def test_eyeriss_isscc16(self):
'''
Reproduce Eyeriss ISSCC'16 paper Fig. 14.5.6, JSSC'17 paper Table V.
'''
network = self.alex_net
batch_size = 4
resource = Resource(
proc_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.PROC),
dram_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DRAM),
src_data_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DRAM),
dst_data_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DRAM),
dim_array=PhyDim2(12, 14),
size_gbuf=108 * 1024 // 2, # 108 kB
size_regf=261, # 225 + 12 + 24
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False,
)
cost = Cost(
mac_op=2e-12,
mem_hier=(460e-12, 15e-12, 4e-12, 1e-12), # pJ/16-b
noc_hop=0,
idl_unit=30e-3 / 200e6) # 30 mW GBUF + REGF
# nnd = NNDataflow(network, batch_size, resource, cost,
# self.map_strategy)
# tops, _ = nnd.schedule_search(self.options)
# self.assertTrue(tops)
# dfsch = tops[0]
# Use tool
from nn_dataflow.tools import nn_dataflow_search as nnds
cmd = '--batch 4 --nodes 1 1 --array 12 14 --regf 522 --gbuf 110592 --mem-type 3D --disable-bypass i o f --hop-cost 0 --op-cost 2e-12 --hier-cost 460e-12 15e-12 4e-12 1e-12 --unit-idle-cost 1.5e-10 alex_net'
args = nnds.argparser().parse_args(cmd.split())
res = nnds.do_scheduling(args)
dfsch = res['schedules']
## Check results.
# Results as stats of the rows in the table.
header = 'Power, Processing Latency, Ops, Active PEs, Filter size'
stats = {}
for layer in ['conv{}'.format(i) for i in range(1, 6)]:
onchip_cost = 0
time = 0
ops = 0
fil_size = 0
for layer_part in network:
if not layer_part or not layer_part.startswith(layer):
continue
sr = dfsch[layer_part]
onchip_cost += sr.total_cost \
- sr.total_accesses[me.DRAM] * cost.mem_hier[me.DRAM]
time += sr.total_time
ops += sr.total_ops
fil_size += network[layer_part].total_filter_size()
power = onchip_cost / (time / 200e6) * 1e3 # mW
active_pes = int(ops / time)
stats[layer] = []
stats[layer].append(power)
stats[layer].append(time / 200.e3) # cycles to ms
stats[layer].append(ops / 1e6) # to MOPs
stats[layer].append(active_pes)
stats[layer].append(fil_size / 1e3) # to k
# Check.
stats_ref = {
'conv1': [332, 16.5, 421.66, 151, 34.8], # Act PE 154
'conv2': [288, 39.2, 895.79, 135, 307.2],
'conv3': [266, 21.8, 598.1, 156, 884.7],
'conv4': [235, 16.0, 448.6, 156, 663.6],
'conv5': [236, 10.0, 299.0, 156, 442.4],
}
for layer in stats:
success = (0.6 * stats_ref[layer][0]
< stats[layer][0]
< stats_ref[layer][0]) \
and (0.8 * stats_ref[layer][1]
< stats[layer][1]
< stats_ref[layer][1]) \
and all(abs(a - b) < 0.1 for a, b
in zip(stats[layer][2:], stats_ref[layer][2:]))
self.assertTrue(
success, 'test_eyeriss_isscc16: '
'stats diff in layer {}.\n'
'header: {}\n'
'actual: {}\nref: {}'.format(layer, header, stats[layer],
stats_ref[layer]))
def test_invalid_origin(self):
''' Invalid origin. '''
with self.assertRaisesRegexp(TypeError, 'NodeRegion: .*origin.*'):
_ = NodeRegion(dim=PhyDim2(4, 4),
origin=(1, 3),
type=NodeRegion.PROC)
def setUp(self):
self.network = Network('test_net')
self.network.set_input_layer(InputLayer(3, 224))
self.network.add('c1', ConvLayer(3, 64, 224, 3))
self.network.add('p1', PoolingLayer(64, 7, 32), prevs='c1')
self.network.add('p2', PoolingLayer(64, 7, 32), prevs='c1')
self.network.add('f1', FCLayer(128, 1000, 7), prevs=['p1', 'p2'])
self.batch_size = 4
input_layer = self.network.input_layer()
self.input_layout = DataLayout(
frngs=(FmapRange((0, 0, 0, 0),
FmapPosition(b=self.batch_size,
n=input_layer.nofm,
h=input_layer.hofm,
w=input_layer.wofm)), ),
regions=(NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(2, 1),
type=NodeRegion.DRAM), ),
parts=(PartitionScheme(order=range(pe.NUM),
pdims=[(1, 1)] * pe.NUM), ))
c1_layer = self.network['c1']
self.c1res = SchedulingResult(
scheme=OrderedDict([
('cost', 1.5),
('time', 200.),
('ops', 4.),
('num_nodes', 4),
('cost_op', 0.5),
('cost_access', 1.),
('cost_noc', 0),
('cost_static', 0),
('proc_time', 200),
('bus_time', 0),
('dram_time', 0),
('access', [[7, 8, 9]] * me.NUM),
('remote_gbuf_access', [0] * 3),
('total_nhops', [4, 5, 6]),
('fetch', [[1, 1, 1], [2, 2, 2]]),
('ti', [2, 2, 3]),
('to', [1, 2, 3]),
('tb', [1, 2, 3]),
('tvals', [[2, 1, 1], [2, 2, 2], [3, 3, 3]]),
('orders', [range(3)] * 2),
]),
ofmap_layout=DataLayout(
frngs=(FmapRange(
(0, 0, 0, 0),
FmapPosition(b=self.batch_size,
n=c1_layer.nofm,
h=c1_layer.hofm,
w=c1_layer.wofm)), ),
regions=(NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 2),
type=NodeRegion.DRAM), ),
parts=(PartitionScheme(order=range(pe.NUM),
pdims=[(1, 1)] * pe.NUM), )),
sched_seq=(0, 0, 0))
p1_layer = self.network['p1']
self.p1res = SchedulingResult(
scheme=OrderedDict([
('cost', 0.6),
('time', 5),
('ops', 0.1),
('num_nodes', 2),
('cost_op', 0.1),
('cost_access', 0.5),
('cost_noc', 0),
('cost_static', 0),
('proc_time', 5),
('bus_time', 0),
('dram_time', 0),
('access', [[.7, .8, .9]] * me.NUM),
('remote_gbuf_access', [0] * 3),
('total_nhops', [.4, .5, .6]),
('fetch', [[1, 1, 1], [2, 2, 2]]),
('ti', [2, 2, 3]),
('to', [1, 2, 3]),
('tb', [1, 2, 3]),
('tvals', [[2, 1, 1], [2, 2, 2], [3, 3, 3]]),
('orders', [range(3)] * 2),
]),
ofmap_layout=DataLayout(
frngs=(FmapRange(
(0, 0, 0, 0),
FmapPosition(b=self.batch_size,
n=p1_layer.nofm,
h=p1_layer.hofm,
w=p1_layer.wofm)), ),
regions=(NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 2),
type=NodeRegion.DRAM), ),
parts=(PartitionScheme(order=range(pe.NUM),
pdims=[(1, 1)] * pe.NUM), )),
sched_seq=(0, 1, 0))
self.p2res = SchedulingResult(scheme=self.p1res.scheme,
ofmap_layout=self.p1res.ofmap_layout,
sched_seq=(0, 2, 0))
self.dtfl = NNDataflowScheme(self.network, self.input_layout)
self.dtfl['c1'] = self.c1res
self.dtfl['p1'] = self.p1res
self.dtfl['p2'] = self.p2res
def test_use_fwd(self):
''' Use access forwarding. '''
layer = self.layers['BASE']
part = PartitionScheme(order=(pe.BATP, pe.INPP, pe.OUTP, pe.OFMP),
pdims=((2, 1), (2, 4), (1, 2), (2, 1)))
nr = NodeRegion(origin=PhyDim2(0, 0),
dim=part.dim(),
type=NodeRegion.PROC)
far_dist = 1000
ilayout = self._make_data_layout(layer.nifm, layer.hifm, layer.wifm,
PhyDim2(-far_dist, 0), (1, 1), (1, 1),
PhyDim2(1, 1))
olayout = self._make_data_layout(layer.nofm, layer.hofm, layer.wofm,
PhyDim2(0, -far_dist), (1, 1), (1, 1),
PhyDim2(1, 1))
filter_nodes = frozenset([PhyDim2(far_dist, 0), PhyDim2(0, far_dist)])
nhops_base = partition.unit_nhops_to_proc_region(
layer, self.batch_size, nr, part, filter_nodes, ilayout, olayout,
self.options['BASE'])
nhops_accfwd = partition.unit_nhops_to_proc_region(
layer, self.batch_size, nr, part, filter_nodes, ilayout, olayout,
self.options['ACCFWD'])
nhops_bufshr = partition.unit_nhops_to_proc_region(
layer, self.batch_size, nr, part, filter_nodes, ilayout, olayout,
self.options['BUFSHR'])
for dce in range(de.NUM):
self.assertEqual(nhops_accfwd[dce], nhops_bufshr[dce])
# In the basic access scheme, FIL and IFM are independently fetched,
# resulting in repeating remote fetch. OFM are merged locally and only
# stored back remotely once.
self.assertGreater(
nhops_base[de.FIL],
layer.total_filter_size() * far_dist * part.size(pe.BATP) *
part.size(pe.OFMP) * 0.8)
self.assertGreater(
nhops_base[de.IFM],
layer.total_ifmap_size(self.batch_size) * far_dist *
part.size(pe.OUTP) * 0.8)
p_layer, p_batch_size, _ = part.part_layer(layer, self.batch_size)
# With forwarding, everyone is only remotely fetched once.
self.assertLess(
nhops_accfwd[de.FIL],
p_layer.total_filter_size() * part.size(pe.INPP, pe.OUTP) *
(far_dist + nr.dim.size()))
self.assertLess(
nhops_accfwd[de.IFM],
p_layer.total_ifmap_size(p_batch_size) *
part.size(pe.INPP, pe.OFMP, pe.BATP) * (far_dist + nr.dim.size()))
self.assertLess(
nhops_accfwd[de.OFM],
p_layer.total_ofmap_size(p_batch_size) *
part.size(pe.OUTP, pe.OFMP, pe.BATP) * (far_dist + nr.dim.size()))
def test_dim(self):
''' Get dim. '''
self.assertEqual(self.ps1.dim(0), PhyDim2(2, 3))
self.assertEqual(self.ps1.dim(1), PhyDim2(3, 1))
self.assertEqual(self.ps1.dim(2), PhyDim2(1, 5))
self.assertEqual(self.ps1.dim(3), PhyDim2(5, 2))
self.assertEqual(self.ps2.dim(0), PhyDim2(2, 2))
self.assertEqual(self.ps2.dim(1), PhyDim2(5, 5))
self.assertEqual(self.ps2.dim(2), PhyDim2(3, 3))
self.assertEqual(self.ps2.dim(3), PhyDim2(1, 1))
self.assertEqual(self.ps1.dim(0, 1, 2),
PhyDim2(2, 3) * PhyDim2(3, 1) * PhyDim2(1, 5))
self.assertEqual(
self.ps1.dim(),
PhyDim2(2, 3) * PhyDim2(3, 1) * PhyDim2(1, 5) * PhyDim2(5, 2))
self.assertEqual(self.ps1.dim(0, 1, 2), self.ps1.dim(1, 2, 0))
def test_hop_dist(self):
''' Get hop distance. '''
dim1 = PhyDim2(14, 12)
dim2 = PhyDim2(5, 20)
self.assertEqual(dim1.hop_dist(dim2), 9 + 8, 'hop_dist')
self.assertEqual(dim2.hop_dist(dim1), 9 + 8, 'hop_dist')
def test_origin(self):
''' Origin. '''
layer = self.layers['BASE']
part = PartitionScheme(order=(pe.BATP, pe.INPP, pe.OUTP, pe.OFMP),
pdims=((1, 1), (1, 1), (1, 1), (1, 1)))
nr = NodeRegion(origin=PhyDim2(3, 3),
dim=part.dim(),
type=NodeRegion.PROC)
ilayout = self._make_data_layout(layer.nifm, layer.hifm, layer.wifm,
PhyDim2(-3, -3), (1, 1), (1, 1),
PhyDim2(1, 1))
olayout = self._make_data_layout(layer.nofm, layer.hofm, layer.wofm,
PhyDim2(3, 3), (1, 1), (1, 1),
PhyDim2(1, 1))
filter_nodes = frozenset([PhyDim2(3, -3)])
nhops_1 = partition.unit_nhops_to_proc_region(layer, self.batch_size,
nr, part, filter_nodes,
ilayout, olayout,
self.options['BASE'])
nr = NodeRegion(origin=PhyDim2(6, 6),
dim=part.dim(),
type=NodeRegion.PROC)
ilayout = self._make_data_layout(layer.nifm, layer.hifm, layer.wifm,
PhyDim2(-6, -6), (1, 1), (1, 1),
PhyDim2(1, 1))
olayout = self._make_data_layout(layer.nofm, layer.hofm, layer.wofm,
PhyDim2(6, 6), (1, 1), (1, 1),
PhyDim2(1, 1))
filter_nodes = frozenset([PhyDim2(6, -6)])
nhops_2 = partition.unit_nhops_to_proc_region(layer, self.batch_size,
nr, part, filter_nodes,
ilayout, olayout,
self.options['BASE'])
self.assertListEqual(nhops_2, [n * 2 for n in nhops_1])
def test_merge(self):
''' Merge. '''
fr = FmapRange((0, ) * 4, (30, ) * 4)
frm = FmapRangeMap()
frm.add(FmapRange((0, 0, 0, 0), (4, 1, 16, 16)),
(PhyDim2(0, 0), PhyDim2(1, 1)))
frm.add(FmapRange((0, 1, 0, 0), (4, 3, 16, 16)),
(PhyDim2(1, 0), PhyDim2(2, 2)))
dly = DataLayout(origin=PhyDim2(-1, -1),
frmap=frm,
type=NodeRegion.DATA)
mdly1 = self.dly.merge('|', dly)
mdly2 = dly.merge('|', self.dly)
self.assertEqual(mdly1.type, self.dly.type, 'merge |: type')
self.assertEqual(mdly2.type, dly.type, 'merge |: type')
self.assertEqual(mdly1.frmap.complete_fmap_range(),
mdly2.frmap.complete_fmap_range(),
'merge |: complete_fmap_range')
self.assertEqual(
mdly1.frmap.complete_fmap_range().size(),
self.frm.complete_fmap_range().size() +
frm.complete_fmap_range().size(),
'merge |: complete_fmap_range: size')
self.assertEqual(mdly1.total_transfer_nhops(fr, PhyDim2(0, 0)),
mdly2.total_transfer_nhops(fr, PhyDim2(0, 0)),
'merge |: nhops')
self.assertEqual(
mdly1.total_transfer_nhops(fr, PhyDim2(0, 0)),
self.dly.total_transfer_nhops(fr, PhyDim2(0, 0)) +
dly.total_transfer_nhops(fr, PhyDim2(0, 0)), 'merge |: nhops')
frm.add(FmapRange((0, 3, 0, 0), (4, 4, 16, 16)),
(PhyDim2(1, 3), PhyDim2(2, 1), PhyDim2(-1, -2)))
# Only type differs from self.dly.
sdly = DataLayout(origin=self.dly.origin,
frmap=self.dly.frmap,
type=NodeRegion.PROC)
dly = DataLayout(origin=PhyDim2(-1, -1),
frmap=frm,
type=NodeRegion.PROC)
mdly1 = sdly.merge('+', dly)
mdly2 = dly.merge('+', sdly)
self.assertEqual(mdly1.type, sdly.type, 'merge +: type')
self.assertEqual(mdly2.type, dly.type, 'merge +: type')
self.assertEqual(mdly1.frmap.complete_fmap_range(),
mdly2.frmap.complete_fmap_range(),
'merge +: complete_fmap_range')
self.assertEqual(mdly1.frmap.complete_fmap_range().size(),
self.frm.complete_fmap_range().size(),
'merge +: complete_fmap_range: size')
self.assertEqual(mdly1.total_transfer_nhops(fr, PhyDim2(0, 0)),
mdly2.total_transfer_nhops(fr, PhyDim2(0, 0)),
'merge +: nhops')
self.assertEqual(
mdly1.total_transfer_nhops(fr, PhyDim2(0, 0)),
sdly.total_transfer_nhops(fr, PhyDim2(0, 0)) +
dly.total_transfer_nhops(fr, PhyDim2(0, 0)), 'merge +: nhops')
def test_merge_invalid_type(self):
''' Merge invalid. '''
with self.assertRaisesRegexp(TypeError, 'DataLayout: .*other.*'):
_ = self.dly.merge('|', self.frm)
with self.assertRaisesRegexp(TypeError, 'DataLayout: .*other.*'):
_ = self.dly.merge('+', PhyDim2(1, 3))
def do_scheduling(args):
'''
Get optimal scheduling for given problem. Return a result schedule.
'''
## Network.
network = import_network(args.net)
batch_size = args.batch
## Resource.
dim_nodes = PhyDim2(*args.nodes)
dim_array = PhyDim2(*args.array)
# Sizes of gbuf and regf are in words.
word = (args.word + 7) // 8
size_gbuf = args.gbuf // word
size_regf = args.regf // word
array_bus_width = args.bus_width // args.word
if not array_bus_width:
array_bus_width = float('inf')
dram_bandwidth = args.dram_bw / word
proc_region = NodeRegion(dim=dim_nodes,
origin=PhyDim2(0, 0),
type=NodeRegion.PROC)
if args.mem_type == '2D':
# Memory nodes are on two sides.
data_region = NodeRegion(dim=PhyDim2(2, 2),
origin=PhyDim2(0, 0),
dist=dim_nodes - PhyDim2(1, 1),
type=NodeRegion.DRAM)
assert data_region.rel2abs(PhyDim2(1, 1)) + PhyDim2(1, 1) \
== proc_region.dim
elif args.mem_type == '3D':
# Memory nodes are on the top.
data_region = NodeRegion(dim=dim_nodes,
origin=PhyDim2(0, 0),
type=NodeRegion.DRAM)
resource = Resource(proc_region=proc_region,
dram_region=data_region,
src_data_region=data_region,
dst_data_region=data_region,
dim_array=dim_array,
size_gbuf=size_gbuf,
size_regf=size_regf,
array_bus_width=array_bus_width,
dram_bandwidth=dram_bandwidth,
no_time_mux=False)
## Cost.
hier_cost = [0] * me.NUM
hier_cost[me.DRAM] = args.hier_cost[0]
hier_cost[me.GBUF] = args.hier_cost[1]
hier_cost[me.ITCN] = args.hier_cost[2]
hier_cost[me.REGF] = args.hier_cost[3]
cost = Cost(mac_op=args.op_cost,
mem_hier=tuple(hier_cost),
noc_hop=args.hop_cost,
idl_unit=args.unit_idle_cost)
## Options.
bypass = [True] * de.NUM
bypass[de.IFM] = 'i' not in args.disable_bypass
bypass[de.OFM] = 'o' not in args.disable_bypass
bypass[de.FIL] = 'f' not in args.disable_bypass
options = Option(
sw_gbuf_bypass=tuple(bypass),
sw_solve_loopblocking=args.solve_loopblocking,
hw_access_forwarding=args.enable_access_forwarding,
hw_gbuf_sharing=args.enable_gbuf_sharing,
hw_gbuf_save_writeback=args.enable_save_writeback,
partition_hybrid=args.hybrid_partition,
partition_batch=args.batch_partition,
partition_ifmaps=args.ifmaps_partition,
partition_interlayer=args.interlayer_partition,
layer_pipeline_time_ovhd=args.layer_pipeline_time_overhead,
layer_pipeline_max_degree=args.layer_pipeline_max_degree,
layer_pipeline_opt=not args.disable_interlayer_opt,
opt_goal=args.goal.lower(),
ntops=args.top,
nprocesses=args.processes,
verbose=args.verbose)
## Search schedules.
nnd = NNDataflow(network, batch_size, resource, cost, MapStrategyEyeriss)
tbeg = time.time()
tops, cache_stats = nnd.schedule_search(options)
tend = time.time()
telapsed = tend - tbeg
if not tops:
sys.stderr.write('No valid dataflow found.\n')
return None
top = tops[0]
## Write results.
res_map = OrderedDict()
res_map['version'] = get_version(with_local=True)
res_map['net'] = args.net
res_map['batch'] = args.batch
res_map['resource'] = resource._asdict()
res_map['cost'] = cost._asdict()
res_map['options'] = options._asdict()
res_map['cache_stats'] = cache_stats
res_map['elapsed'] = telapsed
stats = stats_dict(top, cost)
for key, val in stats.items():
res_map[key] = val
return res_map
def test_view(self):
''' Get view. '''
frm = self.frm.copy()
frm.add(FmapRange((0, 0, 0, 16), (4, 4, 16, 20)),
(PhyDim2(2, 2), PhyDim2(3, 3)))
frm.add(FmapRange((0, 0, 16, 0), (4, 4, 20, 20)),
(PhyDim2(1, 1), PhyDim2(3, 3), PhyDim2(5, 5)))
dly = DataLayout(origin=PhyDim2(1, 1), frmap=frm, type=NodeRegion.DATA)
cfr = dly.frmap.complete_fmap_range()
counters = dly.frmap.rget_counter(cfr)
nhops = dly.total_transfer_nhops(cfr, PhyDim2(1, 2))
dly1 = dly.view(PhyDim2(-1, -1))
self.assertEqual(dly1.origin, PhyDim2(0, 0), 'view: origin')
self.assertEqual(dly1.type, dly.type, 'view: type')
self.assertEqual(dly1.frmap.complete_fmap_range(), cfr,
'view: complete_fmap_range')
self.assertDictEqual(dly1.frmap.rget_counter(cfr), counters,
'view: counter')
self.assertEqual(
dly1.total_transfer_nhops(cfr,
PhyDim2(1, 2) + PhyDim2(-1, -1)), nhops,
'view: nhops')
dly2 = dly.view(PhyDim2(3, 1))
self.assertEqual(dly2.type, dly.type, 'view: type')
self.assertEqual(dly2.frmap.complete_fmap_range(), cfr,
'view: complete_fmap_range')
self.assertDictEqual(dly2.frmap.rget_counter(cfr), counters,
'view: counter')
self.assertEqual(
dly2.total_transfer_nhops(cfr,
PhyDim2(1, 2) + PhyDim2(3, 1)), nhops,
'view: nhops')
def test_invalid_data_regions_type(self):
''' Invalid data_regions type. '''
with self.assertRaisesRegexp(TypeError, 'Resource: .*data_regions.*'):
_ = Resource(
proc_region=NodeRegion(dim=PhyDim2(2, 2),
origin=PhyDim2(0, 0),
type=NodeRegion.PROC),
data_regions=[
NodeRegion(dim=PhyDim2(2, 1),
origin=PhyDim2(0, 0),
type=NodeRegion.DATA),
],
dim_array=PhyDim2(16, 16),
size_gbuf=131072,
size_regf=(512, ),
)
with self.assertRaisesRegexp(TypeError, 'Resource: .*data_regions.*'):
_ = Resource(
proc_region=NodeRegion(dim=PhyDim2(2, 2),
origin=PhyDim2(0, 0),
type=NodeRegion.PROC),
data_regions=NodeRegion(dim=PhyDim2(2, 1),
origin=PhyDim2(0, 0),
type=NodeRegion.DATA),
dim_array=PhyDim2(16, 16),
size_gbuf=131072,
size_regf=(512, ),
)
with self.assertRaisesRegexp(TypeError, 'Resource: .*data_regions.*'):
_ = Resource(
proc_region=NodeRegion(dim=PhyDim2(2, 2),
origin=PhyDim2(0, 0),
type=NodeRegion.PROC),
data_regions=(NodeRegion(dim=PhyDim2(2, 1),
origin=PhyDim2(0, 0),
type=NodeRegion.DATA), PhyDim2(2, 1)),
dim_array=PhyDim2(16, 16),
size_gbuf=131072,
size_regf=(512, ),
)
def test_eyeriss_asplos17(self):
'''
Reproduce TETRIS ASPLOS'17 paper Figure 8.
'''
network = self.alex_net
batch_size = 16
## L-1 configuration.
resource = Resource(
proc_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.PROC),
data_regions=(NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DATA), ),
dim_array=PhyDim2(16, 16),
size_gbuf=576056 // 2, # 576 kB
size_regf=1024 // 2, # 1 kB
)
cost = Cost(
mac_op=2e-12,
mem_hier=(240e-12, 28e-12, 4e-12, 1e-12), # pJ/16-b
noc_hop=0,
unit_static=320e-12)
nnd = NNDataflow(network, batch_size, resource, cost,
self.map_strategy)
tops, _ = nnd.schedule_search(self.options)
self.assertTrue(tops)
dfsch_l1 = tops[0]
## T-16 configuration.
resource = Resource(
proc_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 4),
type=NodeRegion.PROC),
data_regions=(NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 4),
type=NodeRegion.DATA), ),
dim_array=PhyDim2(14, 14),
size_gbuf=133032 // 2, # 133 kB
size_regf=512 // 2, # 512 B
)
cost = Cost(
mac_op=2e-12,
mem_hier=(80e-12, 14e-12, 4e-12, 0.6e-12), # pJ/16-b
noc_hop=40e-12,
unit_static=200e-12)
options = Option(sw_gbuf_bypass=(True, True, True),
sw_solve_loopblocking=True,
partition_hybrid=True)
nnd = NNDataflow(network, batch_size, resource, cost,
self.map_strategy)
tops, _ = nnd.schedule_search(options)
self.assertTrue(tops)
dfsch_t16 = tops[0]
## Check results.
# Same workload.
self.assertAlmostEqual(dfsch_t16.total_ops, dfsch_l1.total_ops)
# Performance of T-16 is proportional to PE resource (20% margin).
self.assertLess(dfsch_t16.total_time,
1.2 * dfsch_l1.total_time * (16 * 16) / (14 * 14 * 16))
# Energy reduced by > 30%.
self.assertLess(dfsch_t16.total_cost, dfsch_l1.total_cost * 0.7)
def test_eyeriss_isscc16(self):
'''
Reproduce Eyeriss ISSCC'16 paper Fig. 14.5.6, JSSC'17 paper Table V.
'''
network = self.alex_net
batch_size = 4
resource = Resource(
proc_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.PROC),
data_regions=(NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DATA), ),
dim_array=PhyDim2(12, 14),
size_gbuf=108 * 1024 // 2, # 108 kB
size_regf=261, # 225 + 12 + 24
)
cost = Cost(
mac_op=2e-12,
mem_hier=(460e-12, 15e-12, 4e-12, 1e-12), # pJ/16-b
noc_hop=0,
unit_static=30e-3 / 200e6) # 30 mW GBUF + REGF
nnd = NNDataflow(network, batch_size, resource, cost,
self.map_strategy)
tops, _ = nnd.schedule_search(self.options)
self.assertTrue(tops)
dfsch = tops[0]
## Check results.
# Results as stats of the rows in the table.
header = 'Power, Processing Latency, Ops, Active PEs, Filter size'
stats = {}
for layer in ['conv{}'.format(i) for i in range(1, 6)]:
onchip_cost = 0
time = 0
ops = 0
fil_size = 0
for layer_part in network:
if not layer_part or not layer_part.startswith(layer):
continue
sr = dfsch[layer_part]
onchip_cost += sr.total_cost \
- sr.total_accesses[me.DRAM] * cost.mem_hier[me.DRAM]
time += sr.total_time
ops += sr.total_ops
fil_size += network[layer_part].total_filter_size()
power = onchip_cost / (time / 200e6) * 1e3 # mW
active_pes = int(ops / time)
stats[layer] = []
stats[layer].append(power)
stats[layer].append(time / 200.e3) # cycles to ms
stats[layer].append(ops / 1e6) # to MOPs
stats[layer].append(active_pes)
stats[layer].append(fil_size / 1e3) # to k
# Check.
stats_ref = {
'conv1': [332, 16.5, 421.66, 151, 34.8], # Act PE 154
'conv2': [288, 39.2, 895.79, 135, 307.2],
'conv3': [266, 21.8, 598.1, 156, 884.7],
'conv4': [235, 16.0, 448.6, 156, 663.6],
'conv5': [236, 10.0, 299.0, 156, 442.4],
}
for layer in stats:
success = (0.6 * stats_ref[layer][0]
< stats[layer][0]
< stats_ref[layer][0]) \
and (0.8 * stats_ref[layer][1]
< stats[layer][1]
< stats_ref[layer][1]) \
and all(abs(a - b) < 0.1 for a, b
in zip(stats[layer][2:], stats_ref[layer][2:]))
self.assertTrue(
success, 'test_eyeriss_isscc16: '
'stats diff in layer {}.\n'
'header: {}\n'
'actual: {}\nref: {}'.format(layer, header, stats[layer],
stats_ref[layer]))
def test_sub(self):
''' Operation sub. '''
dim1 = PhyDim2(14, 12)
dim2 = PhyDim2(5, 3)
self.assertTupleEqual(dim1 - dim2, (9, 9), 'sub')
self.assertTupleEqual(dim1 - 3, (11, 9), 'sub')
def test_invalid_frmap_type(self):
''' Invalid frmap type. '''
with self.assertRaisesRegexp(TypeError, 'DataLayout: .*frmap.*'):
_ = DataLayout(origin=PhyDim2(2, 3),
frmap=FmapRange((0, ) * 4, (1, ) * 4),
type=NodeRegion.DATA)
def test_neg(self):
''' Operation neg. '''
dim1 = PhyDim2(14, 12)
dim2 = PhyDim2(5, 3)
self.assertTupleEqual(-dim1, (-14, -12), 'neg')
self.assertTupleEqual(-dim2, (-5, -3), 'neg')
def test_invalid_type(self):
''' Invalid type. '''
with self.assertRaisesRegexp(ValueError, 'DataLayout: .*type.*'):
_ = DataLayout(origin=PhyDim2(2, 3),
frmap=self.frm,
type=NodeRegion.NUM)
def test_hop_dist_error(self):
''' Get hop distance. '''
dim1 = PhyDim2(14, 12)
with self.assertRaisesRegexp(TypeError, 'hop_dist'):
_ = dim1.hop_dist((5, 20))
def test_is_in_region(self):
''' Whether is in region. '''
nr1 = NodeRegion(dim=PhyDim2(2, 2),
origin=PhyDim2(1, 1),
type=NodeRegion.DATA)
self.assertTrue(self.dly.is_in_region(nr1))
nr2 = NodeRegion(dim=PhyDim2(3, 3),
origin=PhyDim2(0, 0),
type=NodeRegion.DATA)
self.assertTrue(self.dly.is_in_region(nr2))
nr3 = NodeRegion(dim=PhyDim2(2, 2),
origin=PhyDim2(0, 0),
type=NodeRegion.DATA)
self.assertFalse(self.dly.is_in_region(nr3))
nr4 = NodeRegion(dim=PhyDim2(3, 3),
origin=PhyDim2(0, 0),
type=NodeRegion.PROC)
self.assertFalse(self.dly.is_in_region(nr4))
frm = self.frm.copy()
frm.add(FmapRange((0, 0, 0, 16), (4, 4, 16, 20)),
(PhyDim2(2, 2), PhyDim2(3, 3)))
dly = DataLayout(origin=PhyDim2(1, 1), frmap=frm, type=NodeRegion.DATA)
nr4 = NodeRegion(dim=PhyDim2(3, 3),
origin=PhyDim2(1, 1),
type=NodeRegion.DATA)
self.assertFalse(dly.is_in_region(nr4))
nr5 = NodeRegion(dim=PhyDim2(4, 4),
origin=PhyDim2(1, 1),
type=NodeRegion.DATA)
self.assertTrue(dly.is_in_region(nr5))
frm.add(FmapRange((0, 0, 16, 0), (4, 4, 20, 20)),
(PhyDim2(1, 1), PhyDim2(3, 3), PhyDim2(5, 5)))
dly = DataLayout(origin=PhyDim2(1, 1), frmap=frm, type=NodeRegion.DATA)
self.assertFalse(dly.is_in_region(nr5))
nr6 = NodeRegion(dim=PhyDim2(7, 7),
origin=PhyDim2(0, 0),
type=NodeRegion.DATA)
self.assertTrue(dly.is_in_region(nr6))
def test_small(self):
''' Small case with hand calculation. '''
layer = ConvLayer(6, 8, 16, 3)
assert self.batch_size == 8
# i (0, 0), (2, 0): (0, 0, 0, 0) -- (4, 6, 10, 10)
# (0, 1), (2, 1): (0, 0, 0, 8) -- (4, 6, 10, 18)
# (0, 2), (2, 2): (4, 0, 0, 0) -- (8, 6, 10, 10)
# (0, 3), (2, 3): (4, 0, 0, 8) -- (8, 6, 10, 18)
# (1, 0), (3, 0): (0, 0, 8, 0) -- (4, 6, 18, 10)
# (1, 1), (3, 1): (0, 0, 8, 8) -- (4, 6, 18, 18)
# (1, 2), (3, 2): (4, 0, 8, 0) -- (8, 6, 18, 10)
# (1, 3), (3, 3): (4, 0, 8, 8) -- (8, 6, 18, 18)
# o (0, 0): (0, 0, 0, 0) -- (4, 4, 8, 8)
# (0, 1): (0, 0, 0, 8) -- (4, 4, 8, 16)
# (0, 2): (4, 0, 0, 0) -- (8, 4, 8, 8)
# (0, 3): (4, 0, 0, 8) -- (8, 4, 8, 16)
# (1, 0): (0, 0, 8, 0) -- (4, 4, 16, 8)
# (1, 1): (0, 0, 8, 8) -- (4, 4, 16, 16)
# (1, 2): (4, 0, 8, 0) -- (8, 4, 16, 8)
# (1, 3): (4, 0, 8, 8) -- (8, 4, 16, 16)
# (2, 0): (0, 4, 0, 0) -- (4, 8, 8, 8)
# (2, 1): (0, 4, 0, 8) -- (4, 8, 8, 16)
# (2, 2): (4, 4, 0, 0) -- (8, 8, 8, 8)
# (2, 3): (4, 4, 0, 8) -- (8, 8, 8, 16)
# (3, 0): (0, 4, 8, 0) -- (4, 8, 16, 8)
# (3, 1): (0, 4, 8, 8) -- (4, 8, 16, 16)
# (3, 2): (4, 4, 8, 0) -- (8, 8, 16, 8)
# (3, 3): (4, 4, 8, 8) -- (8, 8, 16, 16)
part = PartitionScheme(order=(pe.BATP, pe.INPP, pe.OUTP, pe.OFMP),
pdims=((2, 1), (2, 2), (1, 2), (1, 1)))
nr = NodeRegion(origin=PhyDim2(0, 0),
dim=part.dim(),
type=NodeRegion.PROC)
# (0, 0, 0, 0) -- (4, 6, 18, 9): (-2, -2)
# (0, 0, 0, 9) -- (4, 6, 18, 18): (-2, -1)
# (4, 0, 0, 0) -- (8, 6, 18, 9): (-1, -2)
# (4, 0, 0, 9) -- (8, 6, 18, 18): (-1, -1)
ilayout = self._make_data_layout(layer.nifm, layer.hifm, layer.wifm,
PhyDim2(-2, -2), (2, 1), (1, 1),
PhyDim2(2, 2))
# (0, 0, 0, 0) -- (8, 4, 16, 8): (2, 2)
# (0, 0, 0, 8) -- (8, 4, 16, 16): (2, 3)
# (0, 4, 0, 0) -- (8, 8, 16, 8): (3, 2)
# (0, 4, 0, 8) -- (8, 8, 16, 16): (3, 3)
olayout = self._make_data_layout(layer.nofm, layer.hofm, layer.wofm,
PhyDim2(2, 2), (1, 1), (2, 1),
PhyDim2(2, 2))
filter_nodes = frozenset([PhyDim2(0, 0)])
# filter: (0, 0) -> all, 6 * 4 * 3 * 3
# ifmap: (-2, -2) -> (0, 0), (2, 0): 4 * 6 * 10 * 9
# -> (0, 1), (2, 1): 4 * 6 * 10 * (9 - 8)
# -> (1, 0), (3, 0): 4 * 6 * (18 - 8) * 9
# -> (1, 1), (3, 1): 4 * 6 * (18 - 8) * (9 - 8)
# (-2, -1) -> (0, 0), (2, 0): 4 * 6 * 10 * (10 - 9)
# -> (0, 1), (2, 1): 4 * 6 * 10 * (18 - 9)
# -> (1, 0), (3, 0): 4 * 6 * (18 - 8) * (10 - 9)
# -> (1, 1), (3, 1): 4 * 6 * (18 - 8) * (18 - 9)
# (-1, -2) -> (0, 2), (2, 2): (8 - 4) * 6 * 10 * 9
# -> (0, 3), (2, 3): (8 - 4) * 6 * 10 * (9 - 8)
# -> (1, 2), (3, 2): (8 - 4) * 6 * (18 - 8) * 9
# -> (1, 3), (3, 3): (8 - 4) * 6 * (18 - 8) * (9 - 8)
# (-1, -1) -> (0, 2), (2, 2): (8 - 4) * 6 * 10 * (10 - 9)
# -> (0, 3), (2, 3): (8 - 4) * 6 * 10 * (18 - 9)
# -> (1, 2), (3, 2): (8 - 4) * 6 * (18 - 8) * (10 - 9)
# -> (1, 3), (3, 3): (8 - 4) * 6 * (18 - 8) * (18 - 9)
# ofmap: (2, 2) -> (0, 0):
# -> (0, 2):
# -> (1, 0):
# -> (1, 2): 4 * 4 * 8 * 8
# (2, 3) -> (0/1, 1/3)
# (3, 2) -> (2/3, 0/2)
# (3, 3) -> (2/3, 1/3)
nhops = partition.unit_nhops_to_proc_region(layer, self.batch_size, nr,
part, filter_nodes,
ilayout, olayout,
self.options['BASE'])
self.assertEqual(
nhops[de.FIL],
6 * 4 * 3 * 3 * sum(h + w for h in range(4) for w in range(4)))
self.assertEqual(
nhops[de.IFM], 4 * 6 * 10 *
((4 + 6) * 9 + (5 + 7) * 1 + (5 + 7) * 9 + (6 + 8) * 1 +
(3 + 5) * 1 + (4 + 6) * 9 + (4 + 6) * 1 + (5 + 7) * 9 +
(5 + 7) * 9 + (6 + 8) * 1 + (6 + 8) * 9 + (7 + 9) * 1 +
(4 + 6) * 1 + (5 + 7) * 9 + (5 + 7) * 1 + (6 + 8) * 9))
self.assertEqual(
nhops[de.OFM], 4 * 4 * 8 * 8 * ((4 + 2 + 3 + 1) + (4 + 2 + 3 + 1) +
(3 + 1 + 2 + 0) + (3 + 1 + 2 + 0)))
def setUp(self):
self.net = {}
net = Network('net1')
# Linear.
net.set_input_layer(InputLayer(10, 1))
net.add('0', FCLayer(10, 20))
net.add('1', FCLayer(20, 30))
net.add('1p', PoolingLayer(30, 1, 1))
net.add('2', FCLayer(30, 40))
net.add('3', FCLayer(40, 50))
self.net[net.net_name] = net
net = Network('net2')
# Long linear.
net.set_input_layer(InputLayer(1, 1))
for idx in range(16):
net.add(str(idx), FCLayer(1, 1))
self.net[net.net_name] = net
net = Network('net3')
# Fork.
# /0-2\ /6- 7- 8\
# x 4-5 12
# \1-3/ \9-10-11/
net.set_input_layer(InputLayer(1, 1))
net.add('0', FCLayer(1, 1), prevs=net.INPUT_LAYER_KEY)
net.add('1', FCLayer(1, 1), prevs=net.INPUT_LAYER_KEY)
net.add('2', FCLayer(2, 1), prevs=('0', '1'))
net.add('2p', PoolingLayer(1, 1, 1))
net.add('3', FCLayer(2, 1), prevs=('0', '1'))
net.add('4', FCLayer(2, 1), prevs=('2p', '3'))
net.add('5', FCLayer(1, 1))
net.add('5p', PoolingLayer(1, 1, 1))
net.add('6', FCLayer(1, 1), prevs='5p')
net.add('7', FCLayer(1, 1))
net.add('8', FCLayer(1, 1))
net.add('9', FCLayer(1, 1), prevs='5p')
net.add('10', FCLayer(1, 1))
net.add('11', FCLayer(1, 1))
net.add('12', FCLayer(2, 1), prevs=('8', '11'))
self.net[net.net_name] = net
net = Network('net4')
# Complex fork.
# /5 \
# 0-1-2-3-4-6-7-8-10-14
# \9/
# \11-12 /
# \13 /
net.set_input_layer(InputLayer(1, 1))
net.add('0', FCLayer(1, 1))
net.add('1', FCLayer(1, 1))
net.add('2', FCLayer(1, 1))
net.add('3', FCLayer(1, 1))
net.add('4', FCLayer(1, 1))
net.add('5', FCLayer(1, 1), prevs='4')
net.add('6', FCLayer(1, 1), prevs='4')
net.add('7', FCLayer(1, 1))
net.add('8', FCLayer(1, 1), prevs='7')
net.add('9', FCLayer(1, 1), prevs='7')
net.add('10', FCLayer(1, 1))
net.add('10p', PoolingLayer(2, 1, 1), prevs=('8', '10'))
net.add('11', PoolingLayer(1, 1, 1), prevs='4')
net.add('12', FCLayer(1, 1))
net.add('13', PoolingLayer(1, 1, 1), prevs='4')
net.add('14', FCLayer(5, 1), prevs=('5', '10p', '12', '13'))
self.net[net.net_name] = net
net = Network('net5')
# Corner cases.
# ----\
# //1-2\ 7-8\
# 0-3-4-x 10-11-12
# \ \5/ 9 / \__/
# 6--/
net.set_input_layer(InputLayer(1, 1))
net.add('0', FCLayer(1, 1))
net.add('1', FCLayer(1, 1), prevs='0')
net.add('2', FCLayer(1, 1))
net.add('3', FCLayer(1, 1), prevs='0')
net.add('4', FCLayer(1, 1), prevs='3')
net.add('5', FCLayer(1, 1), prevs='3')
net.add('6', FCLayer(1, 1), prevs='0')
net.add('7', FCLayer(5, 1), prevs=('0', '2', '4', '5', '6'))
net.add('8', FCLayer(1, 1))
net.add('9', FCLayer(5, 1), prevs=('0', '2', '4', '5', '6'))
net.add('10', FCLayer(2, 1), prevs=('8', '9'))
net.add('11', FCLayer(1, 1))
net.add('12', FCLayer(2, 1), prevs=('10', '11'))
self.net[net.net_name] = net
net = Network('net6')
# Fmap sizes.
net.set_input_layer(InputLayer(1, 24))
net.add('0', ConvLayer(1, 1, 24, 3))
net.add('1', ConvLayer(1, 1, 12, 3, strd=2))
net.add('1p', PoolingLayer(1, 6, 2))
net.add('2', ConvLayer(1, 1, 6, 3))
net.add('3', ConvLayer(1, 1, 6, 3, strd=4), prevs=('0'))
self.net[net.net_name] = net
net = Network('net7')
# Topological order: see a visited vertex again.
# /---
# 0-1-\\
# \2--2p
net.set_input_layer(InputLayer(1, 1))
net.add('0', FCLayer(1, 1))
net.add('1', FCLayer(1, 1), prevs='0')
net.add('2', FCLayer(1, 1), prevs='0')
net.add('2p', PoolingLayer(3, 1, 1), prevs=('0', '1', '2'))
self.net[net.net_name] = net
net = Network('net8')
# Forward to the middle.
# /-\
# 0-1-2-2p-4-4p
# \-3------/
net.set_input_layer(InputLayer(1, 1))
net.add('0', FCLayer(1, 1))
net.add('1', FCLayer(1, 1), prevs='0')
net.add('2', FCLayer(1, 1), prevs='1')
net.add('2p', PoolingLayer(2, 1, 1), prevs=('1', '2'))
net.add('3', FCLayer(1, 1), prevs='0')
net.add('4', FCLayer(2, 1), prevs='2p')
net.add('4p', PoolingLayer(2, 1, 1), prevs=('3', '4'))
self.net[net.net_name] = net
net = Network('net9')
# Previous layers include input and others.
net.set_input_layer(InputLayer(1, 1))
net.add('0', FCLayer(1, 1))
net.add('1', FCLayer(2, 1), prevs=(net.INPUT_LAYER_KEY, '0'))
self.net[net.net_name] = net
# Real networks.
for net_name in all_networks():
self.net[net_name] = import_network(net_name)
self.batch_size = 16
self.resource = Resource(
proc_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(8, 8),
type=NodeRegion.PROC),
dram_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(8, 8),
type=NodeRegion.DRAM),
src_data_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(8, 4),
type=NodeRegion.DRAM),
dst_data_region=NodeRegion(origin=PhyDim2(0, 4),
dim=PhyDim2(8, 4),
type=NodeRegion.DRAM),
dim_array=PhyDim2(16, 16),
size_gbuf=65536,
size_regf=64,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
part = PartitionScheme(order=range(pe.NUM), pdims=[(1, 1)] * pe.NUM)
self.ofmap_layout = DataLayout(
frngs=(FmapRange((0, 0, 0, 0), (2, 4, 16, 16)), ),
regions=(NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DRAM), ),
parts=(part, ))
def _make_region(dim):
return NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(*dim),
type=NodeRegion.DRAM)
def test_valid_args(self):
''' Valid arguments. '''
dim = PhyDim2(14, 12)
self.assertEqual(dim.h, 14, 'h')
self.assertEqual(dim.w, 12, 'w')
def eyerissAsplos17(self):
'''
Reproduce TETRIS ASPLOS'17 paper Figure 8.
'''
network = self.alex_net
batch_size = 16
resource = Resource(
proc_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 4),
type=NodeRegion.PROC),
dram_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 4),
type=NodeRegion.DRAM),
src_data_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 4),
type=NodeRegion.DRAM),
dst_data_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 4),
type=NodeRegion.DRAM),
dim_array=PhyDim2(14, 14),
size_gbuf=133032 // 2, # 133 kB
size_regf=512 // 2, # 512 B
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False,
)
cost = Cost(
mac_op=2e-12,
mem_hier=(80e-12, 14e-12, 4e-12, 0.6e-12), # pJ/16-b
noc_hop=40e-12,
idl_unit=200e-12)
options = Option(sw_gbuf_bypass=(True, True, True),
sw_solve_loopblocking=True,
partition_hybrid=True)
pdb.set_trace()
nnd = NNDataflow(network, batch_size, resource, cost,
self.map_strategy)
tops, _ = nnd.schedule_search(options)
self.assertTrue(tops)
dfsch_t16 = tops[0]
## Check results.
# Same workload.
#self.assertAlmostEqual(dfsch_t16.total_ops, dfsch_l1.total_ops)
print('t16 ops: {}'.format(dfsch_t16.total_ops))
# Performance of T-16 is proportional to PE resource (20% margin).
#self.assertLess(dfsch_t16.total_time,
# 1.2 * dfsch_l1.total_time * (16 * 16) / (14 * 14 * 16))
print('t16_time: {}'.format(dfsch_t16.total_time))
# Energy reduced by > 30%.
# self.assertLess(dfsch_t16.total_cost, dfsch_l1.total_cost * 0.7)
# With dimension restriction on partitioning, this is slightly violated.
#self.assertLess(dfsch_t16.total_cost, dfsch_l1.total_cost * 0.72)
print('t16_energy: {}'.format(dfsch_t16.total_cost))
for i in dfsch_t16:
print(str(i) + ',')
## Check results.
# Results as cost for each component:
header = 'ALU, DRAM, Buffer, Array, RF'
cost_bkdn = {}
for layer in dfsch_t16:
layer = str(layer)
op_cost = 0
access_cost = [0] * me.NUM
for layer_part in network:
if not layer_part or not layer_part.startswith(layer):
continue
sr = dfsch_t16[layer_part]
op_cost += sr.total_ops * cost.mac_op
access_cost = [
ac + a * c for ac, a, c in zip(
access_cost, sr.total_accesses, cost.mem_hier)
]
cost_bkdn[layer] = []
# To 1e9.
cost_bkdn[layer].append(op_cost * 1e12 / 1e9)
cost_bkdn[layer].append(access_cost[me.DRAM] * 1e12 / 1e9)
cost_bkdn[layer].append(access_cost[me.GBUF] * 1e12 / 1e9)
cost_bkdn[layer].append(access_cost[me.ITCN] * 1e12 / 1e9)
cost_bkdn[layer].append(access_cost[me.REGF] * 1e12 / 1e9)
for layer in cost_bkdn:
print(cost_bkdn[layer])
def test_size(self):
''' Get size. '''
dim = PhyDim2(14, 12)
self.assertEqual(dim.size(), 14 * 12, 'size')
def test_invalid_type(self):
''' Invalid type. '''
with self.assertRaisesRegexp(ValueError, 'NodeRegion: .*type.*'):
_ = NodeRegion(dim=PhyDim2(4, 4),
origin=PhyDim2(1, 3),
type=NodeRegion.NUM)
def test_add(self):
''' Operation add. '''
dim1 = PhyDim2(14, 12)
dim2 = PhyDim2(5, 3)
self.assertTupleEqual(dim1 + dim2, (19, 15), 'add')
self.assertTupleEqual(dim1 + 3, (17, 15), 'add')
def setUp(self):
# Workload.
self.layer = {}
self.layer['BASE'] = ConvLayer(12, 10, 28, 3)
self.layer['LGFIL'] = ConvLayer(2, 4, 28, 20)
self.layer['POOL'] = PoolingLayer(32, 28, 2)
self.layer['PAR'] = ConvLayer(24, 36, 56, 3)
self.batch_size = 4
# Resource.
self.resource = {}
dim_array = PhyDim2(16, 16)
proc_region = NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.PROC)
data_region = NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DRAM)
# Typical resource.
self.resource['BASE'] = Resource(proc_region=proc_region,
dram_region=data_region,
src_data_region=data_region,
dst_data_region=data_region,
dim_array=dim_array,
size_gbuf=65536,
size_regf=64,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
# Larger resource with sufficient capacity, to make all schemes valid.
self.resource['LG'] = Resource(proc_region=proc_region,
dram_region=data_region,
src_data_region=data_region,
dst_data_region=data_region,
dim_array=dim_array,
size_gbuf=1024**3,
size_regf=1024**3,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
# Small resource.
self.resource['SM'] = Resource(proc_region=proc_region,
dram_region=data_region,
src_data_region=data_region,
dst_data_region=data_region,
dim_array=dim_array,
size_gbuf=4096,
size_regf=16,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
# Multi-node parallel resource.
self.resource['PAR'] = Resource(proc_region=NodeRegion(
origin=PhyDim2(0, 0), dim=PhyDim2(4, 2), type=NodeRegion.PROC),
dram_region=data_region,
src_data_region=data_region,
dst_data_region=data_region,
dim_array=dim_array,
size_gbuf=25000,
size_regf=64,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
# Resource with no data regions.
proc_data_region = NodeRegion(origin=PhyDim2(1, 1),
dim=PhyDim2(1, 1),
type=NodeRegion.PROC)
self.resource['SRCNOTDATA'] = Resource(
proc_region=proc_region,
dram_region=data_region,
src_data_region=proc_data_region,
dst_data_region=data_region,
dim_array=dim_array,
size_gbuf=1024**3,
size_regf=1024**3,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
self.resource['DSTNOTDATA'] = Resource(
proc_region=proc_region,
dram_region=data_region,
src_data_region=data_region,
dst_data_region=proc_data_region,
dim_array=dim_array,
size_gbuf=1024**3,
size_regf=1024**3,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
self.resource['DATALOCAL'] = Resource(proc_region=proc_region,
dram_region=data_region,
src_data_region=proc_region,
dst_data_region=proc_region,
dim_array=dim_array,
size_gbuf=1024**3,
size_regf=1024**3,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False)
# Filter pinning.
self.resource['FILPIN'] = Resource(proc_region=proc_region,
dram_region=data_region,
src_data_region=data_region,
dst_data_region=data_region,
dim_array=dim_array,
size_gbuf=1024**3,
size_regf=1024**3,
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=True)
# Nested loop description after mapping.
self.nld = {}
self.nld['BASE'] = next(
MapStrategyEyeriss(self.layer['BASE'], self.batch_size, 1,
dim_array).gen_nested_loop_desc())
self.nld['LGFIL'] = next(
MapStrategyEyeriss(self.layer['LGFIL'], self.batch_size, 1,
dim_array).gen_nested_loop_desc())
self.nld['POOL'] = next(
MapStrategyEyeriss(self.layer['POOL'], self.batch_size, 1,
dim_array).gen_nested_loop_desc())
# Fake nested loop, with zero filter size.
self.nld['ZERO_FIL'] = NestedLoopDesc(
loopcnt=(12, 10, 4),
usize_gbuf=(0, 1000, 800),
usize_regf=(0, 3, 1),
unit_access=((0, 1000, 800), (0, 1000, 800), (3, 9, 7), (1, 1, 1)),
data_loops=(DataDimLoops(le.IFM,
le.OFM), DataDimLoops(le.IFM, le.BAT),
DataDimLoops(le.OFM, le.BAT)),
unit_ops=1,
unit_time=1)
# Fake nested loop, with zero ifmap size.
self.nld['ZERO_IFM'] = NestedLoopDesc(
loopcnt=(12, 10, 4),
usize_gbuf=(9, 0, 800),
usize_regf=(3, 0, 1),
unit_access=((9, 0, 800), (9, 0, 800), (3, 9, 7), (1, 1, 1)),
data_loops=(DataDimLoops(le.IFM,
le.OFM), DataDimLoops(le.IFM, le.BAT),
DataDimLoops(le.OFM, le.BAT)),
unit_ops=1,
unit_time=1)
# Fake partition scheme.
self.part = PartitionScheme(range(pe.NUM), ((1, 1), ) * pe.NUM)
# Fake buffer sharing scheme.
self.bufshr = BufShrScheme(proc_region, self.part)
# Options.
self.options = {}
# Basic.
self.options['BASE'] = Option(ntops=2**30)
# Multiprocessing.
self.options['MP'] = Option(ntops=2**30, nprocesses=8)
# Limited top schemes.
self.options['NTOPS'] = Option(ntops=10)
# Bypass.
self.options['BYP'] = Option(sw_gbuf_bypass=(True, ) * 3, ntops=2**30)
# Bypass solver.
self.options['BYPSOL'] = Option(sw_gbuf_bypass=(True, ) * 3,
sw_solve_loopblocking=True,
ntops=2**30)
# Access forwarding.
self.options['ACCFWD'] = Option(hw_access_forwarding=True, ntops=2**30)
# Buffer sharing.
self.options['BUFSHR'] = Option(hw_gbuf_sharing=True, ntops=2**30)
# Buffer sharing with bypassing.
self.options['BUFSHR-BYP'] = Option(sw_gbuf_bypass=(True, ) * 3,
hw_gbuf_sharing=True,
ntops=2**30)
# Constraint.
self.none_cstr = SchedulingConstraint()
self.cstr = SchedulingConstraint(topifm=1, topbat=1)
# Cost.
self.cost = Cost(mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=50,
idl_unit=50)
def test_concat_invalid_type(self):
''' Concat invalid type. '''
with self.assertRaisesRegex(TypeError, 'DataLayout: .*concat.*'):
_ = DataLayout.concat(self.dl1, self.frng1)
with self.assertRaisesRegex(TypeError, 'DataLayout: .*concat.*'):
_ = DataLayout.concat(self.dl1, PhyDim2(1, 3))