def __get_synapse_constant_sdram(self, vertex_slice):
""" Get the amount of fixed SDRAM used by synapse parts
:param ~pacman.model.graphs.common.Slice vertex_slice:
The slice of neurons to get the size of
:rtype: ~pacman.model.resources.MultiRegionSDRAM
"""
sdram = MultiRegionSDRAM()
app_vertex = self._governed_app_vertex
sdram.add_cost(
PopulationMachineVertex.SYNAPSE_REGIONS.synapse_params,
app_vertex.get_synapse_params_size())
sdram.add_cost(
PopulationMachineVertex.SYNAPSE_REGIONS.synapse_dynamics,
app_vertex.get_synapse_dynamics_size(vertex_slice))
sdram.add_cost(
PopulationMachineVertex.SYNAPSE_REGIONS.structural_dynamics,
self.__structural_size(vertex_slice))
sdram.add_cost(
PopulationMachineVertex.SYNAPSE_REGIONS.synaptic_matrix,
self.__all_syn_block_size(vertex_slice))
sdram.add_cost(
PopulationMachineVertex.SYNAPSE_REGIONS.direct_matrix,
app_vertex.all_single_syn_size)
sdram.add_cost(
PopulationMachineVertex.SYNAPSE_REGIONS.pop_table,
MasterPopTableAsBinarySearch.get_master_population_table_size(
app_vertex.incoming_projections))
sdram.add_cost(
PopulationMachineVertex.SYNAPSE_REGIONS.connection_builder,
self.__synapse_expander_size())
sdram.merge(self.__bitfield_size())
return sdram
def __get_neuron_resources(self, vertex_slice):
""" Gets the resources of the neurons of a slice of atoms from a given
app vertex.
:param ~pacman.model.graphs.common.Slice vertex_slice: the slice
:rtype: ~pacman.model.resources.ResourceContainer
"""
n_record = len(self._governed_app_vertex.neuron_recordables)
variable_sdram = self._governed_app_vertex.get_neuron_variable_sdram(
vertex_slice)
sdram = MultiRegionSDRAM()
sdram.merge(self._governed_app_vertex.get_common_constant_sdram(
n_record, NeuronProvenance.N_ITEMS + NeuronMainProvenance.N_ITEMS,
PopulationNeuronsMachineVertex.COMMON_REGIONS))
sdram.merge(self._governed_app_vertex.get_neuron_constant_sdram(
vertex_slice, PopulationNeuronsMachineVertex.NEURON_REGIONS))
sdram.add_cost(
PopulationNeuronsMachineVertex.REGIONS.SDRAM_EDGE_PARAMS.value,
NEURONS_SDRAM_PARAMS_SIZE)
sdram.nest(
len(PopulationNeuronsMachineVertex.REGIONS) + 1, variable_sdram)
dtcm = self._governed_app_vertex.get_common_dtcm()
dtcm += self._governed_app_vertex.get_neuron_dtcm(vertex_slice)
cpu_cycles = self._governed_app_vertex.get_common_cpu()
cpu_cycles += self._governed_app_vertex.get_neuron_cpu(vertex_slice)
# set resources required from this object
container = ResourceContainer(
sdram=sdram, dtcm=DTCMResource(dtcm),
cpu_cycles=CPUCyclesPerTickResource(cpu_cycles))
# return the total resources.
return container
def test_sdram(self):
"""
test that adding a SDRAM resource to a resource container works
correctly
"""
const1 = ConstantSDRAM(128)
self.assertEqual(const1.get_total_sdram(None), 128)
const2 = ConstantSDRAM(256)
combo = const1 + const2
self.assertEqual(combo.get_total_sdram(None), 128 + 256)
combo = const1 - const2
self.assertEqual(combo.get_total_sdram(None), 128 - 256)
combo = const2 + const1
self.assertEqual(combo.get_total_sdram(None), 256 + 128)
combo = const2 - const1
self.assertEqual(combo.get_total_sdram(None), 256 - 128)
var1 = VariableSDRAM(124, 8)
self.assertEqual(var1.get_total_sdram(100), 124 + 8 * 100)
combo = var1 + const1
self.assertEqual(combo.get_total_sdram(100), 124 + 8 * 100 + 128)
combo = var1 - const1
self.assertEqual(combo.get_total_sdram(100), 124 + 8 * 100 - 128)
combo = const1 + var1
self.assertEqual(combo.get_total_sdram(100), 128 + 124 + 8 * 100)
combo = const1 - var1
self.assertEqual(combo.get_total_sdram(100), 128 - (124 + 8 * 100))
var2 = VariableSDRAM(234, 6)
combo = var2 + var1
self.assertEqual(combo.get_total_sdram(150), 234 + 124 + (6 + 8) * 150)
combo = var2 - var1
self.assertEqual(combo.get_total_sdram(150), 234 - 124 + (6 - 8) * 150)
multi1 = MultiRegionSDRAM()
multi1.add_cost(1, 100, 4)
multi1.add_cost(2, 50, 3)
multi1.add_cost("overheads", 20)
multi2 = MultiRegionSDRAM()
multi2.add_cost(MockEnum.ZERO, 88)
multi2.add_cost(MockEnum.ONE, 72)
multi2.add_cost("overheads", 22)
combo = multi1 + multi2
self.assertEqual(combo.get_total_sdram(150),
100 + 50 + 20 + 88 + 72 + 22 + (4 + 3) * 150)
multi3 = MultiRegionSDRAM()
multi3.nest("foo", multi1)
multi3.nest("bar", multi2)
multi1.merge(multi2)
self.assertEqual(len(multi1.regions), 5)
self.assertEqual(multi1.regions["overheads"], ConstantSDRAM(20 + 22))
self.assertEqual(multi1.get_total_sdram(150),
100 + 50 + 20 + 88 + 72 + 22 + (4 + 3) * 150)
self.assertEqual(multi1, combo)
self.assertEqual(multi1, multi3)
with tempfile.TemporaryFile(mode="w") as target:
multi3.report(1000, target=target)
def __get_synapse_resources(self, vertex_slice, shared_sdram=None):
""" Get the resources of the synapses of a slice of atoms from a
given app vertex.
:param ~pacman.model.graphs.common.Slice vertex_slice: the slice
:param ~pacman.model.resources.MultiRegionSDRAM shared_sdram:
The SDRAM shared between cores, if this is to be included
:rtype: ~pacman.model.resources.ResourceContainer
"""
n_record = len(self._governed_app_vertex.synapse_recordables)
variable_sdram = self._governed_app_vertex.get_synapse_variable_sdram(
vertex_slice)
sdram = MultiRegionSDRAM()
sdram.merge(
self._governed_app_vertex.get_common_constant_sdram(
n_record, SynapseProvenance.N_ITEMS +
SpikeProcessingFastProvenance.N_ITEMS,
PopulationSynapsesMachineVertexLead.COMMON_REGIONS))
sdram.add_cost(
PopulationSynapsesMachineVertexLead.REGIONS.SDRAM_EDGE_PARAMS.
value, SYNAPSES_SDRAM_PARAMS_SIZE)
sdram.add_cost(
PopulationSynapsesMachineVertexLead.REGIONS.KEY_REGION.value,
KEY_CONFIG_SIZE)
sdram.nest(
len(PopulationSynapsesMachineVertexLead.REGIONS) + 1,
variable_sdram)
if shared_sdram is not None:
sdram.merge(shared_sdram)
dtcm = self._governed_app_vertex.get_common_dtcm()
dtcm += self._governed_app_vertex.get_synapse_dtcm(vertex_slice)
cpu_cycles = self._governed_app_vertex.get_common_cpu()
cpu_cycles += self._governed_app_vertex.get_synapse_cpu(vertex_slice)
# set resources required from this object
container = ResourceContainer(
sdram=sdram,
dtcm=DTCMResource(dtcm),
cpu_cycles=CPUCyclesPerTickResource(cpu_cycles))
# return the total resources.
return container
def get_resources_used_by_atoms(self, vertex_slice):
""" Gets the resources of a slice of atoms
:param ~pacman.model.graphs.common.Slice vertex_slice: the slice
:rtype: ~pacman.model.resources.ResourceContainer
"""
# pylint: disable=arguments-differ
variable_sdram = self.__get_variable_sdram(vertex_slice)
constant_sdram = self.__get_constant_sdram(vertex_slice)
sdram = MultiRegionSDRAM()
sdram.nest(len(PopulationMachineVertex.REGIONS) + 1, variable_sdram)
sdram.merge(constant_sdram)
# set resources required from this object
container = ResourceContainer(
sdram=sdram, dtcm=self.__get_dtcm_cost(vertex_slice),
cpu_cycles=self.__get_cpu_cost(vertex_slice))
# return the total resources.
return container
def __shared_synapse_sdram(self, independent_synapse_sdram,
proj_dependent_sdram, all_syn_block_sz,
structural_sz, dynamics_sz):
""" Get the SDRAM shared between synapse cores
:rtype: ~pacman.model.resources.MultiRegionSDRAM
"""
sdram = MultiRegionSDRAM()
sdram.merge(independent_synapse_sdram)
sdram.merge(proj_dependent_sdram)
sdram.add_cost(
PopulationSynapsesMachineVertexLead.SYNAPSE_REGIONS.
synaptic_matrix, all_syn_block_sz)
sdram.add_cost(
PopulationSynapsesMachineVertexLead.SYNAPSE_REGIONS.direct_matrix,
max(self._governed_app_vertex.all_single_syn_size, BYTES_PER_WORD))
sdram.add_cost(
PopulationSynapsesMachineVertexLead.SYNAPSE_REGIONS.
structural_dynamics, structural_sz)
sdram.add_cost(
PopulationSynapsesMachineVertexLead.SYNAPSE_REGIONS.
synapse_dynamics, dynamics_sz)
return sdram
def __get_constant_sdram(self, vertex_slice):
""" returns the constant sdram used by the vertex slice.
:param ~pacman.model.graphs.common.Slice vertex_slice:
the atoms to get constant sdram of
:rtype: ~pacman.model.resources.MultiRegionSDRAM
"""
n_record = (
len(self._governed_app_vertex.neuron_recordables) +
len(self._governed_app_vertex.synapse_recordables))
n_provenance = (
NeuronProvenance.N_ITEMS + SynapseProvenance.N_ITEMS +
MainProvenance.N_ITEMS + SpikeProcessingProvenance.N_ITEMS)
sdram = MultiRegionSDRAM()
sdram.merge(self._governed_app_vertex.get_common_constant_sdram(
n_record, n_provenance, PopulationMachineVertex.COMMON_REGIONS))
sdram.merge(self._governed_app_vertex.get_neuron_constant_sdram(
vertex_slice, PopulationMachineVertex.NEURON_REGIONS))
sdram.merge(self.__get_synapse_constant_sdram(vertex_slice))
return sdram
def create_machine_vertices(self, resource_tracker, machine_graph):
app_vertex = self._governed_app_vertex
label = app_vertex.label
constraints = get_remaining_constraints(app_vertex)
# Structural plasticity can only be run on a single synapse core
if (isinstance(app_vertex.synapse_dynamics,
AbstractSynapseDynamicsStructural)
and self.__n_synapse_vertices != 1):
raise SynapticConfigurationException(
"The current implementation of structural plasticity can only"
" be run on a single synapse core. Please ensure the number"
" of synapse cores is set to 1")
# Do some checks to make sure everything is likely to fit
atoms_per_core = min(app_vertex.get_max_atoms_per_core(),
app_vertex.n_atoms)
n_synapse_types = app_vertex.neuron_impl.get_n_synapse_types()
if (get_n_bits(atoms_per_core) + get_n_bits(n_synapse_types) +
get_n_bits(self.__get_max_delay)) > MAX_RING_BUFFER_BITS:
raise SynapticConfigurationException(
"The combination of the number of neurons per core ({}), "
"the number of synapse types ({}), and the maximum delay per "
"core ({}) will require too much DTCM. Please reduce one or "
"more of these values.".format(atoms_per_core, n_synapse_types,
self.__get_max_delay))
self.__neuron_vertices = list()
self.__synapse_vertices = list()
self.__synapse_verts_by_neuron = defaultdict(list)
incoming_direct_poisson = self.__handle_poisson_sources(
label, machine_graph)
# Work out the ring buffer shifts based on all incoming things
rb_shifts = app_vertex.get_ring_buffer_shifts(
app_vertex.incoming_projections)
weight_scales = app_vertex.get_weight_scales(rb_shifts)
# Get resources for synapses
independent_synapse_sdram = self.__independent_synapse_sdram()
proj_dependent_sdram = self.__proj_dependent_synapse_sdram(
app_vertex.incoming_projections)
for index, vertex_slice in enumerate(self.__get_fixed_slices()):
# Find the maximum number of cores on any chip available
max_crs = resource_tracker.get_maximum_cores_available_on_a_chip()
if max_crs < (self.__n_synapse_vertices + 1):
raise ConfigurationException(
"No chips remaining with enough cores for"
f" {self.__n_synapse_vertices} synapse cores and a neuron"
" core")
max_crs -= self.__n_synapse_vertices + 1
# Create the neuron vertex for the slice
neuron_vertex, neuron_resources = self.__add_neuron_core(
vertex_slice, label, index, rb_shifts, weight_scales,
machine_graph, constraints)
# Keep track of synapse vertices for each neuron vertex and
# resources used by each core (neuron core is added later)
synapse_vertices = list()
self.__synapse_verts_by_neuron[neuron_vertex] = synapse_vertices
all_resources = []
# Add the first vertex
synapse_references, syn_label = self.__add_lead_synapse_core(
vertex_slice, independent_synapse_sdram, proj_dependent_sdram,
label, rb_shifts, weight_scales, all_resources, machine_graph,
synapse_vertices, neuron_vertex, constraints)
# Do the remaining synapse cores
for i in range(1, self.__n_synapse_vertices):
self.__add_shared_synapse_core(syn_label, i, vertex_slice,
synapse_references,
all_resources, machine_graph,
synapse_vertices, neuron_vertex,
constraints)
# Add resources for Poisson vertices up to core limit
poisson_vertices = incoming_direct_poisson[vertex_slice]
remaining_poisson_vertices = list()
added_poisson_vertices = list()
for poisson_vertex, poisson_edge in poisson_vertices:
if max_crs <= 0:
remaining_poisson_vertices.append(poisson_vertex)
self.__add_poisson_multicast(poisson_vertex,
synapse_vertices,
machine_graph, poisson_edge)
else:
all_resources.append(
(poisson_vertex.resources_required, []))
added_poisson_vertices.append(poisson_vertex)
max_crs -= 1
if remaining_poisson_vertices:
logger.warn(
f"Vertex {label} is using multicast for"
f" {len(remaining_poisson_vertices)} one-to-one Poisson"
" sources as not enough cores exist to put them on the"
" same chip")
# Create an SDRAM edge partition
sdram_label = "SDRAM {} Synapses-->Neurons:{}-{}".format(
label, vertex_slice.lo_atom, vertex_slice.hi_atom)
source_vertices = added_poisson_vertices + synapse_vertices
sdram_partition = SourceSegmentedSDRAMMachinePartition(
SYNAPSE_SDRAM_PARTITION_ID, sdram_label, source_vertices)
machine_graph.add_outgoing_edge_partition(sdram_partition)
neuron_vertex.set_sdram_partition(sdram_partition)
# Add SDRAM edges for synapse vertices
for source_vertex in source_vertices:
edge_label = "SDRAM {}-->{}".format(source_vertex.label,
neuron_vertex.label)
machine_graph.add_edge(
SDRAMMachineEdge(source_vertex, neuron_vertex, edge_label),
SYNAPSE_SDRAM_PARTITION_ID)
source_vertex.set_sdram_partition(sdram_partition)
# Add SDRAM edge requirements to the neuron SDRAM, as the resource
# tracker will otherwise try to add another core for it
extra_sdram = MultiRegionSDRAM()
extra_sdram.merge(neuron_resources.sdram)
extra_sdram.add_cost(
len(extra_sdram.regions) + 1,
sdram_partition.total_sdram_requirements())
neuron_resources_plus = ResourceContainer(
sdram=extra_sdram,
dtcm=neuron_resources.dtcm,
cpu_cycles=neuron_resources.cpu_cycles,
iptags=neuron_resources.iptags,
reverse_iptags=neuron_resources.reverse_iptags)
all_resources.append((neuron_resources_plus, constraints))
# Allocate all the resources to ensure they all fit
resource_tracker.allocate_constrained_group_resources(
all_resources)
return True