import os
import numpy as np
import argparse
import math
import time
import cnn_mapping as cm
import datetime
import json
def basic_optimizer(arch_info, network_info, schedule_info=None, basic=False, verbose=False):
resource = cm.Resource.arch(arch_info)
layer = cm.Layer.layer(network_info)
schedule = cm.Schedule.schedule(schedule_info) if schedule_info != None else None
smallest_cost, best_mapping_point, perf = cm.optimizer.opt_optimizer(resource, layer, schedule, verbose)
level_costs = cm.cost_model.get_cost(resource, best_mapping_point, layer, verbose)
#if verbose or basic:
if not verbose:
print("Optimal_energy_(pJ): ", smallest_cost)
#print("Cost for Each Level (pJ): ", level_costs) #TODO
print("Best_Schedule_(innermost_loop_to_outermost_loop): ")
cm.utils.print_best_schedule(best_mapping_point)
return smallest_cost, perf
def mem_explore_optimizer(arch_info, network_info, schedule_info, verbose=False):
assert "explore_points" in arch_info, "missing explore_points in arch file"
assert "capacity_scale" in arch_info, "missing capacity_scale in arch file"
assert "access_cost_scale" in arch_info, "missing access_cost_scale in arch file"
cwd = os.getcwd()
# output_filename = os.path.join(cwd, "dataset", network_info['layer_name'] + '_128.csv')
explore_points = arch_info["explore_points"]
energy_list = np.zeros(tuple(explore_points))
summary_array = np.zeros([np.product(explore_points), 12])
#TODO support more than two levels of explorations
capacity0 = arch_info["capacity"][0]
capacity1 = arch_info["capacity"][1]
cost0 = arch_info["access_cost"][0]
cost1 = arch_info["access_cost"][1]
i = 0
for x in xrange(explore_points[0]):
arch_info["capacity"][0] = capacity0 * (arch_info["capacity_scale"][0]**x)
arch_info["access_cost"][0] = cost0 * (arch_info["access_cost_scale"][0]**x)
for y in xrange(explore_points[1]):
#if x == 0 and y < 1:
# continue
arch_info["capacity"][1] = capacity1 * (arch_info["capacity_scale"][1]**y)
arch_info["access_cost"][1] = cost1 * (arch_info["access_cost_scale"][1]**y)
print(arch_info)
energy, perf = basic_optimizer(arch_info, network_info, schedule_info, False, verbose)
energy_list[x][y] = energy
cur_point = network_info["layer_info"] + arch_info["capacity"][:-1] + [energy]
summary_array[i] = cur_point
# np.savetxt(output_filename, summary_array, delimiter=",")
i += 1
print(list(energy_list))
print("optimal energy for all memory systems: ", np.min(np.array(energy_list)))
def mac_explore_optimizer(arch_info, network_info, schedule_info, verbose=False):
dataflow_res = []
#TODO check the case when parallel count larger than layer dimension size
dataflow_generator = dataflow_generator_function(arch_info)
for dataflow in dataflow_generator:
energy, perf = basic_optimizer(arch_info, network_info, schedule_info, False, verbose)
dataflow_res.append[energy]
if verbose:
print("optimal energy for all dataflows: ", dataflow_res)
return dataflow_res
def dataflow_explore_optimizer(arch_info, network_info, file_name, verbose=False):
assert arch_info["parallel_count"] > 1, \
"parallel count has to be more than 1 for dataflow exploration"
resource = cm.Resource.arch(arch_info)
layer = cm.Layer.layer(network_info)
dataflow_tb = cm.mapping_point_generator.dataflow_exploration(resource, layer, file_name, verbose)
if verbose:
print("dataflow table done ")
def mem_explore_optimizer_4_level(arch_info, network_info, schedule_info, verbose=False):
assert "explore_points" in arch_info, "missing explore_points in arch file"
assert "capacity_scale" in arch_info, "missing capacity_scale in arch file"
assert "access_cost_scale" in arch_info, "missing access_cost_scale in arch file"
cwd = os.getcwd()
# output_filename = os.path.join(cwd, "dataset", network_info['layer_name'] + '_128.csv')
explore_points = arch_info["explore_points"]
energy_list = np.zeros(tuple(explore_points))
summary_array = np.zeros([np.product(explore_points), 13])
capacity0 = arch_info["capacity"][0]
capacity1 = arch_info["capacity"][1]
capacity2 = arch_info["capacity"][2]
cost0 = arch_info["access_cost"][0]
cost1 = arch_info["access_cost"][1]
cost2 = arch_info["access_cost"][2]
i = 0
for x in xrange(explore_points[0]):
arch_info["capacity"][0] = capacity0 * (arch_info["capacity_scale"][0]**x)
arch_info["access_cost"][0] = cost0 * (arch_info["access_cost_scale"][0]**x)
for y in xrange(explore_points[1]):
arch_info["capacity"][1] = capacity1 * (arch_info["capacity_scale"][1]**y)
arch_info["access_cost"][1] = cost1 * (arch_info["access_cost_scale"][1]**y)
for z in xrange(explore_points[2]):
arch_info["capacity"][2] = capacity2 * (arch_info["capacity_scale"][2]**z)
arch_info["access_cost"][2] = cost2 * (arch_info["access_cost_scale"][2]**z)
print(arch_info)
# TODO: Calculate the area of the design
# total_area = sum(area(MAC)) + sum(2 * area(register file)) + 2 * area(buffer)
area = (2 * (arch_info["capacity"][0] + arch_info["capacity"][1] + arch_info["capacity"][2]) \
+ 2 * arch_info["capacity"][3]) / 1e9
print("Area: ", area)
if area > 2:
energy_list[x][y][z] = np.nan
print("INFO: Architecture does not fulfill area constraint.")
continue
try:
energy, perf = basic_optimizer(arch_info, network_info, schedule_info, False, verbose)
energy_list[x][y][z] = energy
cur_point = network_info["layer_info"] + arch_info["capacity"][:-1] + [energy]
summary_array[i] = cur_point
# np.savetxt(output_filename, summary_array, delimiter=",")
i += 1
except Exception as e:
energy_list[x][y][z] = np.nan
print("WARNING: No valid mapping point found.")
print("=" * 80)
print(list(energy_list))
print("optimal energy for all memory systems: ", np.nanmin(np.array(energy_list)))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("type", choices=["basic", "mem_explore", "dataflow_explore", "mem_explore_4"], help="optimizer type")
parser.add_argument("arch", help="architecture specification")
parser.add_argument("network", help="network specification")
parser.add_argument("-s", "--schedule", help="restriction of the schedule space")
parser.add_argument("-n", "--name", default="dataflow_table", help="name for the dumped pickle file")
parser.add_argument("-v", "--verbose", action='count', help="vebosity")
parser.add_argument("-j", "--json_name", default="result.json", help="result json file name for basic search")
args = parser.parse_args()
if args.json_name == "result.json":
json_name = "result_" + datetime.datetime.now().strftime("%m-%d_%H_%M_%S") + ".json"
else:
json_name = args.json_name
start = time.time()
arch_info, network_info, schedule_info = cm.extract_input.extract_info(args)
if args.type == "basic":
energy, perf = basic_optimizer(arch_info, network_info, schedule_info, True, args.verbose)
json_data = {}
json_data["runtime"] = perf
json_data["energy"] = energy
json_data["file_arch"] = args.arch
json_data["file_layer"] = args.network
with open(json_name, "w") as jf:
json.dump(json_data, jf)
elif args.type == "mem_explore":
mem_explore_optimizer(arch_info, network_info, schedule_info, args.verbose)
elif args.type == "dataflow_explore":
dataflow_explore_optimizer(arch_info, network_info, args.name, args.verbose)
elif args.type == "mem_explore_4":
mem_explore_optimizer_4_level(arch_info, network_info, schedule_info, args.verbose)
end = time.time()
print("Elapsed_Time_(s): ", round((end-start), 2))