File: profiler.py

package info (click to toggle)
pytorch 1.7.1-7
  • links: PTS, VCS
  • area: main
  • in suites: bullseye
  • size: 80,340 kB
  • sloc: cpp: 670,830; python: 343,991; ansic: 67,845; asm: 5,503; sh: 2,924; java: 2,888; xml: 266; makefile: 244; ruby: 148; yacc: 144; objc: 51; lex: 44
file content (48 lines) | stat: -rw-r--r-- 1,228 bytes parent folder | download
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
import tempfile
import contextlib
from . import cudart, check_error


DEFAULT_FLAGS = [
    "gpustarttimestamp",
    "gpuendtimestamp",
    "gridsize3d",
    "threadblocksize",
    "streamid",
    "enableonstart 0",
    "conckerneltrace",
]


def init(output_file, flags=None, output_mode='key_value'):
    rt = cudart()
    if not hasattr(rt, 'cudaOutputMode'):
        raise AssertionError("HIP does not support profiler initialization!")
    flags = DEFAULT_FLAGS if flags is None else flags
    if output_mode == 'key_value':
        output_mode_enum = rt.cudaOutputMode.KeyValuePair
    elif output_mode == 'csv':
        output_mode_enum = rt.cudaOutputMode.CSV
    else:
        raise RuntimeError("supported CUDA profiler output modes are: key_value and csv")
    with tempfile.NamedTemporaryFile(delete=True) as f:
        f.write(b'\n'.join(map(lambda f: f.encode('ascii'), flags)))
        f.flush()
        check_error(rt.cudaProfilerInitialize(f.name, output_file, output_mode_enum))


def start():
    check_error(cudart().cudaProfilerStart())


def stop():
    check_error(cudart().cudaProfilerStop())


@contextlib.contextmanager
def profile():
    try:
        start()
        yield
    finally:
        stop()