File: profiler.py

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (48 lines) | stat: -rw-r--r-- 1,221 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(f.encode('ascii') for f in 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()