def test_access_cost_same_lbs(self):
''' get_access_cost same lbs. '''
lbs = self._lbs(self._make_bl_ts((0, 1, 1), (1, 0, 1), (1, 1, 0)),
rsrckey='LG')
self.assertTrue(lbs.is_valid())
c1 = lbs.get_access_cost(Cost(mac_op=1, mem_hier=(200, 6, 2, 1),
noc_hop=50, idl_unit=50))
c2 = lbs.get_access_cost(Cost(mac_op=-1, mem_hier=(-200, -6, -2, -1),
noc_hop=-50, idl_unit=-50))
self.assertAlmostEqual(c1, -c2)
def test_invalid_mem_hier_type(self):
''' Invalid mem_hier type. '''
with self.assertRaisesRegexp(TypeError, 'Cost: .*mem_hier.*'):
_ = Cost(
mac_op=1,
mem_hier=200,
noc_hop=10,
unit_static=0,
)
with self.assertRaisesRegexp(TypeError, 'Cost: .*mem_hier.*'):
_ = Cost(
mac_op=1,
mem_hier=[200, 6, 2, 1],
noc_hop=10,
unit_static=0,
)
def setUp(self):
self.layers = {}
self.layers['BASE'] = ConvLayer(8, 16, 28, 3)
self.layers['POOL'] = PoolingLayer(16, 28, 2)
self.layers['LR'] = LocalRegionLayer(16, 28, nreg=3, sreg=1)
self.batch_size = 4
self.cost = Cost(mac_op=1, mem_hier=(200, 6, 2, 1),
noc_hop=50, unit_static=50)
self.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, 1),
type=NodeRegion.DATA),),
dim_array=PhyDim2(16, 16), size_gbuf=65536, size_regf=64)
self.options = Option(partition_hybrid=True, partition_batch=True,
partition_ifmaps=True, ntops=10)
self.ifmap_layouts = {}
part = PartitionScheme(order=(pe.INPP, pe.BATP, pe.OUTP, pe.OFMP),
pdims=((1, 2), (2, 1), (1, 2), (2, 1)))
for wlkey in self.layers:
self.ifmap_layouts[wlkey] = partition.get_ofmap_layout(
self.layers[wlkey].input_layer(), self.batch_size, part,
self.resource.src_data_region())
def setUp(self):
self.alex_net = import_network('alex_net')
self.vgg_net = import_network('vgg_net')
self.map_strategy = MapStrategyEyeriss
self.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=128 * 1024 // 2, # 128 kB
size_regf=512 // 2, # 512 B
)
self.cost = Cost(mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=0,
unit_static=0)
self.options = Option()
def setUp(self):
self.layers = {}
self.layers['BASE'] = ConvLayer(8, 16, 28, 3)
self.layers['POOL'] = PoolingLayer(16, 28, 2)
self.layers['LR'] = LocalRegionLayer(16, 28, nreg=3, sreg=1)
self.batch_size = 4
self.cost = Cost(mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=50,
idl_unit=50)
self.none_cstr = SchedulingConstraint()
self.cstr = SchedulingConstraint(topofm=1, topbat=self.batch_size)
self.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, 1),
type=NodeRegion.DRAM),
src_data_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 1),
type=NodeRegion.DRAM),
dst_data_region=NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(4, 1),
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)
self.options = Option(partition_hybrid=True,
partition_batch=True,
partition_ifmaps=True,
ntops=10)
self.ifmap_layouts = {}
part = PartitionScheme(order=(pe.INPP, pe.BATP, pe.OUTP, pe.OFMP),
pdims=((1, 2), (2, 1), (1, 2), (2, 1)))
for wlkey in self.layers:
input_layer = self.layers[wlkey].input_layer()
self.ifmap_layouts[wlkey] = 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=(self.resource.src_data_region, ),
parts=(part.projection(self.resource.src_data_region,
appl2frng=True), ))
self.sched_seq = (2, 0, 1)
def test_invalid_noc_hop(self):
''' Invalid noc_hop. '''
with self.assertRaisesRegexp(TypeError, 'Cost: .*noc_hop.*'):
_ = Cost(
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=[10, 10],
unit_static=0,
)
def test_invalid_mem_hier_len(self):
''' Invalid mem_hier len. '''
with self.assertRaisesRegexp(ValueError, 'Cost: .*mem_hier.*'):
_ = Cost(
mac_op=1,
mem_hier=(200, 6),
noc_hop=10,
unit_static=0,
)
def test_invalid_idl_unit(self):
''' Invalid idl_unit. '''
with self.assertRaisesRegex(TypeError, 'Cost: .*idl_unit.*'):
_ = Cost(
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=10,
idl_unit=set([1, 2]),
)
def test_invalid_mac_op(self):
''' Invalid mac_op. '''
with self.assertRaisesRegex(TypeError, 'Cost: .*mac_op.*'):
_ = Cost(
mac_op=(1, 2),
mem_hier=(200, 6, 2, 1),
noc_hop=10,
idl_unit=0,
)
def test_invalid_unit_static(self):
''' Invalid unit_static. '''
with self.assertRaisesRegexp(TypeError, 'Cost: .*unit_static.*'):
_ = Cost(
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=10,
unit_static=set([1, 2]),
)
def test_mem_hier_at_error(self):
''' Accessor mem_hier error. '''
cost = Cost(
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=10,
unit_static=0,
)
self.assertIsNone(cost.mem_hier_at(me.NUM))
self.assertIsNone(cost.mem_hier_at(None))
def test_valid_args(self):
''' Valid arguments. '''
cost = Cost(
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=10,
idl_unit=0,
)
self.assertEqual(cost.mac_op, 1, 'mac_op')
self.assertEqual(cost.mem_hier, (200, 6, 2, 1), 'mem_hier')
self.assertEqual(cost.noc_hop, 10, 'noc_hop')
self.assertEqual(cost.idl_unit, 0, 'idl_unit')
def test_valid_args(self):
''' Valid arguments. '''
cost = Cost(
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=10,
unit_static=0,
)
self.assertEqual(cost.mac_op, 1, 'mac_op')
self.assertEqual(cost.mem_hier, (200, 6, 2, 1), 'mem_hier')
self.assertEqual(cost.noc_hop, 10, 'noc_hop')
self.assertEqual(cost.unit_static, 0, 'unit_static')
def test_mem_hier_at(self):
''' Accessor mem_hier. '''
cost = Cost(
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=10,
unit_static=0,
)
self.assertEqual(cost.mem_hier_at(me.DRAM), 200, 'mem_hier: DRAM')
self.assertEqual(cost.mem_hier_at(me.GBUF), 6, 'mem_hier: GBUF')
self.assertEqual(cost.mem_hier_at(me.ITCN), 2, 'mem_hier: ITCN')
self.assertEqual(cost.mem_hier_at(me.REGF), 1, 'mem_hier: REGF')
def setUp(self):
self.alex_net = import_network('alex_net')
self.vgg_net = import_network('vgg_net')
net = Network('simple')
net.set_input_layer(InputLayer(4, 2))
net.add('1', ConvLayer(4, 4, 2, 1))
net.add('2', ConvLayer(4, 4, 2, 1))
# Two more layers to avoid single-segment case.
net.add('a1', ConvLayer(4, 1, 1, 1, strd=2))
net.add('a2', ConvLayer(1, 1, 1, 1))
self.simple_net = net
net = Network('complex')
net.set_input_layer(InputLayer(8, 8))
net.add('1', ConvLayer(8, 8, 8, 1))
net.add('2a', ConvLayer(8, 8, 8, 1), prevs=('1', ))
net.add('3a', ConvLayer(8, 8, 8, 1))
net.add('2b', ConvLayer(8, 8, 8, 1), prevs=('1', ))
net.add('3b', ConvLayer(8, 8, 8, 1))
net.add('4', ConvLayer(16, 8, 8, 1), prevs=('3a', '3b'))
self.complex_net = net
self.map_strategy = MapStrategyEyeriss
self.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(16, 16),
size_gbuf=128 * 1024 // 2, # 128 kB
size_regf=512 // 2, # 512 B
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False,
)
self.cost = Cost(mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=0,
idl_unit=0)
self.options = Option()
def __init__(self):
self.alex_net = import_network('alex_net')
self.mock_net = import_network('mock_net')
self.map_strategy = MapStrategyEyeriss
value_mult = {}
value_control = 1
my_weights = {}
self.cost = Cost(value_control=value_control,
value_mult=value_mult,
adder_cost=1,
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=0,
idl_unit=0,
my_weights=my_weights,
mem_cycles=(200, 6, 2, 1))
self.options = Option()
def __init__(self, mlp_network):
self.net = mlp_network #MLP_network(18,32,64,32,2)
self.map_strategy = MapStrategyEyeriss
self.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=128 * 1024 // 2, # 128 kB
size_regf=512 // 2, # 512 B
)
self.cost = Cost(mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=0,
unit_static=0)
self.options = Option()
def __init__(self):
self.alex_net = import_network('alex_net')
self.map_strategy = MapStrategyEyeriss
value_mult = {}
value_control = {}
my_weights = {}
self.cost = Cost(value_control=value_control,
value_mult=value_mult,
mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=0,
idl_unit=0,
my_weights=my_weights)
self.options = Option()
print('mapping is : {}'.format(self.map_strategy))
print('cost is: {}'.format(self.cost))
print('options are: {}'.format(self.options))
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 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 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.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_regions = (NodeRegion(origin=PhyDim2(0, 0),
dim=PhyDim2(1, 1),
type=NodeRegion.DATA), )
# Typical resource.
self.resource['BASE'] = Resource(proc_region=proc_region,
data_regions=data_regions,
dim_array=dim_array,
size_gbuf=65536,
size_regf=64)
# Larger resource with sufficient capacity, to make all schemes valid.
self.resource['LG'] = Resource(proc_region=proc_region,
data_regions=data_regions,
dim_array=dim_array,
size_gbuf=1024**3,
size_regf=1024**3)
# Small resource.
self.resource['SM'] = Resource(proc_region=proc_region,
data_regions=data_regions,
dim_array=dim_array,
size_gbuf=4096,
size_regf=16)
# Nested loop description after mapping.
self.nld = {}
self.nld['BASE'] = next(
MapStrategyEyeriss(self.layer['BASE'], self.batch_size,
dim_array).gen_nested_loop_desc())
self.nld['LGFIL'] = next(
MapStrategyEyeriss(self.layer['LGFIL'], self.batch_size,
dim_array).gen_nested_loop_desc())
self.nld['POOL'] = next(
MapStrategyEyeriss(self.layer['POOL'], self.batch_size,
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)
# 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)
# Cost.
self.cost = Cost(mac_op=1,
mem_hier=(200, 6, 2, 1),
noc_hop=50,
unit_static=50)
# Partition occupation.
self.part_occ = 0.91
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),
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(16, 16),
size_gbuf=576056 // 2, # 576 kB
size_regf=1024 // 2, # 1 kB
array_bus_width=float('inf'),
dram_bandwidth=float('inf'),
no_time_mux=False,
)
cost = Cost(mac_op=2e-12,
mem_hier=(240e-12, 28e-12, 4e-12, 1e-12), # pJ/16-b
noc_hop=0,
idl_unit=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),
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)
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)
# With dimension restriction on partitioning, this is slightly violated.
self.assertLess(dfsch_t16.total_cost, dfsch_l1.total_cost * 0.72)
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]
## 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 eyerissAsplos17(self):
'''
Reproduce TETRIS ASPLOS'17 paper Figure 8.
'''
#network = self.alex_net
network = self.mock_net
batch_size = 1
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,
num_value_pes=256,
)
# model values
print('converting weights')
q_weight_dict = {}
weights_dict = read_weights()
for w_layer in [
'conv1', 'conv2', 'conv3', 'conv4', 'conv5', 'fc6', 'fc7',
'fc8'
]:
array = convertToArray(weights_dict, w_layer)
array_qint8 = quantizeWeights(array, 'qint8')
q_weight_dict[w_layer] = array_qint8
#print('''Hey num weights in conv1 are {} '''.format(len(array_qint8)))
# hardware costs
mult_cost = readValueMult8Cost()
#control_cost = readValueControl8Cost()
print('done converting weights')
#with open('weights.pickle', 'wb') as f:
# pickle.dump(q_weight_dict,f)
#counter = 0
#c = 0
#for m in mult_cost.keys():
# c += mult_cost[m]
# counter += 1
#ave = c/counter
#print('{} '.format(counter))
#print('average = {}'.format(ave))
#print('conv3 weights are')
#for w in q_weight_dict['conv1']:
# print(w)
#exit()
cost = Cost(
value_control=1.92e-13,
value_mult=mult_cost,
mac_op=2e-12,
adder_cost=(1.178e-5) / 200000000,
mem_hier=(80e-12, 14e-12, 4e-12, 0.6e-12), # pj/16-b
noc_hop=40e-12,
idl_unit=200e-12,
my_weights=q_weight_dict,
mem_cycles=(200, 6, 2, 1))
#cost = cost(value_control=control_cost,
# value_mult=mult_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 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 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
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_regions = (NodeRegion(dim=PhyDim2(h=dim_nodes.h, w=1),
origin=PhyDim2(h=0, w=0),
type=NodeRegion.DATA),
NodeRegion(dim=PhyDim2(h=dim_nodes.h, w=1),
origin=PhyDim2(h=0, w=dim_nodes.w - 1),
type=NodeRegion.DATA))
elif args.mem_type == '3D':
# All nodes have memory.
data_regions = (NodeRegion(dim=dim_nodes,
origin=PhyDim2(0, 0),
type=NodeRegion.DATA), )
resource = Resource(proc_region=proc_region,
data_regions=data_regions,
dim_array=dim_array,
size_gbuf=size_gbuf,
size_regf=size_regf)
## 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,
unit_static=args.unit_static_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,
partition_hybrid=args.hybrid_partition,
partition_batch=args.batch_partition,
partition_ifmaps=args.ifmaps_partition,
ntops=args.top,
nprocesses=args.processes,
verbose=args.verbose)
## Search schedules.
nnd = NNDataflow(network, batch_size, resource, cost, MapStrategyEyeriss)
tops, cache_stats = nnd.schedule_search(options)
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
stats = stats_dict(top, cost)
for key, val in stats.items():
res_map[key] = val
return res_map