Merge pull request #5082 from resonantsky/dev

Add MIOpen logging support
This commit is contained in:
Vladimir Mandic
2026-09-11 10:11:23 +02:00
committed by GitHub
2 changed files with 224 additions and 3 deletions
+198
View File
@@ -0,0 +1,198 @@
"""Capture and report native MIOpen convolution selections."""
import atexit
import os
import re
import sys
import threading
import time
from modules.logger import log
_ALGORITHM_PATTERN = re.compile(r"FW Chosen Algorithm:\s*([^,\s]+)")
_CHOSEN_PATTERN = re.compile(r"FW Chosen Algorithm:\s*([^,\s]+)\s*,\s*[^,]*,\s*([0-9.eE+-]+)")
_MIOPEN_PREFIX = "MIOpen(HIP):"
_logging_capture = None
def _forward_stderr(fd, line):
data = line if line.endswith(b"\n") else line + b"\n"
while data:
written = os.write(fd, data)
data = data[written:]
def _process_line(line, saved_stderr):
text = line.decode(errors="replace").rstrip("\r\n")
if not text.lstrip().startswith(_MIOPEN_PREFIX):
_forward_stderr(saved_stderr, line)
return
match = _CHOSEN_PATTERN.search(text)
if match:
log.info(f'MIOpen: algorithm={match.group(1)} time={float(match.group(2)):.3f}')
class MIOpenLogRedirect:
"""Redirect native MIOpen diagnostics into structured informational logs."""
def __init__(self):
self.read_fd = -1
self.saved_stderr = -1
self.saved_python_stderr = sys.stderr
self.safe_stderr = None
self.reader = None
def __enter__(self):
write_fd = -1
try:
self.read_fd, write_fd = os.pipe()
self.saved_stderr = os.dup(2)
self.saved_python_stderr = sys.stderr
self.safe_stderr = os.fdopen(os.dup(self.saved_stderr), "w", encoding=getattr(sys.stderr, "encoding", None) or "utf-8", buffering=1)
os.dup2(write_fd, 2)
os.close(write_fd)
write_fd = -1
sys.stderr = self.safe_stderr
def read_output():
pending = b""
while True:
chunk = os.read(self.read_fd, 4096)
if not chunk:
break
pending += chunk
while b"\n" in pending:
line, pending = pending.split(b"\n", 1)
_process_line(line + b"\n", self.saved_stderr)
if pending:
_process_line(pending, self.saved_stderr)
self.reader = threading.Thread(target=read_output, daemon=True)
self.reader.start()
return self
except Exception:
if write_fd >= 0:
os.close(write_fd)
if self.saved_stderr >= 0:
os.dup2(self.saved_stderr, 2)
sys.stderr = self.saved_python_stderr
if self.safe_stderr is not None:
self.safe_stderr.close()
if self.reader is not None:
self.reader.join()
if self.saved_stderr >= 0:
os.close(self.saved_stderr)
if self.read_fd >= 0:
os.close(self.read_fd)
raise
def __exit__(self, _exc_type, _exc_value, _traceback):
os.dup2(self.saved_stderr, 2)
sys.stderr = self.saved_python_stderr
self.safe_stderr.close()
self.reader.join()
os.close(self.saved_stderr)
os.close(self.read_fd)
return False
def start_miopen_logging():
"""Start filtering native MIOpen diagnostics without changing the environment."""
global _logging_capture # pylint: disable=global-statement
if _logging_capture is None:
try:
_logging_capture = MIOpenLogRedirect()
_logging_capture.__enter__()
except Exception as err:
log.warning(f'MIOpen logging: failed to start: {err}')
_logging_capture = None
def stop_miopen_logging():
"""Stop filtering native MIOpen diagnostics and restore stderr."""
global _logging_capture # pylint: disable=global-statement
if _logging_capture is not None:
_logging_capture.__exit__(None, None, None)
_logging_capture = None
atexit.register(stop_miopen_logging)
class MIOpenLogCapture:
"""Capture one native MIOpen operation and log its selected algorithm and time."""
def __init__(self, operation: str = "convolution", repeats: int = 1):
self.operation = operation
self.repeats = max(1, repeats)
self.lines: list[str] = []
self.elapsed_ms = 0.0
self.algorithms: list[str] = []
self.read_fd = -1
self.saved_stderr = -1
self.saved_python_stderr = sys.stderr
self.safe_stderr = None
self.start = 0.0
self.reader = None
def __enter__(self):
write_fd = -1
try:
self.read_fd, write_fd = os.pipe()
self.saved_stderr = os.dup(2)
self.saved_python_stderr = sys.stderr
self.safe_stderr = os.fdopen(
os.dup(self.saved_stderr),
"w",
encoding=getattr(sys.stderr, "encoding", None) or "utf-8",
buffering=1,
)
os.dup2(write_fd, 2)
os.close(write_fd)
write_fd = -1
sys.stderr = self.safe_stderr
self.lines = []
self.start = time.perf_counter()
def read_output():
chunks = []
while True:
chunk = os.read(self.read_fd, 4096)
if not chunk:
break
chunks.append(chunk)
self.lines.extend(b"".join(chunks).decode(errors="replace").splitlines())
self.reader = threading.Thread(target=read_output, daemon=True)
self.reader.start()
return self
except Exception:
if write_fd >= 0:
os.close(write_fd)
if self.saved_stderr >= 0:
os.dup2(self.saved_stderr, 2)
sys.stderr = self.saved_python_stderr
if self.safe_stderr is not None:
self.safe_stderr.close()
if self.reader is not None:
self.reader.join()
if self.saved_stderr >= 0:
os.close(self.saved_stderr)
if self.read_fd >= 0:
os.close(self.read_fd)
raise
def __exit__(self, _exc_type, _exc_value, _traceback):
os.dup2(self.saved_stderr, 2)
sys.stderr = self.saved_python_stderr
self.safe_stderr.close()
os.close(self.saved_stderr)
self.reader.join()
os.close(self.read_fd)
return False
def capture_miopen(operation: str = "convolution", repeats: int = 1):
"""Return a side-effect-free context manager for one MIOpen operation."""
return MIOpenLogCapture(operation=operation, repeats=repeats)
+26 -3
View File
@@ -11,6 +11,7 @@ from modules.shared import opts
from scripts.rocm.rocm_vars import ROCM_ENV_VARS # pylint: disable=no-name-in-module
from scripts.rocm import rocm_profiles # pylint: disable=no-name-in-module
from scripts.rocm import rocm_log # pylint: disable=no-name-in-module
CONFIG = Path(os.path.abspath(os.path.join('data', 'rocm.json')))
@@ -275,6 +276,16 @@ def apply_env(config: Optional[Dict[str, str]] = None) -> None:
os.environ[var] = "0"
def start_miopen_logging() -> None:
"""Start explicit MIOpen diagnostic capture for a scoped operation."""
rocm_log.start_miopen_logging()
def stop_miopen_logging() -> None:
"""Stop explicit MIOpen diagnostic capture and restore stderr."""
rocm_log.stop_miopen_logging()
def apply_all(names: list, values: list) -> None:
config = load_config().copy()
arch = config.get(_ARCH_KEY, "")
@@ -520,9 +531,21 @@ def info() -> dict:
}
# Apply saved config to os.environ at import time (only when ROCm is present)
if installer.torch_info.get('type', None) == 'rocm' and CONFIG.exists():
def _is_rocm_runtime() -> bool:
if installer.torch_info.get('type', None) == 'rocm':
return True
try:
apply_env()
import torch # pylint: disable=import-outside-toplevel
return bool(getattr(torch.version, 'hip', None))
except Exception:
return False
# Apply saved config to os.environ at import time (only when ROCm is present).
if _is_rocm_runtime():
try:
if CONFIG.exists():
apply_env()
rocm_log.start_miopen_logging()
except Exception as _e:
log.debug(f"[rocm_mgr] Warning: failed to apply env at import: {_e}")