-
Notifications
You must be signed in to change notification settings - Fork 20
/
performance_latency.py
346 lines (271 loc) · 13.6 KB
/
performance_latency.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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
# ====TensorRT Hackathon 2022 Medcare AI Lab====
#
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
#
'''
mst++ benchmark的统计和可视化包括：Latency， ThroughPut, 加速比等（pytorch, onnxruntime, TensorRT ONNXParser, TensorRT API)
'''
import os
import argparse
import ctypes
from glob import glob
import torch
import architecture
import onnxruntime
import numpy as np
import tensorrt as trt
from cuda import cudart
from time import time_ns,sleep
# import seaborn as sns
import matplotlib.pyplot as plt
import matplotlib
matplotlib.use('Agg')
plt.style.use('ggplot')
files = os.listdir("./data")
soFileList = glob("./" + "*.so")
# trt
logger = trt.Logger(trt.Logger.VERBOSE) # trt.Logger.VERBOSE
trt.init_libnvinfer_plugins(logger, '')
PLUGIN_CREATORS = trt.get_plugin_registry().plugin_creator_list
for soFile in soFileList:
ctypes.cdll.LoadLibrary(soFile)
EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
builder = trt.Builder(logger)
network = builder.create_network(EXPLICIT_BATCH)
config = builder.create_builder_config()
config.max_workspace_size = 1 << 30
def time_static_torch(input_data,model_path='./model/mst_plus_plus.pth',batch_size=1,nRound=1000):
#pytorch
torch_model = architecture.MST_Plus_Plus() # 导入模型
check_point = torch.load(model_path,map_location='cpu')
# torch_model.load_state_dict(check_point['state_dict']) # 初始化权重
torch_model.load_state_dict({k.replace('module.', ''): v for k, v in check_point['state_dict'].items()},
strict=True)
# torch_model.half()
torch_model.eval()
torch_model.to("cuda")
torch_input = torch.from_numpy(input_data).cuda()
for i in range(50):
with torch.no_grad():
torch_output = torch_model(torch_input)
torch.cuda.synchronize()
t0 = time_ns()
for i in range(nRound):
with torch.no_grad():
torch_output = torch_model(torch_input)
# torch.cuda.empty_cache()
torch.cuda.synchronize()
t1 = time_ns()
Latency_pytorch = (t1 - t0)/1000/1000/nRound
Throughput_pytorch = 1000/Latency_pytorch*batch_size
# 清空释放显存
del torch_model
torch_input.cpu()
del torch_input
torch.cuda.empty_cache()
return Latency_pytorch, Throughput_pytorch
def time_static_onnx(input_data,onnx_path,batch_size=1,nRound=1000):
ort_session = onnxruntime.InferenceSession(onnx_path,providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
# providers=['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']
ort_inputs = {"input":input_data}
output_name = ort_session.get_outputs()[0].name
for i in range(50):
output = ort_session.run([output_name],ort_inputs)
torch.cuda.synchronize()
t0 = time_ns()
for i in range(nRound):
output = ort_session.run(None,ort_inputs)
torch.cuda.synchronize()
t1 = time_ns()
Latency_onnx = (t1 - t0)/1000/1000/nRound
Throughput_onnx = 1000/Latency_onnx*batch_size
return Latency_onnx, Throughput_onnx
def time_static_trt(input_data,plan_path,batch_size=1,nRound=1000):
with open(plan_path, "rb") as f:
engine = trt.Runtime(logger).deserialize_cuda_engine(f.read())
context = engine.create_execution_context()
inputH0 = np.ascontiguousarray(input_data.ravel())
_, inputD0 = cudart.cudaMalloc(inputH0.nbytes)
cudart.cudaMemcpy(inputD0, inputH0.ctypes.data, inputH0.nbytes, cudart.cudaMemcpyKind.cudaMemcpyHostToDevice)
outputH0 = np.empty(context.get_binding_shape(1),dtype=trt.nptype(engine.get_binding_dtype(1)))
_, outputD0 = cudart.cudaMalloc(outputH0.nbytes)
for i in range(50):
context.execute_v2([int(inputD0),int(outputD0)])
torch.cuda.synchronize()
t0 = time_ns()
for i in range(nRound):
context.execute_v2([int(inputD0),int(outputD0)])
torch.cuda.synchronize()
t1 = time_ns()
Latency_trt = (t1 - t0)/1000/1000/nRound
Throughput_trt = 1000/Latency_trt*batch_size
cudart.cudaMemcpy(outputH0.ctypes.data, outputD0, outputH0.nbytes, cudart.cudaMemcpyKind.cudaMemcpyDeviceToHost)
cudart.cudaFree(inputD0)
cudart.cudaFree(outputD0)
return Latency_trt, Throughput_trt
if __name__ == "__main__":
# # Latency and Throughput
# Pytorch batch size = 32时 out of memory, 因此我们仅对比了batch size是16以下的batch
batch_sizes = [16,8,4,2,1]
# FP32
static_torch_32 = {'batch_size':[],'latency':[],'throughput':[],'LSU':[],"TSU":[]}
static_trt_onnx_32 = {'batch_size':[],'latency':[],'throughput':[],'LSU':[],"TSU":[]}
static_trt_onnxparser_32 = {'batch_size':[],'latency':[],'throughput':[],'LSU':[],"TSU":[]}
static_trt_onnxparser_16 = {'batch_size':[],'latency':[],'throughput':[],'LSU':[],"TSU":[]}
static_trt_api_32 = {'batch_size':[],'latency':[],'throughput':[],'LSU':[],"TSU":[]}
static_trt_api_16 = {'batch_size':[],'latency':[],'throughput':[],'LSU':[],"TSU":[]}
for batch_size in batch_sizes:
print(f"[INFO] 当前测试的batch size为：{batch_size}")
np.random.seed(2022)
input_data = np.random.uniform(-1,1,(batch_size, 3, 512, 482)).astype(np.float32)
# torch
print("[INFO] 正在进行pytorch测试")
l_torch,t_torch = time_static_torch(input_data=input_data,model_path='./model/mst_plus_plus.pth',batch_size=batch_size,nRound=1000)
sleep(10)
# onnxruntime
if batch_size == 16: # batch_size=16
l_onnx = 0
t_onnx = 0
lsu_onnx = 1
tsu_onnx = 1
else:
print("[INFO] 正在进行onnxruntime测试")
l_onnx,t_onnx = time_static_onnx(input_data=input_data,onnx_path=f'./model/mst_plus_plus_b{batch_size}.onnx',batch_size=batch_size,nRound=1000)
lsu_onnx = round( l_torch / l_onnx,2)
tsu_onnx = round( t_onnx/ t_torch ,2)
sleep(10)
# fp32 trt api
print("[INFO] 正在进行trt api FP32测试")
l_trt,t_trt = time_static_trt(input_data=input_data,plan_path=f"./model/mst_plus_plus_b{batch_size}.plan",batch_size=batch_size,nRound=1000)
lsu = round( l_torch / l_trt,2)
tsu = round( t_trt / t_torch ,2)
sleep(10)
# fp32 trt onnxparser
print("[INFO] 正在进行trt onnxparser FP32测试")
l_trt_1,t_trt_1 = time_static_trt(input_data=input_data,plan_path=f"./model/mst_plus_plus_onnxparser_b{batch_size}.plan",batch_size=batch_size,nRound=1000)
lsu_1 = round(l_torch / l_trt_1,2)
tsu_1 = round(t_trt_1 / t_torch ,2)
sleep(10)
# fp16 trt api
print("[INFO] 正在进行trt api FP16测试")
l_trt_16,t_trt_16 = time_static_trt(input_data=input_data,plan_path=f"./model/mst_plus_plus_onnxparser_b{batch_size}_fp16.plan",batch_size=batch_size,nRound=1000)
lsu_16 = round(l_torch /l_trt_16,2)
tsu_16 = round(t_trt_16 /t_torch,2)
sleep(10)
# fp16 trt onnxparser
print("[INFO] 正在进行trt onnxparser FP16测试")
l_trt_16_1,t_trt_16_1 = time_static_trt(input_data=input_data,plan_path=f"./model/mst_plus_plus_b{batch_size}_fp16.plan",batch_size=batch_size,nRound=1000)
lsu_16_1 = round(l_torch / l_trt_16_1,2)
tsu_16_1 = round(t_trt_16_1 / t_torch,2)
sleep(10)
#pytorch
static_torch_32['batch_size'].append(batch_size)
static_torch_32['latency'].append(l_torch)
static_torch_32['throughput'].append(t_torch)
static_torch_32['LSU'].append("1x")
static_torch_32['TSU'].append("1x")
#onnx
static_trt_onnx_32['batch_size'].append(batch_size)
static_trt_onnx_32['latency'].append(l_onnx)
static_trt_onnx_32['throughput'].append(t_onnx)
static_trt_onnx_32['LSU'].append(str(lsu_onnx)+"x")
static_trt_onnx_32['TSU'].append(str(tsu_onnx)+"x")
# trt api
static_trt_api_32['batch_size'].append(batch_size)
static_trt_api_32['latency'].append(l_trt)
static_trt_api_32['throughput'].append(t_trt)
static_trt_api_32['LSU'].append(str(lsu)+"x")
static_trt_api_32['TSU'].append(str(tsu)+"x")
static_trt_api_16['batch_size'].append(batch_size)
static_trt_api_16['latency'].append(l_trt_16)
static_trt_api_16['throughput'].append(t_trt_16)
static_trt_api_16['LSU'].append(str(lsu_16)+"x")
static_trt_api_16['TSU'].append(str(tsu_16)+"x")
#trt onnxparser
static_trt_onnxparser_32['batch_size'].append(batch_size)
static_trt_onnxparser_32['latency'].append(l_trt_1)
static_trt_onnxparser_32['throughput'].append(t_trt_1)
static_trt_onnxparser_32['LSU'].append(str(lsu_1)+"x")
static_trt_onnxparser_32['TSU'].append(str(tsu_1)+"x")
static_trt_onnxparser_16['batch_size'].append(batch_size)
static_trt_onnxparser_16['latency'].append(l_trt_16_1)
static_trt_onnxparser_16['throughput'].append(t_trt_16_1)
static_trt_onnxparser_16['LSU'].append(str(lsu_16_1)+"x")
static_trt_onnxparser_16['TSU'].append(str(tsu_16_1)+"x")
print("-"*50)
print("torch:")
print(static_torch_32)
print("onnxruntime")
print(static_trt_onnx_32)
print("trt api fp32:")
print(static_trt_api_32)
print("trt api fp16:")
print(static_trt_api_16)
print("trt onnxparser fp32:")
print(static_trt_onnxparser_32)
print("trt onnxparser fp16:")
print(static_trt_onnxparser_16)
print("-"*50)
# plot latency vs throughput
torch_x = static_torch_32['latency']
torch_y = static_torch_32['throughput']
onnx_x = static_trt_onnx_32['latency']
onnx_y = static_trt_onnx_32['throughput']
trt_api_32_x = static_trt_api_32['latency']
trt_api_32_y = static_trt_api_32['throughput']
trt_api_16_x = static_trt_api_16['latency']
trt_api_16_y = static_trt_api_16['throughput']
trt_onnx_32_x = static_trt_onnxparser_32['latency']
trt_onnx_32_y = static_trt_onnxparser_32['throughput']
trt_onnx_16_x = static_trt_onnxparser_16['latency']
trt_onnx_16_y = static_trt_onnxparser_16['throughput']
plt.rcParams['figure.figsize'] = (16.0, 9.0)
# pytorch
plt.plot(torch_x, torch_y, 'ro--',label='Pytorch',markersize=7,linewidth=2)
for i,(a, b) in enumerate(zip(torch_x, torch_y)):
# plt.text(a+15,b-0.15,'(%d,%d,%d)'%(batch_size[i],a,b),ha='center', va='bottom',fontdict={'size': 10, 'color': 'r'})
plt.text(a+15,b-0.15,f'Batch:{batch_sizes[i]}',ha='center', va='bottom',fontdict={'size': 10, 'color': 'r'})
# onnx
plt.plot(onnx_x[1:], onnx_y[1:], 'ms:',label='onnxruntime',markersize=7,linewidth=2)
for i,(a, b) in enumerate(zip(onnx_x[1:], onnx_y[1:])):
# plt.text(a+15,b-0.15,'(%d,%d,%d)'%(batch_size[i],a,b),ha='center', va='bottom',fontdict={'size': 10, 'color': 'r'})
plt.text(a+15,b-0.15,f'Batch:{batch_sizes[i+1]}',ha='center', va='bottom',fontdict={'size': 10, 'color': 'm'})
# trt api
plt.plot(trt_api_32_x, trt_api_32_y, 'b^-',label='TensorRT API (FP32)',markersize=7,linewidth=2)
for i,(a, b) in enumerate(zip(trt_api_32_x, trt_api_32_y)):
# plt.text(a+15,b-0.15,'(%d,%d)'%(a,b),ha='center', va='bottom',fontdict={'size': 10, 'color': 'b'})
plt.text(a+15,b-0.15,f'Batch:{batch_sizes[i]}',ha='center', va='bottom',fontdict={'size': 10, 'color': 'b'})
plt.plot(trt_api_16_x, trt_api_16_y, 'g*-.',label='TensorRT API (FP16)',markersize=7,linewidth=2)
for i,(a, b) in enumerate(zip(trt_api_16_x, trt_api_16_y)):
# plt.text(a+15,b-0.15,'(%d,%d)'%(a,b),ha='center', va='bottom',fontdict={'size': 10, 'color': 'g'})
plt.text(a+15,b-0.15,f'Batch:{batch_sizes[i]}',ha='center', va='bottom',fontdict={'size': 10, 'color': 'g'})
if batch_sizes[i] in [4,8]:
plt.annotate(f"({int(a)},{int(b)})",xy=(a,b),xytext=(a*0.9,b*0.9),arrowprops=dict(arrowstyle='->',connectionstyle='arc3,rad=.2'))
# trt onnxparser
plt.plot(trt_onnx_32_x, trt_onnx_32_y, 'yp--',label='TensorRT ONNXParser (FP32)',markersize=7,linewidth=2)
for i,(a, b) in enumerate(zip(trt_onnx_32_x, trt_onnx_32_y)):
# plt.text(a+15,b-0.15,'(%d,%d)'%(a,b),ha='center', va='bottom',fontdict={'size': 10, 'color': 'b'})
plt.text(a+15,b-0.15,f'Batch:{batch_sizes[i]}',ha='center', va='bottom',fontdict={'size': 10, 'color': 'y'})
plt.plot(trt_onnx_16_x, trt_onnx_16_y, 'kh-.',label='TensorRT ONNXParser (FP16)',markersize=7,linewidth=2)
for i,(a, b) in enumerate(zip(trt_onnx_16_x, trt_onnx_16_y)):
# plt.text(a+15,b-0.15,'(%d,%d)'%(a,b),ha='center', va='bottom',fontdict={'size': 10, 'color': 'g'})
plt.text(a+15,b-0.15,f'Batch:{batch_sizes[i]}',ha='center', va='bottom',fontdict={'size': 10, 'color': 'k'})
if batch_sizes[i] in [4,8]:
plt.annotate(f"({int(a)},{int(b)})",xy=(a,b),xytext=(a*0.9,b*0.9),arrowprops=dict(arrowstyle='->',connectionstyle='arc3,rad=.2'))
plt.xlabel('Latency (ms)')
plt.ylabel( r"$Throughput=\frac{1000}{Latency} \times Batch\_Size$")
plt.legend()
plt.savefig("./latency_vs_throughput.png",bbox_inches='tight', pad_inches=0)
plt.close()