FM / cuda

/usr/local/lib64/python3.6/site-packages/torch/cuda/

/usr/local/lib64/python3.6/site-packages/torch/cuda

Name	Size	Mode	Actions
amp/	-	0755	rm
__pycache__/	-	0755	rm
comm.py	293	0644	edit dl rm
error.py	0	0644	edit dl rm
graphs.py	18835	0644	edit dl rm
memory.py	22599	0644	edit dl rm
nccl.py	3942	0644	edit dl rm
nvtx.py	1508	0644	edit dl rm
profiler.py	1221	0644	edit dl rm
random.py	5223	0644	edit dl rm
sparse.py	67	0644	edit dl rm
streams.py	8093	0644	edit dl rm
_utils.py	1845	0644	edit dl rm
__init__.py	24370	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/cuda/profiler.py (1221B)

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()

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()