forked from openvinotoolkit/nncf
-
Notifications
You must be signed in to change notification settings - Fork 0
/
benchmark.py
100 lines (76 loc) · 2.89 KB
/
benchmark.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
import math
import time
import torch
import torch.distributed as dist
import torch.nn as nn
TIME_SCALES = {'ms': 1000}
def warmup(layer, input_, runs, forward_only=False):
for _ in range(runs):
new_i = layer(input_)
if not forward_only:
new_i[0].sum().backward()
def run_wall(layer, input_size_, device, runs, is_print=True):
input_ = torch.randn(input_size_, device=torch.device(device))
# Force CUDA initialization & warm up
warmup(layer, input_, 100)
torch.cuda.synchronize()
start = time.time()
for _ in range(runs):
layer.zero_grad()
new_i = layer(input_)
new_i[0].sum().backward()
torch.cuda.synchronize()
elapsed = time.time() - start
ctime, scale = list(TIME_SCALES.items())[0]
fbtime = elapsed / runs * scale
if is_print:
print('Forward&Backward: {0:.3f} {1}'.format(
fbtime, ctime))
def run_profile(layer, input_size_, device, runs, forward_only=False):
input_ = torch.randn(input_size_, device=torch.device(device))
# Force CUDA initialization & warm up
warmup(layer, input_, 100, forward_only)
start = time.time()
forward_min = math.inf
forward_time = 0
backward_min = math.inf
backward_time = 0
for _ in range(runs):
layer.zero_grad()
torch.cuda.synchronize()
start = time.time()
new_i = layer(input_)
torch.cuda.synchronize()
elapsed = time.time() - start
forward_min = min(forward_min, elapsed)
forward_time += elapsed
if not forward_only:
torch.cuda.synchronize()
start = time.time()
new_i[0].sum().backward()
torch.cuda.synchronize()
elapsed = time.time() - start
backward_min = min(backward_min, elapsed)
backward_time += elapsed
ctime, scale = list(TIME_SCALES.items())[0]
forward_min *= scale
backward_min *= scale
forward_average = forward_time / runs * scale
backward_average = backward_time / runs * scale
print('Forward: min {0:.3f}{4} / avg {1:.3f}{4} | Backward: min {2:.3f}{4} / avg {3:.3f}{4}'.format(
forward_min, forward_average, backward_min, backward_average, ctime))
def run_worker(gpu, world_size, layer, input_size_, runs):
dist.init_process_group(backend='nccl', init_method="tcp://127.0.0.1:8899",
world_size=world_size, rank=gpu)
device = torch.device('cuda:%d' % gpu)
torch.cuda.set_device(gpu)
batch = (int)(input_size_[0] / world_size)
if gpu == 0:
run_size = input_size_.copy()
run_size[0] = input_size_[0] - batch * (world_size - 1)
else:
run_size = input_size_.copy()
run_size[0] = batch
run_model = layer.to(device)
run_model = nn.parallel.DistributedDataParallel(run_model, device_ids=[gpu])
run_wall(run_model, run_size, device, runs, (gpu == 0))