def do_scheduling_efficients():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(arch_file, dataflow_file)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
# batch size = 4
batch_size.init(4)
for net in all_networks():
# Network.
network = import_network(net)
print("\n============================================")
print(network.net_name)
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model, loop_lower_bound, z_fusion=True, womincost=True)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
res_parse(schedule_info_list, resource,
cost_model, sg, network,
loop_lower_bound,
'./result/overall_experiment/efficients', arch_info, True)
def do_scheduling_optimus():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(arch_file, dataflow_file)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
# Network.
for net in ['vggnet', 'squeezenet']:
for bs in [1, 2, 4, 6, 8, 16, 32]:
batch_size.init(bs)
network = import_network(net)
for layer in network.layer_dict:
network[layer].nimg = bs
print("\n============================================")
print('{}, batch size {}:'.format(network.net_name, bs))
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model,
loop_lower_bound)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
energy, access = res_parse(schedule_info_list, resource,
cost_model, sg, network,
loop_lower_bound,
'./result/batch_size/hafs', arch_info)
def do_scheduling(args):
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(args.arch, args.dataflow)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
# Network.
batch_size.init(args.batch)
network = import_network(args.net)
print(network)
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model, loop_lower_bound)
schedule_info_list, _ = sg.schedule_search()
cost, access = res_parse(schedule_info_list, resource, cost_model, sg,
network, loop_lower_bound, args.path, arch_info)
return schedule_info_list, cost
def do_scheduling_woFullSpace1():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(arch_file, dataflow_file)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
# batch size = 4
access_list = []
batch_size.init(4)
for net in all_networks():
# Network.
network = import_network(net)
print("\n============================================")
print(network.net_name)
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model, loop_lower_bound, d_fusion=True)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
_, access = res_parse(schedule_info_list, resource,
cost_model, sg, network,
loop_lower_bound,
'./result/analysis/woFullSpace1', arch_info, True, is_access=True)
access_list.append(int(access))
return access_list
def do_scheduling():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
buffer = [128, 128, 256, 256, 512, 512, 512, 512]
pe_array = [16, 32, 16, 32, 16, 16, 32, 64]
# Network.
batch_size.init(4)
network = import_network("squeezenet")
dataflow_info = extract_dataflow_info(
'./fusion/dataflow/dataflow_Ow_Cout.json')
access_list = []
energy_list = []
for pe, bf in zip(pe_array, buffer):
arch_file = './fusion/arch/3_level_mem_{}KB.json'.format(bf)
arch_info = extract_arch_info(arch_file)
arch_info["parallel_count"][0] = pe**2
if pe == 8:
arch_info["parallel_cost"][0] = 0.05
resource = Resource.arch(arch_info)
# Unroll loop lower bound
dataflow_info["partitioning_size"] = [pe] * len(
dataflow_info["partitioning_size"])
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
print("\n===========================================================")
print('PE-array: {}x{}, buffer size: {}(KB)'.format(pe, pe, bf))
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model, loop_lower_bound)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
energy, access = res_parse(schedule_info_list, resource, cost_model,
sg, network, loop_lower_bound,
'./result/pe_array', arch_info)
energy_list.append(energy)
access_list.append(access)
x = [
"16x16,128", "32x32,128", "16x16,256", "32x32,256", "8x8,512",
"16x16,512", "32x32,512", "64x64,512"
]
energy_list = np.array(energy_list) / energy_list[0]
access_list = np.array(access_list) / access_list[0]
plt.figure(figsize=(8, 2))
plt.plot(x, energy_list, label="Normalized Energy")
plt.plot(x, access_list, label="Normalized DRAM Access")
plt.ylim(0.2, 1.2)
plt.legend()
plt.savefig('./result/pe_array/pe_array.png')
plt.show()
def do_scheduling_dnnvm():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(arch_file, dataflow_file)
num_bytes = arch_info["precision"] / 8
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
# Network.
batch_size.init(4)
network = import_network("squeezenet")
buffers = [
16, 32, 64, 96, 128, 160, 192, 224, 256, 320, 384, 448, 512, 640, 768,
896, 1024
]
access_list = []
for bf in buffers:
if type(arch_info["capacity"][1]) is list:
arch_info["capacity"][1] = [bf * 1024 / num_bytes] * len(
arch_info["capacity"][1])
else:
arch_info["capacity"][1] = bf * 1024 / num_bytes
resource = Resource.arch(arch_info)
print("\n===========================================================")
print('buffer size: {}(KB)'.format(bf))
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network,
resource,
cost_model,
loop_lower_bound,
d_fusion=True,
womincost=True)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
_, access = res_parse(schedule_info_list,
resource,
cost_model,
sg,
network,
loop_lower_bound,
'./result/buffer_size/dnnvm',
arch_info,
True,
is_access=True)
access_list.append(access)
x = [str(bf) for bf in buffers]
access_list = [access / 1024 for access in access_list]
return x, access_list
def do_scheduling_woFusion(arch_file,
dataflow_file,
store_path,
is_shiDianNao=False):
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
# Resource.
arch_info, dataflow_info = extract_info(arch_file, dataflow_file)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
access_list = []
# batch size = 4
batch_size.init(4)
for net in networks:
# Network.
network = import_network(net)
# print("\n============================================")
# print(network.net_name)
# print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network,
resource,
cost_model,
loop_lower_bound,
wofusion=True,
is_shiDianNao=is_shiDianNao)
schedule_info_list, _ = sg.schedule_search()
# print("done!\n\n")
_, access = res_parse(schedule_info_list,
resource,
cost_model,
sg,
network,
loop_lower_bound,
store_path,
arch_info,
True,
is_access=True,
is_print=False)
access_list.append(access)
return access_list
def do_scheduling():
"""
Get optimal scheduling for given problem. Return a result schedule.
"""
dataflow_dir = './fusion/dataflow/'
name = os.listdir(dataflow_dir)
# batch size = 4
batch_size.init(4)
network = import_network("squeezenet")
for rfs in [64, 512]:
print("\n\n" + "*" * 80)
print("\nRFs: {}B/PE".format(rfs))
arch = './fusion/arch/3_level_mem_{}Reg.json'.format(rfs)
for dataflow in name:
if dataflow[-4:] == "json":
# Resource.
arch_info, dataflow_info = extract_info(
arch, dataflow_dir + dataflow)
resource = Resource.arch(arch_info)
# Unroll loop lower bound
loop_lower_bound = LoopLowerBound.dataflow(dataflow_info)
print("\n")
print("=" * 50)
print(dataflow[:-5])
print("waiting...")
cost_model = CostModel(network, resource)
# optimal schedule
sg = ScheduleGenerator(network, resource, cost_model,
loop_lower_bound)
schedule_info_list, _ = sg.schedule_search()
print("done!\n\n")
energy, access = res_parse(schedule_info_list, resource,
cost_model, sg, network,
loop_lower_bound,
'./result/dataflow', arch_info)