/
cupy_memory_profile.py
215 lines (180 loc) · 8.37 KB
/
cupy_memory_profile.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
import collections
import sys
import typing as tp # NOQA
from chainer.backends import cuda
from chainer import function_hook
try:
MemoryHook = cuda.cupy.cuda.memory_hook.MemoryHook # type: tp.Any # to handle https://github.com/python/mypy/issues/2477 # NOQA
memory_hook_available = True
except Exception as e:
_resolution_error = e
MemoryHook = object
memory_hook_available = False
class CupyMemoryProfileHook(function_hook.FunctionHook):
"""Function hook for measuring memory usage of functions in cupy memory pool.
Example:
Code example::
from chainer.function_hooks import CupyMemoryProfileHook
hook = CupyMemoryProfileHook()
with hook:
trainer.run()
hook.print_report()
Output example::
FunctionName UsedBytes AcquiredBytes Occurrence
LinearFunction 5.16GB 179.98MB 3900
ReLU 0.99GB 458.97MB 2600
SoftmaxCrossEntropy 0.01GB 5.08MB 1300
Accuracy 0.00GB 0.35MB 700
where *FunctionName* is the name of function that calls the hook, and
*UsedBytes* is the memory bytes the function used from cupy memory
pool, and *AcquiredBytes* is the actual memory bytes the cupy memory
pool acquired from GPU device on the function call, and *Occurrence*
is the number of calls.
Attributes:
call_history: List of measurement results. It consists of the name of
the function that calls this hook, the memory bytes the function
used from cupy memory pool, and the memory bytes the cupy memory
pool acquired from GPU device on the function call.
"""
name = 'CupyMemoryProfileHook'
_units = ['B', 'KB', 'MB', 'GB', 'TB', 'PB', 'EB', 'ZB']
_table = {u: 1024.0 ** i for i, u in enumerate(_units)}
def __init__(self):
cuda.check_cuda_available()
if not memory_hook_available:
msg = 'CuPy >= 2.0 is required. %s' % str(_resolution_error)
raise RuntimeError(msg)
self.call_history = []
self._memory_hook = CupyMemoryCumulativeHook()
self._running_stack = []
self._total_used_bytes = 0
self._total_acquired_bytes = 0
def added(self, function=None):
self._memory_hook.__enter__()
def deleted(self, function=None):
self._memory_hook.__exit__()
def _preprocess(self):
start_used_bytes = self._memory_hook.used_bytes
start_acquired_bytes = self._memory_hook.acquired_bytes
self._running_stack.append((start_used_bytes, start_acquired_bytes))
def forward_preprocess(self, function, in_data):
self._preprocess()
def backward_preprocess(self, function, in_data, out_grad):
self._preprocess()
def _postprocess(self, function):
start_used_bytes, start_acquired_bytes = self._running_stack.pop()
end_used_bytes = self._memory_hook.used_bytes
end_acquired_bytes = self._memory_hook.acquired_bytes
used_bytes = end_used_bytes - start_used_bytes
acquired_bytes = end_acquired_bytes - start_acquired_bytes
depth = len(self._running_stack)
self.call_history.append(
(function._impl_name, used_bytes, acquired_bytes, depth))
if depth == 0:
self._total_used_bytes += used_bytes
self._total_acquired_bytes += acquired_bytes
def forward_postprocess(self, function, in_data):
self._postprocess(function)
def backward_postprocess(self, function, in_data, out_grad):
self._postprocess(function)
def total_used_bytes(self):
"""Returns total bytes that functions used from cupy memory pool."""
return self._total_used_bytes
def total_acquired_bytes(self):
"""Returns total bytes that cupy memory pool acquired from GPU."""
return self._total_acquired_bytes
def summary(self):
"""Returns a summary of memory profiling in functions.
Returns:
A summarized dictionary whose keys are function names and
values are dictionaries of
``used_bytes``, ``acquired_bytes``, and ``occurrrence``.
"""
# TODO(sonots): PROBLEM: takes count of nested functions duplicately
summary = collections.OrderedDict()
for func_name, used_bytes, acquired_bytes, depth in self.call_history:
if func_name not in summary:
summary[func_name] = {'used_bytes': 0,
'acquired_bytes': 0, 'occurrence': 0}
record = summary[func_name]
record['used_bytes'] += used_bytes
record['acquired_bytes'] += acquired_bytes
record['occurrence'] += 1
return summary
def _choose_unit(self, size):
"""Choose optimal unit.
Returns:
Tuple of denomi (float) and human-readable unit (str).
"""
denomi = 1.0
if size <= 0:
return denomi, self._units[0]
for unit in self._units[:-1]:
if size / (denomi * 1024) < 1:
return denomi, unit
denomi *= 1024
return denomi, self._units[-1]
def print_report(self, unit='auto', file=sys.stdout):
"""Prints a summary report of memory profiling in functions.
Args:
unit (str): Supplementary units used for used memories.
`B`, `KB`, `MB`, `GB`, `TB`, `PB`, `EB`, `ZB`, `auto`(default)
and `auto_foreach` are supported. If `auto`, units of memories
are aligned to the largest values of 'used_bytes' and
'acquired_bytes'. If `auto_foreach`, units of memories are
adjusted for each element.
"""
entries = [[
'FunctionName', 'UsedBytes', 'AcquiredBytes', 'Occurrence']]
if unit == 'auto':
max_used = max(
record['used_bytes'] for record in self.summary().values())
max_acquired = max(
record['acquired_bytes'] for record in self.summary().values())
denomi_used, unit_used = self._choose_unit(max_used)
denomi_acquired, unit_acquired = self._choose_unit(max_acquired)
elif unit != 'auto_foreach':
denomi_used = denomi_acquired = self._table[unit]
unit_used = unit_acquired = unit
for function_name, record in self.summary().items():
used_bytes = record['used_bytes']
acquired_bytes = record['acquired_bytes']
if unit == 'auto_foreach':
denomi_used, unit_used = self._choose_unit(used_bytes)
denomi_acquired, unit_acquired = self._choose_unit(
acquired_bytes)
used_bytes = '%3.2f%s' % (used_bytes / denomi_used, unit_used)
acquired_bytes = '%3.2f%s' % (
acquired_bytes / denomi_acquired, unit_acquired)
occurrence = str(record['occurrence'])
entries.append(
[function_name, used_bytes, acquired_bytes, occurrence])
entry_widths = []
entry_widths.append(max(len(f) for f, _, _, _ in entries))
entry_widths.append(max(len(u) for _, u, _, _ in entries))
entry_widths.append(max(len(a) for _, _, a, _ in entries))
entry_widths.append(max(len(o) for _, _, _, o in entries))
template = ' '.join('{:>%d}' % w for w in entry_widths)
for function_name, used_bytes, acquired_bytes, occurrence in entries:
line = template.format(
function_name, used_bytes, acquired_bytes, occurrence)
file.write(line)
file.write('\n')
if hasattr(file, 'flush'):
file.flush()
class CupyMemoryCumulativeHook(MemoryHook):
"""A simple memory hook for cupy measuring memory usage cumulatively.
Attributes:
used_bytes (int): cumulative bytes that application used from cupy
memory pool.
acquired_bytes (int): cumulative bytes that cupy memory pool acquired
from GPU device.
"""
name = 'CupyMemoryCumulativeHook'
def __init__(self):
self.used_bytes = 0
self.acquired_bytes = 0
def alloc_preprocess(self, **kwargs):
self.acquired_bytes += kwargs['mem_size']
def malloc_preprocess(self, **kwargs):
self.used_bytes += kwargs['mem_size']