mirror of
https://github.com/MobileGL-Dev/MobileGL
synced 2026-11-20 00:11:56 +09:00
[Docs] (Disaggregated): P5e wave 3 - the strict lane goes hard green, run-ahead is armed, the device matrix is written back
- CURRENT_STAGE_PROGRESS.md gains section 2.6: the four packages (pa / mv / gl / ra2), the
admission rule's three disjuncts and which real entry forced each one, coverage going from
116 to 179 entries plus 1357, the race the flip exposed and its root cause, and the finding
that a hard-green strict lane is necessary and NOT sufficient (ID-132). Code head is now
25fba0d5.
- MEASUREMENTS.md gains section 11: the device matrix. The section leads with clock pinning
because the first matrix was run without it and had to be thrown away - monolith's OWN arm
moved 4.43 -> 7.07 ms/frame between two runs, and the SLOWER run was the COOLER one: the
governor had dropped policy0 from 2745 to 748 MHz. It records pin_redmi.sh's write order and
read-back verification, and why pin_device.sh refusing to guess this machine's serial is
correct rather than a gap. Conclusion: at both pinned tiers monolith and run-ahead each hit
the 120 Hz panel ceiling and only lockstep cannot reach it.
- ROADMAP.md's P5e row and status line now say twelve packages landed, run-ahead armed.
The phase is NOT closed: which arm is faster ABOVE the panel ceiling is unanswered (this
device cannot hold a higher clock), and BRIEF-P5E section 3's three negative controls and
section 5's per-package red-once reruns have been reordered by ID-121 to after the lane goes
green, so they have not been executed yet.
This commit is contained in:
@@ -508,3 +508,56 @@ monolith 这一次未被 120 Hz vsync 封顶(与 §9 的 205.8 那次同类)
|
||||
- 派给外部 CLI 的任务,其进程会活过 harness 的完成通知;**派出后不得自己再跑同一件事**,否则两套实验同时对一台手机 force-stop / 清 logcat,得出假失败。
|
||||
- 不得在脚本正被 bash 执行时修改它(边读边执行);要改先冻结副本。
|
||||
- 游戏内 F3 的 `GIT@` 戳记在增量构建下是旧的,**判断库版本要看 APK 里 `.so` 的符号 / 字符串**,不看戳记。
|
||||
|
||||
## 11. P5e wave 3 — run-ahead 武装后的设备矩阵(头 `25fba0d5`)
|
||||
|
||||
Redmi `2f7cbe2e`,FCL fordebug + Minecraft 26.3-rc-3 世界 `test`,VD12,DirectGLES,~850 draws/帧。
|
||||
**CPU 定频**(见下),GPU 钉 pwrlevel 0 且每臂重新断言,风扇开,进世界后 20 s 稳定再取 30 s 窗口,臂交错。
|
||||
|
||||
### 定频,以及为什么这一节必须先讲它
|
||||
|
||||
第一轮矩阵没有定频,于是 **monolith 自己那一臂**的逐帧 CPU 在两次运行之间从 4.43 ms 跳到 7.07 ms,
|
||||
而慢的那次**温度更低**(53.5 °C vs 58.9 °C)——不是热降频,是 walt governor 把 policy0 从 2745 MHz
|
||||
拉到 748 MHz。未定频的逐线程 CPU ms/帧**不能跨臂比较**,那一轮矩阵作废。
|
||||
|
||||
`tools/device_bench/pin_device.sh` 不认识本机 serial,并**明确拒绝猜**("pin path 因 SoC 而异,猜错会静默失败")——
|
||||
这是对的,也正是它没有静默半应用的原因。按本机实测节点另写 `pin_redmi.sh`:写序 min→硬件底、max→目标、
|
||||
min→目标(顺序要紧,否则按 stock 相对目标的位置会半应用),并**回读 `scaling_cur_freq` 确认**,跑完再 check 一次。
|
||||
|
||||
### 结果(每臂两次,离散度 < 2%)
|
||||
|
||||
定频 little 1555200 / big 1958400:
|
||||
|
||||
| 臂 | fps p50 | fps max | client ms/帧 | apply ms/帧 |
|
||||
|---|---|---|---|---|
|
||||
| monolith | 115.1 | 117.0 | 6.19 | — |
|
||||
| inproc run-ahead | 117.6 / 117.9 | 118.3 / 118.6 | 6.71 / 6.63 | 4.30 / 4.32 |
|
||||
| inproc lockstep (`MOBILEGL_IPC_RUN_AHEAD=0`) | 111.1 / 110.7 | 114.9 / 115.1 | 8.27 / 8.09 | 6.44 / 6.41 |
|
||||
|
||||
定频 little 1996800 / big 1958400(更高会被厂商限幅器夹住):
|
||||
|
||||
| 臂 | fps p50 | fps max | client ms/帧 | apply ms/帧 |
|
||||
|---|---|---|---|---|
|
||||
| monolith | 115.6 / 115.4 | 117.2 / 117.5 | 6.45 / 6.25 | — |
|
||||
| inproc run-ahead | 117.6 / 118.0 | 118.5 / 118.6 | 6.92 / 6.87 | 4.43 / 4.35 |
|
||||
| inproc lockstep | 111.6 | 113.7 | 8.16 | 6.53 |
|
||||
|
||||
### 读法
|
||||
|
||||
- **配对 A/B**:`MOBILEGL_IPC_RUN_AHEAD=0` 与默认是同一构建、同一份记录、同一定频,唯一差别是客户端等不等
|
||||
(ID-114:**不得**用 `VERB_BARRIER=0`,它是 `RunAheadArmed()` 的第一个合取项,会顺手关掉 run-ahead,
|
||||
而且实测几秒内即 `Fatal{UnmigratedPipeInput, "IsCapabilityEnabled@ClientWaitSync"}`)。
|
||||
差值:client **−18%**、apply **−33%**、fps p50 **+6%**。
|
||||
- **两个定频档下 monolith 与 run-ahead 都撞到 120 Hz 面板上限**(max 117-118.6),**只有 lockstep 撞不到**
|
||||
(max 113.7-115.1)。所以"inproc 是否追平 monolith"在本机真实负载上的答案是**追平了**;阶段开始时是
|
||||
**0.62-0.74 倍**(见 §10)。
|
||||
- 逐帧 client CPU:run-ahead 6.9 / monolith 6.3,多约 **8%**,两者都在上限之下有余量。
|
||||
- **未回答**:面板上限之上谁更快。本机定不住更高频率;一次未定频的高频窗口给出 monolith p50 201 /
|
||||
run-ahead p50 181(约 1.10 倍),但那次不可配对,只作方向性提示。需要关 vsync 或换上限更高的设备。
|
||||
|
||||
### 侧写(`25fba0d5` 之前,`7c6f6886`,说明 run-ahead 为什么值得做)
|
||||
|
||||
symbolized `simpleperf cpu-cycles`,客户端 GL 线程:`SessionProducer::WaitForAppliedOrEventBacklog`
|
||||
**31.58% 自身时间**,其后第二名只有 3.72%。设备计数器同期显示客户端每帧进入 doorbell 等待约 **916 次**
|
||||
(对 ~849 次 draw),其中 99.8% 靠自旋解决——`MOBILEGL_IPC_SPIN_US=0` 让每次等待都 park,帧率塌到 20.9。
|
||||
所以代价是**会合本身**,不是自旋参数,这也是为什么修法是"不再每次 draw 会合"而不是"把每次会合做便宜"。
|
||||
|
||||
Reference in New Issue
Block a user