def run(self, subgraph, graph_mapper):
new_sub_graph = PartitionedGraph(label=subgraph.label)
new_graph_mapper = GraphMapper(graph_mapper.first_graph_label,
subgraph.label)
# create progress bar
progress_bar = ProgressBar(
len(subgraph.subvertices) + len(subgraph.subedges),
"Filtering edges")
# add the subverts directly, as they wont be pruned.
for subvert in subgraph.subvertices:
new_sub_graph.add_subvertex(subvert)
associated_vertex = graph_mapper.get_vertex_from_subvertex(subvert)
vertex_slice = graph_mapper.get_subvertex_slice(subvert)
new_graph_mapper.add_subvertex(
subvertex=subvert, vertex_slice=vertex_slice,
vertex=associated_vertex)
progress_bar.update()
# start checking subedges to decide which ones need pruning....
for subedge in subgraph.subedges:
if not self._is_filterable(subedge, graph_mapper):
logger.debug("this subedge was not pruned {}".format(subedge))
new_sub_graph.add_subedge(subedge)
associated_edge = graph_mapper.\
get_partitionable_edge_from_partitioned_edge(subedge)
new_graph_mapper.add_partitioned_edge(subedge, associated_edge)
else:
logger.debug("this subedge was pruned {}".format(subedge))
progress_bar.update()
progress_bar.end()
# returned the pruned partitioned_graph and graph_mapper
return new_sub_graph, new_graph_mapper
def reload_application_data(self, reload_application_data_items,
load_data=True):
"""
:param reload_application_data_items: the application data for each
core which needs data to be reloaded to work
:param load_data: a boolean which will not reload if set to false.
:return: None
"""
progress = ProgressBar(len(reload_application_data_items),
"Reloading Application Data")
# fixme need to find a way to remove these private accesses (maybe
# when the dsg in partitioned it will clear up)
for reload_application_data in reload_application_data_items:
if load_data:
data_file = FileDataReader(reload_application_data.data_file)
self._spinnaker_interface._txrx.write_memory(
reload_application_data.chip_x,
reload_application_data.chip_y,
reload_application_data.base_address, data_file,
reload_application_data.data_size)
data_file.close()
user_0_register_address = self._spinnaker_interface._txrx.\
get_user_0_register_address_from_core(
reload_application_data.chip_x,
reload_application_data.chip_y,
reload_application_data.processor_id)
self._spinnaker_interface._txrx.write_memory(
reload_application_data.chip_x, reload_application_data.chip_y,
user_0_register_address, reload_application_data.base_address)
progress.update()
self._total_processors += 1
progress.end()
def load_initial_buffers(self):
""" Load the initial buffers for the senders using mem writes
"""
progress_bar = ProgressBar(len(self._sender_vertices),
"on loading buffer dependant vertices")
for vertex in self._sender_vertices:
for region in vertex.get_regions():
self._send_initial_messages(vertex, region)
progress_bar.update()
progress_bar.end()
def _load_application_data(
self,
placements,
vertex_to_subvertex_mapper,
processor_to_app_data_base_address,
hostname,
app_data_folder,
verify=False,
):
# go through the placements and see if there's any application data to
# load
progress_bar = ProgressBar(len(list(placements.placements)), "Loading application data onto the machine")
for placement in placements.placements:
associated_vertex = vertex_to_subvertex_mapper.get_vertex_from_subvertex(placement.subvertex)
if isinstance(associated_vertex, AbstractDataSpecableVertex):
logger.debug("loading application data for vertex {}".format(associated_vertex.label))
key = (placement.x, placement.y, placement.p)
start_address = processor_to_app_data_base_address[key]["start_address"]
memory_written = processor_to_app_data_base_address[key]["memory_written"]
file_path_for_application_data = associated_vertex.get_application_data_file_path(
placement.x, placement.y, placement.p, hostname, app_data_folder
)
application_data_file_reader = SpinnmanFileDataReader(file_path_for_application_data)
logger.debug("writing application data for vertex {}".format(associated_vertex.label))
self._txrx.write_memory(
placement.x, placement.y, start_address, application_data_file_reader, memory_written
)
application_data_file_reader.close()
if verify:
application_data_file_reader = SpinnmanFileDataReader(file_path_for_application_data)
all_data = application_data_file_reader.readall()
read_data = self._txrx.read_memory(placement.x, placement.y, start_address, memory_written)
if read_data != all_data:
raise Exception(
"Miswrite of {}, {}, {}, {}".format(placement.x, placement.y, placement.p, start_address)
)
application_data_file_reader.close()
# update user 0 so that it points to the start of the
# applications data region on sdram
logger.debug("writing user 0 address for vertex {}".format(associated_vertex.label))
user_o_register_address = self._txrx.get_user_0_register_address_from_core(
placement.x, placement.y, placement.p
)
self._txrx.write_memory(placement.x, placement.y, user_o_register_address, start_address)
# add lines to rerun_script if requested
if self._reports_states.transciever_report:
self._reload_script.add_application_data(file_path_for_application_data, placement, start_address)
progress_bar.update()
progress_bar.end()
def get_gsyn(self, label, n_atoms, transceiver, region,
n_machine_time_steps, placements, graph_mapper,
partitionable_vertex):
ms_per_tick = self._machine_time_step / 1000.0
subvertices = \
graph_mapper.get_subvertices_from_vertex(partitionable_vertex)
tempfilehandle = tempfile.NamedTemporaryFile()
data = numpy.memmap(
tempfilehandle.file, shape=(n_machine_time_steps, n_atoms),
dtype="float64,float64,float64,float64")
data["f0"] = (numpy.arange(
n_atoms * n_machine_time_steps) % n_atoms).reshape(
(n_machine_time_steps, n_atoms))
data["f1"] = numpy.repeat(numpy.arange(
0, n_machine_time_steps * ms_per_tick, ms_per_tick),
n_atoms).reshape((n_machine_time_steps, n_atoms))
progress_bar = ProgressBar(
len(subvertices), "Getting conductance for {}".format(label))
for subvertex in subvertices:
vertex_slice = graph_mapper.get_subvertex_slice(subvertex)
placement = placements.get_placement_of_subvertex(subvertex)
region_size = recording_utils.get_recording_region_size_in_bytes(
n_machine_time_steps, 8 * vertex_slice.n_atoms)
neuron_param_region_data = recording_utils.get_data(
transceiver, placement, region, region_size)
numpy_data = (numpy.asarray(
neuron_param_region_data, dtype="uint8").view(dtype="<i4") /
32767.0).reshape(
(n_machine_time_steps, vertex_slice.n_atoms * 2))
data["f2"][:, vertex_slice.lo_atom:vertex_slice.hi_atom + 1] =\
numpy_data[:, 0::2]
data["f3"][:, vertex_slice.lo_atom:vertex_slice.hi_atom + 1] =\
numpy_data[:, 1::2]
progress_bar.update()
progress_bar.end()
data.shape = n_atoms * n_machine_time_steps
# Sort the data - apparently, using lexsort is faster, but it might
# consume more memory, so the option is left open for sort-in-place
order = numpy.lexsort((data["f1"], data["f0"]))
# data.sort(order=['f0', 'f1'], axis=0)
result = data.view(dtype="float64").reshape(
(n_atoms * n_machine_time_steps, 4))[order]
return result
def get_synaptic_list_from_machine(self, graph_mapper, partitioned_graph,
placements, transceiver, routing_infos):
"""
Get synaptic data for all connections in this Projection from the
machine.
"""
if self._stored_synaptic_data_from_machine is None:
timer = None
if conf.config.getboolean("Reports", "outputTimesForSections"):
timer = Timer()
timer.start_timing()
subedges = \
graph_mapper.get_partitioned_edges_from_partitionable_edge(
self)
if subedges is None:
subedges = list()
synaptic_list = copy.copy(self._synapse_list)
synaptic_list_rows = synaptic_list.get_rows()
progress_bar = ProgressBar(
len(subedges),
"Reading back synaptic matrix for edge between"
" {} and {}".format(self._pre_vertex.label,
self._post_vertex.label))
for subedge in subedges:
n_rows = subedge.get_n_rows(graph_mapper)
pre_vertex_slice = \
graph_mapper.get_subvertex_slice(subedge.pre_subvertex)
post_vertex_slice = \
graph_mapper.get_subvertex_slice(subedge.post_subvertex)
sub_edge_post_vertex = \
graph_mapper.get_vertex_from_subvertex(
subedge.post_subvertex)
rows = sub_edge_post_vertex.get_synaptic_list_from_machine(
placements, transceiver, subedge.pre_subvertex, n_rows,
subedge.post_subvertex,
self._synapse_row_io, partitioned_graph,
routing_infos, subedge.weight_scales).get_rows()
for i in range(len(rows)):
synaptic_list_rows[
i + pre_vertex_slice.lo_atom].set_slice_values(
rows[i], vertex_slice=post_vertex_slice)
progress_bar.update()
progress_bar.end()
self._stored_synaptic_data_from_machine = synaptic_list
if conf.config.getboolean("Reports", "outputTimesForSections"):
logger.info("Time to read matrix: {}".format(
timer.take_sample()))
return self._stored_synaptic_data_from_machine
def get_synaptic_list_from_machine(self, graph_mapper, partitioned_graph,
placements, transceiver, routing_infos):
"""
Get synaptic data for all connections in this Projection from the
machine.
"""
if self._stored_synaptic_data_from_machine is None:
timer = None
if conf.config.getboolean("Reports", "outputTimesForSections"):
timer = Timer()
timer.start_timing()
logger.debug(
"Reading synapse data for edge between {} and {}".format(
self._pre_vertex.label, self._post_vertex.label))
subedges = \
graph_mapper.get_partitioned_edges_from_partitionable_edge(
self)
if subedges is None:
subedges = list()
synaptic_list = copy.copy(self._synapse_list)
synaptic_list_rows = synaptic_list.get_rows()
progress_bar = ProgressBar(
len(subedges), "progress on reading back synaptic matrix")
for subedge in subedges:
n_rows = subedge.get_n_rows(graph_mapper)
pre_vertex_slice = \
graph_mapper.get_subvertex_slice(subedge.pre_subvertex)
post_vertex_slice = \
graph_mapper.get_subvertex_slice(subedge.post_subvertex)
sub_edge_post_vertex = \
graph_mapper.get_vertex_from_subvertex(
subedge.post_subvertex)
rows = sub_edge_post_vertex.get_synaptic_list_from_machine(
placements, transceiver, subedge.pre_subvertex, n_rows,
subedge.post_subvertex, self._synapse_row_io,
partitioned_graph, routing_infos,
subedge.weight_scales).get_rows()
for i in range(len(rows)):
synaptic_list_rows[
i + pre_vertex_slice.lo_atom].set_slice_values(
rows[i], vertex_slice=post_vertex_slice)
progress_bar.update()
progress_bar.end()
self._stored_synaptic_data_from_machine = synaptic_list
if conf.config.getboolean("Reports", "outputTimesForSections"):
timer.take_sample()
return self._stored_synaptic_data_from_machine
def generate_data_specifications(self):
""" generates the dsg for the graph.
:return:
"""
# iterate though subvertexes and call generate_data_spec for each
# vertex
executable_targets = ExecutableTargets()
# create a progress bar for end users
progress_bar = ProgressBar(len(list(self._placements.placements)),
"Generating data specifications")
for placement in self._placements.placements:
associated_vertex =\
self._graph_mapper.get_vertex_from_subvertex(
placement.subvertex)
# if the vertex can generate a DSG, call it
if isinstance(associated_vertex, AbstractDataSpecableVertex):
ip_tags = self._tags.get_ip_tags_for_vertex(
placement.subvertex)
reverse_ip_tags = self._tags.get_reverse_ip_tags_for_vertex(
placement.subvertex)
associated_vertex.generate_data_spec(
placement.subvertex, placement, self._partitioned_graph,
self._partitionable_graph, self._routing_infos,
self._hostname, self._graph_mapper,
self._report_default_directory, ip_tags, reverse_ip_tags,
self._write_text_specs, self._app_data_runtime_folder)
# Get name of binary from vertex
binary_name = associated_vertex.get_binary_file_name()
# Attempt to find this within search paths
binary_path = executable_finder.get_executable_path(
binary_name)
if binary_path is None:
raise exceptions.ExecutableNotFoundException(binary_name)
if not executable_targets.has_binary(binary_path):
executable_targets.add_binary(binary_path)
executable_targets.add_processor(
binary_path, placement.x, placement.y, placement.p)
progress_bar.update()
# finish the progress bar
progress_bar.end()
return executable_targets
def generate_data_specifications(self):
""" generates the dsg for the graph.
:return:
"""
# iterate though subvertexes and call generate_data_spec for each
# vertex
executable_targets = ExecutableTargets()
# create a progress bar for end users
progress_bar = ProgressBar(len(list(self._placements.placements)),
"on generating data specifications")
for placement in self._placements.placements:
associated_vertex =\
self._graph_mapper.get_vertex_from_subvertex(
placement.subvertex)
# if the vertex can generate a DSG, call it
if isinstance(associated_vertex, AbstractDataSpecableVertex):
ip_tags = self._tags.get_ip_tags_for_vertex(
placement.subvertex)
reverse_ip_tags = self._tags.get_reverse_ip_tags_for_vertex(
placement.subvertex)
associated_vertex.generate_data_spec(
placement.subvertex, placement, self._partitioned_graph,
self._partitionable_graph, self._routing_infos,
self._hostname, self._graph_mapper,
self._report_default_directory, ip_tags, reverse_ip_tags,
self._writeTextSpecs, self._app_data_runtime_folder)
progress_bar.update()
# Get name of binary from vertex
binary_name = associated_vertex.get_binary_file_name()
# Attempt to find this within search paths
binary_path = executable_finder.get_executable_path(
binary_name)
if binary_path is None:
raise exceptions.ExecutableNotFoundException(binary_name)
if not executable_targets.has_binary(binary_path):
executable_targets.add_binary(binary_path)
executable_targets.add_processor(binary_path, placement.x,
placement.y, placement.p)
# finish the progress bar
progress_bar.end()
return executable_targets
def load_executable_images(self, executable_targets, app_id):
""" Go through the executable targets and load each binary to \
everywhere and then send a start request to the cores that \
actually use it
"""
progress_bar = ProgressBar(executable_targets.total_processors, "Loading executables onto the machine")
for executable_target_key in executable_targets.binary_paths():
file_reader = SpinnmanFileDataReader(executable_target_key)
core_subset = executable_targets.retrieve_cores_for_a_executable_target(executable_target_key)
statinfo = os.stat(executable_target_key)
size = statinfo.st_size
# TODO there is a need to parse the binary and see if its
# ITCM and DTCM requirements are within acceptable params for
# operating on spinnaker. Currnently there jsut a few safety
# checks which may not be accurate enough.
if size > constants.MAX_SAFE_BINARY_SIZE:
logger.warn(
"The size of this binary is large enough that its"
" possible that the binary may be larger than what is"
" supported by spinnaker currently. Please reduce the"
" binary size if it starts to behave strangely, or goes"
" into the wdog state before starting."
)
if size > constants.MAX_POSSIBLE_BINARY_SIZE:
raise exceptions.ConfigurationException(
"The size of the binary is too large and therefore"
" will very likely cause a WDOG state. Until a more"
" precise measurement of ITCM and DTCM can be produced"
" this is deemed as an error state. Please reduce the"
" size of your binary or circumvent this error check."
)
self._txrx.execute_flood(core_subset, file_reader, app_id, size)
if self._reports_states.transciever_report:
self._reload_script.add_binary(executable_target_key, core_subset)
acutal_cores_loaded = 0
for chip_based in core_subset.core_subsets:
for _ in chip_based.processor_ids:
acutal_cores_loaded += 1
progress_bar.update(amount_to_add=acutal_cores_loaded)
progress_bar.end()
def load_routing_tables(self, router_tables, app_id):
progress_bar = ProgressBar(len(list(router_tables.routing_tables)), "Loading routing data onto the machine")
# load each router table that is needed for the application to run into
# the chips sdram
for router_table in router_tables.routing_tables:
if not self._machine.get_chip_at(router_table.x, router_table.y).virtual:
self._txrx.clear_multicast_routes(router_table.x, router_table.y)
self._txrx.clear_router_diagnostic_counters(router_table.x, router_table.y)
if len(router_table.multicast_routing_entries) > 0:
self._txrx.load_multicast_routes(
router_table.x, router_table.y, router_table.multicast_routing_entries, app_id=app_id
)
if self._reports_states.transciever_report:
self._reload_script.add_routing_table(router_table)
progress_bar.update()
progress_bar.end()
def _set_up_send_buffering(self):
progress_bar = ProgressBar(
len(self.partitioned_graph.subvertices),
"on initialising the buffer managers for vertices which require"
" buffering")
# Create the buffer manager
self._send_buffer_manager = BufferManager(
self._placements, self._routing_infos, self._tags, self._txrx)
for partitioned_vertex in self.partitioned_graph.subvertices:
if isinstance(partitioned_vertex,
AbstractSendsBuffersFromHostPartitionedVertex):
# Add the vertex to the managed vertices
self._send_buffer_manager.add_sender_vertex(
partitioned_vertex)
progress_bar.update()
progress_bar.end()
def get_spikes(self, label, transceiver, region, placements, graph_mapper,
partitionable_vertex):
results = list()
ms_per_tick = self._machine_time_step / 1000.0
subvertices = \
graph_mapper.get_subvertices_from_vertex(partitionable_vertex)
progress_bar = ProgressBar(len(subvertices),
"Getting spikes for {}".format(label))
for subvertex in subvertices:
placement = placements.get_placement_of_subvertex(subvertex)
subvertex_slice = graph_mapper.get_subvertex_slice(subvertex)
# Read the spikes
spike_data = recording_utils.get_data(
transceiver, placement, region, subvertex.region_size)
number_of_bytes_written = len(spike_data)
offset = 0
while offset < number_of_bytes_written:
eieio_header = EIEIODataHeader.from_bytestring(
spike_data, offset)
offset += eieio_header.size
timestamp = eieio_header.payload_base * ms_per_tick
timestamps = numpy.repeat([timestamp], eieio_header.count)
keys = numpy.frombuffer(
spike_data, dtype="<u4", count=eieio_header.count,
offset=offset)
neuron_ids = \
(keys - subvertex.base_key) + subvertex_slice.lo_atom
offset += eieio_header.count * 4
results.append(numpy.dstack((neuron_ids, timestamps))[0])
progress_bar.update()
progress_bar.end()
result = numpy.vstack(results)
result = result[numpy.lexsort((result[:, 1], result[:, 0]))]
return result
def get_spikes(self, label, transceiver, region, n_machine_time_steps,
placements, graph_mapper, partitionable_vertex):
spike_times = list()
spike_ids = list()
ms_per_tick = self._machine_time_step / 1000.0
subvertices = \
graph_mapper.get_subvertices_from_vertex(partitionable_vertex)
progress_bar = ProgressBar(len(subvertices),
"Getting spikes for {}".format(label))
for subvertex in subvertices:
placement = placements.get_placement_of_subvertex(subvertex)
subvertex_slice = graph_mapper.get_subvertex_slice(subvertex)
lo_atom = subvertex_slice.lo_atom
# Read the spikes
n_bytes = int(math.ceil(subvertex_slice.n_atoms / 32.0)) * 4
region_size = recording_utils.get_recording_region_size_in_bytes(
n_machine_time_steps, n_bytes)
spike_data = recording_utils.get_data(
transceiver, placement, region, region_size)
numpy_data = numpy.asarray(spike_data, dtype="uint8").view(
dtype="uint32").byteswap().view("uint8")
bits = numpy.fliplr(numpy.unpackbits(numpy_data).reshape(
(-1, 32))).reshape((-1, n_bytes * 8))
times, indices = numpy.where(bits == 1)
times = times * ms_per_tick
indices = indices + lo_atom
spike_ids.append(indices)
spike_times.append(times)
progress_bar.update()
progress_bar.end()
spike_ids = numpy.hstack(spike_ids)
spike_times = numpy.hstack(spike_times)
result = numpy.dstack((spike_ids, spike_times))[0]
return result[numpy.lexsort((spike_times, spike_ids))]
def set_up_send_buffering(self, partitioned_graph, placements, tags):
"""
interface for buffered vertices
:param partitioned_graph: the partitioned graph object
:param placements: the placements object
:param tags: the tags object
:return: None
"""
progress_bar = ProgressBar(len(partitioned_graph.subvertices), "Initialising buffers")
# Create the buffer manager
self._send_buffer_manager = BufferManager(
placements, tags, self._txrx, self._reports_states, self._app_data_folder, self._reload_script
)
for partitioned_vertex in partitioned_graph.subvertices:
if isinstance(partitioned_vertex, AbstractSendsBuffersFromHostPartitionedVertex):
# Add the vertex to the managed vertices
self._send_buffer_manager.add_sender_vertex(partitioned_vertex)
progress_bar.update()
progress_bar.end()
def run(self, subgraph, graph_mapper):
new_sub_graph = PartitionedGraph(label=subgraph.label)
new_graph_mapper = GraphMapper(graph_mapper.first_graph_label,
subgraph.label)
# create progress bar
progress_bar = \
ProgressBar(len(subgraph.subvertices) + len(subgraph.subedges),
"on checking which subedges are filterable given "
"heuristics")
# add the subverts directly, as they wont be pruned.
for subvert in subgraph.subvertices:
new_sub_graph.add_subvertex(subvert)
associated_vertex = graph_mapper.get_vertex_from_subvertex(subvert)
vertex_slice = graph_mapper.get_subvertex_slice(subvert)
new_graph_mapper.add_subvertex(subvertex=subvert,
vertex_slice=vertex_slice,
vertex=associated_vertex)
progress_bar.update()
# start checking subedges to decide which ones need pruning....
for subedge in subgraph.subedges:
if not self._is_filterable(subedge, graph_mapper):
logger.debug("this subedge was not pruned {}".format(subedge))
new_sub_graph.add_subedge(subedge)
associated_edge = graph_mapper.\
get_partitionable_edge_from_partitioned_edge(subedge)
new_graph_mapper.add_partitioned_edge(subedge, associated_edge)
else:
logger.debug("this subedge was pruned {}".format(subedge))
progress_bar.update()
progress_bar.end()
# returned the pruned partitioned_graph and graph_mapper
return new_sub_graph, new_graph_mapper
def get_spikes(
self, txrx, placements, graph_mapper, compatible_output=False):
"""
Return a 2-column numpy array containing cell ids and spike times for
recorded cells. This is read directly from the memory for the board.
:param transceiver:
:param placements:
:param graph_mapper:
:param compatible_output:
"""
logger.info("Getting spikes for {}".format(self._label))
# Find all the sub-vertices that this pynn_population.py exists on
subvertices = graph_mapper.get_subvertices_from_vertex(self)
progress_bar = ProgressBar(len(subvertices), "Getting spikes")
results = list()
for subvertex in subvertices:
placement = placements.get_placement_of_subvertex(subvertex)
(x, y, p) = placement.x, placement.y, placement.p
subvertex_slice = graph_mapper.get_subvertex_slice(subvertex)
lo_atom = subvertex_slice.lo_atom
hi_atom = subvertex_slice.hi_atom
logger.debug("Reading spikes from chip {}, {}, core {}, "
"lo_atom {} hi_atom {}".format(
x, y, p, lo_atom, hi_atom))
# Get the App Data for the core
app_data_base_address = \
txrx.get_cpu_information_from_core(x, y, p).user[0]
# Get the position of the spike buffer
spike_region_base_address_offset = \
dsg_utility_calls.get_region_base_address_offset(
app_data_base_address,
self._SPIKE_SOURCE_REGIONS
.SPIKE_DATA_RECORDED_REGION.value)
spike_region_base_address_buf = buffer(txrx.read_memory(
x, y, spike_region_base_address_offset, 4))
spike_region_base_address = struct.unpack_from(
"<I", spike_region_base_address_buf)[0]
spike_region_base_address += app_data_base_address
# Read the spike data size
number_of_bytes_written_buf = buffer(txrx.read_memory(
x, y, spike_region_base_address, 4))
number_of_bytes_written = struct.unpack_from(
"<I", number_of_bytes_written_buf)[0]
# check that the number of spikes written is smaller or the same as
# the size of the memory region we allocated for spikes
send_buffer = self._get_spike_send_buffer(subvertex_slice)
if number_of_bytes_written > send_buffer.total_region_size:
raise exceptions.MemReadException(
"the amount of memory written ({}) was larger than was "
"allocated for it ({})"
.format(number_of_bytes_written,
send_buffer.total_region_size))
# Read the spikes
logger.debug("Reading {} ({}) bytes starting at {} + 4"
.format(number_of_bytes_written,
hex(number_of_bytes_written),
hex(spike_region_base_address)))
spike_data_block = txrx.read_memory(
x, y, spike_region_base_address + 4, number_of_bytes_written)
# translate block of spikes into EIEIO messages
offset = 0
while offset <= number_of_bytes_written - 4:
eieio_header = EIEIODataHeader.from_bytestring(
spike_data_block, offset)
offset += eieio_header.size
timestamps = numpy.repeat([eieio_header.payload_base],
eieio_header.count)
keys = numpy.frombuffer(
spike_data_block, dtype="<u4", count=eieio_header.count,
offset=offset)
neuron_ids = ((keys - subvertex.base_key) +
subvertex_slice.lo_atom)
offset += eieio_header.count * 4
results.append(numpy.dstack((neuron_ids, timestamps))[0])
# complete the buffer
progress_bar.update()
progress_bar.end()
result = numpy.vstack(results)
result = result[numpy.lexsort((result[:, 1], result[:, 0]))]
return result
def run(self, run_time):
"""
:param run_time:
:return:
"""
# sort out config param to be valid types
width = config.get("Machine", "width")
height = config.get("Machine", "height")
if width == "None":
width = None
else:
width = int(width)
if height == "None":
height = None
else:
height = int(height)
number_of_boards = config.get("Machine", "number_of_boards")
if number_of_boards == "None":
number_of_boards = None
self.setup_interfaces(
hostname=self._hostname,
bmp_details=config.get("Machine", "bmp_names"),
downed_chips=config.get("Machine", "down_chips"),
downed_cores=config.get("Machine", "down_cores"),
board_version=config.getint("Machine", "version"),
number_of_boards=number_of_boards, width=width, height=height,
is_virtual=config.getboolean("Machine", "virtual_board"),
virtual_has_wrap_arounds=config.getboolean(
"Machine", "requires_wrap_arounds"),
auto_detect_bmp=config.getboolean("Machine", "auto_detect_bmp"))
# adds extra stuff needed by the reload script which cannot be given
# directly.
if self._reports_states.transciever_report:
self._reload_script.runtime = run_time
self._reload_script.time_scale_factor = self._time_scale_factor
# create network report if needed
if self._reports_states is not None:
reports.network_specification_partitionable_report(
self._report_default_directory, self._partitionable_graph,
self._hostname)
# calculate number of machine time steps
if run_time is not None:
self._no_machine_time_steps =\
int((run_time * 1000.0) / self._machine_time_step)
ceiled_machine_time_steps = \
math.ceil((run_time * 1000.0) / self._machine_time_step)
if self._no_machine_time_steps != ceiled_machine_time_steps:
raise common_exceptions.ConfigurationException(
"The runtime and machine time step combination result in "
"a factional number of machine runable time steps and "
"therefore spinnaker cannot determine how many to run for")
for vertex in self._partitionable_graph.vertices:
if isinstance(vertex, AbstractDataSpecableVertex):
vertex.set_no_machine_time_steps(
self._no_machine_time_steps)
else:
self._no_machine_time_steps = None
logger.warn("You have set a runtime that will never end, this may"
"cause the neural models to fail to partition "
"correctly")
for vertex in self._partitionable_graph.vertices:
if (isinstance(vertex, AbstractPopulationRecordableVertex) and
vertex.record):
raise common_exceptions.ConfigurationException(
"recording a population when set to infinite runtime "
"is not currently supportable in this tool chain."
"watch this space")
do_timing = config.getboolean("Reports", "outputTimesForSections")
if do_timing:
timer = Timer()
else:
timer = None
self.set_runtime(run_time)
logger.info("*** Running Mapper *** ")
if do_timing:
timer.start_timing()
self.map_model()
if do_timing:
timer.take_sample()
# add database generation if requested
needs_database = self._auto_detect_database(self._partitioned_graph)
user_create_database = config.get("Database", "create_database")
if ((user_create_database == "None" and needs_database) or
user_create_database == "True"):
wait_on_confirmation = config.getboolean(
"Database", "wait_on_confirmation")
self._database_interface = SpynnakerDataBaseInterface(
self._app_data_runtime_folder, wait_on_confirmation,
self._database_socket_addresses)
self._database_interface.add_system_params(
self._time_scale_factor, self._machine_time_step,
self._runtime)
self._database_interface.add_machine_objects(self._machine)
self._database_interface.add_partitionable_vertices(
self._partitionable_graph)
self._database_interface.add_partitioned_vertices(
self._partitioned_graph, self._graph_mapper,
self._partitionable_graph)
self._database_interface.add_placements(self._placements,
self._partitioned_graph)
self._database_interface.add_routing_infos(
self._routing_infos, self._partitioned_graph)
self._database_interface.add_routing_tables(self._router_tables)
self._database_interface.add_tags(self._partitioned_graph,
self._tags)
execute_mapping = config.getboolean(
"Database", "create_routing_info_to_neuron_id_mapping")
if execute_mapping:
self._database_interface.create_neuron_to_key_mapping(
graph_mapper=self._graph_mapper,
partitionable_graph=self._partitionable_graph,
partitioned_graph=self._partitioned_graph,
routing_infos=self._routing_infos)
# if using a reload script, add if that needs to wait for
# confirmation
if self._reports_states.transciever_report:
self._reload_script.wait_on_confirmation = wait_on_confirmation
for socket_address in self._database_socket_addresses:
self._reload_script.add_socket_address(socket_address)
self._database_interface.send_read_notification()
# execute data spec generation
if do_timing:
timer.start_timing()
logger.info("*** Generating Output *** ")
logger.debug("")
executable_targets = self.generate_data_specifications()
if do_timing:
timer.take_sample()
# execute data spec execution
if do_timing:
timer.start_timing()
processor_to_app_data_base_address = \
self.execute_data_specification_execution(
config.getboolean("SpecExecution", "specExecOnHost"),
self._hostname, self._placements, self._graph_mapper,
write_text_specs=config.getboolean(
"Reports", "writeTextSpecs"),
runtime_application_data_folder=self._app_data_runtime_folder,
machine=self._machine)
if self._reports_states is not None:
reports.write_memory_map_report(self._report_default_directory,
processor_to_app_data_base_address)
if do_timing:
timer.take_sample()
if (not isinstance(self._machine, VirtualMachine) and
config.getboolean("Execute", "run_simulation")):
if do_timing:
timer.start_timing()
logger.info("*** Loading tags ***")
self.load_tags(self._tags)
if self._do_load is True:
logger.info("*** Loading data ***")
self._load_application_data(
self._placements, self._graph_mapper,
processor_to_app_data_base_address, self._hostname,
app_data_folder=self._app_data_runtime_folder,
verify=config.getboolean("Mode", "verify_writes"))
self.load_routing_tables(self._router_tables, self._app_id)
logger.info("*** Loading executables ***")
self.load_executable_images(executable_targets, self._app_id)
logger.info("*** Loading buffers ***")
self.set_up_send_buffering(self._partitioned_graph,
self._placements, self._tags)
# end of entire loading setup
if do_timing:
timer.take_sample()
if self._do_run is True:
logger.info("*** Running simulation... *** ")
if do_timing:
timer.start_timing()
# every thing is in sync0. load the initial buffers
self._send_buffer_manager.load_initial_buffers()
if do_timing:
timer.take_sample()
wait_on_confirmation = config.getboolean(
"Database", "wait_on_confirmation")
send_start_notification = config.getboolean(
"Database", "send_start_notification")
self.wait_for_cores_to_be_ready(executable_targets,
self._app_id)
# wait till external app is ready for us to start if required
if (self._database_interface is not None and
wait_on_confirmation):
self._database_interface.wait_for_confirmation()
self.start_all_cores(executable_targets, self._app_id)
if (self._database_interface is not None and
send_start_notification):
self._database_interface.send_start_notification()
if self._runtime is None:
logger.info("Application is set to run forever - exiting")
else:
self.wait_for_execution_to_complete(
executable_targets, self._app_id, self._runtime,
self._time_scale_factor)
self._has_ran = True
if self._retrieve_provance_data:
progress = ProgressBar(self._placements.n_placements + 1,
"getting provenance data")
# retrieve provence data from central
file_path = os.path.join(self._report_default_directory,
"provance_data")
# check the directory doesnt already exist
if not os.path.exists(file_path):
os.mkdir(file_path)
# write provanence data
self.write_provenance_data_in_xml(file_path, self._txrx)
progress.update()
# retrieve provenance data from any cores that provide data
for placement in self._placements.placements:
if isinstance(placement.subvertex,
AbstractProvidesProvenanceData):
core_file_path = os.path.join(
file_path,
"Provanence_data_for_{}_{}_{}_{}.xml".format(
placement.subvertex.label,
placement.x, placement.y, placement.p))
placement.subvertex.write_provenance_data_in_xml(
core_file_path, self.transceiver, placement)
progress.update()
progress.end()
elif isinstance(self._machine, VirtualMachine):
logger.info(
"*** Using a Virtual Machine so no simulation will occur")
else:
logger.info("*** No simulation requested: Stopping. ***")
def _get_spikes(
self, graph_mapper, placements, transceiver, compatible_output,
spike_recording_region, sub_vertex_out_spike_bytes_function):
"""
Return a 2-column numpy array containing cell ids and spike times for
recorded cells. This is read directly from the memory for the board.
"""
spike_times = list()
spike_ids = list()
ms_per_tick = self._machine_time_step / 1000.0
# Find all the sub-vertices that this pynn_population.py exists on
subvertices = graph_mapper.get_subvertices_from_vertex(self)
progress_bar = ProgressBar(
len(subvertices), "Getting spikes for {}".format(self._label))
for subvertex in subvertices:
placement = placements.get_placement_of_subvertex(subvertex)
(x, y, p) = placement.x, placement.y, placement.p
subvertex_slice = graph_mapper.get_subvertex_slice(subvertex)
lo_atom = subvertex_slice.lo_atom
hi_atom = subvertex_slice.hi_atom
logger.debug("Reading spikes from chip {}, {}, core {}, "
"lo_atom {} hi_atom {}".format(
x, y, p, lo_atom, hi_atom))
# Get the App Data for the core
app_data_base_address = \
transceiver.get_cpu_information_from_core(x, y, p).user[0]
# Get the position of the spike buffer
spike_region_base_address_offset = \
dsg_utility_calls.get_region_base_address_offset(
app_data_base_address, spike_recording_region)
spike_region_base_address_buf = buffer(transceiver.read_memory(
x, y, spike_region_base_address_offset, 4))
spike_region_base_address = struct.unpack_from(
"<I", spike_region_base_address_buf)[0]
spike_region_base_address += app_data_base_address
# Read the spike data size
number_of_bytes_written_buf = buffer(transceiver.read_memory(
x, y, spike_region_base_address, 4))
number_of_bytes_written = struct.unpack_from(
"<I", number_of_bytes_written_buf)[0]
# check that the number of spikes written is smaller or the same as
# the size of the memory region we allocated for spikes
out_spike_bytes = sub_vertex_out_spike_bytes_function(
subvertex, subvertex_slice)
size_of_region = self.get_recording_region_size(out_spike_bytes)
if number_of_bytes_written > size_of_region:
raise exceptions.MemReadException(
"the amount of memory written ({}) was larger than was "
"allocated for it ({})"
.format(number_of_bytes_written, size_of_region))
# Read the spikes
logger.debug("Reading {} ({}) bytes starting at {} + 4"
.format(number_of_bytes_written,
hex(number_of_bytes_written),
hex(spike_region_base_address)))
spike_data = transceiver.read_memory(
x, y, spike_region_base_address + 4, number_of_bytes_written)
numpy_data = numpy.asarray(spike_data, dtype="uint8").view(
dtype="uint32").byteswap().view("uint8")
bits = numpy.fliplr(numpy.unpackbits(numpy_data).reshape(
(-1, 32))).reshape((-1, out_spike_bytes * 8))
times, indices = numpy.where(bits == 1)
times = times * ms_per_tick
indices = indices + lo_atom
spike_ids.append(indices)
spike_times.append(times)
progress_bar.update()
progress_bar.end()
spike_ids = numpy.hstack(spike_ids)
spike_times = numpy.hstack(spike_times)
result = numpy.dstack((spike_ids, spike_times))[0]
return result[numpy.lexsort((spike_times, spike_ids))]
def get_neuron_parameter(
self, region, compatible_output, has_ran, graph_mapper, placements,
txrx, machine_time_step):
if not has_ran:
raise exceptions.SpynnakerException(
"The simulation has not yet ran, therefore neuron param "
"cannot be retrieved")
times = numpy.zeros(0)
ids = numpy.zeros(0)
values = numpy.zeros(0)
ms_per_tick = self._machine_time_step / 1000.0
# Find all the sub-vertices that this pynn_population.py exists on
subvertices = graph_mapper.get_subvertices_from_vertex(self)
progress_bar = ProgressBar(len(subvertices), "Getting recorded data")
for subvertex in subvertices:
placment = placements.get_placement_of_subvertex(subvertex)
(x, y, p) = placment.x, placment.y, placment.p
# Get the App Data for the core
app_data_base_address = txrx.\
get_cpu_information_from_core(x, y, p).user[0]
# Get the position of the value buffer
neuron_param_region_base_address_offset = \
dsg_utility_calls.get_region_base_address_offset(
app_data_base_address, region)
neuron_param_region_base_address_buf = str(list(txrx.read_memory(
x, y, neuron_param_region_base_address_offset, 4))[0])
neuron_param_region_base_address = \
struct.unpack("<I", neuron_param_region_base_address_buf)[0]
neuron_param_region_base_address += app_data_base_address
# Read the size
number_of_bytes_written_buf = \
str(list(txrx.read_memory(
x, y, neuron_param_region_base_address, 4))[0])
number_of_bytes_written = \
struct.unpack_from("<I", number_of_bytes_written_buf)[0]
# Read the values
logger.debug("Reading {} ({}) bytes starting at {}".format(
number_of_bytes_written, hex(number_of_bytes_written),
hex(neuron_param_region_base_address + 4)))
neuron_param_region_data = txrx.read_memory(
x, y, neuron_param_region_base_address + 4,
number_of_bytes_written)
vertex_slice = graph_mapper.get_subvertex_slice(subvertex)
n_atoms = (vertex_slice.hi_atom - vertex_slice.lo_atom) + 1
bytes_per_time_step = n_atoms * 4
number_of_time_steps_written = \
number_of_bytes_written / bytes_per_time_step
logger.debug("Processing {} timesteps"
.format(number_of_time_steps_written))
data_list = bytearray()
for data in neuron_param_region_data:
data_list.extend(data)
numpy_data = numpy.asarray(data_list, dtype="uint8").view(
dtype="<i4") / 32767.0
values = numpy.append(values, numpy_data)
times = numpy.append(
times, numpy.repeat(range(numpy_data.size / n_atoms),
n_atoms) * ms_per_tick)
ids = numpy.append(ids, numpy.add(
numpy.arange(numpy_data.size) % n_atoms, vertex_slice.lo_atom))
progress_bar.update()
progress_bar.end()
result = numpy.dstack((ids, times, values))[0]
result = result[numpy.lexsort((times, ids))]
return result
def _get_gsyn(
self, region, compatible_output, has_ran, graph_mapper, placements,
txrx, machine_time_step, runtime):
if not has_ran:
raise exceptions.SpynnakerException(
"The simulation has not yet ran, therefore neuron param "
"cannot be retrieved")
ms_per_tick = self._machine_time_step / 1000.0
n_timesteps = runtime / ms_per_tick
tempfilehandle = tempfile.NamedTemporaryFile()
data = numpy.memmap(
tempfilehandle.file, shape=(n_timesteps, self._n_atoms),
dtype="float64,float64,float64,float64")
data["f0"] = (numpy.arange(self._n_atoms * n_timesteps) %
self._n_atoms).reshape((n_timesteps, self._n_atoms))
data["f1"] = numpy.repeat(numpy.arange(0, n_timesteps * ms_per_tick,
ms_per_tick), self._n_atoms).reshape(
(n_timesteps, self._n_atoms))
# Find all the sub-vertices that this pynn_population.py exists on
subvertices = graph_mapper.get_subvertices_from_vertex(self)
progress_bar = ProgressBar(
len(subvertices), "Getting recorded gsyn for {}".format(
self._label))
for subvertex in subvertices:
placment = placements.get_placement_of_subvertex(subvertex)
(x, y, p) = placment.x, placment.y, placment.p
# Get the App Data for the core
app_data_base_address = txrx.\
get_cpu_information_from_core(x, y, p).user[0]
# Get the position of the value buffer
neuron_param_region_base_address_offset = \
dsg_utility_calls.get_region_base_address_offset(
app_data_base_address, region)
neuron_param_region_base_address_buf = buffer(txrx.read_memory(
x, y, neuron_param_region_base_address_offset, 4))
neuron_param_region_base_address = struct.unpack_from(
"<I", neuron_param_region_base_address_buf)[0]
neuron_param_region_base_address += app_data_base_address
# Read the size
number_of_bytes_written_buf = buffer(txrx.read_memory(
x, y, neuron_param_region_base_address, 4))
number_of_bytes_written = struct.unpack_from(
"<I", number_of_bytes_written_buf)[0]
# Read the values
logger.debug("Reading {} ({}) bytes starting at {}".format(
number_of_bytes_written, hex(number_of_bytes_written),
hex(neuron_param_region_base_address + 4)))
neuron_param_region_data = txrx.read_memory(
x, y, neuron_param_region_base_address + 4,
number_of_bytes_written)
vertex_slice = graph_mapper.get_subvertex_slice(subvertex)
bytes_per_time_step = vertex_slice.n_atoms * 4
number_of_time_steps_written = \
number_of_bytes_written / bytes_per_time_step
logger.debug("Processing {} timesteps"
.format(number_of_time_steps_written))
numpy_data = (numpy.asarray(
neuron_param_region_data, dtype="uint8").view(dtype="<i4") /
32767.0).reshape((n_timesteps, vertex_slice.n_atoms * 2))
data["f2"][:, vertex_slice.lo_atom:vertex_slice.hi_atom + 1] =\
numpy_data[:, 0::2]
data["f3"][:, vertex_slice.lo_atom:vertex_slice.hi_atom + 1] =\
numpy_data[:, 1::2]
progress_bar.update()
progress_bar.end()
data.shape = self._n_atoms * n_timesteps
# Sort the data - apparently, using lexsort is faster, but it might
# consume more memory, so the option is left open for sort-in-place
order = numpy.lexsort((data["f1"], data["f0"]))
# data.sort(order=['f0', 'f1'], axis=0)
result = data.view(dtype="float64").reshape(
(self._n_atoms * n_timesteps, 4))[order]
return result
def host_based_data_specification_execution(
self, hostname, placements, graph_mapper, write_text_specs, application_data_runtime_folder, machine
):
"""
:param hostname:
:param placements:
:param graph_mapper:
:param write_text_specs:
:param application_data_runtime_folder:
:param machine:
:return:
"""
next_position_tracker = dict()
space_available_tracker = dict()
processor_to_app_data_base_address = dict()
# create a progress bar for end users
progress_bar = ProgressBar(len(list(placements.placements)), "Executing data specifications")
for placement in placements.placements:
associated_vertex = graph_mapper.get_vertex_from_subvertex(placement.subvertex)
# if the vertex can generate a DSG, call it
if isinstance(associated_vertex, AbstractDataSpecableVertex):
data_spec_file_path = associated_vertex.get_data_spec_file_path(
placement.x, placement.y, placement.p, hostname, application_data_runtime_folder
)
app_data_file_path = associated_vertex.get_application_data_file_path(
placement.x, placement.y, placement.p, hostname, application_data_runtime_folder
)
data_spec_reader = FileDataReader(data_spec_file_path)
data_writer = FileDataWriter(app_data_file_path)
# locate current memory requirement
chip = machine.get_chip_at(placement.x, placement.y)
next_position = chip.sdram.user_base_address
space_available = chip.sdram.size
placement_key = (placement.x, placement.y)
if placement_key in next_position_tracker:
next_position = next_position_tracker[placement_key]
space_available = space_available_tracker[placement_key]
# generate a file writer for dse report (app pointer table)
report_writer = None
if write_text_specs:
new_report_directory = os.path.join(self._report_default_directory, "data_spec_text_files")
if not os.path.exists(new_report_directory):
os.mkdir(new_report_directory)
file_name = "{}_DSE_report_for_{}_{}_{}.txt".format(hostname, placement.x, placement.y, placement.p)
report_file_path = os.path.join(new_report_directory, file_name)
report_writer = FileDataWriter(report_file_path)
# generate data spec executor
host_based_data_spec_executor = DataSpecificationExecutor(
data_spec_reader, data_writer, space_available, report_writer
)
# update memory calc and run data spec executor
bytes_used_by_spec = 0
bytes_written_by_spec = 0
try:
bytes_used_by_spec, bytes_written_by_spec = host_based_data_spec_executor.execute()
except DataSpecificationException as e:
logger.error("Error executing data specification for {}".format(associated_vertex))
raise e
# update base address mapper
processor_mapping_key = (placement.x, placement.y, placement.p)
processor_to_app_data_base_address[processor_mapping_key] = {
"start_address": next_position,
"memory_used": bytes_used_by_spec,
"memory_written": bytes_written_by_spec,
}
next_position_tracker[placement_key] = next_position + bytes_used_by_spec
space_available_tracker[placement_key] = space_available - bytes_used_by_spec
# update the progress bar
progress_bar.update()
# close the progress bar
progress_bar.end()
return processor_to_app_data_base_address
def get_synaptic_list_from_machine(self, graph_mapper, partitioned_graph,
placements, transceiver, routing_infos):
"""
Get synaptic data for all connections in this Projection from the
machine.
:param graph_mapper:
:param partitioned_graph:
:param placements:
:param transceiver:
:param routing_infos:
:return:
"""
if self._stored_synaptic_data_from_machine is None:
logger.debug("Reading synapse data for edge between {} and {}"
.format(self._pre_vertex.label,
self._post_vertex.label))
timer = None
if conf.config.getboolean("Reports", "outputTimesForSections"):
timer = Timer()
timer.start_timing()
subedges = \
graph_mapper.get_partitioned_edges_from_partitionable_edge(
self)
if subedges is None:
subedges = list()
synaptic_list = copy.copy(self._synapse_list)
synaptic_list_rows = synaptic_list.get_rows()
progress_bar = ProgressBar(
len(subedges), "progress on reading back synaptic matrix")
for subedge in subedges:
n_rows = subedge.get_n_rows(graph_mapper)
pre_vertex_slice = \
graph_mapper.get_subvertex_slice(subedge.pre_subvertex)
post_vertex_slice = \
graph_mapper.get_subvertex_slice(subedge.post_subvertex)
sub_edge_post_vertex = \
graph_mapper.get_vertex_from_subvertex(
subedge.post_subvertex)
rows = sub_edge_post_vertex.get_synaptic_list_from_machine(
placements, transceiver, subedge.pre_subvertex, n_rows,
subedge.post_subvertex,
self._synapse_row_io, partitioned_graph,
routing_infos, subedge.weight_scales).get_rows()
for i in range(len(rows)):
delay_stage = math.floor(
float(i) / float(pre_vertex_slice.n_atoms)) + 1
min_delay = (delay_stage *
self.pre_vertex.max_delay_per_neuron)
max_delay = (min_delay +
self.pre_vertex.max_delay_per_neuron - 1)
synaptic_list_rows[
(i % pre_vertex_slice.n_atoms) +
pre_vertex_slice.lo_atom].set_slice_values(
rows[i], post_vertex_slice, min_delay, max_delay)
progress_bar.update()
progress_bar.end()
self._stored_synaptic_data_from_machine = synaptic_list
if conf.config.getboolean("Reports", "outputTimesForSections"):
timer.take_sample()
return self._stored_synaptic_data_from_machine
def _get_spikes(self, graph_mapper, placements, transceiver,
compatible_output, spike_recording_region,
sub_vertex_out_spike_bytes_function):
"""
Return a 2-column numpy array containing cell ids and spike times for
recorded cells. This is read directly from the memory for the board.
"""
logger.info("Getting spikes for {}".format(self._label))
spike_times = list()
spike_ids = list()
ms_per_tick = self._machine_time_step / 1000.0
# Find all the sub-vertices that this pynn_population.py exists on
subvertices = graph_mapper.get_subvertices_from_vertex(self)
progress_bar = ProgressBar(len(subvertices), "Getting spikes")
for subvertex in subvertices:
placement = placements.get_placement_of_subvertex(subvertex)
(x, y, p) = placement.x, placement.y, placement.p
subvertex_slice = graph_mapper.get_subvertex_slice(subvertex)
lo_atom = subvertex_slice.lo_atom
hi_atom = subvertex_slice.hi_atom
logger.debug("Reading spikes from chip {}, {}, core {}, "
"lo_atom {} hi_atom {}".format(
x, y, p, lo_atom, hi_atom))
# Get the App Data for the core
app_data_base_address = \
transceiver.get_cpu_information_from_core(x, y, p).user[0]
# Get the position of the spike buffer
spike_region_base_address_offset = \
dsg_utility_calls.get_region_base_address_offset(
app_data_base_address, spike_recording_region)
spike_region_base_address_buf = transceiver.read_memory(
x, y, spike_region_base_address_offset, 4)
spike_region_base_address = struct.unpack_from(
"<I", spike_region_base_address_buf)[0]
spike_region_base_address += app_data_base_address
# Read the spike data size
number_of_bytes_written_buf = transceiver.read_memory(
x, y, spike_region_base_address, 4)
number_of_bytes_written = struct.unpack_from(
"<I", number_of_bytes_written_buf)[0]
# check that the number of spikes written is smaller or the same as
# the size of the memory region we allocated for spikes
out_spike_bytes = sub_vertex_out_spike_bytes_function(
subvertex, subvertex_slice)
size_of_region = self.get_recording_region_size(out_spike_bytes)
if number_of_bytes_written > size_of_region:
raise exceptions.MemReadException(
"the amount of memory written ({}) was larger than was "
"allocated for it ({})".format(number_of_bytes_written,
size_of_region))
# Read the spikes
logger.debug("Reading {} ({}) bytes starting at {} + 4".format(
number_of_bytes_written, hex(number_of_bytes_written),
hex(spike_region_base_address)))
spike_data = transceiver.read_memory(x, y,
spike_region_base_address + 4,
number_of_bytes_written)
numpy_data = numpy.asarray(
spike_data,
dtype="uint8").view(dtype="uint32").byteswap().view("uint8")
bits = numpy.fliplr(
numpy.unpackbits(numpy_data).reshape((-1, 32))).reshape(
(-1, out_spike_bytes * 8))
times, indices = numpy.where(bits == 1)
times = times * ms_per_tick
indices = indices + lo_atom
spike_ids.append(indices)
spike_times.append(times)
progress_bar.update()
progress_bar.end()
spike_ids = numpy.hstack(spike_ids)
spike_times = numpy.hstack(spike_times)
result = numpy.dstack((spike_ids, spike_times))[0]
return result[numpy.lexsort((spike_times, spike_ids))]
def get_neuron_parameter(self, region, compatible_output, has_ran,
graph_mapper, placements, txrx, machine_time_step,
runtime):
if not has_ran:
raise exceptions.SpynnakerException(
"The simulation has not yet ran, therefore neuron param "
"cannot be retrieved")
ms_per_tick = self._machine_time_step / 1000.0
n_timesteps = runtime / ms_per_tick
tempfilehandle = tempfile.NamedTemporaryFile()
data = numpy.memmap(tempfilehandle.file,
shape=(n_timesteps, self._n_atoms),
dtype="float64,float64,float64")
data["f0"] = (numpy.arange(self._n_atoms * n_timesteps) %
self._n_atoms).reshape((n_timesteps, self._n_atoms))
data["f1"] = numpy.repeat(
numpy.arange(0, n_timesteps * ms_per_tick, ms_per_tick),
self._n_atoms).reshape((n_timesteps, self._n_atoms))
# Find all the sub-vertices that this pynn_population.py exists on
subvertices = graph_mapper.get_subvertices_from_vertex(self)
progress_bar = ProgressBar(len(subvertices), "Getting recorded data")
for subvertex in subvertices:
placment = placements.get_placement_of_subvertex(subvertex)
(x, y, p) = placment.x, placment.y, placment.p
# Get the App Data for the core
app_data_base_address = txrx.\
get_cpu_information_from_core(x, y, p).user[0]
# Get the position of the value buffer
neuron_param_region_base_address_offset = \
dsg_utility_calls.get_region_base_address_offset(
app_data_base_address, region)
neuron_param_region_base_address_buf = txrx.read_memory(
x, y, neuron_param_region_base_address_offset, 4)
neuron_param_region_base_address = struct.unpack_from(
"<I", neuron_param_region_base_address_buf)[0]
neuron_param_region_base_address += app_data_base_address
# Read the size
number_of_bytes_written_buf = txrx.read_memory(
x, y, neuron_param_region_base_address, 4)
number_of_bytes_written = struct.unpack_from(
"<I", number_of_bytes_written_buf)[0]
# Read the values
logger.debug("Reading {} ({}) bytes starting at {}".format(
number_of_bytes_written, hex(number_of_bytes_written),
hex(neuron_param_region_base_address + 4)))
neuron_param_region_data = txrx.read_memory(
x, y, neuron_param_region_base_address + 4,
number_of_bytes_written)
vertex_slice = graph_mapper.get_subvertex_slice(subvertex)
bytes_per_time_step = vertex_slice.n_atoms * 4
number_of_time_steps_written = \
number_of_bytes_written / bytes_per_time_step
logger.debug(
"Processing {} timesteps".format(number_of_time_steps_written))
numpy_data = (numpy.asarray(neuron_param_region_data,
dtype="uint8").view(dtype="<i4") /
32767.0).reshape((n_timesteps, vertex_slice.n_atoms))
data["f2"][:, vertex_slice.lo_atom:vertex_slice.hi_atom + 1] =\
numpy_data
progress_bar.update()
progress_bar.end()
data.shape = self._n_atoms * n_timesteps
# Sort the data - apparently, using lexsort is faster, but it might
# consume more memory, so the option is left open for sort-in-place
order = numpy.lexsort((data["f1"], data["f0"]))
# data.sort(order=['f0', 'f1'], axis=0)
result = data.view(dtype="float64").reshape(
(self._n_atoms * n_timesteps, 3))[order]
return result
def run(self, run_time):
"""
:param run_time:
:return:
"""
# sort out config param to be valid types
width = config.get("Machine", "width")
height = config.get("Machine", "height")
if width == "None":
width = None
else:
width = int(width)
if height == "None":
height = None
else:
height = int(height)
number_of_boards = config.get("Machine", "number_of_boards")
if number_of_boards == "None":
number_of_boards = None
self.setup_interfaces(
hostname=self._hostname,
bmp_details=config.get("Machine", "bmp_names"),
downed_chips=config.get("Machine", "down_chips"),
downed_cores=config.get("Machine", "down_cores"),
board_version=config.getint("Machine", "version"),
number_of_boards=number_of_boards,
width=width,
height=height,
is_virtual=config.getboolean("Machine", "virtual_board"),
virtual_has_wrap_arounds=config.getboolean(
"Machine", "requires_wrap_arounds"),
auto_detect_bmp=config.getboolean("Machine", "auto_detect_bmp"))
# adds extra stuff needed by the reload script which cannot be given
# directly.
if self._reports_states.transciever_report:
self._reload_script.runtime = run_time
self._reload_script.time_scale_factor = self._time_scale_factor
# create network report if needed
if self._reports_states is not None:
reports.network_specification_partitionable_report(
self._report_default_directory, self._partitionable_graph,
self._hostname)
# calculate number of machine time steps
if run_time is not None:
self._no_machine_time_steps =\
int((run_time * 1000.0) / self._machine_time_step)
ceiled_machine_time_steps = \
math.ceil((run_time * 1000.0) / self._machine_time_step)
if self._no_machine_time_steps != ceiled_machine_time_steps:
raise common_exceptions.ConfigurationException(
"The runtime and machine time step combination result in "
"a factional number of machine runable time steps and "
"therefore spinnaker cannot determine how many to run for")
for vertex in self._partitionable_graph.vertices:
if isinstance(vertex, AbstractDataSpecableVertex):
vertex.set_no_machine_time_steps(
self._no_machine_time_steps)
else:
self._no_machine_time_steps = None
logger.warn("You have set a runtime that will never end, this may"
"cause the neural models to fail to partition "
"correctly")
for vertex in self._partitionable_graph.vertices:
if (isinstance(vertex, AbstractPopulationRecordableVertex)
and vertex.record):
raise common_exceptions.ConfigurationException(
"recording a population when set to infinite runtime "
"is not currently supportable in this tool chain."
"watch this space")
do_timing = config.getboolean("Reports", "outputTimesForSections")
if do_timing:
timer = Timer()
else:
timer = None
self.set_runtime(run_time)
logger.info("*** Running Mapper *** ")
if do_timing:
timer.start_timing()
self.map_model()
if do_timing:
timer.take_sample()
# add database generation if requested
needs_database = self._auto_detect_database(self._partitioned_graph)
user_create_database = config.get("Database", "create_database")
if ((user_create_database == "None" and needs_database)
or user_create_database == "True"):
wait_on_confirmation = config.getboolean("Database",
"wait_on_confirmation")
self._database_interface = SpynnakerDataBaseInterface(
self._app_data_runtime_folder, wait_on_confirmation,
self._database_socket_addresses)
self._database_interface.add_system_params(self._time_scale_factor,
self._machine_time_step,
self._runtime)
self._database_interface.add_machine_objects(self._machine)
self._database_interface.add_partitionable_vertices(
self._partitionable_graph)
self._database_interface.add_partitioned_vertices(
self._partitioned_graph, self._graph_mapper,
self._partitionable_graph)
self._database_interface.add_placements(self._placements,
self._partitioned_graph)
self._database_interface.add_routing_infos(self._routing_infos,
self._partitioned_graph)
self._database_interface.add_routing_tables(self._router_tables)
self._database_interface.add_tags(self._partitioned_graph,
self._tags)
execute_mapping = config.getboolean(
"Database", "create_routing_info_to_neuron_id_mapping")
if execute_mapping:
self._database_interface.create_neuron_to_key_mapping(
graph_mapper=self._graph_mapper,
partitionable_graph=self._partitionable_graph,
partitioned_graph=self._partitioned_graph,
routing_infos=self._routing_infos)
# if using a reload script, add if that needs to wait for
# confirmation
if self._reports_states.transciever_report:
self._reload_script.wait_on_confirmation = wait_on_confirmation
for socket_address in self._database_socket_addresses:
self._reload_script.add_socket_address(socket_address)
self._database_interface.send_read_notification()
# execute data spec generation
if do_timing:
timer.start_timing()
logger.info("*** Generating Output *** ")
logger.debug("")
executable_targets = self.generate_data_specifications()
if do_timing:
timer.take_sample()
# execute data spec execution
if do_timing:
timer.start_timing()
processor_to_app_data_base_address = \
self.execute_data_specification_execution(
config.getboolean("SpecExecution", "specExecOnHost"),
self._hostname, self._placements, self._graph_mapper,
write_text_specs=config.getboolean(
"Reports", "writeTextSpecs"),
runtime_application_data_folder=self._app_data_runtime_folder,
machine=self._machine)
if self._reports_states is not None:
reports.write_memory_map_report(
self._report_default_directory,
processor_to_app_data_base_address)
if do_timing:
timer.take_sample()
if (not isinstance(self._machine, VirtualMachine)
and config.getboolean("Execute", "run_simulation")):
if do_timing:
timer.start_timing()
logger.info("*** Loading tags ***")
self.load_tags(self._tags)
if self._do_load is True:
logger.info("*** Loading data ***")
self._load_application_data(
self._placements,
self._graph_mapper,
processor_to_app_data_base_address,
self._hostname,
app_data_folder=self._app_data_runtime_folder,
verify=config.getboolean("Mode", "verify_writes"))
self.load_routing_tables(self._router_tables, self._app_id)
logger.info("*** Loading executables ***")
self.load_executable_images(executable_targets, self._app_id)
logger.info("*** Loading buffers ***")
self.set_up_send_buffering(self._partitioned_graph,
self._placements, self._tags)
# end of entire loading setup
if do_timing:
timer.take_sample()
if self._do_run is True:
logger.info("*** Running simulation... *** ")
if do_timing:
timer.start_timing()
# every thing is in sync0. load the initial buffers
self._send_buffer_manager.load_initial_buffers()
if do_timing:
timer.take_sample()
wait_on_confirmation = config.getboolean(
"Database", "wait_on_confirmation")
send_start_notification = config.getboolean(
"Database", "send_start_notification")
self.wait_for_cores_to_be_ready(executable_targets,
self._app_id)
# wait till external app is ready for us to start if required
if (self._database_interface is not None
and wait_on_confirmation):
self._database_interface.wait_for_confirmation()
self.start_all_cores(executable_targets, self._app_id)
if (self._database_interface is not None
and send_start_notification):
self._database_interface.send_start_notification()
if self._runtime is None:
logger.info("Application is set to run forever - exiting")
else:
self.wait_for_execution_to_complete(
executable_targets, self._app_id, self._runtime,
self._time_scale_factor)
self._has_ran = True
if self._retrieve_provance_data:
progress = ProgressBar(self._placements.n_placements + 1,
"getting provenance data")
# retrieve provence data from central
file_path = os.path.join(self._report_default_directory,
"provance_data")
# check the directory doesnt already exist
if not os.path.exists(file_path):
os.mkdir(file_path)
# write provanence data
self.write_provenance_data_in_xml(file_path, self._txrx)
progress.update()
# retrieve provenance data from any cores that provide data
for placement in self._placements.placements:
if isinstance(placement.subvertex,
AbstractProvidesProvenanceData):
core_file_path = os.path.join(
file_path,
"Provanence_data_for_{}_{}_{}_{}.xml".format(
placement.subvertex.label, placement.x,
placement.y, placement.p))
placement.subvertex.write_provenance_data_in_xml(
core_file_path, self.transceiver, placement)
progress.update()
progress.end()
elif isinstance(self._machine, VirtualMachine):
logger.info(
"*** Using a Virtual Machine so no simulation will occur")
else:
logger.info("*** No simulation requested: Stopping. ***")