Files
javis_bot/scripts/m0/m0lib/frames.py
javis-bot cec52165b2 fix(m0): judge game frames by what the screen showed and drop spoilt windows
- Frame rate, 1 % low and hitches come from msBetweenDisplayChange of the
  frames that reached the screen; frames never shown are counted apart and
  the presented rate is kept as presented_fps. A game presenting 60 fps
  while the screen changes every 167 ms now shows as the stutter it is.
- A window whose capture failed, a busy window where the LLM finished no
  answer, and every busy window when the STT load failed are not counted.
- A yielding window counts only when every Jarvis process got both its
  CPU and its GPU priority lowered.
- The "game hardly drawing" floor applies only with Jarvis off or idle, so
  a game Jarvis slows that far still counts as harm.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-11 09:12:15 +09:00

515 lines
23 KiB
Python

"""Stage part: does Jarvis make the user's game drop frames, and does yielding to the game help?
The GPU cap is gone (2026-11-11): pausing Jarvis for 5 % of every 200 ms did not lower nvidia-smi's
utilisation, cost 6 to 10 % of the LLM's output, and a YouTube video next to a busy Jarvis did not
stutter with or without it. What is left unmeasured is a game. So this stage measures what the user
would notice, the game's own frame times, instead of a GPU-wide number:
- PresentMon (Intel's open-source tool, pinned by hash) records when the game presents each frame
through Windows' event tracing. It sees no picture; the kit keeps only the frame-time numbers and
deletes its CSV (which names the game) right after reading it. Event tracing needs administrator
rights, so this stage runs elevated (run_m0_game_frames.bat asks once, before the game is started).
- Windows: the game alone, Jarvis loaded and idle, Jarvis busy in the game placement the app uses
(the LLM answering, the STT doing only finals, as decided for a game), and busy while Jarvis yields:
its processes at below-normal CPU priority and below-normal GPU scheduling priority. Unlike the cap
nothing is paused; Windows hands the game the cores and the GPU first only when it wants them. The
busy rows run A B B A (plain, yield, yield, plain) so a slow drift does not count for one side.
"""
from __future__ import annotations
import contextlib
import csv
import hashlib
import io
import subprocess
import threading
import time
import urllib.request
from pathlib import Path
from typing import Any
from m0lib import gpucap, llm
from m0lib.common import IS_WINDOWS, StageResult, gpu_memory_mib, interactive, ask, say, stats, wait_enter
from m0lib.llm_bench import KOREAN_PROMPTS, NO_THINK, VOICE_SYSTEM
from m0lib.llm_client import ChatClient
PRESENTMON_VERSION = "2.6.0"
PRESENTMON_URL = (f"https://github.com/GameTechDev/PresentMon/releases/download/v{PRESENTMON_VERSION}/"
f"PresentMon-{PRESENTMON_VERSION}-x64.exe")
PRESENTMON_SHA256 = "b2a706bc6ad475749e3b7e3409263aa1e6906d45bdcf993f6dbc0f660188f1af"
WINDOW_S = 60
SETTLE_S = 5 # Jarvis busy (or yielding) before a window starts
HITCH_MS = (50, 100) # a frame this long is a visible hitch
MIN_FRAMES = 300 # under this in a minute (5 fps) with Jarvis off or idle, the game was not really drawing
ROUNDS = ("busy", "busy_yield", "busy_yield_2", "busy_2")
# The foreground window belongs to one of these when the user did not get back to the game.
NOT_A_GAME = {"windowsterminal.exe", "openconsole.exe", "conhost.exe", "cmd.exe", "powershell.exe",
"pwsh.exe", "explorer.exe"}
# Windows' values: process priority classes and D3DKMT_SCHEDULINGPRIORITYCLASS.
CPU_NORMAL, CPU_BELOW_NORMAL = 0x20, 0x4000
GPU_NORMAL, GPU_BELOW_NORMAL = 2, 1
GAME_LEVEL = next(lv for lv in llm.TOGETHER if lv["key"] == "game_experts_2g_16k_stt_gpu")
# --------------------------------------------------------------------------
# Windows: rights, the game's process, priorities.
def is_admin() -> bool:
if not IS_WINDOWS:
return False
import ctypes
try:
return bool(ctypes.windll.shell32.IsUserAnAdmin())
except Exception:
return False
def foreground_process() -> tuple[int, str] | None:
"""(pid, exe file name in lower case) of the window in front, or None."""
if not IS_WINDOWS:
return None
import ctypes
from ctypes import wintypes
user32 = ctypes.WinDLL("user32")
k32 = ctypes.WinDLL("kernel32")
user32.GetForegroundWindow.restype = ctypes.c_void_p
user32.GetWindowThreadProcessId.argtypes = [ctypes.c_void_p, ctypes.POINTER(wintypes.DWORD)]
k32.OpenProcess.restype = ctypes.c_void_p
k32.OpenProcess.argtypes = [wintypes.DWORD, wintypes.BOOL, wintypes.DWORD]
k32.QueryFullProcessImageNameW.argtypes = [ctypes.c_void_p, wintypes.DWORD, wintypes.LPWSTR,
ctypes.POINTER(wintypes.DWORD)]
k32.CloseHandle.argtypes = [ctypes.c_void_p]
hwnd = user32.GetForegroundWindow()
if not hwnd:
return None
pid = wintypes.DWORD()
user32.GetWindowThreadProcessId(hwnd, ctypes.byref(pid))
h = k32.OpenProcess(0x1000, False, pid.value) # PROCESS_QUERY_LIMITED_INFORMATION
name = ""
if h:
try:
buf = ctypes.create_unicode_buffer(1024)
n = wintypes.DWORD(len(buf))
if k32.QueryFullProcessImageNameW(h, 0, buf, ctypes.byref(n)):
name = Path(buf.value).name.lower()
finally:
k32.CloseHandle(h)
return pid.value, name
_API: dict = {}
def _api() -> dict:
import ctypes
from ctypes import wintypes
if not _API:
k32 = ctypes.WinDLL("kernel32", use_last_error=True)
gdi = ctypes.WinDLL("gdi32")
k32.OpenProcess.restype = ctypes.c_void_p
k32.OpenProcess.argtypes = [wintypes.DWORD, wintypes.BOOL, wintypes.DWORD]
k32.CloseHandle.argtypes = [ctypes.c_void_p]
k32.GetPriorityClass.argtypes = [ctypes.c_void_p]
k32.GetPriorityClass.restype = wintypes.DWORD
k32.SetPriorityClass.argtypes = [ctypes.c_void_p, wintypes.DWORD]
gdi.D3DKMTGetProcessSchedulingPriorityClass.argtypes = [ctypes.c_void_p, ctypes.POINTER(ctypes.c_int)]
gdi.D3DKMTGetProcessSchedulingPriorityClass.restype = ctypes.c_long
gdi.D3DKMTSetProcessSchedulingPriorityClass.argtypes = [ctypes.c_void_p, ctypes.c_int]
gdi.D3DKMTSetProcessSchedulingPriorityClass.restype = ctypes.c_long
_API.update(k32=k32, gdi=gdi, ctypes=ctypes)
return _API
def get_priority(pid: int) -> tuple[int | None, int | None]:
"""(CPU priority class, GPU scheduling priority class) of a process; None where unreadable."""
a = _api()
h = a["k32"].OpenProcess(0x0200 | 0x1000, False, pid) # SET_INFORMATION | QUERY_LIMITED
if not h:
return None, None
try:
cpu = a["k32"].GetPriorityClass(h) or None
val = a["ctypes"].c_int()
gpu = val.value if a["gdi"].D3DKMTGetProcessSchedulingPriorityClass(h, a["ctypes"].byref(val)) == 0 else None
return cpu, gpu
finally:
a["k32"].CloseHandle(h)
def set_priority(pid: int, cpu: int | None, gpu: int | None) -> tuple[bool, bool]:
"""Set what is given; (CPU done, GPU done)."""
a = _api()
h = a["k32"].OpenProcess(0x0200 | 0x1000, False, pid)
if not h:
return False, False
try:
ok_cpu = cpu is None or bool(a["k32"].SetPriorityClass(h, cpu))
ok_gpu = gpu is None or a["gdi"].D3DKMTSetProcessSchedulingPriorityClass(h, gpu) == 0
return ok_cpu, ok_gpu
finally:
a["k32"].CloseHandle(h)
class Yield:
"""Jarvis's processes at below-normal CPU and GPU priority while the block runs, put back after.
Each lowering is read back, so a priority Windows ignored does not count as lowered."""
def __init__(self, pids: list[int]):
self.pids = list(pids)
self.before: dict[int, tuple[int | None, int | None]] = {}
self.cpu_lowered = self.gpu_lowered = self.restored = 0
self.not_restored: list[int] = []
def __enter__(self) -> "Yield":
for pid in self.pids:
self.before[pid] = get_priority(pid)
set_priority(pid, CPU_BELOW_NORMAL, GPU_BELOW_NORMAL)
cpu, gpu = get_priority(pid)
self.cpu_lowered += cpu == CPU_BELOW_NORMAL
self.gpu_lowered += gpu == GPU_BELOW_NORMAL
return self
def __exit__(self, *exc) -> None:
for pid, (cpu, gpu) in self.before.items():
ok = set_priority(pid, cpu or CPU_NORMAL, GPU_NORMAL if gpu is None else gpu)
if all(ok):
self.restored += 1
else:
self.not_restored.append(pid)
def summary(self) -> dict:
return {"processes": len(self.pids), "cpu_lowered": self.cpu_lowered, "gpu_lowered": self.gpu_lowered,
"restored": self.restored, "not_restored": len(self.not_restored)}
def yield_problems(summary: dict) -> list[str]:
"""Why a yielding window did not yield: no process, a CPU or GPU priority left normal on any of
Jarvis's processes (yielding with some of them is not the yield being measured), or one not put back."""
out = []
n = summary.get("processes") or 0
if not n:
out.append("yield found no Jarvis process")
for key, what in (("cpu_lowered", "CPU"), ("gpu_lowered", "GPU")):
if n and summary.get(key, 0) < n:
out.append(f"yield lowered the {what} priority of only {summary.get(key, 0)} of {n} processes")
if summary.get("not_restored"):
out.append(f"{summary['not_restored']} process(es) kept the lowered priority")
return out
# --------------------------------------------------------------------------
# PresentMon and the frame times.
def ensure_presentmon(work: Path) -> Path:
exe = work / "tools" / f"PresentMon-{PRESENTMON_VERSION}-x64.exe"
if exe.exists() and hashlib.sha256(exe.read_bytes()).hexdigest() == PRESENTMON_SHA256:
return exe
exe.parent.mkdir(parents=True, exist_ok=True)
part = exe.with_suffix(".part")
with urllib.request.urlopen(PRESENTMON_URL, timeout=120) as resp:
part.write_bytes(resp.read())
got = hashlib.sha256(part.read_bytes()).hexdigest()
if got != PRESENTMON_SHA256:
part.unlink()
raise RuntimeError(f"PresentMon download has the wrong hash ({got[:12]})")
part.replace(exe)
return exe
def presentmon_cmd(exe: Path, pid: int, seconds: int, out_csv: Path) -> list[str]:
return [str(exe), "--process_id", str(pid), "--output_file", str(out_csv), "--timed", str(seconds),
"--terminate_after_timed", "--no_console_stats", "--v1_metrics", "--no_track_input",
"--session_name", "JarvisM0", "--stop_existing_session"]
def capture(exe: Path, pid: int, seconds: int, out_csv: Path, log: Path) -> tuple[list[dict], str | None]:
"""The game's frames over ``seconds`` (rows of PresentMon's CSV) and an error, if any. The CSV
names the game, so it is deleted as soon as it is read."""
out_csv.unlink(missing_ok=True)
with open(log, "a", encoding="utf-8") as f:
try:
p = subprocess.run(presentmon_cmd(exe, pid, seconds, out_csv), stdout=f, stderr=subprocess.STDOUT,
timeout=seconds + 60)
code = p.returncode
except subprocess.TimeoutExpired:
code = "timeout"
try:
text = out_csv.read_text(encoding="utf-8", errors="replace")
except OSError:
return [], f"PresentMon wrote no frames (exit {code})"
finally:
out_csv.unlink(missing_ok=True)
return parse_csv(text), None if code == 0 else f"PresentMon exited with {code}"
def parse_csv(text: str) -> list[dict]:
"""PresentMon's rows with lower-case column names (1.x writes msBetweenPresents, 2.x MsBetweenPresents)."""
return [{(k or "").strip().lower(): (v or "").strip() for k, v in row.items()}
for row in csv.DictReader(io.StringIO(text))]
def _num(value: str) -> float | None:
try:
return float(value)
except (TypeError, ValueError):
return None
def frame_stats(rows: list[dict], seconds: float, min_frames: int = 0) -> dict:
"""What the screen showed of the game's main swap chain (the one with the most frames): the rate of
shown frames, the 1 % low (the 99th-percentile time between screen updates as a rate) and hitches,
all from the time each shown frame stayed before the next (msBetweenDisplayChange). A frame the
game presented but the screen never showed is counted in ``not_shown``; the rate the game
presented at is kept apart (``presented_fps``), since a game can present at 60 fps while the
screen changes far less often. ``min_frames`` (per minute) marks a window where the game was
hardly drawing; it is only given with Jarvis off or idle, so a game Jarvis slows down that much
still counts as the harm it is."""
chains: dict[str, list[dict]] = {}
for x in rows:
chains.setdefault(x.get("swapchainaddress", ""), []).append(x)
main = (max(chains.values(), key=len) if chains else [])[1:] # the first reaches back before the capture
presents = [v for v in (_num(x.get("msbetweenpresents")) for x in main) if v is not None and v > 0]
shown = [x for x in main if x.get("dropped") != "1"]
ms = [v for v in (_num(x.get("msbetweendisplaychange")) for x in shown) if v is not None and v > 0][1:]
out: dict[str, Any] = {"seconds": round(seconds, 1), "frames": len(ms), "presented": len(presents),
"not_shown": len(main) - len(shown)}
if presents:
out["presented_fps"] = round(len(presents) / (sum(presents) / 1000), 1)
if not ms:
why = "the game drew no frames" if not presents else "PresentMon saw none of the frames reach the screen"
return {**out, "valid": False, "invalid_because": [why]}
srt = sorted(ms)
p99 = srt[min(len(srt) - 1, int(len(srt) * 0.99))]
out.update(fps=round(len(ms) / (sum(ms) / 1000), 1), low1_fps=round(1000 / p99, 1),
p99_ms=round(p99, 2), max_ms=round(srt[-1], 1),
hitches_over_ms={str(h): sum(v > h for v in ms) for h in HITCH_MS})
reasons = []
if len(ms) < min_frames * seconds / 60:
reasons.append(f"the screen showed only {len(ms)} of the game's frames")
out["valid"] = not reasons
if reasons:
out["invalid_because"] = reasons
return out
def compare(rows: list[dict]) -> dict:
"""The game alone, next to busy Jarvis and next to yielding Jarvis, over the valid windows, with
the change against the game alone in per cent."""
def pool(kind: str) -> dict:
valid = [x["game"] for x in rows if x.get("kind") == kind and (x.get("game") or {}).get("valid")]
if not valid:
return {"windows": 0}
n = len(valid)
return {"windows": n, "fps": round(sum(v["fps"] for v in valid) / n, 1),
"low1_fps": round(sum(v["low1_fps"] for v in valid) / n, 1),
**{f"hitches_over_{h}ms": sum(v["hitches_over_ms"][str(h)] for v in valid) for h in HITCH_MS}}
out = {k: pool(k) for k in ("off", "busy", "busy_yield")}
off = out["off"]
for k in ("busy", "busy_yield"):
if off["windows"] and out[k]["windows"]:
out[k]["vs_off_pct"] = {m: round((out[k][m] - off[m]) / off[m] * 100, 1) for m in ("fps", "low1_fps")}
return out
# --------------------------------------------------------------------------
# Jarvis busy for as long as a window lasts.
class BusyLlm:
"""The LLM answering back to back (a long answer, then a voice answer) until the block ends."""
def __init__(self, c: ChatClient):
self.c = c
self.stop = threading.Event()
self.decode: list[float] = []
self.voice: list[float] = []
self.errors: list[str] = []
self.thread = threading.Thread(target=self._loop, daemon=True)
def _loop(self) -> None:
voice = [p for kind, p in KOREAN_PROMPTS if kind == "voice"]
i = 0
while not self.stop.is_set():
try:
a = self.c.chat([{"role": "user", "content": f"({i}) 한국의 사계절을 각각 세 문장씩 자세히 설명해 줘."}],
max_tokens=320, extra=NO_THINK)
if a.get("decode_tok_s"):
self.decode.append(a["decode_tok_s"])
if voice and not self.stop.is_set():
a = self.c.chat([{"role": "system", "content": VOICE_SYSTEM},
{"role": "user", "content": voice[i % len(voice)]}], max_tokens=120, extra=NO_THINK)
if a.get("first_sentence") is not None:
self.voice.append(a["first_sentence"])
except Exception as e:
self.errors.append(f"{type(e).__name__}")
if len(self.errors) > 20:
return
time.sleep(1)
i += 1
def __enter__(self) -> "BusyLlm":
self.thread.start()
return self
def __exit__(self, *exc) -> None:
self.stop.set()
self.thread.join(timeout=120)
def summary(self) -> dict:
return {"decode_tok_s": stats(self.decode), "voice_first_sentence": stats(self.voice),
"errors": sorted(set(self.errors))}
# --------------------------------------------------------------------------
# The stage.
def ask_for_a_game() -> bool:
if not interactive():
return False
say(" 게임을 켠 채로 게임 화면의 프레임을 잽니다. 10~15분 걸립니다.")
if ask(" 지금 게임을 켜고 재시겠습니까?") != "y":
return False
wait_enter(" 게임을 켜고 플레이 화면까지 들어간 뒤 Enter를 누르고, 15초 안에 게임으로 돌아가 주세요. "
"끝날 때까지 캐릭터를 한자리에 세워 두고 마우스·키보드에 손대지 말아 주세요. 끝나면 알림음이 울립니다:")
time.sleep(15)
return True
def _spoil(x: dict, why: str) -> None:
x["game"]["valid"] = False
x["game"]["invalid_because"] = x["game"].get("invalid_because", []) + [why]
def measure(r: StageResult, s: "llm.Strata", preset: dict, context: int, stt_python: str) -> None:
work = s.log.parent.parent
pm_log = s.log.with_name("presentmon.log")
rows: list[dict] = []
r["frame_rows"] = rows
if not is_admin():
r.error("admin", "게임 프레임은 관리자 권한이 있어야 잴 수 있습니다. run_m0_game_frames.bat으로 실행해 주세요.")
return
try:
pm = ensure_presentmon(work)
except Exception as e:
r.error("presentmon", e)
return
r["presentmon"] = PRESENTMON_VERSION
if not ask_for_a_game():
r["skipped"] = "no game started"
return
fg = foreground_process()
if fg is None or fg[1] in NOT_A_GAME:
r.error("game", "게임 창이 앞에 없었습니다. 게임으로 돌아간 뒤 다시 실행해 주세요.")
return
game = fg[0]
r.note("measured while the user's game was running (its name is not kept)")
base = f"http://127.0.0.1:{llm.PORT}/v1"
def window(key: str, kind: str) -> dict:
x: dict = {"key": key, "kind": kind}
used = gpu_memory_mib()
if used:
x["vram_free_mib"] = used[1] - used[0]
with gpucap.UtilSampler() as util:
frames, err = capture(pm, game, WINDOW_S, work / "cache" / "frames.csv", pm_log)
g = frame_stats(frames, WINDOW_S, MIN_FRAMES if kind in ("off", "idle") else 0)
g["gpu_util"] = util.summary()
now = foreground_process()
if now is not None and now[0] != game:
g["valid"] = False
g["invalid_because"] = g.get("invalid_because", []) + ["the game was not in front at the end"]
if err:
# Frames from a capture that failed or was cut off do not cover the window.
r.error(f"frames {key}", err)
g["valid"] = False
g["invalid_because"] = g.get("invalid_because", []) + [f"the capture failed ({err})"]
x["game"] = g
rows.append(x)
if not g["valid"]:
r.note(f"{key}: not counted ({'; '.join(g['invalid_because'])})")
say(f" {key}: {g.get('fps')} fps, 1% low {g.get('low1_fps')} fps, "
f"{HITCH_MS[0]}ms 넘는 프레임 {(g.get('hitches_over_ms') or {}).get(str(HITCH_MS[0]))}개")
r.save()
return x
stt = None
try:
say(" - 게임만")
window("game_only", "off")
say(" - Jarvis 켜기 (게임 중 배치: 전문가 2GB, 맥락 16K, 음성 인식 GPU)")
s.stop()
stt = llm.SttLoad(stt_python, work, "cuda", s.log.with_name("stt_load.log"), pace=llm._FINALS_ONLY)
load = stt.start()
r["stt_load"] = load
if load.get("status") != "loaded":
r.error("frames stt", "; ".join(load.get("errors") or []) or "the STT did not load")
return
lv = GAME_LEVEL
r["placement"] = {"key": lv["key"], "sets": lv["sets"], "context": lv["context"], "parallel": lv["parallel"],
"vision": lv["vision"], "stt_pace": llm._FINALS_ONLY}
launch = s.launch(preset, lv["context"], parallel=lv["parallel"], vision=lv["vision"],
extra_sets=list(lv["sets"]), timeout=3600)
if not launch.get("ready"):
r.error("frames llm", "the game placement did not start")
return
c = ChatClient(base)
r["server"] = {"model": c.model}
c.chat([{"role": "user", "content": "안녕"}], max_tokens=8)
say(" - Jarvis 켜 두고 쉬는 중")
window("jarvis_idle", "idle")
stt.go()
for key in ROUNDS:
yielding = "yield" in key
say(f" - Jarvis가 답하고 음성 인식이 최종 전사를 하는 중{' (게임에 양보)' if yielding else ''}")
pids = llm.cap_pids(s, stt) if yielding else []
y = Yield(pids)
with y if yielding else contextlib.nullcontext(), BusyLlm(c) as busy:
time.sleep(SETTLE_S)
x = window(key, "busy_yield" if yielding else "busy")
x["llm"] = busy.summary()
for err in x["llm"]["errors"]:
r.error(f"frames {key} llm", err)
if not x["llm"]["decode_tok_s"].get("n"):
# Jarvis was not busy, so the window measured the game next to an idle Jarvis.
r.error(f"frames {key} llm", "no answer finished during the window")
_spoil(x, "the LLM finished no answer during the window")
if yielding:
x["yield"] = y.summary()
problems = yield_problems(x["yield"])
for why in problems:
r.error(f"frames {key} yield", why)
for why in problems:
_spoil(x, why)
r.save()
busy_stt = stt.finish()
r["stt_busy"] = {k: busy_stt.get(k) for k in ("status", "single_s", "errors", "final_wait_s")
if k in busy_stt}
for err in llm.stt_gaps(busy_stt, llm._FINALS_ONLY):
r.error("frames stt", err)
# The STT's part of the load is not known to have run, in any busy window.
for x in rows:
if x["kind"].startswith("busy"):
_spoil(x, f"the STT load failed ({err})")
stt.close()
stt = None
s.stop()
say(" - Jarvis 끄고 게임만 다시")
window("game_only_again", "off")
r["comparison"] = compare(rows)
except Exception as e:
r.error("frames", e)
finally:
if stt:
stt.close()
s.stop()
say("\a 게임 프레임 측정이 끝났습니다. 게임을 꺼도 됩니다.")
try:
import winsound
winsound.MessageBeep()
except Exception:
pass