def tryCreatePartition(numCoresPerAxis, coreShape, postLayerPartition, layer,
logdir):
"""Try creating a partition of the layer.
Fails if proposed partition exceeds one of the Loihi limits.
:param np.ndarray | list | tuple numCoresPerAxis: Number of cores along
each layer dimension.
:param np.ndarray | list | tuple coreShape: The shape of the largest
core.
:param Layer postLayerPartition: The subsequent partitioned layer.
:param KerasLayer | NxConv2D layer: The layer to partition.
:param str logdir: Where to save plots.
:return: Valid partition candidate.
:rtype: Layer
"""
outputShape = layer.output_shape[1:]
# When using signed spikes the number of channels in the output
# is doubled.
if hasattr(layer, 'signed'):
if layer.signed:
outputShape = outputShape[:-1] + (2 * outputShape[-1],)
coreIdMap = getCoreIdMapFromCoreShape(coreShape, outputShape,
numCoresPerAxis)
coreOccupancy = getCoreOccupancy(coreIdMap, numCoresPerAxis)
if np.any(coreOccupancy > layer.maxNumCompartments):
return
multiplicityMap = layer.getMultiplicityMap(coreIdMap)
partitionCandidate = Layer(layer.name, layer.__class__.__name__,
layer.compartmentKwargs, layer.connectionKwargs,
coreIdMap, multiplicityMap, postLayerPartition)
# Pass coreOccupancy to partitionCandidate here only to be able to plot it
# later when the partition has been stored to disk.
partitionCandidate.coreOccupancy = coreOccupancy
partitionCandidate = layer.compile(partitionCandidate)
if partitionCandidate is None:
print('.', end='', flush=True)
return
layer.validatePartition(partitionCandidate)
layer.visualizePartition(logdir, partitionCandidate, coreIdMap,
coreOccupancy, multiplicityMap=multiplicityMap)
print('x', end='', flush=True)
return partitionCandidate
def test_mapCompartments_2(self):
"""Check mapping of CompartmentGroups with soma traces."""
verbose = False
c = CompartmentGroup(cxIds=np.arange(15),
biasMant=np.concatenate(
[range(10, 0, -1),
range(5, 0, -1)]),
biasExp=np.zeros(15, int),
relToAbsDestCxIdxMap=np.array([], int))
layer = Layer(layerId=0,
layerType='',
compartmentKwargs={
"compartmentCurrentDecay": 4096,
"compartmentVoltageDecay": 10,
"refractoryDelay": 30,
"vThMant": 40,
"enableSomaTrace": 1,
"threshOp": 0
},
connectionKwargs={},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=None)
p = Partition(partitionId=0,
chipCounter=0,
sizeInterleaved=-1,
parentLayer=layer)
p.addCompartmentGroup(c)
layer.addPartition(p)
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, 0)[0]
c._mapCompartments(p, core) # pylint: disable=protected-access
if verbose:
c.printCore(core, compartments=True)
self.assertEqual(len(core.somaState.modified), 15)
self.assertEqual(core.somaState[0].vth, 40)
self.assertEqual(core.vthProfileCfg[0].dynamicCfg.enableHomeostasis, 1)
self.assertEqual(core.vthProfileCfg[0].dynamicCfg.aMax, 127)
def getDummyLayer(shape):
"""Create a dummy layer, typically as postLayer of the output layer.
:param list | tuple | np.ndarray shape: Shape of layer.
:return: Dummy layer.
:rtype: Layer
"""
return Layer('DummyPartitionFinalLayer', '', {}, {}, np.array([]),
np.ones(shape, int))
def test_compressConnections(self):
"""Check axon compression."""
inputShapes = np.array([[7, 7, 2], [32, 32, 3],
[32, 32, 3], [27, 26, 2], [32, 30, 2],
[30, 35, 2], [7, 7, 2], [20, 25, 2], [7, 7, 2],
[30, 35, 2], [7, 7, 2], [30, 35, 2]])
layerArgs = np.array([[2, 3], [2, 3], [2, 3], [2, 1], [2, 2], [2, 3],
[2, 3], [2, 3], [2, 3], [2, 3], [2, 3], [2, 3]])
layerKwargs = np.array([
{
'synapseEncoding': 'sparse'
},
{
'synapseEncoding': 'sparse'
},
{
'synapseEncoding': 'sparse',
'strides': 2
},
{
'synapseEncoding': 'sparse',
'strides': 2
},
{
'synapseEncoding': 'sparse',
'strides': 2
},
{
'synapseEncoding': 'sparse',
'strides': 2,
'padding': 'same'
},
{
'synapseEncoding': 'runlength'
},
{
'synapseEncoding': 'runlength',
'strides': 2,
'padding': 'same'
},
{
'synapseEncoding': 'dense1'
},
{
'synapseEncoding': 'dense1',
'strides': 2,
'padding': 'same'
},
{
'synapseEncoding': 'dense2'
},
{
'synapseEncoding': 'dense2',
'strides': 2,
'padding': 'same'
},
])
coreShapes = np.array([[5, 5, 2], [30, 10, 2], [10, 5, 2], [7, 13, 2],
[16, 5, 2], [8, 9, 2], [5, 5, 2], [6, 9, 2],
[5, 5, 2], [15, 10, 2], [5, 5, 2], [6, 9, 2]])
# numOutputAxons (second column) is expected to be zero because we are
# looking at the output layer.
expectedResults = np.array([[14, 0, 900, 281, 378],
[108, 0, 29160, 6873, 10080],
[198, 0, 4050, 2430, 2592],
[52, 0, 56, 42, 28],
[60, 0, 768, 528, 384],
[144, 0, 1408, 808, 868],
[14, 0, 900, 288, 371],
[104, 0, 928, 524, 580],
[14, 0, 900, 660, 427],
[72, 0, 1408, 1226, 840],
[14, 0, 3587, 290, 700],
[216, 0, 2920, 874, 1052]])
doTest = np.array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], bool)
np.random.seed(0)
for inputShape, args, kwargs, coreShape, expectedResult in \
zip(inputShapes[doTest],
layerArgs[doTest],
layerKwargs[doTest],
coreShapes[doTest],
expectedResults[doTest]):
with self.subTest(inputShape=inputShape,
args=args,
kwargs=kwargs,
coreShape=coreShape,
expectedResult=expectedResult):
# The Keras layer initializer does not know how to handle
# np.int types. Convert to builtin int manually.
args = list(args)
args[1] = int(args[1])
inputLayer = NxInputLayer(tuple(inputShape))
layer = NxConv2D(*args, **kwargs)
layer(inputLayer.input)
# Assumes that this layer will never go beyond one chip.
layer.coreCounter = 0
layerShape = layer.output_shape[1:]
numCoresPerAxis = np.ceil(layerShape / coreShape).astype(int)
coreIdMap = getCoreIdMapFromCoreShape(coreShape, layerShape,
numCoresPerAxis)
multiplicityMapPre = np.ones(layer.input_shape[1:-1], int)
postLayerPartition = getDummyLayer(layer.output_shape[1:])
partitionCandidate = Layer(layer.name, '',
layer.compartmentKwargs,
layer.connectionKwargs, coreIdMap,
multiplicityMapPre,
postLayerPartition)
partitionCandidate = layer.compile(partitionCandidate)
try:
layer.validatePartition(partitionCandidate)
finally:
layer.deleteKernelIdMap()
layer.deleteKernelIdMap()
result = [
partitionCandidate.numInputAxons,
partitionCandidate.numOutputAxons,
partitionCandidate.numSyn,
partitionCandidate.numSynEntries,
partitionCandidate.numSynMemWords
]
self.verbose = True
if self.verbose:
print(result)
for er, r in zip(expectedResult, result):
self.assertEqual(er, r)
def test_mapOutputAxons(self):
"""Check mapping of OutputAxonGroups."""
verbose = False
# Create destination layer
lDst = Layer(layerId=0,
layerType='foo',
compartmentKwargs={},
connectionKwargs={},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=getDummyLayer((1, )))
# Create destination partition with InputAxonGroups
pDst = Partition(partitionId=0,
chipCounter=10,
sizeInterleaved=-1,
parentLayer=lDst)
sgDst = SynapseGroup(groupId=0, synEntries=[[]])
iag0 = InputAxonGroup(srcNodeIds=np.array([10, 11, 12]),
multiplicity=np.array([1, 1, 1]),
synGroup=sgDst,
cxBase=1,
parentPartition=pDst)
iag1 = InputAxonGroup(srcNodeIds=np.array([20, 21, 22]),
multiplicity=np.array([1, 2, 1]),
synGroup=sgDst,
cxBase=2,
parentPartition=pDst)
iag2 = InputAxonGroup(srcNodeIds=np.array([30, 31, 32]),
multiplicity=np.array([2, 3, 4]),
synGroup=sgDst,
cxBase=3,
parentPartition=pDst)
pDst.addInputAxonGroup(iag0)
pDst.addInputAxonGroup(iag1)
pDst.addInputAxonGroup(iag2)
lDst.addPartition(pDst)
# Create source layer
lSrc = Layer(layerId=0,
layerType='foo',
compartmentKwargs={},
connectionKwargs={},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=lDst)
# Create source partition with output axons
pSrc = Partition(partitionId=1,
chipCounter=8,
sizeInterleaved=-1,
parentLayer=lSrc)
oag0 = OutputAxonGroup(cxIds=np.array([0, 1, 2]),
multiplicity=np.array([1, 1, 1]),
relSrcIds=np.array([0, 1, 2]),
inAxGrp=iag0,
parentPartition=pSrc)
oag1 = OutputAxonGroup(cxIds=np.array([3, 4, 5]),
multiplicity=np.array([1, 2, 1]),
relSrcIds=np.array([0, 1, 2]),
inAxGrp=iag1,
parentPartition=pSrc)
oag2 = OutputAxonGroup(cxIds=np.array([6, 7, 8]),
multiplicity=np.array([2, 3, 4]),
relSrcIds=np.array([0, 1, 2]),
inAxGrp=iag2,
parentPartition=pSrc)
pSrc.addOutputAxonGroup(oag0)
pSrc.addOutputAxonGroup(oag1)
pSrc.addOutputAxonGroup(oag2)
lSrc.addPartition(pSrc)
# Initialize board and mapper
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, 0)[0]
# Create fake inAxMap as if it was created by mapInputAxons
c._inAxMap[iag0.id] = np.array([[10, 10, 1], [10, 10, 1], [10, 10, 1]],
int)
c._inAxMap[iag1.id] = np.array([[11, 10, 1], [12, 10, 1], [11, 10, 1]],
int)
c._inAxMap[iag2.id] = np.array([[13, 10, 1], [14, 10, 1], [15, 10, 1]],
int)
# First call: nothing will happen.
c._mapOutputAxons(pDst, core) # pylint: disable=protected-access
c._mapOutputAxons(pSrc, core) # pylint: disable=protected-access
# Validate output axons
# OutputAxonGroup 0
self.assertEqual(core.axons[0].srcCxId, 0)
self.assertEqual(core.axons[0].srcRelCxId, 0)
self.assertEqual(core.axons[0].dstChipId, 10)
self.assertEqual(core.axons[0].dstCoreId, 1)
self.assertEqual(core.axons[0].dstSynMapId, 10)
self.assertEqual(core.axons[0].axonType, 2) # pop32
self.assertEqual(core.axons[0].popId, 0)
self.assertEqual(core.axons[1].srcCxId, 1)
self.assertEqual(core.axons[1].srcRelCxId, 1)
self.assertEqual(core.axons[1].dstChipId, 10)
self.assertEqual(core.axons[1].dstCoreId, 1)
self.assertEqual(core.axons[1].dstSynMapId, 10)
self.assertEqual(core.axons[1].axonType, 2) # pop32
self.assertEqual(core.axons[1].popId, 1)
self.assertEqual(core.axons[2].srcCxId, 2)
self.assertEqual(core.axons[2].srcRelCxId, 2)
self.assertEqual(core.axons[2].dstChipId, 10)
self.assertEqual(core.axons[2].dstCoreId, 1)
self.assertEqual(core.axons[2].dstSynMapId, 10)
self.assertEqual(core.axons[2].axonType, 2) # pop32
self.assertEqual(core.axons[2].popId, 2)
# OutputAxonGroup 1
self.assertEqual(core.axons[3].srcCxId, 3)
self.assertEqual(core.axons[3].srcRelCxId, 0)
self.assertEqual(core.axons[3].dstChipId, 10)
self.assertEqual(core.axons[3].dstCoreId, 1)
self.assertEqual(core.axons[3].dstSynMapId, 11)
self.assertEqual(core.axons[3].axonType, 2) # pop32
self.assertEqual(core.axons[3].popId, 3)
self.assertEqual(core.axons[4].srcCxId, 4)
self.assertEqual(core.axons[4].dstChipId, 10)
self.assertEqual(core.axons[4].dstCoreId, 1)
self.assertEqual(core.axons[4].dstSynMapId, 12)
self.assertEqual(core.axons[4].axonType, 0) # discrete
self.assertEqual(core.axons[5].srcCxId, 5)
self.assertEqual(core.axons[5].srcRelCxId, 2)
self.assertEqual(core.axons[5].dstChipId, 10)
self.assertEqual(core.axons[5].dstCoreId, 1)
self.assertEqual(core.axons[5].dstSynMapId, 11)
self.assertEqual(core.axons[5].axonType, 2) # pop32
self.assertEqual(core.axons[5].popId, 4)
# OutputAxonGroup 1
self.assertEqual(core.axons[6].srcCxId, 6)
self.assertEqual(core.axons[6].dstChipId, 10)
self.assertEqual(core.axons[6].dstCoreId, 1)
self.assertEqual(core.axons[6].dstSynMapId, 13)
self.assertEqual(core.axons[6].axonType, 0) # discrete
self.assertEqual(core.axons[7].srcCxId, 7)
self.assertEqual(core.axons[7].dstChipId, 10)
self.assertEqual(core.axons[7].dstCoreId, 1)
self.assertEqual(core.axons[7].dstSynMapId, 14)
self.assertEqual(core.axons[7].axonType, 0) # discrete
self.assertEqual(core.axons[8].srcCxId, 8)
self.assertEqual(core.axons[8].dstChipId, 10)
self.assertEqual(core.axons[8].dstCoreId, 1)
self.assertEqual(core.axons[8].dstSynMapId, 15)
self.assertEqual(core.axons[8].axonType, 0) # discrete
if verbose:
c.printCore(core, outputAxons=True)
def test_mapCompartments_1(self):
"""Check mapping of CompartmentGroups without soma traces."""
verbose = False
c = CompartmentGroup(cxIds=np.arange(15),
biasMant=np.concatenate(
[range(10, 0, -1),
range(5, 0, -1)]),
biasExp=np.zeros(15, int),
relToAbsDestCxIdxMap=np.array([], int))
layer = Layer(layerId=0,
layerType='',
compartmentKwargs={
"compartmentCurrentDecay": 4096,
"compartmentVoltageDecay": 10,
"refractoryDelay": 30,
"vThMant": 40,
"enableSomaTrace": 0,
"threshOp": 0
},
connectionKwargs={},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=None)
p = Partition(partitionId=0,
chipCounter=0,
sizeInterleaved=-1,
parentLayer=layer)
p.addCompartmentGroup(c)
layer.addPartition(p)
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, 0)[0]
c._mapCompartments(p, core) # pylint: disable=protected-access
if verbose:
c.printCore(core, compartments=True)
self.assertEqual(len(core.cxCfg.modified), 15)
self.assertEqual(core.numUpdates.numUpdates, 4)
self.assertEqual(len(core.cxMetaState.modified), 4)
self.assertEqual(core.cxMetaState[3].phase2, 2)
self.assertEqual(core.cxMetaState[3].somaOp2, 0)
self.assertEqual(core.cxMetaState[3].phase3, 2)
self.assertEqual(core.cxMetaState[3].somaOp3, 0)
self.assertEqual(core.cxMetaState[4].phase2, 0)
self.assertEqual(core.cxMetaState[4].somaOp2, 0)
self.assertEqual(core.dendriteSharedCfg[0].dsOffset, 1)
self.assertEqual(core.dendriteSharedCfg[0].dmOffsets, 0)
self.assertEqual(core.cxProfileCfg[0].refractDelay, 30)
self.assertEqual(core.cxProfileCfg[0].decayU, 4095)
self.assertEqual(core.cxProfileCfg[0].decayV, 10)
self.assertEqual(core.vthProfileCfg[0].staticCfg.vth, 40)
def test_mapSynapses_3(self):
"""Check generation of synGrpMap."""
verbose = False
layer = Layer(layerId=0,
layerType='',
compartmentKwargs={},
connectionKwargs={"weightExponent": 2},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=None)
sf = SynFmt(synFmtId=0,
cIdxOffset=0,
cIdxMult=2,
numIdxBits=6,
numSkipBits=0,
numWgtBits=8,
compression=0,
signMode=1)
se0 = SynEntry(prefixOffset=0,
idxs=np.array([0, 1, 2]),
weights=np.array([6, 7, 8]),
synFmt=sf)
se1 = SynEntry(prefixOffset=10,
idxs=np.array([0, 1]),
weights=np.array([16, 17]),
synFmt=sf)
se2 = SynEntry(prefixOffset=20,
idxs=np.array([0, 1, 2]),
weights=np.array([26, 27, 28]),
synFmt=sf)
se3 = SynEntry(prefixOffset=30,
idxs=np.array([0, 1]),
weights=np.array([36, 37]),
synFmt=sf)
sg0 = SynapseGroup(groupId=0, synEntries=[[se0, se1, se2], [se3]])
sg1 = SynapseGroup(groupId=1, synEntries=[[se0], [se1], [se2, se3]])
p = Partition(partitionId=0,
chipCounter=0,
sizeInterleaved=-1,
parentLayer=layer)
p.addSynFmt(sf)
p.addSynapseGroup(sg0)
p.addSynapseGroup(sg1)
layer.addPartition(p)
numSyn = 10 * 2
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, numSyn)[0]
c._mapSynapses(p, core) # pylint: disable=protected-access
if verbose:
print(c._synGrpMap)
self.assertEqual(len(c._synGrpMap), 5)
self.assertEqual(c._synGrpMap[(sg0, 0)], (0, 8))
self.assertEqual(c._synGrpMap[(sg0, 1)], (8, 2))
self.assertEqual(c._synGrpMap[(sg1, 0)], (10, 3))
self.assertEqual(c._synGrpMap[(sg1, 1)], (13, 2))
self.assertEqual(c._synGrpMap[(sg1, 2)], (15, 5))
def test_mapSynapses_2(self):
"""Check switching between synFmt for successive synEntries if
synFmtId does not change."""
verbose = False
layer = Layer(layerId=0,
layerType='',
compartmentKwargs={},
connectionKwargs={"weightExponent": 2},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=None)
sf0 = SynFmt(synFmtId=0,
cIdxOffset=0,
cIdxMult=2,
numIdxBits=6,
numSkipBits=0,
numWgtBits=8,
compression=0,
signMode=1)
sf1 = SynFmt(synFmtId=1,
cIdxOffset=0,
cIdxMult=2,
numIdxBits=7,
numSkipBits=0,
numWgtBits=8,
compression=1,
signMode=1)
se00 = SynEntry(prefixOffset=0,
idxs=np.array([0, 1, 2]),
weights=np.array([6, 7, 8]),
synFmt=sf0)
se01 = SynEntry(prefixOffset=10,
idxs=np.array([0, 1]),
weights=np.array([16, 17]),
synFmt=sf0)
se02 = SynEntry(prefixOffset=20,
idxs=np.array([0, 1, 2]),
weights=np.array([26, 27, 28]),
synFmt=sf0)
se10 = SynEntry(prefixOffset=30,
idxs=np.array([0, 1]),
weights=np.array([36, 37]),
synFmt=sf1)
sg = SynapseGroup(groupId=0, synEntries=[[se00, se01, se02], [se10]])
p = Partition(partitionId=0,
chipCounter=0,
sizeInterleaved=-1,
parentLayer=layer)
p.addSynFmt(sf0)
p.addSynFmt(sf1)
p.addSynapseGroup(sg)
layer.addPartition(p)
numSyn = 10
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, numSyn)[0]
c._mapSynapses(p, core) # pylint: disable=protected-access
if verbose:
c.printCore(core, synapses=True)
cIdx = [0, 1, 2, 10, 11, 20, 21, 22, 30, 31]
wgts = [6, 7, 8, 16, 17, 26, 27, 28, 36, 37]
synFmtIds = [1, 1, 1, 2, 2, 1, 1, 1, 3, 3]
for i in range(numSyn):
self.assertEqual(core.synapses[i].CIdx, cIdx[i])
self.assertEqual(core.synapses[i].Wgt, wgts[i])
self.assertEqual(core.synapses[i].synFmtId, synFmtIds[i])
def test_mapSynapses_1(self):
"""Check allocation of synFmts."""
verbose = False
layer = Layer(layerId=0,
layerType='',
compartmentKwargs={},
connectionKwargs={"weightExponent": 2},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=None)
sf0 = SynFmt(synFmtId=0,
cIdxOffset=0,
cIdxMult=2,
numIdxBits=6,
numSkipBits=0,
numWgtBits=8,
compression=0,
signMode=2)
sf1 = SynFmt(synFmtId=1,
cIdxOffset=0,
cIdxMult=2,
numIdxBits=6,
numSkipBits=0,
numWgtBits=6,
compression=0,
signMode=3)
p = Partition(partitionId=0,
chipCounter=0,
sizeInterleaved=-1,
parentLayer=layer)
p.addSynFmt(sf0)
p.addSynFmt(sf1)
layer.addPartition(p)
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, 0)[0]
c._mapSynapses(p, core) # pylint: disable=protected-access
if verbose:
c.printCore(core, synFmts=True)
# Validate number of synFmts
self.assertEqual(len(core.synapseFmt.modified), 4)
# Validate duplication of synFmts by checking unique fanoutTypes
self.assertEqual(core.synapseFmt[1].fanoutType, 2)
self.assertEqual(core.synapseFmt[2].fanoutType, 2)
self.assertEqual(core.synapseFmt[3].fanoutType, 3)
self.assertEqual(core.synapseFmt[4].fanoutType, 3)
# Validate mapping of numWgtBits -> wgtBits
self.assertEqual(core.synapseFmt[1].wgtBits, 7)
self.assertEqual(core.synapseFmt[3].wgtBits, 6)
# Validate wgtExp
self.assertEqual(core.synapseFmt[1].wgtExp, 2)
def test_mapInputAxons_1(self):
"""Check mapping of InputAxonGroup with all input nodes having
multiplicity 1."""
verbose = False
layer = Layer(layerId=0,
layerType='',
compartmentKwargs={},
connectionKwargs={"weightExponent": 2},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=None)
sf = SynFmt(synFmtId=0,
cIdxOffset=0,
cIdxMult=2,
numIdxBits=6,
numSkipBits=0,
numWgtBits=8,
compression=0,
signMode=1)
se0 = SynEntry(prefixOffset=0,
idxs=np.array([0, 1, 2]),
weights=np.array([6, 7, 8]),
synFmt=sf)
se1 = SynEntry(prefixOffset=10,
idxs=np.array([0, 1]),
weights=np.array([16, 17]),
synFmt=sf)
se2 = SynEntry(prefixOffset=20,
idxs=np.array([0, 1, 2]),
weights=np.array([26, 27, 28]),
synFmt=sf)
se3 = SynEntry(prefixOffset=30,
idxs=np.array([0, 1]),
weights=np.array([36, 37]),
synFmt=sf)
sg = SynapseGroup(groupId=0,
synEntries=[[se0, se1], [se1, se2, se3], [se3]])
p = Partition(partitionId=0,
chipCounter=0,
sizeInterleaved=-1,
parentLayer=layer)
p.addSynFmt(sf)
p.addSynapseGroup(sg)
iag = InputAxonGroup(srcNodeIds=np.array([10, 11, 12]),
multiplicity=np.array([1, 1, 1]),
synGroup=sg,
cxBase=2,
parentPartition=p)
p.addInputAxonGroup(iag)
layer.addPartition(p)
numSyn = 14
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, numSyn)[0]
c._mapSynapses(p, core) # pylint: disable=protected-access
c._mapInputAxons(p, core) # pylint: disable=protected-access
# Validate input axon
sMap = core.synapseMap[0]
self.assertEqual(sMap.synapsePtr, 0)
self.assertEqual(sMap.synapseLen, [5, 7, 2])
self.assertEqual(sMap.popSize, 3)
self.assertEqual(sMap.population32MapEntry.ptr, 0)
self.assertEqual(sMap.population32MapEntry.length, 3)
self.assertEqual(sMap.population32MapEntry.cxBase, 2)
# Validate inAxMap
self.assertTrue(
np.array_equal(c._inAxMap[iag.id],
np.array([[0, 0, 4], [0, 0, 4], [0, 0, 4]], int)))
if verbose:
c.printCore(core, inputAxons=True)
def test_mapInputAxons_4(self):
"""Check mapping of multiple InputAxonGroups."""
verbose = False
layer = Layer(layerId=0,
layerType='',
compartmentKwargs={},
connectionKwargs={"weightExponent": 2},
coreIdMap=np.array([]),
multiplicityMap=np.array([]),
postLayer=None)
sf = SynFmt(synFmtId=0,
cIdxOffset=0,
cIdxMult=2,
numIdxBits=6,
numSkipBits=0,
numWgtBits=8,
compression=0,
signMode=1)
se0 = SynEntry(prefixOffset=0,
idxs=np.array([0, 1, 2]),
weights=np.array([6, 7, 8]),
synFmt=sf)
se1 = SynEntry(prefixOffset=10,
idxs=np.array([0, 1]),
weights=np.array([16, 17]),
synFmt=sf)
se2 = SynEntry(prefixOffset=20,
idxs=np.array([0, 1, 2]),
weights=np.array([26, 27, 28]),
synFmt=sf)
se3 = SynEntry(prefixOffset=30,
idxs=np.array([0, 1]),
weights=np.array([36, 37]),
synFmt=sf)
sg0 = SynapseGroup(groupId=0,
synEntries=[[se0, se1], [se1, se2, se3], [se3]])
sg1 = SynapseGroup(groupId=1,
synEntries=[[se0, se1], [se2], [se1, se2, se3]])
p = Partition(partitionId=0,
chipCounter=0,
sizeInterleaved=-1,
parentLayer=layer)
p.addSynFmt(sf)
p.addSynapseGroup(sg0)
p.addSynapseGroup(sg1)
iag0 = InputAxonGroup(srcNodeIds=np.array([10, 11, 12]),
multiplicity=np.array([1, 1, 1]),
synGroup=sg0,
cxBase=1,
parentPartition=p)
iag1 = InputAxonGroup(srcNodeIds=np.array([10, 11, 12]),
multiplicity=np.array([1, 2, 1]),
synGroup=sg0,
cxBase=2,
parentPartition=p)
iag2 = InputAxonGroup(srcNodeIds=np.array([10, 11, 12]),
multiplicity=np.array([1, 2, 3]),
synGroup=sg1,
cxBase=3,
parentPartition=p)
p.addInputAxonGroup(iag0)
p.addInputAxonGroup(iag1)
p.addInputAxonGroup(iag2)
layer.addPartition(p)
board = N2Board(0)
c = DnnMapper(board)
core = board.allocateCores(1, p.numSyn)[0]
c._mapSynapses(p, core) # pylint: disable=protected-access
c._mapInputAxons(p, core) # pylint: disable=protected-access
# Validate axons of iag0
sMapShared = core.synapseMap[0]
self.assertEqual(sMapShared.synapsePtr, 0)
self.assertEqual(sMapShared.synapseLen, [5, 7, 2])
self.assertEqual(sMapShared.popSize, 3)
self.assertEqual(sMapShared.population32MapEntry.ptr, 0)
self.assertEqual(sMapShared.population32MapEntry.length, 3)
self.assertEqual(sMapShared.population32MapEntry.cxBase, 1)
# Validate axons of iag1
sMapShared = core.synapseMap[1]
self.assertEqual(sMapShared.synapsePtr, 0)
self.assertEqual(sMapShared.synapseLen, [5, 7, 2])
self.assertEqual(sMapShared.popSize, 3)
self.assertEqual(sMapShared.population32MapEntry.ptr, 0)
self.assertEqual(sMapShared.population32MapEntry.length, 3)
self.assertEqual(sMapShared.population32MapEntry.cxBase, 2)
sMapDiscrete = core.synapseMap[2]
self.assertEqual(sMapDiscrete.synapsePtr, 5)
self.assertEqual(sMapDiscrete.synapseLen, 7)
self.assertEqual(sMapDiscrete.discreteMapEntry.ptr, 3)
self.assertEqual(sMapDiscrete.discreteMapEntry.length, 3)
self.assertEqual(sMapDiscrete.discreteMapEntry.cxBase, 2)
# Validate axons of iag2
sMapShared = core.synapseMap[3]
self.assertEqual(sMapShared.synapsePtr, 14)
self.assertEqual(sMapShared.synapseLen, [5, 3, 7])
self.assertEqual(sMapShared.popSize, 3)
self.assertEqual(sMapShared.population32MapEntry.ptr, 9)
self.assertEqual(sMapShared.population32MapEntry.length, 3)
self.assertEqual(sMapShared.population32MapEntry.cxBase, 3)
sMapDiscrete = core.synapseMap[4]
self.assertEqual(sMapDiscrete.synapsePtr, 19)
self.assertEqual(sMapDiscrete.synapseLen, 3)
self.assertEqual(sMapDiscrete.discreteMapEntry.ptr, 12)
self.assertEqual(sMapDiscrete.discreteMapEntry.length, 1)
self.assertEqual(sMapDiscrete.discreteMapEntry.cxBase, 3)
sMapDiscrete = core.synapseMap[5]
self.assertEqual(sMapDiscrete.synapsePtr, 22)
self.assertEqual(sMapDiscrete.synapseLen, 7)
self.assertEqual(sMapDiscrete.discreteMapEntry.ptr, 15)
self.assertEqual(sMapDiscrete.discreteMapEntry.length, 3)
self.assertEqual(sMapDiscrete.discreteMapEntry.cxBase, 3)
# Validate inAxMap
self.assertTrue(
np.array_equal(c._inAxMap[iag0.id],
np.array([[0, 0, 4], [0, 0, 4], [0, 0, 4]], int)))
self.assertTrue(
np.array_equal(c._inAxMap[iag1.id],
np.array([[1, 0, 4], [2, 0, 4], [1, 0, 4]], int)))
self.assertTrue(
np.array_equal(c._inAxMap[iag2.id],
np.array([[3, 0, 4], [4, 0, 4], [5, 0, 4]], int)))
if verbose:
c.printCore(core, inputAxons=True)