def test_retrieve_synaptic_block(self):
default_config_paths = os.path.join(
os.path.dirname(abstract_spinnaker_common.__file__),
AbstractSpiNNakerCommon.CONFIG_FILE_NAME)
config = conf_loader.load_config(
AbstractSpiNNakerCommon.CONFIG_FILE_NAME, default_config_paths)
key = 0
synaptic_manager = SynapticManager(
synapse_type=None, ring_buffer_sigma=5.0, spikes_per_second=100.0,
config=config,
population_table_type=MockMasterPopulationTable(
{key: [(1, 0, False)]}),
synapse_io=MockSynapseIO())
transceiver = MockTransceiverRawData(bytearray(16))
placement = Placement(None, 0, 0, 1)
first_block, row_len_1 = synaptic_manager._retrieve_synaptic_block(
transceiver=transceiver, placement=placement,
master_pop_table_address=0, indirect_synapses_address=0,
direct_synapses_address=0, key=key, n_rows=1, index=0,
using_extra_monitor_cores=False)
same_block, row_len_1_2 = synaptic_manager._retrieve_synaptic_block(
transceiver=transceiver, placement=placement,
master_pop_table_address=0, indirect_synapses_address=0,
direct_synapses_address=0, key=key, n_rows=1, index=0,
using_extra_monitor_cores=False)
synaptic_manager.clear_connection_cache()
different_block, row_len_2 = synaptic_manager._retrieve_synaptic_block(
transceiver=transceiver, placement=placement,
master_pop_table_address=0, indirect_synapses_address=0,
direct_synapses_address=0, key=key, n_rows=1, index=0,
using_extra_monitor_cores=False)
# Check that the row lengths are all the same
assert row_len_1 == row_len_1_2
assert row_len_1 == row_len_2
# Check that the block retrieved twice without reset is cached
assert id(first_block) == id(same_block)
# Check that the block after reset is not a copy
assert id(first_block) != id(different_block)
class AbstractPopulationVertex(
ApplicationVertex, AbstractGeneratesDataSpecification,
AbstractHasAssociatedBinary, AbstractContainsUnits,
AbstractSpikeRecordable, AbstractNeuronRecordable,
AbstractProvidesOutgoingPartitionConstraints,
AbstractProvidesIncomingPartitionConstraints,
AbstractPopulationInitializable, AbstractPopulationSettable,
AbstractChangableAfterRun, AbstractHasGlobalMaxAtoms,
AbstractRewritesDataSpecification, AbstractReadParametersBeforeSet,
AbstractAcceptsIncomingSynapses, ProvidesKeyToAtomMappingImpl):
""" Underlying vertex model for Neural Populations.
"""
__slots__ = [
"_additional_input", "_binary", "_buffer_size_before_receive",
"_change_requires_mapping",
"_change_requires_neuron_parameters_reload",
"_incoming_spike_buffer_size", "_input_type",
"_maximum_sdram_for_buffering", "_minimum_buffer_sdram", "_model_name",
"_n_atoms", "_n_profile_samples", "_neuron_model", "_neuron_recorder",
"_receive_buffer_host", "_receive_buffer_port", "_spike_recorder",
"_synapse_manager", "_threshold_type", "_time_between_requests",
"_units", "_using_auto_pause_and_resume"
]
BASIC_MALLOC_USAGE = 2
# recording region ids
SPIKE_RECORDING_REGION = 0
V_RECORDING_REGION = 1
GSYN_EXCITATORY_RECORDING_REGION = 2
GSYN_INHIBITORY_RECORDING_REGION = 3
RECORDING_REGION = {"spikes": 0, "v": 1, "gsyn_exc": 2, "gsyn_inh": 3}
VARIABLE_LONG = {
"spikes": "spikes",
"v": "membrane voltage",
"gsyn_exc": "gsyn_excitatory",
"gsyn_inh": "gsyn_inhibitory"
}
N_RECORDING_REGIONS = 4
# the size of the runtime SDP port data region
RUNTIME_SDP_PORT_SIZE = 4
# 6 elements before the start of global parameters
BYTES_TILL_START_OF_GLOBAL_PARAMETERS = 24
_n_vertices = 0
non_pynn_default_parameters = {
'spikes_per_second': None,
'ring_buffer_sigma': None,
'incoming_spike_buffer_size': None,
'constraints': None,
'label': None
}
def __init__(self,
n_neurons,
binary,
label,
max_atoms_per_core,
spikes_per_second,
ring_buffer_sigma,
incoming_spike_buffer_size,
model_name,
neuron_model,
input_type,
synapse_type,
threshold_type,
additional_input=None,
constraints=None):
# pylint: disable=too-many-arguments, too-many-locals
super(AbstractPopulationVertex, self).__init__(label, constraints,
max_atoms_per_core)
self._units = {
'spikes': 'spikes',
'v': 'mV',
'gsyn_exc': "uS",
'gsyn_inh': "uS"
}
self._binary = binary
self._n_atoms = n_neurons
# buffer data
self._incoming_spike_buffer_size = incoming_spike_buffer_size
# get config from simulator
config = globals_variables.get_simulator().config
if incoming_spike_buffer_size is None:
self._incoming_spike_buffer_size = config.getint(
"Simulation", "incoming_spike_buffer_size")
self._model_name = model_name
self._neuron_model = neuron_model
self._input_type = input_type
self._threshold_type = threshold_type
self._additional_input = additional_input
# Set up for recording
self._neuron_recorder = NeuronRecorder(
["spikes", "v", "gsyn_exc", "gsyn_inh"], n_neurons)
self._time_between_requests = config.getint("Buffers",
"time_between_requests")
self._minimum_buffer_sdram = config.getint("Buffers",
"minimum_buffer_sdram")
self._using_auto_pause_and_resume = config.getboolean(
"Buffers", "use_auto_pause_and_resume")
self._receive_buffer_host = config.get("Buffers",
"receive_buffer_host")
self._receive_buffer_port = helpful_functions.read_config_int(
config, "Buffers", "receive_buffer_port")
# If live buffering is enabled, set a maximum on the buffer sizes
spike_buffer_max_size = 0
v_buffer_max_size = 0
gsyn_buffer_max_size = 0
self._buffer_size_before_receive = None
if config.getboolean("Buffers", "enable_buffered_recording"):
spike_buffer_max_size = config.getint("Buffers",
"spike_buffer_size")
v_buffer_max_size = config.getint("Buffers", "v_buffer_size")
gsyn_buffer_max_size = config.getint("Buffers", "gsyn_buffer_size")
self._buffer_size_before_receive = config.getint(
"Buffers", "buffer_size_before_receive")
self._maximum_sdram_for_buffering = [
spike_buffer_max_size, v_buffer_max_size, gsyn_buffer_max_size,
gsyn_buffer_max_size
]
# Set up synapse handling
self._synapse_manager = SynapticManager(synapse_type,
ring_buffer_sigma,
spikes_per_second, config)
# bool for if state has changed.
self._change_requires_mapping = True
self._change_requires_neuron_parameters_reload = False
# Set up for profiling
self._n_profile_samples = helpful_functions.read_config_int(
config, "Reports", "n_profile_samples")
@property
@overrides(ApplicationVertex.n_atoms)
def n_atoms(self):
return self._n_atoms
@inject_items({
"graph": "MemoryApplicationGraph",
"n_machine_time_steps": "TotalMachineTimeSteps",
"machine_time_step": "MachineTimeStep"
})
@overrides(ApplicationVertex.get_resources_used_by_atoms,
additional_arguments={
"graph", "n_machine_time_steps", "machine_time_step"
})
def get_resources_used_by_atoms(self, vertex_slice, graph,
n_machine_time_steps, machine_time_step):
# pylint: disable=arguments-differ
# set resources required from this object
container = ResourceContainer(
sdram=SDRAMResource(
self.get_sdram_usage_for_atoms(vertex_slice, graph,
machine_time_step)),
dtcm=DTCMResource(self.get_dtcm_usage_for_atoms(vertex_slice)),
cpu_cycles=CPUCyclesPerTickResource(
self.get_cpu_usage_for_atoms(vertex_slice)))
recording_sizes = recording_utilities.get_recording_region_sizes(
self._get_buffered_sdram(vertex_slice, n_machine_time_steps),
self._minimum_buffer_sdram, self._maximum_sdram_for_buffering,
self._using_auto_pause_and_resume)
container.extend(
recording_utilities.get_recording_resources(
recording_sizes, self._receive_buffer_host,
self._receive_buffer_port))
# return the total resources.
return container
@property
@overrides(AbstractChangableAfterRun.requires_mapping)
def requires_mapping(self):
return self._change_requires_mapping
@overrides(AbstractChangableAfterRun.mark_no_changes)
def mark_no_changes(self):
self._change_requires_mapping = False
def _get_buffered_sdram_per_timestep(self, vertex_slice):
return [
self._neuron_recorder.get_buffered_sdram_per_timestep(
"spikes", vertex_slice),
self._neuron_recorder.get_buffered_sdram_per_timestep(
"v", vertex_slice),
self._neuron_recorder.get_buffered_sdram_per_timestep(
"gsyn_exc", vertex_slice),
self._neuron_recorder.get_buffered_sdram_per_timestep(
"gsyn_inh", vertex_slice)
]
def _get_buffered_sdram(self, vertex_slice, n_machine_time_steps):
return [
self._neuron_recorder.get_buffered_sdram("spikes", vertex_slice,
n_machine_time_steps),
self._neuron_recorder.get_buffered_sdram("v", vertex_slice,
n_machine_time_steps),
self._neuron_recorder.get_buffered_sdram("gsyn_exc", vertex_slice,
n_machine_time_steps),
self._neuron_recorder.get_buffered_sdram("gsyn_inh", vertex_slice,
n_machine_time_steps)
]
@inject_items({"n_machine_time_steps": "TotalMachineTimeSteps"})
@overrides(ApplicationVertex.create_machine_vertex,
additional_arguments={"n_machine_time_steps"})
def create_machine_vertex(self,
vertex_slice,
resources_required,
n_machine_time_steps,
label=None,
constraints=None):
# pylint: disable=too-many-arguments, arguments-differ
is_recording = len(self._neuron_recorder.recording_variables) > 0
buffered_sdram_per_timestep = self._get_buffered_sdram_per_timestep(
vertex_slice)
buffered_sdram = self._get_buffered_sdram(vertex_slice,
n_machine_time_steps)
minimum_buffer_sdram = recording_utilities.get_minimum_buffer_sdram(
buffered_sdram, self._minimum_buffer_sdram)
overflow_sdram = self._neuron_recorder.get_sampling_overflow_sdram(
vertex_slice)
vertex = PopulationMachineVertex(resources_required, is_recording,
minimum_buffer_sdram,
buffered_sdram_per_timestep, label,
constraints, overflow_sdram)
AbstractPopulationVertex._n_vertices += 1
# return machine vertex
return vertex
def get_cpu_usage_for_atoms(self, vertex_slice):
per_neuron_cycles = (
_NEURON_BASE_N_CPU_CYCLES_PER_NEURON +
self._neuron_model.get_n_cpu_cycles_per_neuron() +
self._input_type.get_n_cpu_cycles_per_neuron(
self._synapse_manager.synapse_type.get_n_synapse_types()) +
self._threshold_type.get_n_cpu_cycles_per_neuron())
if self._additional_input is not None:
per_neuron_cycles += \
self._additional_input.get_n_cpu_cycles_per_neuron()
return (_NEURON_BASE_N_CPU_CYCLES + _C_MAIN_BASE_N_CPU_CYCLES +
(per_neuron_cycles * vertex_slice.n_atoms) +
self._neuron_recorder.get_n_cpu_cycles(vertex_slice.n_atoms) +
self._synapse_manager.get_n_cpu_cycles())
def get_dtcm_usage_for_atoms(self, vertex_slice):
per_neuron_usage = (
self._neuron_model.get_dtcm_usage_per_neuron_in_bytes() +
self._input_type.get_dtcm_usage_per_neuron_in_bytes() +
self._threshold_type.get_dtcm_usage_per_neuron_in_bytes())
if self._additional_input is not None:
per_neuron_usage += \
self._additional_input.get_dtcm_usage_per_neuron_in_bytes()
return (_NEURON_BASE_DTCM_USAGE_IN_BYTES +
(per_neuron_usage * vertex_slice.n_atoms) +
self._neuron_recorder.get_dtcm_usage_in_bytes(vertex_slice) +
self._synapse_manager.get_dtcm_usage_in_bytes())
def _get_sdram_usage_for_neuron_params_per_neuron(self):
per_neuron_usage = (
self._input_type.get_sdram_usage_per_neuron_in_bytes() +
self._threshold_type.get_sdram_usage_per_neuron_in_bytes() +
self._neuron_recorder.get_sdram_usage_per_neuron_in_bytes() +
self._neuron_model.get_sdram_usage_per_neuron_in_bytes())
if self._additional_input is not None:
per_neuron_usage += \
self._additional_input.get_sdram_usage_per_neuron_in_bytes()
return per_neuron_usage
def _get_sdram_usage_for_neuron_params(self, vertex_slice):
""" calculates the sdram usage for just the neuron parameters region
:param vertex_slice: the slice of atoms.
:return: The sdram required for the neuron region
"""
per_neuron_usage = \
self._get_sdram_usage_for_neuron_params_per_neuron()
return (self.BYTES_TILL_START_OF_GLOBAL_PARAMETERS +
self._neuron_model.
get_sdram_usage_for_global_parameters_in_bytes() +
self._neuron_recorder.
get_sdram_usage_for_global_parameters_in_bytes() +
(per_neuron_usage * vertex_slice.n_atoms))
def get_sdram_usage_for_atoms(self, vertex_slice, graph,
machine_time_step):
sdram_requirement = (
common_constants.SYSTEM_BYTES_REQUIREMENT +
self._get_sdram_usage_for_neuron_params(vertex_slice) +
recording_utilities.get_recording_header_size(
self.N_RECORDING_REGIONS) +
PopulationMachineVertex.get_provenance_data_size(
PopulationMachineVertex.N_ADDITIONAL_PROVENANCE_DATA_ITEMS) +
self._synapse_manager.get_sdram_usage_in_bytes(
vertex_slice, graph.get_edges_ending_at_vertex(self),
machine_time_step) +
(self._get_number_of_mallocs_used_by_dsg() *
common_constants.SARK_PER_MALLOC_SDRAM_USAGE) +
profile_utils.get_profile_region_size(self._n_profile_samples))
return sdram_requirement
def _get_number_of_mallocs_used_by_dsg(self):
extra_mallocs = len(self._neuron_recorder.recording_variables)
return (self.BASIC_MALLOC_USAGE +
self._synapse_manager.get_number_of_mallocs_used_by_dsg() +
extra_mallocs)
def _reserve_memory_regions(self, spec, vertex_slice, vertex):
spec.comment("\nReserving memory space for data regions:\n\n")
# Reserve memory:
spec.reserve_memory_region(
region=constants.POPULATION_BASED_REGIONS.SYSTEM.value,
size=common_constants.SYSTEM_BYTES_REQUIREMENT,
label='System')
self._reserve_neuron_params_data_region(spec, vertex_slice)
spec.reserve_memory_region(
region=constants.POPULATION_BASED_REGIONS.RECORDING.value,
size=recording_utilities.get_recording_header_size(
self.N_RECORDING_REGIONS))
profile_utils.reserve_profile_region(
spec, constants.POPULATION_BASED_REGIONS.PROFILING.value,
self._n_profile_samples)
vertex.reserve_provenance_data_region(spec)
def _reserve_neuron_params_data_region(self, spec, vertex_slice):
""" reserve the neuron parameter data region
:param spec: the spec to write the dsg region to
:param vertex_slice: the slice of atoms from the application vertex
:return: None
"""
params_size = self._get_sdram_usage_for_neuron_params(vertex_slice)
spec.reserve_memory_region(
region=constants.POPULATION_BASED_REGIONS.NEURON_PARAMS.value,
size=params_size,
label='NeuronParams')
def _write_neuron_parameters(self, spec, key, vertex_slice,
machine_time_step, time_scale_factor):
# pylint: disable=too-many-arguments
n_atoms = (vertex_slice.hi_atom - vertex_slice.lo_atom) + 1
spec.comment(
"\nWriting Neuron Parameters for {} Neurons:\n".format(n_atoms))
# Set the focus to the memory region 2 (neuron parameters):
spec.switch_write_focus(
region=constants.POPULATION_BASED_REGIONS.NEURON_PARAMS.value)
# Write the random back off value
spec.write_value(
random.randint(0, AbstractPopulationVertex._n_vertices))
# Write the number of microseconds between sending spikes
time_between_spikes = ((machine_time_step * time_scale_factor) /
(n_atoms * 2.0))
spec.write_value(data=int(time_between_spikes))
# Write whether the key is to be used, and then the key, or 0 if it
# isn't to be used
if key is None:
spec.write_value(data=0)
spec.write_value(data=0)
else:
spec.write_value(data=1)
spec.write_value(data=key)
# Write the number of neurons in the block:
spec.write_value(data=n_atoms)
# Write the size of the incoming spike buffer
spec.write_value(data=self._incoming_spike_buffer_size)
# Write the recording rates and sizes
record_globals = self._neuron_recorder.get_global_parameters(
vertex_slice)
for param in record_globals:
spec.write_value(data=param.get_value(),
data_type=param.get_dataspec_datatype())
# Write the index parameters
indexes = self._neuron_recorder.get_index_parameters(vertex_slice)
utility_calls.write_parameters_per_neuron(spec,
vertex_slice,
indexes,
slice_paramaters=True)
# Write the global parameters
global_params = self._neuron_model.get_global_parameters()
for param in global_params:
spec.write_value(data=param.get_value(),
data_type=param.get_dataspec_datatype())
# Write the neuron parameters
utility_calls.write_parameters_per_neuron(
spec, vertex_slice, self._neuron_model.get_neural_parameters())
# Write the input type parameters
utility_calls.write_parameters_per_neuron(
spec, vertex_slice, self._input_type.get_input_type_parameters())
# Write the additional input parameters
if self._additional_input is not None:
utility_calls.write_parameters_per_neuron(
spec, vertex_slice, self._additional_input.get_parameters())
# Write the threshold type parameters
utility_calls.write_parameters_per_neuron(
spec, vertex_slice,
self._threshold_type.get_threshold_parameters())
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos"
})
@overrides(AbstractRewritesDataSpecification.regenerate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info"
})
def regenerate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
routing_info):
# pylint: disable=too-many-arguments, arguments-differ
vertex_slice = graph_mapper.get_slice(placement.vertex)
# reserve the neuron parameters data region
self._reserve_neuron_params_data_region(
spec, graph_mapper.get_slice(placement.vertex))
# write the neuron params into the new dsg region
self._write_neuron_parameters(
key=routing_info.get_first_key_from_pre_vertex(
placement.vertex, constants.SPIKE_PARTITION_ID),
machine_time_step=machine_time_step,
spec=spec,
time_scale_factor=time_scale_factor,
vertex_slice=vertex_slice)
self._synapse_manager.regenerate_data_specification(
spec, placement, machine_time_step, time_scale_factor,
vertex_slice)
# close spec
spec.end_specification()
@overrides(AbstractRewritesDataSpecification.
requires_memory_regions_to_be_reloaded)
def requires_memory_regions_to_be_reloaded(self):
return self._change_requires_neuron_parameters_reload
@overrides(AbstractRewritesDataSpecification.mark_regions_reloaded)
def mark_regions_reloaded(self):
self._change_requires_neuron_parameters_reload = False
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"application_graph": "MemoryApplicationGraph",
"machine_graph": "MemoryMachineGraph",
"routing_info": "MemoryRoutingInfos",
"tags": "MemoryTags",
"n_machine_time_steps": "TotalMachineTimeSteps"
})
@overrides(AbstractGeneratesDataSpecification.generate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"application_graph", "machine_graph", "routing_info",
"tags", "n_machine_time_steps"
})
def generate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
application_graph, machine_graph,
routing_info, tags, n_machine_time_steps):
# pylint: disable=too-many-arguments, arguments-differ
vertex = placement.vertex
spec.comment("\n*** Spec for block of {} neurons ***\n".format(
self._model_name))
vertex_slice = graph_mapper.get_slice(vertex)
# Reserve memory regions
self._reserve_memory_regions(spec, vertex_slice, vertex)
# Declare random number generators and distributions:
# TODO add random distribution stuff
# self.write_random_distribution_declarations(spec)
# Get the key
key = routing_info.get_first_key_from_pre_vertex(
vertex, constants.SPIKE_PARTITION_ID)
# Write the setup region
spec.switch_write_focus(
constants.POPULATION_BASED_REGIONS.SYSTEM.value)
spec.write_array(
simulation_utilities.get_simulation_header_array(
self.get_binary_file_name(), machine_time_step,
time_scale_factor))
# Write the recording region
spec.switch_write_focus(
constants.POPULATION_BASED_REGIONS.RECORDING.value)
ip_tags = tags.get_ip_tags_for_vertex(vertex)
recorded_region_sizes = recording_utilities.get_recorded_region_sizes(
self._get_buffered_sdram(vertex_slice, n_machine_time_steps),
self._maximum_sdram_for_buffering)
spec.write_array(
recording_utilities.get_recording_header_array(
recorded_region_sizes, self._time_between_requests,
self._buffer_size_before_receive, ip_tags))
# Write the neuron parameters
self._write_neuron_parameters(spec, key, vertex_slice,
machine_time_step, time_scale_factor)
# write profile data
profile_utils.write_profile_region_data(
spec, constants.POPULATION_BASED_REGIONS.PROFILING.value,
self._n_profile_samples)
# allow the synaptic matrix to write its data spec-able data
self._synapse_manager.write_data_spec(spec, self, vertex_slice, vertex,
placement, machine_graph,
application_graph, routing_info,
graph_mapper, self._input_type,
machine_time_step)
# End the writing of this specification:
spec.end_specification()
@overrides(AbstractHasAssociatedBinary.get_binary_file_name)
def get_binary_file_name(self):
# Split binary name into title and extension
binary_title, binary_extension = os.path.splitext(self._binary)
# Reunite title and extension and return
return (binary_title + self._synapse_manager.vertex_executable_suffix +
binary_extension)
@overrides(AbstractHasAssociatedBinary.get_binary_start_type)
def get_binary_start_type(self):
return ExecutableType.USES_SIMULATION_INTERFACE
@overrides(AbstractSpikeRecordable.is_recording_spikes)
def is_recording_spikes(self):
return self._neuron_recorder.is_recording("spikes")
@overrides(AbstractSpikeRecordable.set_recording_spikes)
def set_recording_spikes(self,
new_state=True,
sampling_interval=None,
indexes=None):
self.set_recording("spikes", new_state, sampling_interval, indexes)
@overrides(AbstractSpikeRecordable.get_spikes)
def get_spikes(self, placements, graph_mapper, buffer_manager,
machine_time_step):
return self._neuron_recorder.get_spikes(self.label, buffer_manager,
self.SPIKE_RECORDING_REGION,
placements, graph_mapper, self,
machine_time_step)
@overrides(AbstractNeuronRecordable.get_recordable_variables)
def get_recordable_variables(self):
return self._neuron_recorder.get_recordable_variables()
@overrides(AbstractNeuronRecordable.is_recording)
def is_recording(self, variable):
return self._neuron_recorder.is_recording(variable)
@overrides(AbstractNeuronRecordable.set_recording)
def set_recording(self,
variable,
new_state=True,
sampling_interval=None,
indexes=None):
self._change_requires_mapping = not self.is_recording(variable)
self._neuron_recorder.set_recording(variable, new_state,
sampling_interval, indexes)
@overrides(AbstractNeuronRecordable.get_data)
def get_data(self, variable, n_machine_time_steps, placements,
graph_mapper, buffer_manager, machine_time_step):
# pylint: disable=too-many-arguments
return self._neuron_recorder.get_matrix_data(
self.label, buffer_manager, self.RECORDING_REGION[variable],
placements, graph_mapper, self, variable, n_machine_time_steps)
@overrides(AbstractNeuronRecordable.get_neuron_sampling_interval)
def get_neuron_sampling_interval(self, variable):
return self._neuron_recorder.get_neuron_sampling_interval(variable)
@overrides(AbstractSpikeRecordable.get_spikes_sampling_interval)
def get_spikes_sampling_interval(self):
return self._neuron_recorder.get_neuron_sampling_interval("spikes")
@overrides(AbstractPopulationInitializable.initialize)
def initialize(self, variable, value):
initialize_attr = getattr(self._neuron_model,
"initialize_%s" % variable, None)
if initialize_attr is None or not callable(initialize_attr):
raise Exception("Vertex does not support initialisation of"
" parameter {}".format(variable))
initialize_attr(value)
self._change_requires_neuron_parameters_reload = True
@property
def input_type(self):
return self._input_type
@overrides(AbstractPopulationSettable.get_value)
def get_value(self, key):
""" Get a property of the overall model
"""
for obj in [
self._neuron_model, self._input_type, self._threshold_type,
self._synapse_manager.synapse_type, self._additional_input,
self
]:
if hasattr(obj, key):
return getattr(obj, key)
raise Exception("Population {} does not have parameter {}".format(
self._model_name, key))
@overrides(AbstractPopulationSettable.set_value)
def set_value(self, key, value):
""" Set a property of the overall model
"""
for obj in [
self._neuron_model, self._input_type, self._threshold_type,
self._synapse_manager.synapse_type, self._additional_input
]:
if hasattr(obj, key):
setattr(obj, key, value)
self._change_requires_neuron_parameters_reload = True
return
raise InvalidParameterType("Type {} does not have parameter {}".format(
type(self), key))
@overrides(AbstractReadParametersBeforeSet.read_parameters_from_machine)
def read_parameters_from_machine(self, transceiver, placement,
vertex_slice):
# locate sdram address to where the neuron parameters are stored
neuron_region_sdram_address = \
helpful_functions.locate_memory_region_for_placement(
placement,
constants.POPULATION_BASED_REGIONS.NEURON_PARAMS.value,
transceiver)
# shift past the extra stuff before neuron parameters that we don't
# need to read
neuron_parameters_sdram_address = (
neuron_region_sdram_address +
self.BYTES_TILL_START_OF_GLOBAL_PARAMETERS)
# get size of neuron params
size_of_region = self._get_sdram_usage_for_neuron_params(vertex_slice)
size_of_region -= self.BYTES_TILL_START_OF_GLOBAL_PARAMETERS
# get data from the machine
byte_array = transceiver.read_memory(placement.x, placement.y,
neuron_parameters_sdram_address,
size_of_region)
# Skip the recorder globals as these are not change on machione
# Just written out in case data is changed and written back
offset = self._neuron_recorder.get_size_of_global_parameters(
vertex_slice)
# update python neuron parameters with the data
# handle global params (only once, so given a slice of 0 to 0)
global_params, offset = utility_calls.translate_parameters(
self._neuron_model.get_global_parameter_types(), byte_array,
offset, Slice(0, 0))
self._neuron_model.set_global_parameters(global_params)
# handle state params for a neuron
neuron_params, offset = utility_calls.translate_parameters(
self._neuron_model.get_neural_parameter_types(), byte_array,
offset, vertex_slice)
self._neuron_model.set_neural_parameters(neuron_params, vertex_slice)
# handle input params
input_params, offset = utility_calls.translate_parameters(
self._input_type.get_input_type_parameter_types(), byte_array,
offset, vertex_slice)
self._input_type.set_input_type_parameters(input_params, vertex_slice)
# handle additional input params, if they exist
if self._additional_input is not None:
additional_params, offset = utility_calls.translate_parameters(
self._additional_input.get_parameter_types(), byte_array,
offset, vertex_slice)
self._additional_input.set_parameters(additional_params,
vertex_slice)
# handle threshold type params
threshold_params, offset = utility_calls.translate_parameters(
self._threshold_type.get_threshold_parameter_types(), byte_array,
offset, vertex_slice)
self._threshold_type.set_threshold_parameters(threshold_params,
vertex_slice)
# Read synapse parameters
self._synapse_manager.read_parameters_from_machine(
transceiver, placement, vertex_slice)
@property
def weight_scale(self):
return self._input_type.get_global_weight_scale()
@property
def ring_buffer_sigma(self):
return self._synapse_manager.ring_buffer_sigma
@ring_buffer_sigma.setter
def ring_buffer_sigma(self, ring_buffer_sigma):
self._synapse_manager.ring_buffer_sigma = ring_buffer_sigma
@property
def spikes_per_second(self):
return self._synapse_manager.spikes_per_second
@spikes_per_second.setter
def spikes_per_second(self, spikes_per_second):
self._synapse_manager.spikes_per_second = spikes_per_second
@property
def synapse_dynamics(self):
return self._synapse_manager.synapse_dynamics
def set_synapse_dynamics(self, synapse_dynamics):
self._synapse_manager.synapse_dynamics = synapse_dynamics
def add_pre_run_connection_holder(self, connection_holder, edge,
synapse_info):
# pylint: disable=arguments-differ
self._synapse_manager.add_pre_run_connection_holder(
connection_holder, edge, synapse_info)
@overrides(AbstractAcceptsIncomingSynapses.get_connections_from_machine)
def get_connections_from_machine(
self,
transceiver,
placement,
edge,
graph_mapper,
routing_infos,
synapse_information,
machine_time_step,
using_extra_monitor_cores,
placements=None,
data_receiver=None,
sender_extra_monitor_core_placement=None,
extra_monitor_cores_for_router_timeout=None,
handle_time_out_configuration=True,
fixed_routes=None):
# pylint: disable=too-many-arguments
return self._synapse_manager.get_connections_from_machine(
transceiver, placement, edge, graph_mapper, routing_infos,
synapse_information, machine_time_step, using_extra_monitor_cores,
placements, data_receiver, sender_extra_monitor_core_placement,
extra_monitor_cores_for_router_timeout,
handle_time_out_configuration, fixed_routes)
def clear_connection_cache(self):
self._synapse_manager.clear_connection_cache()
@property
def synapse_type(self):
return self._synapse_manager.synapse_type
def get_maximum_delay_supported_in_ms(self, machine_time_step):
return self._synapse_manager.get_maximum_delay_supported_in_ms(
machine_time_step)
@overrides(AbstractProvidesIncomingPartitionConstraints.
get_incoming_partition_constraints)
def get_incoming_partition_constraints(self, partition):
""" Gets the constraints for partitions going into this vertex
:param partition: partition that goes into this vertex
:return: list of constraints
"""
return self._synapse_manager.get_incoming_partition_constraints()
@overrides(AbstractProvidesOutgoingPartitionConstraints.
get_outgoing_partition_constraints)
def get_outgoing_partition_constraints(self, partition):
""" Gets the constraints for partitions going out of this vertex
:param partition: the partition that leaves this vertex
:return: list of constraints
"""
return [ContiguousKeyRangeContraint()]
@overrides(AbstractNeuronRecordable.clear_recording)
def clear_recording(self, variable, buffer_manager, placements,
graph_mapper):
self._clear_recording_region(buffer_manager, placements, graph_mapper,
self.RECORDING_REGION[variable])
@overrides(AbstractSpikeRecordable.clear_spike_recording)
def clear_spike_recording(self, buffer_manager, placements, graph_mapper):
self._clear_recording_region(
buffer_manager, placements, graph_mapper,
AbstractPopulationVertex.SPIKE_RECORDING_REGION)
def _clear_recording_region(self, buffer_manager, placements, graph_mapper,
recording_region_id):
""" clears a recorded data region from the buffer manager
:param buffer_manager: the buffer manager object
:param placements: the placements object
:param graph_mapper: the graph mapper object
:param recording_region_id: the recorded region id for clearing
:rtype: None
"""
machine_vertices = graph_mapper.get_machine_vertices(self)
for machine_vertex in machine_vertices:
placement = placements.get_placement_of_vertex(machine_vertex)
buffer_manager.clear_recorded_data(placement.x, placement.y,
placement.p,
recording_region_id)
@overrides(AbstractContainsUnits.get_units)
def get_units(self, variable):
# search the model components for the variable
for obj in [
self._neuron_model, self._input_type, self._threshold_type,
self._synapse_manager.synapse_type, self._additional_input
]:
if (hasattr(obj, variable)
and isinstance(obj, AbstractContainsUnits)):
return obj.unit(variable)
# if not in the components, must be within myself, so call my own units
if variable in self._units:
return self._units[variable]
else:
raise InvalidParameterType(
"The parameter {} does not exist in this input "
"conductance component".format(variable))
def describe(self):
"""
Returns a human-readable description of the cell or synapse type.
The output may be customised by specifying a different template
together with an associated template engine
(see ``pyNN.descriptions``).
If template is None, then a dictionary containing the template context
will be returned.
"""
parameters = dict()
for parameter_name in self.default_parameters:
parameters[parameter_name] = self.get_value(parameter_name)
context = {
"name": self._model_name,
"default_parameters": self.default_parameters,
"default_initial_values": self.default_parameters,
"parameters": parameters,
}
return context
def __str__(self):
return "{} with {} atoms".format(self.label, self.n_atoms)
def __repr__(self):
return self.__str__()
class SpiNNakEarApplicationVertex(
ApplicationVertex, AbstractAcceptsIncomingSynapses,
SimplePopulationSettable, HandOverToVertex, AbstractChangableAfterRun,
AbstractSpikeRecordable, AbstractNeuronRecordable,
AbstractControlsDestinationOfEdges, AbstractControlsSourceOfEdges,
AbstractSendsOutgoingSynapses, AbstractCanReset, AbstractContainsUnits,
AbstractApplicationSupportsAutoPauseAndResume,
AbstractProvidesNKeysForPartition):
__slots__ = [
# pynn model
'_model',
# bool flag for neuron param changes
'_remapping_required',
# ihcan vertices
"_ihcan_vertices",
# drnl vertices
"_drnl_vertices",
# final agg verts (outgoing atoms)
"_final_agg_vertices",
# storing synapse dynamics
"_synapse_dynamics",
# fibres per.... something
"_n_fibres_per_ihc",
# the seed for the inner hair fibre
"_ihcan_fibre_random_seed",
# the number of columns / rows for aggregation tree
"_n_group_tree_rows",
# the synaptic manager to manage projections into drnl verts.
"__synapse_manager",
# the number of drnls there are.
"_n_dnrls",
# the number of agg verts which are final aggregation verts.
"_n_final_agg_groups",
# the pole frequencies
"_pole_freqs",
# The timer period for the fast components
"_timer_period"
]
# NOTES IHC = inner hair cell
# IHCan = inner hair channel
# DRNL = middle ear filter
# OME ear fluid
# error message for frequency
FREQUENCY_ERROR = (
"The input sampling frequency is too high for the chosen simulation "
"time scale. Please reduce Fs or increase the time scale factor in "
"the config file")
# error message for get units
GET_UNIT_ERROR = "do not know what to do with variable {} for get units"
# error message if getting source outside aggregation verts
PRE_SLICE_ERROR = (
"Why are you asking for a source outside of aggregation verts?!")
# error message if getting destination verts outside drnls.
POST_SLICE_ERROR = (
"why you asking for a destination atoms outside of the drnl verts!?")
# error for processing plastic synapses
PLASTIC_SYNAPSE_ERROR = (
"The SpiNNaear cannot handle plastic synapses at the moment, "
"complain to the SpiNNaker software team if this is a problem.")
# error message for being asked to clear recording of a param we dont know
CLEAR_RECORDING_ERROR = "Spinnakear does not support recording of {}"
# error message for set recording of a variable we dont know about
RECORDING_ERROR = "Spinnakear does not support recording of {}"
# error message for sampling interval
SAMPLING_INTERVAL_ERROR = "do not know how to handle variable {}"
# error message for incorrect neurons map
N_NEURON_ERROR = (
"the number of neurons {} and the number of atoms {} do not match")
# app edge mc partition id
MC_APP_EDGE_PARTITION_ID = "internal_mc"
# app edge sdram partition id
SDRAM_APP_EDGE_PARTITION_ID = "internal_sdram"
# green wood function from https://en.wikipedia.org/wiki/Greenwood_function
# constant below and mapped to variable names
# green wood constants for human cochlea hearing frequency mapping
# A is a scaling constant between the characteristic frequency and the
# upper frequency limit of the species
GREEN_WOOD_HUMAN_CONSTANT_A = 165.4
# a is the slope of the straight-line portion of the frequency-position
# curve, which has shown to be conserved throughout all investigated
# species after scaling the length of the cochlea
GREEN_WOOD_HUMAN_CONSTANT_ALPHA = 2.1
# K is a constant of integration that represents the divergence from the
# log nature of the curve and is determined by the lower frequency
# audible limit in the species.
GREEN_WOOD_HUMAN_CONSTANT_K = 0.88
# n recording regions
_N_POPULATION_RECORDING_REGIONS = 1
# random numbers
_FINAL_ROW_N_ATOMS = 256
MAX_TIME_SCALE_FACTOR_RATIO = 22050
# flags for sorting out random fibres. might be a enum
HSR_FLAG = 2
MSR_FLAG = 1
LSR_FLAG = 0
# how many synapse types this binary supports
N_SYNAPSE_TYPES = 2
# these curve values are built from profiling the IHCAN cores to deduce
# performance.
CURVE_ONE = 18.12
CURVE_TWO = 10.99
# max audio frequency supported
DEFAULT_MAX_AUDIO_FREQUENCY = 20000
# biggest number of neurons for the ear model
FULL_EAR_HAIR_FIBERS = 30000.0
# min audio frequency supported
DEFAULT_MIN_AUDIO_FREQUENCY = 30
def __init__(self, n_neurons, constraints, label, model, profile,
time_scale_factor):
# Superclasses
ApplicationVertex.__init__(self, label, constraints)
AbstractAcceptsIncomingSynapses.__init__(self)
SimplePopulationSettable.__init__(self)
HandOverToVertex.__init__(self)
AbstractChangableAfterRun.__init__(self)
AbstractSpikeRecordable.__init__(self)
AbstractNeuronRecordable.__init__(self)
AbstractProvidesNKeysForPartition.__init__(self)
self._model = model
self._profile = profile
self._remapping_required = True
self._synapse_dynamics = None
self._n_fibres_per_ihc = None
self._n_group_tree_rows = None
self._ihcan_vertices = list()
self._drnl_vertices = list()
self._final_agg_vertices = list()
self.__synapse_manager = SynapticManager(
self.N_SYNAPSE_TYPES, None, None,
globals_variables.get_simulator().config)
# calculate n fibres per ihcan core
sample_time = time_scale_factor / self._model.fs
# how many channels
self._n_channels = int(self.get_out_going_size() /
self._model.n_fibres_per_ihc)
# process pole freqs
self._pole_freqs = self._process_pole_freqs()
# how many fibres / atoms ran on each ihcan core
self._n_fibres_per_ihcan_core = self.fibres_per_ihcan_core(
sample_time, self._model.n_fibres_per_ihc)
# process all the other internal numbers
atoms_per_row = self.process_internal_numbers()
# read in param file if needed
self._process_param_file(atoms_per_row)
# recording stuff
self._drnl_neuron_recorder = NeuronRecorder(
DRNLMachineVertex.RECORDABLES,
DRNLMachineVertex.get_matrix_scalar_data_types(),
DRNLMachineVertex.get_matrix_output_data_types(), self._n_dnrls)
self._ihcan_neuron_recorder = NeuronRecorder(
IHCANMachineVertex.RECORDABLES,
IHCANMachineVertex.get_matrix_scalar_data_types(),
IHCANMachineVertex.get_matrix_output_data_types(),
self._n_dnrls * self._n_fibres_per_ihc)
# bool for if state has changed.
self._change_requires_mapping = True
self._change_requires_neuron_parameters_reload = False
self._change_requires_data_generation = False
self._has_reset_last = True
# safety check
if self._n_atoms != n_neurons:
raise ConfigurationException(
self.N_NEURON_ERROR.format(n_neurons, self._n_atoms))
# safety stuff
if (self._model.fs / time_scale_factor >
self.MAX_TIME_SCALE_FACTOR_RATIO):
raise Exception(self.FREQUENCY_ERROR)
# write timer period
self._timer_period = (MICRO_TO_SECOND_CONVERSION *
(self._model.seq_size / self._model.fs))
@overrides(AbstractProvidesNKeysForPartition.get_n_keys_for_partition)
def get_n_keys_for_partition(self, partition, graph_mapper):
return partition.pre_vertex.get_n_keys_for_partition(
partition, graph_mapper)
@overrides(AbstractNeuronRecordable.get_expected_n_rows)
def get_expected_n_rows(self, current_run_timesteps_map, sampling_rate,
vertex, variable):
if isinstance(vertex, DRNLMachineVertex):
return int(
(self._drnl_neuron_recorder.expected_rows_for_a_run_time(
current_run_timesteps_map, vertex, sampling_rate)) *
self._model.seq_size)
else:
return int(
self._ihcan_neuron_recorder.expected_rows_for_a_run_time(
current_run_timesteps_map, vertex, sampling_rate))
@staticmethod
def fibres_per_ihcan_core(sample_time, n_fibres_per_ihc):
# how many fibras / atoms ran on each ihcan core
max_possible = abs(
int(
math.floor(((sample_time * MICRO_TO_SECOND_CONVERSION) -
SpiNNakEarApplicationVertex.CURVE_ONE) /
SpiNNakEarApplicationVertex.CURVE_TWO)))
return min(n_fibres_per_ihc, max_possible, 2)
@overrides(AbstractAcceptsIncomingSynapses.gen_on_machine)
def gen_on_machine(self, vertex_slice):
return self.__synapse_manager.gen_on_machine(vertex_slice)
@overrides(AbstractApplicationSupportsAutoPauseAndResume.
my_variable_local_time_period)
def my_variable_local_time_period(self, default_machine_time_step,
variable):
if variable == DRNLMachineVertex.MOC:
return default_machine_time_step
else:
return self._timer_period
def reset_to_first_timestep(self):
# Mark that reset has been done, and reload state variables
self._has_reset_last = True
self._change_requires_neuron_parameters_reload = False
# If synapses change during the run,
if self._synapse_manager.synapse_dynamics.changes_during_run:
self._change_requires_data_generation = True
def get_units(self, variable):
if variable in DRNLMachineVertex.RECORDABLES:
return DRNLMachineVertex.RECORDABLE_UNITS[variable]
elif variable in IHCANMachineVertex.RECORDABLES:
return IHCANMachineVertex.RECORDABLE_UNITS[variable]
else:
raise Exception(self.GET_UNIT_ERROR.format(variable))
def _process_param_file(self, atoms_per_row):
if self._model.param_file is not None:
try:
pre_gen_vars = numpy.load(self._model.param_file)
self._n_atoms = pre_gen_vars['n_atoms']
self._mv_index_list = pre_gen_vars['mv_index_list']
self._parent_index_list = pre_gen_vars['parent_index_list']
self._edge_index_list = pre_gen_vars['edge_index_list']
self._ihc_seeds = pre_gen_vars['ihc_seeds']
self._ome_indices = pre_gen_vars['ome_indices']
except Exception:
self._n_atoms, self._n_dnrls, self._n_final_agg_groups = \
self.calculate_n_atoms_for_each_vertex_type(
atoms_per_row, self._n_channels,
self._model.n_fibres_per_ihc, self._model.seq_size)
# save fixed param file
self._save_pre_gen_vars(self._model.param_file)
else:
self._n_atoms, self._n_dnrls, self._n_final_agg_groups = \
self.calculate_n_atoms_for_each_vertex_type(
atoms_per_row, self._n_channels,
self._model.n_fibres_per_ihc, self._model.seq_size)
def process_internal_numbers(self):
# ear hair frequency bits in total per inner ear channel
self._n_fibres_per_ihc = (self._model.n_lsr_per_ihc +
self._model.n_msr_per_ihc +
self._model.n_hsr_per_ihc)
# number of columns needed for the aggregation tree
atoms_per_row = self.calculate_atoms_per_row(
self._n_channels, self._n_fibres_per_ihc,
self._n_fibres_per_ihcan_core,
self._model.max_input_to_aggregation_group)
# if no rows, then just add 1 row with 1 vertex.
if atoms_per_row == 0:
self._n_group_tree_rows = 1
return self._n_fibres_per_ihcan_core
else:
# figure how many atoms per aggregation element per row
max_n_atoms_per_group_tree_row = (
(self._model.max_input_to_aggregation_group**numpy.arange(
1, atoms_per_row + 1)) * self._n_fibres_per_ihcan_core)
# filter rows max atoms so that its capped at 256
max_n_atoms_per_group_tree_row = \
max_n_atoms_per_group_tree_row[
max_n_atoms_per_group_tree_row <= min(
self._FINAL_ROW_N_ATOMS,
self._n_channels * self._n_fibres_per_ihc)]
self._n_group_tree_rows = max_n_atoms_per_group_tree_row.size
return atoms_per_row
def _process_pole_freqs(self):
if self._model.pole_freqs is None:
if self._model.fs > 2 * self.DEFAULT_MAX_AUDIO_FREQUENCY: # use
# the greenwood mapping
pole_freqs = (numpy.flipud([
self.GREEN_WOOD_HUMAN_CONSTANT_A *
(10**(self.GREEN_WOOD_HUMAN_CONSTANT_ALPHA *
numpy.linspace([0], [1], self._n_channels)) -
self.GREEN_WOOD_HUMAN_CONSTANT_K)
]))
# don't want alias frequencies so we use a capped log scale map
else:
max_power = min([
numpy.log10(self.fs / 2.),
numpy.log10(self.DEFAULT_MAX_AUDIO_FREQUENCY)
])
pole_freqs = numpy.flipud(
numpy.logspace(
numpy.log10(self.DEFAULT_MIN_AUDIO_FREQUENCY),
max_power, self._n_channels))
else:
pole_freqs = self._model.pole_freqs
return pole_freqs[0]
@overrides(AbstractSendsOutgoingSynapses.get_out_going_size)
def get_out_going_size(self):
return (int(self.FULL_EAR_HAIR_FIBERS * float(self._model.scale) /
self._model.n_fibres_per_ihc) *
self._model.n_fibres_per_ihc)
@overrides(AbstractControlsSourceOfEdges.get_out_going_slices)
def get_out_going_slices(self):
slices = list()
starter = 0
for agg_vertex in self._final_agg_vertices:
slices.append(agg_vertex.connection_slice)
starter += agg_vertex.n_atoms
return slices
@overrides(AbstractControlsDestinationOfEdges.get_in_coming_slices)
def get_in_coming_slices(self):
slices = list()
starter = 0
for _ in self._drnl_vertices:
slices.append(Slice(starter, starter))
starter += 1
return slices
@overrides(AbstractControlsSourceOfEdges.get_pre_slice_for)
def get_pre_slice_for(self, machine_vertex):
if isinstance(machine_vertex, ANGroupMachineVertex):
if machine_vertex.is_final_row:
return machine_vertex.connection_slice
raise Exception(self.PRE_SLICE_ERROR)
@overrides(AbstractControlsDestinationOfEdges.get_post_slice_for)
def get_post_slice_for(self, machine_vertex):
if isinstance(machine_vertex, DRNLMachineVertex):
return Slice(machine_vertex.drnl_index, machine_vertex.drnl_index)
raise Exception(self.POST_SLICE_ERROR)
@overrides(AbstractAcceptsIncomingSynapses.get_in_coming_size)
def get_in_coming_size(self):
return self._n_dnrls
@overrides(AbstractAcceptsIncomingSynapses.get_synapse_id_by_target)
def get_synapse_id_by_target(self, target):
if target == "excitatory":
return 0
elif target == "inhibitory":
return 1
return None
@overrides(
AbstractControlsDestinationOfEdges.get_destinations_for_edge_from)
def get_destinations_for_edge_from(self, app_edge, partition_id,
graph_mapper,
original_source_machine_vertex):
if ((app_edge.pre_vertex != self and app_edge.post_vertex == self)
and not isinstance(original_source_machine_vertex,
OMEMachineVertex)):
return self._drnl_vertices
else:
return []
@overrides(AbstractControlsSourceOfEdges.get_sources_for_edge_from)
def get_sources_for_edge_from(self, app_edge, partition_id, graph_mapper,
original_source_machine_vertex):
if ((app_edge.pre_vertex == self and app_edge.post_vertex != self)
and isinstance(original_source_machine_vertex,
ANGroupMachineVertex)
and original_source_machine_vertex.is_final_row):
return [original_source_machine_vertex]
else:
return []
@overrides(
AbstractAcceptsIncomingSynapses.get_maximum_delay_supported_in_ms)
def get_maximum_delay_supported_in_ms(self, default_machine_time_step):
return self.__synapse_manager.get_maximum_delay_supported_in_ms(
default_machine_time_step)
@overrides(AbstractAcceptsIncomingSynapses.add_pre_run_connection_holder)
def add_pre_run_connection_holder(self, connection_holder, projection_edge,
synapse_information):
self.__synapse_manager.add_pre_run_connection_holder(
connection_holder, projection_edge, synapse_information)
def _save_pre_gen_vars(self, file_path):
""" saves params into a numpy file.
:param file_path: path to file to store stuff into
:rtype: None
"""
numpy.savez_compressed(file_path,
n_atoms=self._n_atoms,
mv_index_list=self._mv_index_list,
parent_index_list=self._parent_index_list,
edge_index_list=self._edge_index_list,
ihc_seeds=self._ihc_seeds,
ome_indices=self._ome_indices)
@overrides(AbstractAcceptsIncomingSynapses.set_synapse_dynamics)
def set_synapse_dynamics(self, synapse_dynamics):
if not isinstance(synapse_dynamics, SynapseDynamicsStatic):
raise Exception(self.PLASTIC_SYNAPSE_ERROR)
self._synapse_dynamics = synapse_dynamics
@overrides(AbstractAcceptsIncomingSynapses.get_connections_from_machine)
def get_connections_from_machine(self,
transceiver,
placement,
edge,
graph_mapper,
routing_infos,
synapse_information,
local_time_step_map,
using_extra_monitor_cores,
placements=None,
monitor_api=None,
monitor_placement=None,
monitor_cores=None,
handle_time_out_configuration=True,
fixed_routes=None):
return self.__synapse_manager.get_connections_from_machine(
transceiver, placement, edge, graph_mapper, routing_infos,
synapse_information, local_time_step_map,
using_extra_monitor_cores,
DRNLMachineVertex.REGIONS.POPULATION_TABLE.value,
DRNLMachineVertex.REGIONS.SYNAPTIC_MATRIX.value,
DRNLMachineVertex.REGIONS.DIRECT_MATRIX.value, placements,
monitor_api, monitor_placement, monitor_cores,
handle_time_out_configuration, fixed_routes)
@overrides(AbstractAcceptsIncomingSynapses.clear_connection_cache)
def clear_connection_cache(self):
self.__synapse_manager.clear_connection_cache()
@overrides(SimplePopulationSettable.set_value)
def set_value(self, key, value):
SimplePopulationSettable.set_value(self, key, value)
self._remapping_required = True
def describe(self):
""" Returns a human-readable description of the cell or synapse type.
The output may be customised by specifying a different template\
together with an associated template engine\
(see ``pyNN.descriptions``).
If template is None, then a dictionary containing the template\
context will be returned.
"""
parameters = dict()
for parameter_name in self._model.default_parameters:
parameters[parameter_name] = self.get_value(parameter_name)
context = {
"name": self._model.model_name,
"default_parameters": self._model.default_parameters,
"default_initial_values": self._model.default_parameters,
"parameters": parameters,
}
return context
@overrides(AbstractPopulationSettable.get_value)
def get_value(self, key):
if hasattr(self._model, key):
return getattr(self._model, key)
raise Exception("Population {} does not have parameter {}".format(
self, key))
def _add_to_graph_components(self, machine_graph, graph_mapper, slice,
vertex, resource_tracker):
""" adds the vertex to all the graph components and resources
:param machine_graph: machine graph
:param graph_mapper: graph mapper
:param slice: slice
:param vertex: machien vertex
:param resource_tracker: resource tracker
:rtype: None
"""
machine_graph.add_vertex(vertex)
graph_mapper.add_vertex_mapping(vertex, slice, self)
resource_tracker.allocate_constrained_resources(
vertex.resources_required, vertex.constraints)
def _build_ome_vertex(self, machine_graph, graph_mapper, lo_atom,
resource_tracker, timer_period):
""" builds the ome vertex
:param machine_graph: machine graph
:param graph_mapper: graph mapper
:param lo_atom: lo atom to put into graph mapper slice
:param resource_tracker: the resource tracker
:param timer_period: the timer period for all machine verts based on\
the ear vertex
:return: the ome vertex and the new low atom
"""
# build the ome machine vertex
ome_vertex = OMEMachineVertex(self._model.audio_input, self._model.fs,
self._n_channels, self._model.seq_size,
timer_period, self._profile)
# allocate resources and updater graphs
self._add_to_graph_components(machine_graph, graph_mapper,
Slice(lo_atom, lo_atom), ome_vertex,
resource_tracker)
return ome_vertex, lo_atom + 1
def _build_drnl_verts(self, machine_graph, graph_mapper, new_low_atom,
resource_tracker, ome_vertex, timer_period):
""" build the drnl verts
:param machine_graph: machine graph
:param graph_mapper: graph mapper
:param new_low_atom: the current low atom count for the graph mapper
:param resource_tracker: the resource tracker for placement
:param ome_vertex: the ome vertex to tie edges to
:param timer_period: the timer period for all machine verts based on\
the ear vertex
:return: new low atom count
"""
pole_index = 0
for _ in range(self._n_channels):
drnl_vertex = DRNLMachineVertex(
self._pole_freqs[pole_index], self._model.fs,
ome_vertex.n_data_points, pole_index, self._profile,
self._model.seq_size, self.__synapse_manager, self,
self._model.n_buffers_in_sdram_total,
self._drnl_neuron_recorder, timer_period)
pole_index += 1
self._add_to_graph_components(machine_graph, graph_mapper,
Slice(new_low_atom, new_low_atom),
drnl_vertex, resource_tracker)
new_low_atom += 1
self._drnl_vertices.append(drnl_vertex)
return new_low_atom
def _build_edges_between_ome_drnls(self, ome_vertex, machine_graph,
app_edge, graph_mapper):
""" adds edges between the ome and the drnl vertices
:param ome_vertex: the ome vertex
:param machine_graph: the machine graph
:param app_edge: the app edge covering all these edges
:param graph_mapper: the graph mapper
:rtype: None
"""
for drnl_vert in self._drnl_vertices:
edge = SpiNNakEarMachineEdge(ome_vertex, drnl_vert)
machine_graph.add_edge(edge, ome_vertex.OME_PARTITION_ID)
graph_mapper.add_edge_mapping(edge, app_edge)
def _build_ihcan_vertices_and_sdram_edges(self, machine_graph,
graph_mapper, new_low_atom,
resource_tracker, app_edge,
sdram_app_edge, timer_period):
""" builds the ihcan verts and adds edges from drnl to them
:param machine_graph: machine graph
:param graph_mapper: the graph mapper
:param new_low_atom: the lo atom sued to keep the graph mapper happy
:param resource_tracker: the resource tracker for placement
:param app_edge: the app edge to link all mc machine edges to
:param sdram_app_edge: the application sdram edge between drnl and \
inchan to link all sdram machine edges to.
:param timer_period: the timer period for all machine verts based on\
the ear vertex
:return: iterable of ihcan verts
"""
ihcans = list()
# generate ihc seeds
n_ihcans = self._n_channels * self._model.n_fibres_per_ihc
seed_index = 0
random_range = numpy.arange(
n_ihcans * IHCANMachineVertex.N_SEEDS_PER_IHCAN_VERTEX,
dtype=numpy.uint32)
numpy.random.seed(self._model.ihc_seeds_seed)
ihc_seeds = numpy.random.choice(
random_range,
int(n_ihcans * IHCANMachineVertex.N_SEEDS_PER_IHCAN_VERTEX),
replace=False)
ihcan_recording_index = 0
for drnl_vertex in self._drnl_vertices:
machine_graph.add_outgoing_edge_partition(
ConstantSDRAMMachinePartition(
drnl_vertex.DRNL_SDRAM_PARTITION_ID, drnl_vertex,
"sdram edge between drnl vertex {} and its "
"IHCANS".format(drnl_vertex.drnl_index)))
fibres = []
for _ in range(self._model.n_hsr_per_ihc):
fibres.append(self.HSR_FLAG)
for __ in range(self._model.n_msr_per_ihc):
fibres.append(self.MSR_FLAG)
for ___ in range(self._model.n_lsr_per_ihc):
fibres.append(self.LSR_FLAG)
random.seed(self._model.ihcan_fibre_random_seed)
random.shuffle(fibres)
for _ in range(
int(self._model.n_fibres_per_ihc /
self._n_fibres_per_ihcan_core)):
# randomly pick fibre types
chosen_indices = [
fibres.pop() for _ in range(self._n_fibres_per_ihcan_core)
]
ihcan_slice = Slice(
new_low_atom, new_low_atom +
(self._n_fibres_per_ihcan_core * self._model.seq_size) - 1)
ihcan_recording_slice = Slice(
ihcan_recording_index, ihcan_recording_index +
(self._n_fibres_per_ihcan_core * self._model.seq_size) - 1)
ihcan_recording_index += ihcan_recording_slice.n_atoms
vertex = IHCANMachineVertex(
self._model.resample_factor,
ihc_seeds[seed_index:seed_index +
IHCANMachineVertex.N_SEEDS_PER_IHCAN_VERTEX],
self._n_fibres_per_ihcan_core, self._model.ear_index,
self._profile, self._model.fs,
chosen_indices.count(self.LSR_FLAG),
chosen_indices.count(self.MSR_FLAG),
chosen_indices.count(self.HSR_FLAG),
self._model.n_buffers_in_sdram_total, self._model.seq_size,
self._ihcan_neuron_recorder, ihcan_recording_slice,
timer_period)
# update indexes
new_low_atom += ihcan_slice.n_atoms
seed_index += IHCANMachineVertex.N_SEEDS_PER_IHCAN_VERTEX
# add to list of ihcans
ihcans.append(vertex)
self._add_to_graph_components(machine_graph, graph_mapper,
ihcan_slice, vertex,
resource_tracker)
# multicast
mc_edge = SpiNNakEarMachineEdge(drnl_vertex, vertex,
EdgeTrafficType.MULTICAST)
machine_graph.add_edge(mc_edge, drnl_vertex.DRNL_PARTITION_ID)
graph_mapper.add_edge_mapping(mc_edge, app_edge)
# sdram edge
sdram_edge = SDRAMMachineEdge(
drnl_vertex, vertex, drnl_vertex.sdram_edge_size,
"sdram between {} and {}".format(drnl_vertex, vertex))
machine_graph.add_edge(sdram_edge,
drnl_vertex.DRNL_SDRAM_PARTITION_ID)
graph_mapper.add_edge_mapping(sdram_edge, sdram_app_edge)
return ihcans, new_low_atom
def _build_aggregation_group_vertices_and_edges(self, machine_graph,
graph_mapper, new_low_atom,
resource_tracker,
app_edge):
to_process = list()
to_process.extend(self._ihcan_vertices)
n_child_per_group = self._model.max_input_to_aggregation_group
for row in range(self._n_group_tree_rows):
aggregation_verts = list()
n_row_angs = int(
numpy.ceil(float(len(to_process)) / n_child_per_group))
for an in range(n_row_angs):
final_row_lo_atom = 0
child_verts = to_process[an * n_child_per_group:an *
n_child_per_group + n_child_per_group]
# deduce n atoms of the ag node
n_atoms = 0
for child in child_verts:
n_atoms += child.n_atoms
# build silce for an node
an_slice = Slice(new_low_atom, new_low_atom + n_atoms - 1)
new_low_atom += n_atoms
# build vert
final_row = row == self._n_group_tree_rows - 1
final_row_slice = None
if final_row:
final_row_slice = Slice(final_row_lo_atom,
final_row_lo_atom + n_atoms)
final_row_lo_atom += n_atoms
ag_vertex = ANGroupMachineVertex(n_atoms, len(child_verts),
final_row, row,
self._model.ear_index,
final_row_slice)
# only store it in the agg array if its in the final row
if final_row:
self._final_agg_vertices.append(ag_vertex)
# store for the next cycle
aggregation_verts.append(ag_vertex)
# update stuff
self._add_to_graph_components(machine_graph, graph_mapper,
an_slice, ag_vertex,
resource_tracker)
# add edges
for child_vert in child_verts:
# sort out partition id
partition_id = IHCANMachineVertex.IHCAN_PARTITION_ID
if isinstance(child_vert, ANGroupMachineVertex):
partition_id = \
ANGroupMachineVertex.AN_GROUP_PARTITION_IDENTIFIER
# add edge and mapping
mc_edge = SpiNNakEarMachineEdge(child_vert, ag_vertex)
machine_graph.add_edge(mc_edge, partition_id)
graph_mapper.add_edge_mapping(mc_edge, app_edge)
to_process = aggregation_verts
@inject_items({"application_graph": "MemoryApplicationGraph"})
@overrides(HandOverToVertex.create_and_add_to_graphs_and_resources,
additional_arguments={"application_graph"})
def create_and_add_to_graphs_and_resources(self, resource_tracker,
machine_graph, graph_mapper,
application_graph):
mc_app_edge = ApplicationEdge(self, self)
sdram_app_edge = ApplicationEdge(self, self, EdgeTrafficType.SDRAM)
application_graph.add_edge(mc_app_edge, self.MC_APP_EDGE_PARTITION_ID)
application_graph.add_edge(sdram_app_edge,
self.SDRAM_APP_EDGE_PARTITION_ID)
# atom tracker
current_atom_count = 0
timer_period = (MICRO_TO_SECOND_CONVERSION * self._model.seq_size /
self._model.fs)
# ome vertex
ome_vertex, current_atom_count = self._build_ome_vertex(
machine_graph, graph_mapper, current_atom_count, resource_tracker,
timer_period)
# handle the drnl verts
current_atom_count = self._build_drnl_verts(machine_graph,
graph_mapper,
current_atom_count,
resource_tracker,
ome_vertex, timer_period)
# handle edges between ome and drnls
self._build_edges_between_ome_drnls(ome_vertex, machine_graph,
mc_app_edge, graph_mapper)
# build the ihcan verts.
self._ihcan_vertices, current_atom_count = (
self._build_ihcan_vertices_and_sdram_edges(
machine_graph, graph_mapper, current_atom_count,
resource_tracker, mc_app_edge, sdram_app_edge, timer_period))
# build aggregation group verts and edges
self._build_aggregation_group_vertices_and_edges(
machine_graph, graph_mapper, current_atom_count, resource_tracker,
mc_app_edge)
@property
@overrides(ApplicationVertex.n_atoms)
def n_atoms(self):
return self._n_atoms
@staticmethod
def calculate_n_atoms_for_each_vertex_type(n_group_tree_rows, n_channels,
n_ihc, seq_size):
# ome atom
n_atoms = 1
# dnrl atoms
n_atoms += n_channels
# ihcan atoms
n_angs = n_channels * n_ihc
n_atoms += (n_angs * seq_size)
# an group atoms
for row_index in range(n_group_tree_rows):
n_atoms += n_angs
return n_atoms, n_channels, n_angs
@staticmethod
def calculate_atoms_per_row(n_channels, n_fibres_per_ihc,
n_fibres_per_ihcan,
max_input_to_aggregation_group):
return math.ceil(
numpy.ceil(
math.log((n_channels * n_fibres_per_ihc) / n_fibres_per_ihcan,
max_input_to_aggregation_group)))
@overrides(AbstractChangableAfterRun.mark_no_changes)
def mark_no_changes(self):
self._remapping_required = False
@property
@overrides(AbstractChangableAfterRun.requires_mapping)
def requires_mapping(self):
return self._remapping_required
@overrides(AbstractSpikeRecordable.is_recording_spikes)
def is_recording_spikes(self):
return self._ihcan_neuron_recorder.is_recording(
IHCANMachineVertex.SPIKES)
@overrides(AbstractSpikeRecordable.get_spikes_sampling_interval)
def get_spikes_sampling_interval(self, graph_mapper,
local_time_period_map):
if graph_mapper is None or local_time_period_map is None:
return self.my_variable_local_time_period(
get_simulator().default_machine_time_step,
IHCANMachineVertex.SPIKES)
return self._ihcan_neuron_recorder.get_neuron_sampling_interval(
IHCANMachineVertex.SPIKES, self._ihcan_vertices[0],
local_time_period_map)
@overrides(AbstractSpikeRecordable.clear_spike_recording)
def clear_spike_recording(self, buffer_manager, placements, graph_mapper):
for ihcan_vertex in self._ihcan_vertices:
placement = placements.get_placement_of_vertex(ihcan_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
IHCANMachineVertex.SPIKE_RECORDING_REGION_ID)
@overrides(AbstractSpikeRecordable.get_spikes)
def get_spikes(self, placements, graph_mapper, buffer_manager,
local_timer_period_map):
return self._ihcan_neuron_recorder.get_spikes(
self._label, buffer_manager, IHCANMachineVertex.RECORDING_REGIONS.
SPIKE_RECORDING_REGION_ID.value, placements, graph_mapper, self,
local_timer_period_map)
@overrides(AbstractSpikeRecordable.set_recording_spikes)
def set_recording_spikes(self,
default_machine_time_step,
new_state=True,
sampling_interval=None,
indexes=None):
self.set_recording(IHCANMachineVertex.SPIKES, self._timer_period,
new_state, sampling_interval, indexes)
@overrides(AbstractNeuronRecordable.get_recordable_variables)
def get_recordable_variables(self):
recordables = list()
# don't take the drnl spikes, as there's only 1 api for spikes, and the
# drnl spikes are only there for the recording limitations
recordables.append(DRNLMachineVertex.MOC)
recordables.extend(IHCANMachineVertex.RECORDABLES)
return recordables
@overrides(AbstractSpikeRecordable.get_spike_machine_vertices)
def get_spike_machine_vertices(self, graph_mapper):
return self._ihcan_vertices
@overrides(AbstractNeuronRecordable.get_machine_vertices_for)
def get_machine_vertices_for(self, variable, graph_mapper):
if variable == DRNLMachineVertex.MOC:
return self._drnl_vertices
else:
return self._ihcan_vertices
@overrides(AbstractNeuronRecordable.clear_recording)
def clear_recording(self, variable, buffer_manager, placements,
graph_mapper):
if variable == DRNLMachineVertex.MOC:
for drnl_vertex in self._drnl_vertices:
placement = placements.get_placement_of_vertex(drnl_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
DRNLMachineVertex.MOC_RECORDING_REGION_ID.value)
if variable == self.SPIKES:
for ihcan_vertex in self._ihcan_vertices:
placement = placements.get_placement_of_vertex(ihcan_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
(IHCANMachineVertex.RECORDING_REGIONS.
SPIKE_RECORDING_REGION_ID.value))
if variable == self.SPIKE_PROB:
for ihcan_vertex in self._ihcan_vertices:
placement = placements.get_placement_of_vertex(ihcan_vertex)
buffer_manager.clear_recorded_data(
placement.x, placement.y, placement.p,
(IHCANMachineVertex.RECORDING_REGIONS.
SPIKE_PROBABILITY_REGION_ID.value))
else:
raise ConfigurationException(
self.CLEAR_RECORDING_ERROR.format(variable))
@overrides(AbstractNeuronRecordable.get_neuron_sampling_interval)
def get_neuron_sampling_interval(self, variable, graph_mapper,
local_time_period_map):
if graph_mapper is None or local_time_period_map is None:
return self.my_variable_local_time_period(
get_simulator().default_machine_time_step, variable)
if variable == DRNLMachineVertex.MOC:
return self._drnl_neuron_recorder.get_neuron_sampling_interval(
variable, self._drnl_vertices[0], local_time_period_map)
elif variable in IHCANMachineVertex.RECORDABLES:
return self._ihcan_neuron_recorder.get_neuron_sampling_interval(
variable, self._ihcan_vertices[0], local_time_period_map)
else:
raise Exception(self.SAMPLING_INTERVAL_ERROR.format(variable))
@overrides(AbstractNeuronRecordable.set_recording)
def set_recording(self,
variable,
default_machine_time_step,
new_state=True,
sampling_interval=None,
indexes=None):
self._change_requires_mapping = not self.is_recording(variable)
if variable == DRNLMachineVertex.MOC:
self._drnl_neuron_recorder.set_recording(
variable, sampling_interval, indexes, self,
default_machine_time_step, new_state)
elif variable in IHCANMachineVertex.RECORDABLES:
self._ihcan_neuron_recorder.set_recording(
variable, sampling_interval, indexes, self,
default_machine_time_step, new_state)
else:
raise ConfigurationException(self.RECORDING_ERROR.format(variable))
@overrides(AbstractNeuronRecordable.is_recording)
def is_recording(self, variable):
if variable == DRNLMachineVertex.MOC:
return self._drnl_neuron_recorder.is_recording(variable)
elif variable in IHCANMachineVertex.RECORDABLES:
return self._ihcan_neuron_recorder.is_recording(variable)
else:
raise ConfigurationException(self.RECORDING_ERROR.format(variable))
@overrides(AbstractNeuronRecordable.get_recording_slice)
def get_recording_slice(self, graph_mapper, vertex):
return vertex.recorded_slice()
@overrides(AbstractNeuronRecordable.get_data)
def get_data(self, variable, run_time, placements, graph_mapper,
buffer_manager, local_time_period_map):
if variable == DRNLMachineVertex.MOC:
return self._drnl_neuron_recorder.get_matrix_data(
self._label, buffer_manager,
DRNLMachineVertex.MOC_RECORDABLE_REGION_ID, placements,
graph_mapper, self, variable, run_time, local_time_period_map)
elif variable == IHCANMachineVertex.SPIKE_PROB:
matrix_data = self._ihcan_neuron_recorder.get_matrix_data(
self._label, buffer_manager, IHCANMachineVertex.
RECORDING_REGIONS.SPIKE_PROBABILITY_REGION_ID.value,
placements, graph_mapper, self, variable, run_time,
local_time_period_map)
# convert to n fibers per time step.
new_matrix_data = list()
for element in matrix_data[0]:
seq_elements = list()
for seq_index in range(0, self._model.seq_size):
seq_elements.append(
element[0 + seq_index::self._model.seq_size])
for time_step in seq_elements:
new_matrix_data.append(time_step)
return new_matrix_data, matrix_data[1][0:10], matrix_data[2]
elif variable == IHCANMachineVertex.SPIKES:
return self._ihcan_neuron_recorder.get_spikes(
self._label, buffer_manager, IHCANMachineVertex.
RECORDING_REGIONS.SPIKE_RECORDING_REGION_ID.value, placements,
graph_mapper, self, run_time)
else:
raise ConfigurationException(self.RECORDING_ERROR.format(variable))
def get_sampling_interval(self, sample_size_window):
return ((self._timer_period * sample_size_window) *
MICRO_TO_SECOND_CONVERSION)
class AbstractPopulationVertex(
ApplicationVertex, AbstractGeneratesDataSpecification,
AbstractHasAssociatedBinary, AbstractContainsUnits,
AbstractSpikeRecordable, AbstractNeuronRecordable,
AbstractProvidesOutgoingPartitionConstraints,
AbstractProvidesIncomingPartitionConstraints,
AbstractPopulationInitializable, AbstractPopulationSettable,
AbstractChangableAfterRun, AbstractRewritesDataSpecification,
AbstractReadParametersBeforeSet, AbstractAcceptsIncomingSynapses,
ProvidesKeyToAtomMappingImpl):
""" Underlying vertex model for Neural Populations.
"""
__slots__ = [
"_buffer_size_before_receive", "_change_requires_mapping",
"_change_requires_neuron_parameters_reload",
"_incoming_spike_buffer_size", "_maximum_sdram_for_buffering",
"_minimum_buffer_sdram", "_n_atoms", "_n_profile_samples",
"_neuron_impl", "_neuron_recorder", "_parameters", "_pynn_model",
"_receive_buffer_host", "_receive_buffer_port", "_state_variables",
"_synapse_manager", "_time_between_requests", "_units",
"_using_auto_pause_and_resume"
]
BASIC_MALLOC_USAGE = 2
# recording region IDs
SPIKE_RECORDING_REGION = 0
# the size of the runtime SDP port data region
RUNTIME_SDP_PORT_SIZE = 4
# 8 elements before the start of global parameters
BYTES_TILL_START_OF_GLOBAL_PARAMETERS = 32
_n_vertices = 0
def __init__(self, n_neurons, label, constraints, max_atoms_per_core,
spikes_per_second, ring_buffer_sigma,
incoming_spike_buffer_size, neuron_impl, pynn_model):
# pylint: disable=too-many-arguments, too-many-locals
super(AbstractPopulationVertex, self).__init__(label, constraints,
max_atoms_per_core)
self._n_atoms = n_neurons
# buffer data
self._incoming_spike_buffer_size = incoming_spike_buffer_size
# get config from simulator
config = globals_variables.get_simulator().config
if incoming_spike_buffer_size is None:
self._incoming_spike_buffer_size = config.getint(
"Simulation", "incoming_spike_buffer_size")
self._neuron_impl = neuron_impl
self._pynn_model = pynn_model
self._parameters = SpynnakerRangeDictionary(n_neurons)
self._state_variables = SpynnakerRangeDictionary(n_neurons)
self._neuron_impl.add_parameters(self._parameters)
self._neuron_impl.add_state_variables(self._state_variables)
# Set up for recording
recordables = ["spikes"]
recordables.extend(self._neuron_impl.get_recordable_variables())
self._neuron_recorder = NeuronRecorder(recordables, n_neurons)
self._time_between_requests = config.getint("Buffers",
"time_between_requests")
self._minimum_buffer_sdram = config.getint("Buffers",
"minimum_buffer_sdram")
self._using_auto_pause_and_resume = config.getboolean(
"Buffers", "use_auto_pause_and_resume")
self._receive_buffer_host = config.get("Buffers",
"receive_buffer_host")
self._receive_buffer_port = helpful_functions.read_config_int(
config, "Buffers", "receive_buffer_port")
# If live buffering is enabled, set a maximum on the buffer sizes
spike_buffer_max_size = 0
variable_buffer_max_size = 0
self._buffer_size_before_receive = None
if config.getboolean("Buffers", "enable_buffered_recording"):
spike_buffer_max_size = config.getint("Buffers",
"spike_buffer_size")
variable_buffer_max_size = config.getint("Buffers",
"variable_buffer_size")
self._maximum_sdram_for_buffering = [spike_buffer_max_size]
for _ in self._neuron_impl.get_recordable_variables():
self._maximum_sdram_for_buffering.append(variable_buffer_max_size)
# Set up synapse handling
self._synapse_manager = SynapticManager(
self._neuron_impl.get_n_synapse_types(), ring_buffer_sigma,
spikes_per_second, config)
# bool for if state has changed.
self._change_requires_mapping = True
self._change_requires_neuron_parameters_reload = False
# Set up for profiling
self._n_profile_samples = helpful_functions.read_config_int(
config, "Reports", "n_profile_samples")
@property
@overrides(ApplicationVertex.n_atoms)
def n_atoms(self):
return self._n_atoms
@inject_items({
"graph": "MemoryApplicationGraph",
"n_machine_time_steps": "TotalMachineTimeSteps",
"machine_time_step": "MachineTimeStep"
})
@overrides(ApplicationVertex.get_resources_used_by_atoms,
additional_arguments={
"graph", "n_machine_time_steps", "machine_time_step"
})
def get_resources_used_by_atoms(self, vertex_slice, graph,
n_machine_time_steps, machine_time_step):
# pylint: disable=arguments-differ
# set resources required from this object
container = ResourceContainer(
sdram=SDRAMResource(
self.get_sdram_usage_for_atoms(vertex_slice, graph,
machine_time_step)),
dtcm=DTCMResource(self.get_dtcm_usage_for_atoms(vertex_slice)),
cpu_cycles=CPUCyclesPerTickResource(
self.get_cpu_usage_for_atoms(vertex_slice)))
recording_sizes = recording_utilities.get_recording_region_sizes(
self._get_buffered_sdram(vertex_slice, n_machine_time_steps),
self._minimum_buffer_sdram, self._maximum_sdram_for_buffering,
self._using_auto_pause_and_resume)
container.extend(
recording_utilities.get_recording_resources(
recording_sizes, self._receive_buffer_host,
self._receive_buffer_port))
# return the total resources.
return container
@property
@overrides(AbstractChangableAfterRun.requires_mapping)
def requires_mapping(self):
return self._change_requires_mapping
@overrides(AbstractChangableAfterRun.mark_no_changes)
def mark_no_changes(self):
self._change_requires_mapping = False
def _get_buffered_sdram_per_timestep(self, vertex_slice):
values = [
self._neuron_recorder.get_buffered_sdram_per_timestep(
"spikes", vertex_slice)
]
for variable in self._neuron_impl.get_recordable_variables():
values.append(
self._neuron_recorder.get_buffered_sdram_per_timestep(
variable, vertex_slice))
return values
def _get_buffered_sdram(self, vertex_slice, n_machine_time_steps):
values = [
self._neuron_recorder.get_buffered_sdram("spikes", vertex_slice,
n_machine_time_steps)
]
for variable in self._neuron_impl.get_recordable_variables():
values.append(
self._neuron_recorder.get_buffered_sdram(
variable, vertex_slice, n_machine_time_steps))
return values
@inject_items({"n_machine_time_steps": "TotalMachineTimeSteps"})
@overrides(ApplicationVertex.create_machine_vertex,
additional_arguments={"n_machine_time_steps"})
def create_machine_vertex(self,
vertex_slice,
resources_required,
n_machine_time_steps,
label=None,
constraints=None):
# pylint: disable=too-many-arguments, arguments-differ
is_recording = len(self._neuron_recorder.recording_variables) > 0
buffered_sdram_per_timestep = self._get_buffered_sdram_per_timestep(
vertex_slice)
buffered_sdram = self._get_buffered_sdram(vertex_slice,
n_machine_time_steps)
minimum_buffer_sdram = recording_utilities.get_minimum_buffer_sdram(
buffered_sdram, self._minimum_buffer_sdram)
overflow_sdram = self._neuron_recorder.get_sampling_overflow_sdram(
vertex_slice)
vertex = PopulationMachineVertex(resources_required, is_recording,
minimum_buffer_sdram,
buffered_sdram_per_timestep, label,
constraints, overflow_sdram)
AbstractPopulationVertex._n_vertices += 1
# return machine vertex
return vertex
def get_cpu_usage_for_atoms(self, vertex_slice):
return (_NEURON_BASE_N_CPU_CYCLES + _C_MAIN_BASE_N_CPU_CYCLES +
(_NEURON_BASE_N_CPU_CYCLES_PER_NEURON * vertex_slice.n_atoms) +
self._neuron_recorder.get_n_cpu_cycles(vertex_slice.n_atoms) +
self._neuron_impl.get_n_cpu_cycles(vertex_slice.n_atoms) +
self._synapse_manager.get_n_cpu_cycles())
def get_dtcm_usage_for_atoms(self, vertex_slice):
return (
_NEURON_BASE_DTCM_USAGE_IN_BYTES +
self._neuron_impl.get_dtcm_usage_in_bytes(vertex_slice.n_atoms) +
self._neuron_recorder.get_dtcm_usage_in_bytes(vertex_slice) +
self._synapse_manager.get_dtcm_usage_in_bytes())
def _get_sdram_usage_for_neuron_params(self, vertex_slice):
""" Calculate the SDRAM usage for just the neuron parameters region.
:param vertex_slice: the slice of atoms.
:return: The SDRAM required for the neuron region
"""
return (
self.BYTES_TILL_START_OF_GLOBAL_PARAMETERS +
self._neuron_recorder.get_sdram_usage_in_bytes(vertex_slice) +
self._neuron_impl.get_sdram_usage_in_bytes(vertex_slice.n_atoms))
def get_sdram_usage_for_atoms(self, vertex_slice, graph,
machine_time_step):
sdram_requirement = (
common_constants.SYSTEM_BYTES_REQUIREMENT +
self._get_sdram_usage_for_neuron_params(vertex_slice) +
recording_utilities.get_recording_header_size(
len(self._neuron_impl.get_recordable_variables()) + 1) +
PopulationMachineVertex.get_provenance_data_size(
PopulationMachineVertex.N_ADDITIONAL_PROVENANCE_DATA_ITEMS) +
self._synapse_manager.get_sdram_usage_in_bytes(
vertex_slice, graph.get_edges_ending_at_vertex(self),
machine_time_step) +
(self._get_number_of_mallocs_used_by_dsg() *
common_constants.SARK_PER_MALLOC_SDRAM_USAGE) +
profile_utils.get_profile_region_size(self._n_profile_samples))
return sdram_requirement
def _get_number_of_mallocs_used_by_dsg(self):
extra_mallocs = len(self._neuron_recorder.recording_variables)
return (self.BASIC_MALLOC_USAGE +
self._synapse_manager.get_number_of_mallocs_used_by_dsg() +
extra_mallocs)
def _reserve_memory_regions(self, spec, vertex_slice, vertex):
spec.comment("\nReserving memory space for data regions:\n\n")
# Reserve memory:
spec.reserve_memory_region(
region=constants.POPULATION_BASED_REGIONS.SYSTEM.value,
size=common_constants.SYSTEM_BYTES_REQUIREMENT,
label='System')
self._reserve_neuron_params_data_region(spec, vertex_slice)
spec.reserve_memory_region(
region=constants.POPULATION_BASED_REGIONS.RECORDING.value,
size=recording_utilities.get_recording_header_size(
len(self._neuron_impl.get_recordable_variables()) + 1))
profile_utils.reserve_profile_region(
spec, constants.POPULATION_BASED_REGIONS.PROFILING.value,
self._n_profile_samples)
vertex.reserve_provenance_data_region(spec)
def _reserve_neuron_params_data_region(self, spec, vertex_slice):
""" Reserve the neuron parameter data region.
:param spec: the spec to write the DSG region to
:param vertex_slice: the slice of atoms from the application vertex
:return: None
"""
params_size = self._get_sdram_usage_for_neuron_params(vertex_slice)
spec.reserve_memory_region(
region=constants.POPULATION_BASED_REGIONS.NEURON_PARAMS.value,
size=params_size,
label='NeuronParams')
def _write_neuron_parameters(self, spec, key, vertex_slice,
machine_time_step, time_scale_factor):
# pylint: disable=too-many-arguments
n_atoms = vertex_slice.n_atoms
spec.comment(
"\nWriting Neuron Parameters for {} Neurons:\n".format(n_atoms))
# Set the focus to the memory region 2 (neuron parameters):
spec.switch_write_focus(
region=constants.POPULATION_BASED_REGIONS.NEURON_PARAMS.value)
# Write the random back off value
spec.write_value(
random.randint(0, AbstractPopulationVertex._n_vertices))
# Write the number of microseconds between sending spikes
time_between_spikes = ((machine_time_step * time_scale_factor) /
(n_atoms * 2.0))
spec.write_value(data=int(time_between_spikes))
# Write whether the key is to be used, and then the key, or 0 if it
# isn't to be used
if key is None:
spec.write_value(data=0)
spec.write_value(data=0)
else:
spec.write_value(data=1)
spec.write_value(data=key)
# Write the number of neurons in the block:
spec.write_value(data=n_atoms)
# Write the number of synapse types
spec.write_value(data=self._neuron_impl.get_n_synapse_types())
# Write the size of the incoming spike buffer
spec.write_value(data=self._incoming_spike_buffer_size)
# Write the number of variables that can be recorded
spec.write_value(
data=len(self._neuron_impl.get_recordable_variables()))
# Write the recording data
recording_data = self._neuron_recorder.get_data(vertex_slice)
spec.write_array(recording_data)
# Write the neuron parameters
neuron_data = self._neuron_impl.get_data(self._parameters,
self._state_variables,
vertex_slice)
spec.write_array(neuron_data)
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"routing_info": "MemoryRoutingInfos"
})
@overrides(AbstractRewritesDataSpecification.regenerate_data_specification,
additional_arguments={
"machine_time_step", "time_scale_factor", "graph_mapper",
"routing_info"
})
def regenerate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
routing_info):
# pylint: disable=too-many-arguments, arguments-differ
vertex_slice = graph_mapper.get_slice(placement.vertex)
# reserve the neuron parameters data region
self._reserve_neuron_params_data_region(
spec, graph_mapper.get_slice(placement.vertex))
# write the neuron params into the new DSG region
self._write_neuron_parameters(
key=routing_info.get_first_key_from_pre_vertex(
placement.vertex, constants.SPIKE_PARTITION_ID),
machine_time_step=machine_time_step,
spec=spec,
time_scale_factor=time_scale_factor,
vertex_slice=vertex_slice)
# close spec
spec.end_specification()
@overrides(AbstractRewritesDataSpecification.
requires_memory_regions_to_be_reloaded)
def requires_memory_regions_to_be_reloaded(self):
return self._change_requires_neuron_parameters_reload
@overrides(AbstractRewritesDataSpecification.mark_regions_reloaded)
def mark_regions_reloaded(self):
self._change_requires_neuron_parameters_reload = False
@inject_items({
"machine_time_step": "MachineTimeStep",
"time_scale_factor": "TimeScaleFactor",
"graph_mapper": "MemoryGraphMapper",
"application_graph": "MemoryApplicationGraph",
"machine_graph": "MemoryMachineGraph",
"routing_info": "MemoryRoutingInfos",
"tags": "MemoryTags",
"n_machine_time_steps": "TotalMachineTimeSteps",
"placements": "MemoryPlacements",
})
@overrides(AbstractGeneratesDataSpecification.generate_data_specification,
additional_arguments={
"machine_time_step",
"time_scale_factor",
"graph_mapper",
"application_graph",
"machine_graph",
"routing_info",
"tags",
"n_machine_time_steps",
"placements",
})
def generate_data_specification(self, spec, placement, machine_time_step,
time_scale_factor, graph_mapper,
application_graph, machine_graph,
routing_info, tags, n_machine_time_steps,
placements):
# pylint: disable=too-many-arguments, arguments-differ
vertex = placement.vertex
spec.comment("\n*** Spec for block of {} neurons ***\n".format(
self._neuron_impl.model_name))
vertex_slice = graph_mapper.get_slice(vertex)
# Reserve memory regions
self._reserve_memory_regions(spec, vertex_slice, vertex)
# Declare random number generators and distributions:
# TODO add random distribution stuff
# self.write_random_distribution_declarations(spec)
# Get the key
key = routing_info.get_first_key_from_pre_vertex(
vertex, constants.SPIKE_PARTITION_ID)
# Write the setup region
spec.switch_write_focus(
constants.POPULATION_BASED_REGIONS.SYSTEM.value)
spec.write_array(
simulation_utilities.get_simulation_header_array(
self.get_binary_file_name(), machine_time_step,
time_scale_factor))
# Write the recording region
spec.switch_write_focus(
constants.POPULATION_BASED_REGIONS.RECORDING.value)
ip_tags = tags.get_ip_tags_for_vertex(vertex)
recorded_region_sizes = recording_utilities.get_recorded_region_sizes(
self._get_buffered_sdram(vertex_slice, n_machine_time_steps),
self._maximum_sdram_for_buffering)
spec.write_array(
recording_utilities.get_recording_header_array(
recorded_region_sizes, self._time_between_requests,
self._buffer_size_before_receive, ip_tags))
# Write the neuron parameters
self._write_neuron_parameters(spec, key, vertex_slice,
machine_time_step, time_scale_factor)
# write profile data
profile_utils.write_profile_region_data(
spec, constants.POPULATION_BASED_REGIONS.PROFILING.value,
self._n_profile_samples)
# Get the weight_scale value from the appropriate location
weight_scale = self._neuron_impl.get_global_weight_scale()
# allow the synaptic matrix to write its data spec-able data
self._synapse_manager.write_data_spec(spec, self, vertex_slice, vertex,
placement, machine_graph,
application_graph, routing_info,
graph_mapper, weight_scale,
machine_time_step, placements)
# End the writing of this specification:
spec.end_specification()
@overrides(AbstractHasAssociatedBinary.get_binary_file_name)
def get_binary_file_name(self):
# Split binary name into title and extension
binary_title, binary_extension = os.path.splitext(
self._neuron_impl.binary_name)
# Reunite title and extension and return
return (binary_title + self._synapse_manager.vertex_executable_suffix +
binary_extension)
@overrides(AbstractHasAssociatedBinary.get_binary_start_type)
def get_binary_start_type(self):
return ExecutableType.USES_SIMULATION_INTERFACE
@overrides(AbstractSpikeRecordable.is_recording_spikes)
def is_recording_spikes(self):
return self._neuron_recorder.is_recording("spikes")
@overrides(AbstractSpikeRecordable.set_recording_spikes)
def set_recording_spikes(self,
new_state=True,
sampling_interval=None,
indexes=None):
self.set_recording("spikes", new_state, sampling_interval, indexes)
@overrides(AbstractSpikeRecordable.get_spikes)
def get_spikes(self, placements, graph_mapper, buffer_manager,
machine_time_step):
return self._neuron_recorder.get_spikes(self.label, buffer_manager,
self.SPIKE_RECORDING_REGION,
placements, graph_mapper, self,
machine_time_step)
@overrides(AbstractNeuronRecordable.get_recordable_variables)
def get_recordable_variables(self):
return self._neuron_recorder.get_recordable_variables()
@overrides(AbstractNeuronRecordable.is_recording)
def is_recording(self, variable):
return self._neuron_recorder.is_recording(variable)
@overrides(AbstractNeuronRecordable.set_recording)
def set_recording(self,
variable,
new_state=True,
sampling_interval=None,
indexes=None):
self._change_requires_mapping = not self.is_recording(variable)
self._neuron_recorder.set_recording(variable, new_state,
sampling_interval, indexes)
@overrides(AbstractNeuronRecordable.get_data)
def get_data(self, variable, n_machine_time_steps, placements,
graph_mapper, buffer_manager, machine_time_step):
# pylint: disable=too-many-arguments
index = 0
if variable != "spikes":
index = 1 + self._neuron_impl.get_recordable_variable_index(
variable)
return self._neuron_recorder.get_matrix_data(self.label,
buffer_manager, index,
placements, graph_mapper,
self, variable,
n_machine_time_steps)
@overrides(AbstractNeuronRecordable.get_neuron_sampling_interval)
def get_neuron_sampling_interval(self, variable):
return self._neuron_recorder.get_neuron_sampling_interval(variable)
@overrides(AbstractSpikeRecordable.get_spikes_sampling_interval)
def get_spikes_sampling_interval(self):
return self._neuron_recorder.get_neuron_sampling_interval("spikes")
@overrides(AbstractPopulationInitializable.initialize)
def initialize(self, variable, value):
if variable not in self._state_variables:
raise KeyError("Vertex does not support initialisation of"
" parameter {}".format(variable))
self._state_variables.set_value(variable, value)
self._change_requires_neuron_parameters_reload = True
@property
def initialize_parameters(self):
return self._pynn_model.default_initial_values.keys()
def _get_parameter(self, variable):
if variable.endswith("_init"):
# method called with "V_init"
key = variable[:-5]
if variable in self._state_variables:
# variable is v and parameter is v_init
return variable
elif key in self._state_variables:
# Oops neuron defines v and not v_init
return key
else:
# method called with "v"
if variable + "_init" in self._state_variables:
# variable is v and parameter is v_init
return variable + "_init"
if variable in self._state_variables:
# Oops neuron defines v and not v_init
return variable
# parameter not found for this variable
raise KeyError("No variable {} found in {}".format(
variable, self._neuron_impl.model_name))
@overrides(AbstractPopulationInitializable.get_initial_value)
def get_initial_value(self, variable, selector=None):
parameter = self._get_parameter(variable)
ranged_list = self._state_variables[parameter]
if selector is None:
return ranged_list
return ranged_list.get_values(selector)
@overrides(AbstractPopulationInitializable.set_initial_value)
def set_initial_value(self, variable, value, selector=None):
parameter = self._get_parameter(variable)
ranged_list = self._state_variables[parameter]
ranged_list.set_value_by_selector(selector, value)
@property
def conductance_based(self):
return self._neuron_impl.is_conductance_based
@overrides(AbstractPopulationSettable.get_value)
def get_value(self, key):
""" Get a property of the overall model.
"""
if key not in self._parameters:
raise InvalidParameterType(
"Population {} does not have parameter {}".format(
self._neuron_impl.model_name, key))
return self._parameters[key]
@overrides(AbstractPopulationSettable.set_value)
def set_value(self, key, value):
""" Set a property of the overall model.
"""
if key not in self._parameters:
raise InvalidParameterType(
"Population {} does not have parameter {}".format(
self._neuron_impl.model_name, key))
self._parameters.set_value(key, value)
self._change_requires_neuron_parameters_reload = True
@overrides(AbstractReadParametersBeforeSet.read_parameters_from_machine)
def read_parameters_from_machine(self, transceiver, placement,
vertex_slice):
# locate SDRAM address to where the neuron parameters are stored
neuron_region_sdram_address = \
helpful_functions.locate_memory_region_for_placement(
placement,
constants.POPULATION_BASED_REGIONS.NEURON_PARAMS.value,
transceiver)
# shift past the extra stuff before neuron parameters that we don't
# need to read
neuron_parameters_sdram_address = (
neuron_region_sdram_address +
self.BYTES_TILL_START_OF_GLOBAL_PARAMETERS)
# get size of neuron params
size_of_region = self._get_sdram_usage_for_neuron_params(vertex_slice)
size_of_region -= self.BYTES_TILL_START_OF_GLOBAL_PARAMETERS
# get data from the machine
byte_array = transceiver.read_memory(placement.x, placement.y,
neuron_parameters_sdram_address,
size_of_region)
# Skip the recorder globals as these are not change on machine
# Just written out in case data is changed and written back
offset = self._neuron_recorder.get_sdram_usage_in_bytes(vertex_slice)
# update python neuron parameters with the data
self._neuron_impl.read_data(byte_array, offset, vertex_slice,
self._parameters, self._state_variables)
@property
def weight_scale(self):
return self._neuron_impl.get_global_weight_scale()
@property
def ring_buffer_sigma(self):
return self._synapse_manager.ring_buffer_sigma
@ring_buffer_sigma.setter
def ring_buffer_sigma(self, ring_buffer_sigma):
self._synapse_manager.ring_buffer_sigma = ring_buffer_sigma
@property
def spikes_per_second(self):
return self._synapse_manager.spikes_per_second
@spikes_per_second.setter
def spikes_per_second(self, spikes_per_second):
self._synapse_manager.spikes_per_second = spikes_per_second
@property
def synapse_dynamics(self):
return self._synapse_manager.synapse_dynamics
def set_synapse_dynamics(self, synapse_dynamics):
self._synapse_manager.synapse_dynamics = synapse_dynamics
def add_pre_run_connection_holder(self, connection_holder, edge,
synapse_info):
# pylint: disable=arguments-differ
self._synapse_manager.add_pre_run_connection_holder(
connection_holder, edge, synapse_info)
@overrides(AbstractAcceptsIncomingSynapses.get_connections_from_machine)
def get_connections_from_machine(
self,
transceiver,
placement,
edge,
graph_mapper,
routing_infos,
synapse_information,
machine_time_step,
using_extra_monitor_cores,
placements=None,
data_receiver=None,
sender_extra_monitor_core_placement=None,
extra_monitor_cores_for_router_timeout=None,
handle_time_out_configuration=True,
fixed_routes=None):
# pylint: disable=too-many-arguments
return self._synapse_manager.get_connections_from_machine(
transceiver, placement, edge, graph_mapper, routing_infos,
synapse_information, machine_time_step, using_extra_monitor_cores,
placements, data_receiver, sender_extra_monitor_core_placement,
extra_monitor_cores_for_router_timeout,
handle_time_out_configuration, fixed_routes)
def clear_connection_cache(self):
self._synapse_manager.clear_connection_cache()
def get_maximum_delay_supported_in_ms(self, machine_time_step):
return self._synapse_manager.get_maximum_delay_supported_in_ms(
machine_time_step)
@overrides(AbstractProvidesIncomingPartitionConstraints.
get_incoming_partition_constraints)
def get_incoming_partition_constraints(self, partition):
""" Gets the constraints for partitions going into this vertex.
:param partition: partition that goes into this vertex
:return: list of constraints
"""
return self._synapse_manager.get_incoming_partition_constraints()
@overrides(AbstractProvidesOutgoingPartitionConstraints.
get_outgoing_partition_constraints)
def get_outgoing_partition_constraints(self, partition):
""" Gets the constraints for partitions going out of this vertex.
:param partition: the partition that leaves this vertex
:return: list of constraints
"""
return [ContiguousKeyRangeContraint()]
@overrides(AbstractNeuronRecordable.clear_recording)
def clear_recording(self, variable, buffer_manager, placements,
graph_mapper):
index = 0
if variable != "spikes":
index = 1 + self._neuron_impl.get_recordable_variable_index(
variable)
self._clear_recording_region(buffer_manager, placements, graph_mapper,
index)
@overrides(AbstractSpikeRecordable.clear_spike_recording)
def clear_spike_recording(self, buffer_manager, placements, graph_mapper):
self._clear_recording_region(
buffer_manager, placements, graph_mapper,
AbstractPopulationVertex.SPIKE_RECORDING_REGION)
def _clear_recording_region(self, buffer_manager, placements, graph_mapper,
recording_region_id):
""" Clear a recorded data region from the buffer manager.
:param buffer_manager: the buffer manager object
:param placements: the placements object
:param graph_mapper: the graph mapper object
:param recording_region_id: the recorded region ID for clearing
:rtype: None
"""
machine_vertices = graph_mapper.get_machine_vertices(self)
for machine_vertex in machine_vertices:
placement = placements.get_placement_of_vertex(machine_vertex)
buffer_manager.clear_recorded_data(placement.x, placement.y,
placement.p,
recording_region_id)
@overrides(AbstractContainsUnits.get_units)
def get_units(self, variable):
if self._neuron_impl.is_recordable(variable):
return self._neuron_impl.get_recordable_units(variable)
if variable not in self._parameters:
raise Exception("Population {} does not have parameter {}".format(
self._neuron_impl.model_name, variable))
return self._neuron_impl.get_units(variable)
def describe(self):
""" Get a human-readable description of the cell or synapse type.
The output may be customised by specifying a different template\
together with an associated template engine\
(see ``pyNN.descriptions``).
If template is None, then a dictionary containing the template context\
will be returned.
"""
parameters = dict()
for parameter_name in self._pynn_model.default_parameters:
parameters[parameter_name] = self.get_value(parameter_name)
context = {
"name": self._neuron_impl.model_name,
"default_parameters": self._pynn_model.default_parameters,
"default_initial_values": self._pynn_model.default_parameters,
"parameters": parameters,
}
return context
def get_synapse_id_by_target(self, target):
return self._neuron_impl.get_synapse_id_by_target(target)
def __str__(self):
return "{} with {} atoms".format(self.label, self.n_atoms)
def __repr__(self):
return self.__str__()
def gen_on_machine(self, vertex_slice):
return self._synapse_manager.gen_on_machine(vertex_slice)