151 lines
11 KiB
Python
Executable File
151 lines
11 KiB
Python
Executable File
#!/usr/bin/env python3
|
||
# 慢帧归因:输出慢帧清单,每帧拆成 CPU 侧 / 等待侧 / GPU 侧。两种输入:
|
||
# · 宿主日志(boot.txt 或经日志端口实时读回的流,开 DM_SLOWFRAME=1):只含掉帧帧的「[slowframe]\t帧号\t@名=值…」行
|
||
# 与每 300 帧一行分母,读数阶段分段用同一份日志里的 UIMSG;用法 slowframe.py boot.txt 或 … | slowframe.py -
|
||
# · 逐帧文件(DM_NVN_TSV,可加 DM_NVN_TSV_CALLS=0)+ 可选 boot 日志:slowframe.py frames.tsv [boot.txt]
|
||
# --all 连启动 / 读取阶段的慢帧也逐帧列出
|
||
#
|
||
# 慢帧口径:上屏间隔 @presented_ms > 16.7×1.5 ms 或 −1(系统丢弃)记一次掉帧;进局前的帧单列。
|
||
# 无上屏时刻的日志(Mac 无窗口)退回用 @frame_ms > 25 ms,并在表头注明。
|
||
# 各分项(ms,均为该帧号,即上一次出图调用到本次出图调用之间):
|
||
# 间隔 @frame_ms 出图线程两次调 QueuePresent 的间隔
|
||
# CPU @cpu_ms 出图线程本帧在 CPU 上的时间(−1 = 换了线程);其中 @sub_cpu_ms 提交回放、@upload_ms 贴图上传、
|
||
# @lib_new / @pso_new 同步建库 / 建管线个数;@proc_cpu_ms 全进程所有线程 CPU(看是否别的线程吃满核)
|
||
# 等待 @wait_ms 取图时按 60 Hz 节拍等(主动让出,不是问题);@pq_wait_ms 等出图队列归还交换链贴图;
|
||
# 离 CPU 其余 = 间隔 − CPU − 节拍等 − 队列等(锁等待、被抢占、客体同步等)
|
||
# GPU @gpu_ms 本帧命令缓冲 GPU 忙时之和;@gpu_end_ms − @call_ms = 出图调用后多久 GPU 做完;
|
||
# @pres_at_ms − @gpu_end_ms = GPU 做完到上屏;@drawable_ms 出图线程等 nextDrawable
|
||
# Slow-frame attribution: outputs a list of slow frames, each broken into CPU side / wait side / GPU side. Two kinds of input:
|
||
# · Host log (boot.txt or the stream read back live via the log port, with DM_SLOWFRAME=1): contains only "[slowframe]\tframe number\t@name=value..." lines for dropped frames
|
||
# and one denominator line per 300 frames; segmentation by reading phase uses the UIMSG in the same log; usage slowframe.py boot.txt or ... | slowframe.py -
|
||
# · Per-frame file (DM_NVN_TSV, optionally with DM_NVN_TSV_CALLS=0) + optional boot log: slowframe.py frames.tsv [boot.txt]
|
||
# --all also lists slow frames in the launch / loading phases frame by frame
|
||
# Slow-frame definition: a present interval @presented_ms > 16.7×1.5 ms or −1 (dropped by the system) counts as one dropped frame; frames before entering the game are listed separately.
|
||
# Logs without present times (Mac without a window) fall back to @frame_ms > 25 ms, noted in the table header.
|
||
# Components (ms, all for that frame number, i.e. between the previous frame-output call and this frame-output call):
|
||
# Interval @frame_ms interval between the frame-output thread's two calls to QueuePresent
|
||
# CPU @cpu_ms the frame-output thread's time on CPU this frame (−1 = the thread changed); of which @sub_cpu_ms is submission replay, @upload_ms texture upload,
|
||
# @lib_new / @pso_new the number of synchronous library / pipeline builds; @proc_cpu_ms CPU of all threads in the whole process (to see whether other threads saturate the cores)
|
||
# Wait @wait_ms waiting on the 60 Hz tick when acquiring the image (a voluntary yield, not a problem); @pq_wait_ms waiting for the frame-output queue to return a swapchain texture;
|
||
# other off-CPU = interval − CPU − tick wait − queue wait (lock waits, preemption, guest synchronization, etc.)
|
||
# GPU @gpu_ms sum of GPU busy time of this frame's command buffers; @gpu_end_ms − @call_ms = how long after the frame-output call the GPU finished;
|
||
# @pres_at_ms − @gpu_end_ms = from GPU finish to present; @drawable_ms the frame-output thread waiting on nextDrawable
|
||
import re, sys, collections
|
||
|
||
PERIOD = 1000 / 60
|
||
DROP = PERIOD * 1.5
|
||
|
||
def load(lines):
|
||
"""返回 (逐帧字段, 日志分母行 (judged, slow, 判据) 或 None, 原始行列表供分段) | Returns (per-frame fields, log denominator line (judged, slow, criterion) or None, list of raw lines for segmentation)"""
|
||
F = collections.defaultdict(dict); den = None; raw = []
|
||
for l in lines:
|
||
raw.append(l)
|
||
p = l.rstrip("\n").split("\t")
|
||
if p[0].endswith("[slowframe]") and len(p) > 2: # 日志行(前面可能带日志前缀) | Log line (may be preceded by a log prefix)
|
||
if p[1] == "#":
|
||
kv = dict(x.split("=", 1) for x in p[3:] if "=" in x)
|
||
den = (int(kv["judged"]), int(kv["slow"]), kv.get("判据", "?"))
|
||
elif p[1].isdigit():
|
||
for x in p[2:]:
|
||
k, _, v = x.partition("=")
|
||
try: F[int(p[1])][k] = float(v)
|
||
except ValueError: pass
|
||
continue
|
||
if len(p) != 3 or not p[0].isdigit() or not p[1].startswith("@"): continue # 末行可能被截断 | The last line may be truncated
|
||
try: F[int(p[0])][p[1]] = float(p[2])
|
||
except ValueError: pass
|
||
return F, den, raw
|
||
|
||
def phases(lines):
|
||
"""UIMSG 读取画面区间 → (第一次打开读取画面的时刻, [(开, 关)…]);没有任何 UIMSG 行返回 None | UIMSG loading screen intervals → (time the loading screen was first opened, [(open, close)...]); returns None if there are no UIMSG lines at all"""
|
||
iv, o, first, seen = [], None, None, False
|
||
for l in lines:
|
||
if "UIMSG" in l: seen = True
|
||
m = re.search(r"UIMSG \[([\d.]+)s\] PanelManager (OpenPanel|UnloadPanel) LoadScreenPanel", l)
|
||
if not m: continue
|
||
t = float(m.group(1))
|
||
if m.group(2) == "OpenPanel": o = t; first = t if first is None else first
|
||
elif o is not None: iv.append((o, t)); o = None
|
||
if o is not None: iv.append((o, 1e18))
|
||
return (first, iv) if seen else None
|
||
|
||
def phase_of(d, ph):
|
||
if ph is None or "@t_s" not in d: return "未分段"
|
||
first, iv = ph
|
||
t = d["@t_s"]
|
||
if first is None or t < first: return "前端"
|
||
return "读取" if any(a <= t <= b for a, b in iv) else "局内"
|
||
|
||
def g(d, k):
|
||
return d.get(k)
|
||
|
||
def attribute(d):
|
||
"""返回 (CPU 侧 ms, 等待侧 ms, 离 CPU 其余 ms, GPU 侧描述, 判定) | Returns (CPU-side ms, wait-side ms, other off-CPU ms, GPU-side description, verdict)"""
|
||
fr, cpu = g(d, "@frame_ms"), g(d, "@cpu_ms")
|
||
wait, pq = g(d, "@wait_ms") or 0.0, g(d, "@pq_wait_ms") or 0.0
|
||
cpu_ok = cpu is not None and cpu >= 0
|
||
other = fr - cpu - wait - pq if fr is not None and cpu_ok else None
|
||
call, g1, pres = g(d, "@call_ms"), g(d, "@gpu_end_ms"), g(d, "@pres_at_ms")
|
||
gpu_after = g1 - call if g1 is not None and call is not None else None
|
||
disp = pres - g1 if pres is not None and g1 is not None else None
|
||
# 判定:出图调用间隔没超(CPU 按时交了帧)→ GPU / 显示侧;超了 → 看间隔里最大的一块
|
||
# Verdict: present call interval not exceeded (the CPU delivered the frame on time) → GPU / display side; exceeded → look at the largest piece within the interval
|
||
if fr is None: verdict = "缺 @frame_ms"
|
||
elif fr <= DROP:
|
||
if gpu_after is not None and gpu_after > PERIOD: verdict = "GPU 侧(出图调用后 %.1f ms 才做完)" % gpu_after
|
||
elif (g(d, "@drawable_ms") or 0) > PERIOD / 2: verdict = "显示侧(等 drawable %.1f ms)" % d["@drawable_ms"]
|
||
else: verdict = "CPU 按时、GPU 按时:看前几帧积压 / 系统丢弃"
|
||
elif not cpu_ok: verdict = "间隔超时,缺 @cpu_ms(旧日志或换线程),无法拆"
|
||
else:
|
||
parts = {"CPU": cpu, "队列等": pq, "离 CPU 其余(锁 / 抢占)": other}
|
||
k = max(parts, key=parts.get)
|
||
verdict = "%s 主导 %.1f / %.1f ms" % (k, parts[k], fr)
|
||
return cpu if cpu_ok else None, wait + pq, other, gpu_after, disp, verdict
|
||
|
||
def fmt(v, w=6):
|
||
return ("%*.1f" % (w, v)) if v is not None else " " * (w - 1) + "-"
|
||
|
||
def main(argv):
|
||
args = [a for a in argv if not a.startswith("--")]
|
||
if not args: sys.exit(__doc__ or "用法:slowframe.py frames.tsv [boot.txt] [--all]")
|
||
show_all = "--all" in argv
|
||
src = sys.stdin if args[0] == "-" else open(args[0], errors="replace")
|
||
F, den, raw = load(src)
|
||
ph = phases(open(args[1], errors="replace") if len(args) > 1 else raw)
|
||
frames = sorted(F)
|
||
has_pres = any("@presented_ms" in F[f] for f in frames)
|
||
if den is not None: # 日志输入:只有掉帧帧的行,分母来自宿主每 300 帧的统计行 | Log input: only lines for dropped frames; the denominator comes from the host's stats line every 300 frames
|
||
print(f"# 日志输入(DM_SLOWFRAME):宿主判据 {den[2]},宿主统计 可判 {den[0]} 帧、掉帧 {den[1]} 帧(截至最后一行分母);日志里掉帧行 {len(frames)} 帧")
|
||
key = "@presented_ms" if has_pres else "@frame_ms"
|
||
def slow(d):
|
||
v = d.get(key)
|
||
return v is not None and (v > DROP or (has_pres and v < 0))
|
||
judged = [f for f in frames if key in F[f]]
|
||
S = [f for f in judged if slow(F[f])]
|
||
n_cpu = sum(1 for f in frames if F[f].get("@cpu_ms", -1) >= 0)
|
||
print(f"# 帧 {len(frames)}({frames[0] if frames else '-'}–{frames[-1] if frames else '-'});判据 {key} > {DROP:.1f} ms"
|
||
+ (" 或 −1" if has_pres else "(无上屏时刻,退回出图调用间隔)")
|
||
+ f";可判 {len(judged)} 帧;有 @cpu_ms {n_cpu} 帧;分段 {'boot 日志' if ph else '无'}")
|
||
by = collections.Counter(phase_of(F[f], ph) for f in judged)
|
||
sl = collections.Counter(phase_of(F[f], ph) for f in S)
|
||
if den is None: print("# 各段 慢帧 / 可判帧:" + ";".join(f"{p} {sl[p]}/{by[p]}" for p in ("前端", "读取", "局内", "未分段") if by[p]))
|
||
else: print("# 各段 慢帧:" + (";".join(f"{p} {sl[p]}" for p in ("前端", "读取", "局内", "未分段") if sl[p]) or "无"))
|
||
rows = [f for f in S if show_all or phase_of(F[f], ph) in ("局内", "未分段")]
|
||
print("帧\tt_s\t阶段\t上屏ms\t间隔ms\tCPU\t其中提交\t上传\t建库/管线\t进程CPU\t节拍+队列等\t离CPU其余\tGPU忙\tGPU完成-调用\t完成-上屏\t等drawable\t前3帧最大间隔\t判定")
|
||
verdicts = collections.Counter()
|
||
for f in rows:
|
||
d = F[f]
|
||
cpu, waits, other, gafter, disp, verdict = attribute(d)
|
||
prev = d.get("@prev3_max_frame_ms", max((F[p].get("@frame_ms", 0) for p in range(f - 3, f) if p in F), default=None))
|
||
verdicts[verdict.split("(")[0].split(" 主导")[0]] += 1
|
||
cr = "%d/%d" % (d.get("@lib_new", 0), d.get("@pso_new", 0)) if "@lib_new" in d else "-"
|
||
print("\t".join([str(f), fmt(d.get("@t_s"), 1).strip(), phase_of(d, ph), fmt(d.get("@presented_ms")).strip(), fmt(d.get("@frame_ms")).strip(),
|
||
fmt(cpu).strip(), fmt(d.get("@sub_cpu_ms")).strip(), fmt(d.get("@upload_ms")).strip(), cr, fmt(d.get("@proc_cpu_ms")).strip(),
|
||
fmt(waits).strip(), fmt(other).strip(), fmt(d.get("@gpu_ms")).strip(), fmt(gafter).strip(), fmt(disp).strip(),
|
||
fmt(d.get("@drawable_ms")).strip(), fmt(prev).strip(), verdict]))
|
||
print("# 列出 %d 帧%s;判定汇总:%s" % (len(rows), "" if show_all else "(只列局内 / 未分段,--all 看全部)",
|
||
";".join(f"{k} {v}" for k, v in verdicts.most_common()) or "无"))
|
||
|
||
if __name__ == "__main__":
|
||
main(sys.argv[1:])
|