-
Notifications
You must be signed in to change notification settings - Fork 0
/
analyse.py
237 lines (199 loc) · 9.21 KB
/
analyse.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
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
import argparse
import enum
import re
import sys
from types import prepare_class
'''--------------------------------------------------------------
Tools to analyse tflite benchmark_model profiling output csv file.
--------------------------------------------------------------'''
def _replace_flex(name: str, type: str):
op = name.split(':')[0].split('/')[-1]
op = op.lower()
if type == 'swin':
if 'transpose' in op: return 'TRANSPOSE'
if 'add' in op: return 'ADDv2'
if 'roll' in op: return 'ROLL'
if 'erf' in op: return 'ERF'
if type == 't2t_vit':
if 'einsum' in op: return 'EINSUM'
if 'extractimagepatches' in op: return 'EXTRACTIMAGEPATCHES'
return 'TFFLEXDELEGATE'
def _find_op_wise_line_range(rows):
schema = {}
for begin_line in range(len(rows)):
row = rows[begin_line]
if len(row) == 1 and 'Operator-wise Profiling Info for Regular Benchmark Run' in row[0]:
schema_row = rows[begin_line + 2]
schema = {schema_row[i].strip(): i for i in range(len(schema_row))}
begin_line += 3
break
end_line = begin_line
while True:
if len(rows[end_line]) < len(schema):
break
end_line += 1
return begin_line, end_line, schema
def _read_rows(file_path):
import csv
rows = []
with open(file_path) as csv_file:
csv_reader = csv.reader(csv_file, delimiter=',')
for row in csv_reader:
rows.append(row)
return rows
def analyse_op(parser: argparse.ArgumentParser):
parser.add_argument('--file', type=str, required=True, help='csv profile result file')
parser.add_argument('--type', choices=['swin', 't2t_vit'], required=True, help='transformer model type')
args = parser.parse_args()
rows = _read_rows(args.file)
begin_line, end_line, schema = _find_op_wise_line_range(rows)
print(f'Schema: {schema}')
result_table = {}
for row in rows[begin_line: end_line]:
node_type = row[schema['node type']]
if 'TfLiteFlexDelegate' in node_type:
node_type = _replace_flex(row[schema['name']], args.type)
if node_type in result_table.keys():
result_table[node_type]['latency'] += float(row[schema['avg_ms']])
result_table[node_type]['percent'] += float(row[schema['%']][:-1])
else:
result_table[node_type] = {}
result_table[node_type]['latency'] = float(row[schema['avg_ms']])
result_table[node_type]['percent'] = float(row[schema['%']][:-1])
for k, v in result_table.items():
print(f'{k} {v["latency"]: .2f} {v["percent"]: .2f}')
def analyse_gelu_ln(parser: argparse.ArgumentParser):
import csv
parser.add_argument('--file', type=str, required=True, help='csv profile result file')
parser.add_argument('--type', choices=['deit', 'swin', 't2t_vit'], required=True, help='the transformer model type')
args = parser.parse_args()
rows = _read_rows(args.file)
begin_line, end_line, schema = _find_op_wise_line_range(rows)
print(f'Schema: {schema}')
gelu_latency = 0
gelu_percent = 0
ln_latency = 0
ln_percent = 0
hit_gelu = 0
hit_ln = 0
for i in range(begin_line, end_line):
row = rows[i]
node_type = row[schema['node type']]
if args.type == 'deit':
if 'POW' in node_type:
hit_gelu += 1
for j in range(8):
gelu_latency += float(rows[i + j][schema['avg_ms']])
gelu_percent += float(rows[i + j][schema['%']][:-1])
if 'FULLY_CONNECTED' not in node_type and 'RESHAPE' not in node_type and 'layer_normalization' in rows[i][schema['name']]:
hit_ln += 1
ln_latency += float(rows[i][schema['avg_ms']])
ln_percent += float(rows[i][schema['%']][:-1])
elif args.type == 'swin':
if 'gelu' in rows[i][schema['name']].lower():
hit_gelu += 1
gelu_latency += float(rows[i][schema['avg_ms']])
gelu_percent += float(rows[i][schema['%']][:-1])
if 'norm' in rows[i][schema['name']].lower():
hit_ln += 1
ln_latency += float(rows[i][schema['avg_ms']])
ln_percent += float(rows[i][schema['%']][:-1])
elif args.type == 't2t_vit':
if 'POW' in node_type:
hit_gelu += 1
for j in range(8):
gelu_latency += float(rows[i + j][schema['avg_ms']])
gelu_percent += float(rows[i + j][schema['%']][:-1])
if 'layer_normalization' in rows[i][schema['name']].lower():
hit_ln += 1
ln_latency += float(rows[i][schema['avg_ms']])
ln_percent += float(rows[i][schema['%']][:-1])
print('hit_gelu {} hit_ln {} gelu_latency {:.2f} gelu_percent {:.2f} ln_latency {:.2f} ln_percent {:.2f}'.format(
hit_gelu, hit_ln, gelu_latency, gelu_percent, ln_latency, ln_percent))
def analyse_attn_ffn(parser: argparse.ArgumentParser):
import csv
parser.add_argument('--file', type=str, required=True, help='csv profile result file')
parser.add_argument('--type', choices=['deit', 'swin', 't2t_vit'], required=True, help='the transformer model type')
args = parser.parse_args()
rows = _read_rows(args.file)
begin_line, end_line, schema = _find_op_wise_line_range(rows)
rows = sorted(rows[begin_line: end_line], key=lambda row: float(row[schema['start']]))
print(f'Schema: {schema}')
attn_percent = 0
attn_latency = 0
ffn_percent = 0
ffn_latency = 0
pre_post_processing_percent = 0
pre_post_processing_latency = 0
if args.type == 'deit' or args.type == 't2t_vit':
pre_ln_str = 'Null'
is_ffn = 1
for row in rows:
if 'transformer_encoder_block' not in row[schema['name']]:
pre_post_processing_latency += float(row[schema['avg_ms']])
pre_post_processing_percent += float(row[schema['%']][:-1])
else:
ln_str = re.match(r'.*/(layer_norm_?\d*)/.*', row[schema['name']]).groups()[0]
if ln_str != pre_ln_str:
pre_ln_str = ln_str
is_ffn = (is_ffn + 1) % 2
if is_ffn == 0:
attn_latency += float(row[schema['avg_ms']])
attn_percent += float(row[schema['%']][:-1])
else:
ffn_latency += float(row[schema['avg_ms']])
ffn_percent += float(row[schema['%']][:-1])
else: # swin
for row in rows:
if 'swin_transformer_block' not in row[schema['name']]:
pre_post_processing_latency += float(row[schema['avg_ms']])
pre_post_processing_percent += float(row[schema['%']][:-1])
elif 'window_attention' in row[schema['name']] or 'norm1' in row[schema['name']]:
attn_latency += float(row[schema['avg_ms']])
attn_percent += float(row[schema['%']][:-1])
elif 'mlp' in row[schema['name']] or 'norm2' in row[schema['name']]:
ffn_latency += float(row[schema['avg_ms']])
ffn_percent += float(row[schema['%']][:-1])
elif 'norm' not in row[schema['name']]:
pre_post_processing_latency += float(row[schema['avg_ms']])
pre_post_processing_percent += float(row[schema['%']][:-1])
else:
raise RuntimeError()
print(f'{args.type} | attn (percent, latency) = ({attn_percent:.2f}, {attn_latency:.2f}) | ' +
f'ffn (percent, latency) = ({ffn_percent:.2f}, {ffn_latency:.2f}) | ' +
f'pre & post-processing (percent, latency) = ({pre_post_processing_percent:.2f}, {pre_post_processing_latency:.2f})')
def fetch_all_op_latency(parser: argparse.ArgumentParser):
import csv
parser.add_argument('--file', type=str, required=True, help='csv profile result file')
parser.add_argument('--op', choices=['conv', 'dwconv', 'dense'], required=True, help='op type to fetch latency')
args = parser.parse_args()
OP_NAME_DICT = {
'conv': 'CONV_2D',
'dwconv': 'DEPTHWISE_CONV_2D',
'dense': 'FULLY_CONNECTED'
}
rows = _read_rows(args.file)
begin_line, end_line, schema = _find_op_wise_line_range(rows)
latency_list = []
rows = sorted(rows[begin_line: end_line], key=lambda row: float(row[schema['start']]))
for row in rows:
node_type = row[schema['node type']]
if node_type == OP_NAME_DICT[args.op]:
latency_list.append(round(float(row[schema['avg_ms']]), 2))
print(f'{args.op} count = {len(latency_list)}')
print(latency_list)
function_dict = {
'analyse_op': analyse_op,
'analyse_gelu_ln': analyse_gelu_ln,
'analyse_attn_ffn': analyse_attn_ffn,
'fetch_all_op_latency': fetch_all_op_latency
}
if __name__ == '__main__':
assert len(sys.argv) > 1
parser = argparse.ArgumentParser()
parser.add_argument('func', type=str, help='specify the work to do')
func = sys.argv[1]
if func in function_dict.keys():
function_dict[func](parser)
else:
raise ValueError(f'Function {func} not support. Supported functions: {list(function_dict.keys())}')