def set_field_states(self, h_field_states):
"""Copy field states from host to device memory."""
if not self.is_ready():
raise Exception('CUDA has not been initialised')
else:
float_t = get_float_type(self.params.code)
if str(h_field_states.dtype) != float_t:
# TODO: ERROR!!
pass
set_field_states_fn = self._mod.get_function('set_field_states')
timer = Timer("set_fs_cpy")
cuda.memcpy_htod(self._d_field_states, h_field_states)
timer = Timer("set_fs_fn")
set_field_states_fn(self._d_field_states,
self._d_states,
block=self._get_block(),
grid=self._get_grid())
def set_field_states(self, h_field_states):
if not self.is_ready():
raise Exception("CUDA has not been initialised")
else:
float_t = get_float_type(self.params.code)
if str(h_field_states.dtype) != float_t:
# TODO: ERROR!!
pass
set_field_states_fn = self._mod.get_function("set_field_states")
timer = Timer("set_fs_cpy") # noqa: F841
cuda.memcpy_htod(self._d_field_states, h_field_states)
timer = Timer("set_fs_fn") # noqa: F841
set_field_states_fn(
self._d_field_states,
self._d_states,
block=self._get_block(),
grid=self._get_grid(),
)
def forward(self, t, dt, update_host_states=False, synchronize=True):
if not self.is_ready():
raise Exception("CUDA has not been initialised")
else:
timer = Timer("calculate CUDA forward")
args = [self._d_states, t, dt, self._d_parameters]
field_parameters = self.params.code.parameters.field_parameters
# FIXME: modelparameters needs a ListParam
if len(field_parameters) != 1 or field_parameters[0] != "":
args.append(self._d_field_parameters)
args.append(np.uint32(self.num_nodes))
self._forward_fn(*args,
block=self._get_block(),
grid=self._get_grid())
if synchronize:
self.ctx.synchronize()
if update_host_states:
timer = Timer("update host states") # noqa: F841
cuda.memcpy_dtoh(self._h_states, self._d_states)
def forward(self, t, dt, update_host_states=False, synchronize=True):
"""Solve one time step of the ODE system on GPU"""
if not self.is_ready():
raise Exception('CUDA has not been initialised')
else:
timer = Timer("calculate CUDA forward")
args = [self._d_states, t, dt, self._d_parameters]
field_parameters = self.params.code.parameters.field_parameters
# FIXME: modelparameters needs a ListParam
if not (len(field_parameters)==0 or (len(field_parameters) == 1 \
and field_parameters[0] == "")):
args.append(self._d_field_parameters)
args.append(np.uint32(self.num_nodes))
self._forward_fn(*args,
block=self._get_block(),
grid=self._get_grid())
if synchronize:
self.ctx.synchronize()
if update_host_states:
timer = Timer("update host states")
cuda.memcpy_dtoh(self._h_states, self._d_states)
def set_field_states(self, field_states=None):
timer = Timer("set field states") # noqa: F841
field_states = field_states if field_states is not None else self.field_states
if field_states is not None:
self._cudahandler.set_field_states(field_states)
def set_field_states(self, field_states=None):
"""Copy current host field states onto the device."""
timer = Timer("set field states")
field_states = field_states if field_states is not None else self.field_states
if field_states is not None:
self._cudahandler.set_field_states(field_states)