forked from hsharma35/bitfusion
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtop.py
170 lines (139 loc) · 6.44 KB
/
top.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
import pandas
import ConfigParser
import os
import sys
import numpy as np
from bitfusion.graph_plot.barchart import BarChart
import matplotlib
import warnings
warnings.filterwarnings('ignore')
import bitfusion.src.benchmarks.benchmarks as benchmarks
from bitfusion.src.simulator.stats import Stats
from bitfusion.src.simulator.simulator import Simulator
from bitfusion.src.sweep.sweep import SimulatorSweep, check_pandas_or_run
from bitfusion.src.utils.utils import *
from bitfusion.src.optimizer.optimizer import optimize_for_order, get_stats_fast
import dnnweaver2
batch_size = int(sys.argv[2])
results_dir = './results'
if not os.path.exists(results_dir):
os.makedirs(results_dir)
fig_dir = './fig'
if not os.path.exists(fig_dir):
os.makedirs(fig_dir)
# BitFusion configuration file
#config_file = 'bf_e_conf.ini'
config_file = sys.argv[1]
print("Using config file: {}".format(config_file))
# Create simulator object
verbose = False
bf_e_sim = Simulator(config_file, verbose)
bf_e_energy_costs = bf_e_sim.get_energy_cost()
print(bf_e_sim)
energy_tuple = bf_e_energy_costs
print('')
print('*'*50)
print(energy_tuple)
sim_sweep_columns = ['N', 'M',
'Max Precision (bits)', 'Min Precision (bits)',
'Network', 'Layer',
'Cycles', 'Memory wait cycles',
'WBUF Read', 'WBUF Write',
'OBUF Read', 'OBUF Write',
'IBUF Read', 'IBUF Write',
'DRAM Read', 'DRAM Write',
'Bandwidth (bits/cycle)',
'WBUF Size (bits)', 'OBUF Size (bits)', 'IBUF Size (bits)',
'Batch size']
filename = 'bitfusion-eyeriss-sim-sweep'+bf_e_sim.suffix+"_b"+str(batch_size)+'.csv'
bf_e_sim_sweep_csv = os.path.join(results_dir, filename)
if os.path.exists(bf_e_sim_sweep_csv):
print("{} does not exist. Good".format(filename))
bf_e_sim_sweep_df = pandas.read_csv(bf_e_sim_sweep_csv)
else:
bf_e_sim_sweep_df = pandas.DataFrame(columns=sim_sweep_columns)
print('Got BitFusion Eyeriss, Numbers')
bf_e_results = check_pandas_or_run(bf_e_sim, bf_e_sim_sweep_df,
bf_e_sim_sweep_csv, batch_size=batch_size, config_file=config_file)
bf_e_results = bf_e_results.groupby('Network',as_index=False).agg(np.sum)
area_stats = bf_e_sim.get_area()
print("Done")
def get_eyeriss_energy(df):
eyeriss_energy_per_mac = 16 * 0.2 * 1.e-3 #energy in nJ
eyeriss_energy_alu = float(df['ALU'])
eyeriss_energy_dram = float(df['DRAM']) * 0.15 #Scaling due to technology
eyeriss_energy_buffer = float(df['Buffer'])
eyeriss_energy_array = float(df['Array'])
eyeriss_energy_rf = float(df['RF'])
eyeriss_energy = eyeriss_energy_alu + eyeriss_energy_dram + eyeriss_energy_buffer + eyeriss_energy_array + eyeriss_energy_rf
eyeriss_energy *= eyeriss_energy_per_mac
return eyeriss_energy
def get_eyeriss_energy_breakdown(df):
eyeriss_energy_per_mac = 16 * 0.2 * 1.e-3 #energy in nJ
eyeriss_energy_alu = float(df['ALU'])
eyeriss_energy_dram = float(df['DRAM'])
eyeriss_energy_buffer = float(df['Buffer'])
eyeriss_energy_array = float(df['Array'])
eyeriss_energy_rf = float(df['RF'])
eyeriss_energy = [eyeriss_energy_alu+eyeriss_energy_array, eyeriss_energy_buffer, eyeriss_energy_rf, eyeriss_energy_dram]
eyeriss_energy = [x * eyeriss_energy_per_mac for x in eyeriss_energy]
return eyeriss_energy
def df_to_stats(df):
stats = Stats()
stats.total_cycles = float(df['Cycles'])
stats.mem_stall_cycles = float(df['Memory wait cycles'])
stats.reads['act'] = float(df['IBUF Read'])
stats.reads['out'] = float(df['OBUF Read'])
stats.reads['wgt'] = float(df['WBUF Read'])
stats.reads['dram'] = float(df['DRAM Read'])
stats.writes['act'] = float(df['IBUF Write'])
stats.writes['out'] = float(df['OBUF Write'])
stats.writes['wgt'] = float(df['WBUF Write'])
stats.writes['dram'] = float(df['DRAM Write'])
return stats
# Eyeriss-Simulator
eyeriss_data = pandas.read_csv(os.path.join(results_dir, 'eyeriss_results.csv'))
eyeriss_data_bench = eyeriss_data.groupby('Benchmark', as_index=False).agg(np.sum)
eyeriss_data_bench['Platform'] = 'Eyeriss (16-bit)'
print('BitFusion-Eyeriss comparison')
eyeriss_area = 3.5*3.5*45*45/65./65.
print('Area budget = {}'.format(eyeriss_area))
print(area_stats)
if abs(sum(area_stats)-eyeriss_area)/eyeriss_area > 0.1:
print('Warning: BitFusion Area is outside 10% of eyeriss')
print('total_area = {}, budget = {}'.format(sum(area_stats), eyeriss_area))
bf_e_area = sum(area_stats)
baseline_data = []
for bench in benchmarks.benchlist:
lookup_dict = {'Benchmark': bench}
eyeriss_cycles = float(lookup_pandas_dataframe(eyeriss_data_bench, lookup_dict)['time(ms)'])
eyeriss_time = eyeriss_cycles / 500.e3 / 16
eyeriss_energy = get_eyeriss_energy(lookup_pandas_dataframe(eyeriss_data_bench, lookup_dict))
eyeriss_power = eyeriss_energy / eyeriss_time * 1.e-9
eyeriss_speedup = eyeriss_time / eyeriss_time
eyeriss_energy_efficiency = eyeriss_energy / eyeriss_energy
eyeriss_ppa = eyeriss_speedup / eyeriss_area / (eyeriss_speedup / eyeriss_area)
eyeriss_ppw = eyeriss_speedup / eyeriss_power / (eyeriss_speedup / eyeriss_power)
bf_e_stats = df_to_stats(bf_e_results.loc[bf_e_results['Network'] == bench])
bf_e_cycles = bf_e_stats.total_cycles * (batch_size / 16.)
bf_e_time = bf_e_cycles / (bf_e_sim.accelerator.frequency / 1e3)
bf_e_energy = bf_e_stats.get_energy(bf_e_sim.get_energy_cost()) * (batch_size / 16.)
bf_e_power = (bf_e_energy / bf_e_time) * 1e-3
bf_e_speedup = eyeriss_time / bf_e_time
bf_e_energy_efficiency = eyeriss_energy / bf_e_energy
bf_e_ppa = bf_e_speedup / bf_e_area / (eyeriss_speedup / eyeriss_area)
bf_e_ppw = bf_e_speedup / bf_e_power / (eyeriss_speedup / eyeriss_power)
baseline_data.append(['Performance', bench, bf_e_speedup])
baseline_data.append(['Energy reduction', bench, bf_e_energy_efficiency])
baseline_data.append(['Performance-per-Watt', bench, bf_e_ppw])
baseline_data.append(['Performance-per-Area', bench, bf_e_ppa])
print('*'*50)
print('Benchmark: {}'.format(bench))
#print('Eyeriss time: {} ms'.format(eyeriss_time))
print('BitFusion time: {} ms'.format(bf_e_time))
#print('Eyeriss power: {} mWatt'.format(eyeriss_power*1.e3*16))
print('BitFusion power: {} mWatt'.format(bf_e_power))
print('BitFusion energy: {} nJ'.format(bf_e_energy))
print('BitFusion total_cycles: {} '.format(bf_e_cycles))
print('*'*50)
eyeriss_comparison_df = pandas.DataFrame(baseline_data, columns=['Metric', 'Network', 'Value'])