self.post_step_callbacks:
with profile_ctx('Solver.post_step_callback'):
for callback in self.post_step_callbacks:
callback(self)
# update time and iteration counters if successfully
# integrated
self.t += self.dt
self.count += 1
self._epsilon = EPSILON*self.tf*self.count
# Compute the next timestep.
self.dt = self._get_timestep()
# Note: this may adjust dt to land at a desired time.
self._dump_output_if_needed()
# update progress bar
bar.update(self.t)
# update the time for all arrays
self.update_particle_time()
if reorder_freq > 0 and (self.count % reorder_freq == 0):
self.reorder_particles()
if self.execute_commands is not None:
if self.count % self.command_interval == 0:
self.execute_commands(self)
# close the progress bar
bar.finish()
# final output save
self.dump_output()
def update_particle_time(self):
for array in self.particles:
array.set_time(self.t)
@profile
def dump_output(self):
"""Dump the simulation results to file
The arrays used for printing are determined by the particle
array's `output_property_arrays` data attribute. For debugging
it is sometimes nice to have all the arrays (including
accelerations) saved. This can be chosen from using the
command line option `--detailed-output`
Output data Format:
A single file named as: <fname>_<rank>_<iteration_count>.npz
The data is saved as a Python dictionary with two keys:
`solver_data` : Solver meta data like time, dt and iteration number
`arrays` : A dictionary keyed on particle array names and with
particle properties as value.
Example:
You can load the data output by PySPH like so:
>>> from pysph.solver.utils import load
>>> data = load('output_directory/filename_x_xxx.npz')
>>> solver_data = data['solver_data']
>>> arrays = data['arrays']
>>> fluid = arrays['fluid']
>>> ...
In the above example, it is assumed that the output file
contained an array named fluid.
"""
if self.disable_output:
return
if self.rank == 0:
msg = 'Writing output at time %g, iteration %d, dt = %g' % (
self.t, self.count, self.dt)
logger.info(msg)
fname = os.path.join(self.output_directory,
'%s_%05d' % (self.fname, self.count))
comm = None
if self.parallel_output_mode == "collected" and self.in_parallel:
comm = self.comm
dump(fname, self.particles, self._get_solver_data(),
detailed_output=self.detailed_output,
only_real=self.output_only_real, mpi_comm=comm,
compress=self.compress_output)
[docs] def load_output(self, count):
"""Load particle data from dumped output file.
Parameters
----------
count : str
The iteration time from which to load the data. If time is '?' then
list of available data files is returned else the latest available
data file is used
Notes
-----
Data is loaded from the :py:attr:`output_directory` using the same
format as stored by the :py:meth:`dump_output` method. Proper
functioning required that all the relevant properties of arrays be
dumped.
"""
# get the list of available files
available_files = [i.rsplit('_', 1)[1][:-4]
for i in os.listdir(self.output_directory)
if i.startswith(self.fname) and i.endswith('.npz')]
if count == '?':
return sorted(set(available_files), key=int)
else:
if count not in available_files:
msg = "File with iteration count `%s` does not exist" % (count)
msg += "\nValid iteration counts are %s" % (
sorted(set(available_files), key=int))
raise IOError(msg)
array_names = [pa.name for pa in self.particles]
# load the output file
data = load(os.path.join(self.output_directory,
self.fname+'_'+str(count)+'.npz'))
arrays = [data["arrays"][i] for i in array_names]
# set the Particle's arrays
self.particles = arrays
solver_data = data['solver_data']
self.t = float(solver_data['t'])
self.dt = float(solver_data['dt'])
self.count = int(solver_data['count'])
[docs] def get_options(self, arg_parser):
""" Implement this to add additional options for the application """
pass
[docs] def setup_solver(self, options=None):
""" Implement the basic solvers here
All subclasses of Solver may implement this function to add the
necessary operations for the problem at hand.
Parameters
----------
options : dict
options set by the user using commandline (there is no guarantee
of existence of any key)
"""
pass
##########################################################################
# Non-public interface.
##########################################################################
def _compute_timestep(self):
undamped_dt = self._get_undamped_timestep()
if self.adaptive_timestep:
# locally stable time step
dt = self.integrator.compute_time_step(undamped_dt, self.cfl)
# set the globally stable time step across all processors
if self.in_parallel:
if dt is None:
# For some reason this processor does not have an adaptive
# timestep constraint so we set it to a large number so the
# timestep is determined by the other processors.
dt = 1e20
dt = self.pm.update_time_steps(dt)
else:
if dt is None:
dt = undamped_dt
else:
dt = undamped_dt
return dt
def _damp_timestep(self, dt):
"""Damp the timestep initially to prevent transient errors at startup.
This basically damps the initial timesteps by the factor
0.5 (sin(pi*(-0.5 + count/n_damp)) + 1)
Where n_damp is the number of iterations to damp the timestep for and
count is the number of iterations.
"""
n_damp = self.n_damp
if self.count < n_damp and n_damp > 0:
iter_fraction = (self.count+1)/float(n_damp)
fac = 0.5*(numpy.sin(numpy.pi*(-0.5 + iter_fraction)) + 1.0)
self._damping_factor = fac
else:
self._damping_factor = 1.0
return dt*self._damping_factor
def _dump_output_if_needed(self):
"""Dump output if needed while solve is running.
This is called by `solve`.
Warning
-------
This will adjust `dt` if the user has asked for output at a
non-integral multiple of dt.
"""
if abs(self.t - self.tf) < self._epsilon:
return
# dump output if the iteration number is a multiple of the printing
# frequency.
dump = self.count % self.pfreq == 0
# Consider the other cases if user has requested output at a specified
# time.
output_at_times = self.output_at_times
dt = self.dt
# adjust dt to land on specific output times or dump output if we have
# reached a desired time.
if len(output_at_times) > 0:
tdiff = output_at_times - self.t
if numpy.any(numpy.abs(tdiff) < self._epsilon):
dump = True
# Our next step may exceed a required timestep so we adjust the
# timestep.
timestep_too_big = (tdiff > 0.0) & (tdiff < dt)
if numpy.any(timestep_too_big):
indices = numpy.where(timestep_too_big)[0]
index = indices[0]
output_time = output_at_times[index]
if ((abs(output_time - self.t) < self._epsilon) and
(len(indices) > 1)):
index = indices[1]
output_time = output_at_times[index]
if abs(output_time - self.t) > self._epsilon:
# It sometimes happens that the current time is just
# shy of the requested output time which results in a
# ridiculously small dt so we skip that case.
# Compute the new time-step to fall on the specified output
# time instant and save the previous dt value.
self._prev_dt = dt
self.dt = float(output_time - self.t)
if dump:
self.dump_output()
self.barrier()
def _get_solver_data(self):
if self._prev_dt is not None:
dt = self._prev_dt/self._damping_factor
else:
dt = self._get_undamped_timestep()
return {'dt': dt, 't': self.t, 'count': self.count}
@profile
def _get_timestep(self):
if abs(self.tf - self.t) < self._epsilon:
# We have reached the end, so no need to adjust the timestep
# anymore.
return self.dt
if self._prev_dt is not None and \
abs(self._prev_dt - self.dt) > self._epsilon:
# if the _prev_dt was set then we need to use it as the current dt
# was set to print at an intermediate time.
self.dt = self._prev_dt
self._prev_dt = None
dt = self._compute_timestep()
dt = self._damp_timestep(dt)
# adjust dt to land exactly on final time
if (self.t + dt) > (self.tf - self._epsilon):
dt = self.tf - self.t
return dt
def _get_undamped_timestep(self):
return self.dt/self._damping_factor
def _post_stage_callback(self, time, dt, stage):
if self.post_stage_callbacks:
with profile_ctx('Solver.post_stage_callback'):
for callback in self.post_stage_callbacks:
callback(time, dt, stage)
############################################################################