def __halo_end_exchange(self, dim):
"""End a halo exchange along a given Dimension."""
for d, i, payload, req in list(self._in_flight):
if d == dim:
status = MPI.Status()
req.Wait(status=status)
if payload is not None and status.source != MPI.PROC_NULL:
# The MPI.Request `req` originated from a `comm.Irecv`
# Now need to scatter the data to the right place
self._data_in_region(HALO, d, i)[:] = payload
self._in_flight.remove((d, i, payload, req))
def timer_on(self, name, comm=None):
"""
Measure the execution time of a Python-level code region.
Parameters
----------
name : str
A representative string for the timed region.
comm : MPI communicator, optional
If provided, the global execution time is derived by a single MPI
rank, with timers started and stopped right after an MPI barrier.
"""
if comm and comm is not MPI.COMM_NULL:
comm.Barrier()
tic = MPI.Wtime()
yield
comm.Barrier()
toc = MPI.Wtime()
else:
tic = seq_time()
yield
toc = seq_time()
self.py_timers[name] = toc - tic
def _halo_exchange(self):
"""Perform the halo exchange with the neighboring processes."""
if not MPI.Is_initialized() or MPI.COMM_WORLD.size == 1:
# Nothing to do
return
if MPI.COMM_WORLD.size > 1 and self._distributor is None:
raise RuntimeError("`%s` cannot perform a halo exchange as it has "
"no Grid attached" % self.name)
if self._in_flight:
raise RuntimeError("`%s` cannot initiate a halo exchange as previous "
"exchanges are still in flight" % self.name)
for i in self.space_dimensions:
self.__halo_begin_exchange(i)
self.__halo_end_exchange(i)
self._is_halo_dirty = False
assert not self._in_flight
class MPIMsg(CompositeObject):
_C_field_bufs = 'bufs'
_C_field_bufg = 'bufg'
_C_field_sizes = 'sizes'
_C_field_rrecv = 'rrecv'
_C_field_rsend = 'rsend'
if MPI._sizeof(MPI.Request) == sizeof(c_int):
c_mpirequest_p = type('MPI_Request', (c_int, ), {})
else:
c_mpirequest_p = type('MPI_Request', (c_void_p, ), {})
def __init__(self, name, function, halos, fields=None):
self._function = function
self._halos = halos
fields = (fields or []) + [
(MPIMsg._C_field_bufs, c_void_p),
(MPIMsg._C_field_bufg, c_void_p),
(MPIMsg._C_field_sizes, POINTER(c_int)),
(MPIMsg._C_field_rrecv, MPIMsg.c_mpirequest_p),
(MPIMsg._C_field_rsend, MPIMsg.c_mpirequest_p),
]
super(MPIMsg, self).__init__(name, 'msg', fields)
# Required for buffer allocation/deallocation before/after jumping/returning
# to/from C-land
self._allocator = default_allocator()
self._memfree_args = []
def __del__(self):
self._C_memfree()
def _C_memfree(self):
# Deallocate the MPI buffers
for i in self._memfree_args:
self._allocator.free(*i)
self._memfree_args[:] = []
def __value_setup__(self, dtype, value):
# We eventually produce an array of `struct msg` that is as big as
# the number of peers we have to communicate with
return (dtype._type_ * self.npeers)()
@property
def function(self):
return self._function
@property
def halos(self):
return self._halos
@property
def npeers(self):
return len(self._halos)
def _arg_defaults(self, alias=None):
function = alias or self.function
for i, halo in enumerate(self.halos):
entry = self.value[i]
# Buffer size for this peer
shape = []
for dim, side in zip(*halo):
try:
shape.append(getattr(function._size_owned[dim], side.name))
except AttributeError:
assert side is CENTER
shape.append(function._size_domain[dim])
entry.sizes = (c_int * len(shape))(*shape)
# Allocate the send/recv buffers
size = reduce(mul, shape)
ctype = dtype_to_ctype(function.dtype)
entry.bufg, bufg_memfree_args = self._allocator._alloc_C_libcall(
size, ctype)
entry.bufs, bufs_memfree_args = self._allocator._alloc_C_libcall(
size, ctype)
# The `memfree_args` will be used to deallocate the buffer upon returning
# from C-land
self._memfree_args.extend([bufg_memfree_args, bufs_memfree_args])
return {self.name: self.value}
def _arg_values(self, args=None, **kwargs):
return self._arg_defaults(
alias=kwargs.get(self.function.name, self.function))
def _arg_apply(self, *args, **kwargs):
self._C_memfree()
# Pickling support
_pickle_args = ['name', 'function', 'halos']
clear_cache()
gflopss, oi, timings, _ = self.func(*args, **kwargs)
for key in timings.keys():
self.register(gflopss[key], measure="gflopss", event=key.name)
self.register(oi[key], measure="oi", event=key.name)
self.register(timings[key], measure="timings", event=key.name)
return DevitoExecutor(func)
if __name__ == "__main__":
# If running with MPI, we emit logging messages from rank0 only
try:
MPI.Init() # Devito starts off with MPI disabled!
set_log_level('DEBUG', comm=MPI.COMM_WORLD)
if MPI.COMM_WORLD.size > 1 and not configuration['mpi']:
warning(
"It seems that you're running over MPI with %d processes, but "
"DEVITO_MPI is unset. Setting `DEVITO_MPI=basic`..." %
MPI.COMM_WORLD.size)
configuration['mpi'] = 'basic'
except TypeError:
# MPI not available
pass
# Profiling at max level
configuration['profiling'] = 'advanced'
last_res = None
for params in sweep(kwargs, keys=sweep_options):
kwargs.update(params)
_, _, _, res = run(**kwargs)
if last_res is None:
last_res = res
else:
for i in range(len(res)):
assert np.isclose(res[i], last_res[i])
if __name__ == "__main__":
# If running with MPI, we emit logging messages from rank0 only
try:
MPI.Init() # Devito starts off with MPI disabled!
set_log_level('DEBUG', comm=MPI.COMM_WORLD)
if MPI.COMM_WORLD.size > 1 and not configuration['mpi']:
warning(
"It seems that you're running over MPI with %d processes, but "
"DEVITO_MPI is unset. Setting `DEVITO_MPI=basic`..." %
MPI.COMM_WORLD.size)
configuration['mpi'] = 'basic'
except TypeError:
# MPI not available
pass
# Benchmarking cannot be done at basic level
if configuration['profiling'] == 'basic':
configuration['profiling'] = 'advanced'