swung0x48andClaude Fable 5.1
bf3e6f8ad3
Gate 8's last two numbers (SEG_STAGE bytes/frame, records/frame post- chunking) had no producer at all: the summary line's byte classes are backend-to-driver bytes and maxrec is a max, not a per-frame. Two push- guarded counters answer them at their single choke points - ByteClass:: StageSegmentBytes at PipeWireCodec's StageAllocate (the only allocation behind StageBytes, so every buffer walk, texture slab and CSO archive lands there and emitters cannot double-count) and CallClass::WireRecords on EncodeRecord's committed path (post-chunking an emitter no longer knows how many records its one call produced). A staged-blob histogram rides along. G1 holds: symbols 0/0/0/0, .text byte-identical, pull build carries no MG_Remote symbols. The spawn server never counted anything: PipeStats::Init() is a step of MobileGL::Initialize(), which ServerMain never reaches, so every Enabled() site in that process - including ServerLoop's own ServerWaits/ServerParks publish - stayed false and the server log carried no summary line at all (a zero that means 'another process' printed in the shape of one that means 'never waited' is the milder misreading; the line was simply absent). ServerMain now arms and closes PipeStats at its role boundary - after MG_ConfigLoader::Init (Init latches from the parsed config), after the server-role latch (the log sink's role), and before ::_exit(0). Red-once: 0 server summary lines before, 29 after, SSIM 1.0 unchanged. MOBILEGL_PIPE_STATS_FILE now derives per role through the log sink's exported RoleLogPath rule - run.client.json / run.server.json, the bare name resolves to no file - because both roles reach FormatJson() and ofstream trunc made the second writer replace the first. G1 caught three separate leaks on the way (a string temporary in the monolith body, one resized Shutdown() from unequal #if arms, two rewritten MGLOG_W literals in .rodata); the pull arm is statement-for-statement the original. Device tooling for the paired A/B lands with it: pin_device.sh gains a measured 2f7cbe2e entry (this board clamps the GPU at 1050 MHz - thermal_pwrlevel writes 0 and reads back 1, max_gpuclk is read-only; absolute times are not comparable with the 1100 MHz-era rows), and tools/device_bench/p6/ carries the session collector and renderer beside run_android_retrace_local.py's MOBILEGL_TRACE_PACKAGE override. Unit 2367/2367 (PipeStats 25/25), spawn integration 102/102. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
device_bench — in-game FPS benchmark harness
Scripted, repeatable in-game FPS measurement for MobileGL's two Android backends (Espryt/DirectGLES and Magma/DirectVulkan) plus a MobileGlues reference run, driven through the FCL fordebug flavor. Intended for A/B performance work and release regression gates on real devices.
How it measures
FCL's in-game FPS overlay counts eglSwapBuffers calls natively (renderer-
agnostic, not vsync-capped when the game runs with vsync off). When the overlay
is enabled, FCL's FPS thread logs one FCLFPS: <n> logcat line per second;
bench.sh collects those lines during the measurement window and reports
mean / median / min / max / stdev, alongside GPU busy%, SoC temperature, and
frequency-pin integrity.
One-time setup (per device / world)
-
Install the FCL fordebug flavor (
com.tungsten.fcl.mgdebug.debug). Its splash auto-launches the selected profile into the prepared world after a 5 s countdown. -
In-game menu: enable show FPS (persists in
files/menu_setting.json). -
Prepare the benchmark world: fixed camera position, gamerules
doMobSpawning/doDaylightCycle/doWeatherCycle=false, then save & quit once.bench.shalwaysam force-stops the game (never saves), so every run replays the same state. -
options.txt: desiredrenderDistance,enableVsync:false, highmaxFps,inactivityFpsLimit:"minimized"(the "afk" default locks 30 fps after 60 s without input and ruins the window). -
Root required (frequency pinning, GPU busy sampling).
-
Write a device profile under
devices/(seedevices/odinlite.env).A profile carries
PROFILE_VERIFIED=1only once its sysfs nodes and OPPs have been read off that device and one pinned window has been checked against them (big_cur/little_cur/gpu_cur_khzin the result JSON must match the pins). Until then it saysPROFILE_VERIFIED=0andbench.sh/session.shrefuse to run against it unless--allow-unverified-profileis passed, which labels the run unpinned in the warning. A profile that omits the key entirely is refused the same way - the guard defaults to unverified, so copying a verified profile and editing the serial cannot inherit its verdict. Only the file can answer: both scripts resetPROFILE_VERIFIED=0immediately before sourcing it, soPROFILE_VERIFIED=1exported in your shell does not re-open the hole. Nor does a profile path that cannot be read get mistaken for an unverified one - it is reported as unreadable, and a path relative to the directory you ran the script from is resolved. (profile.shpins nothing - it records a simpleperf profile - so it carries no such guard.)That refusal exists because the pin path is silent when it is wrong: the harness writes through
/proc/ppm/policy/hard_userlimit_*and/proc/gpufreq/gpufreq_opp_freq, which are MediaTek nodes, andsu -c 'echo ... > /proc/...'against a device that has neither fails without a non-zero exit. The run then reports numbers it believes were taken under a pin.
Devices
| profile | device | verified |
|---|---|---|
devices/odinlite.env |
AYN Odin Lite, MT6877 / Mali-G68 | yes |
devices/xiaomi-adreno830.env |
Xiaomi, Snapdragon 8 Elite / Adreno 830 (35d0befa) |
no - Qualcomm pin path not yet taught to bench.sh |
devices/oppo-mali.env |
Oppo / ColorOS, MediaTek + Mali (3B159D009VZ00000) |
no - OPPs and thermal zone not yet read off the device |
Usage
./bench.sh --device devices/odinlite.env --backend magma # 30 samples, 180 s warmup
./bench.sh --device devices/odinlite.env --backend espryt --label after-fix-X
./bench.sh --device devices/odinlite.env --backend mobileglues # reference
Results append to results/results.jsonl; per-run screenshots (pre.png,
post.png) land in results/<timestamp>-<backend>[-label]/ — always eyeball
them: the pre/post pair must show the same scene, or the run is invalid.
Protocol discipline (hard-won, do not skip)
- Thermal gate: the script waits for the profile's start-temperature threshold. Runs started hot are not comparable to runs started cool.
- Warmup 180 s: ART JIT takes ~3 min to plateau (62→67→84 fps ramp was measured); short warmups underestimate by 10-20%.
- Pins can be overridden by the thermal engine. The result JSON records
big_cur/little_cur/gpu_cur_khzsampled at window end — discard the run if they do not match the profile pins. - Paired runs: absolute FPS drifts across sessions (camera angle, world state). A/B comparisons must be back-to-back runs in the same session.
- F3 off for standard numbers (the F3 debug overlay multiplies per-draw overhead and skews backends differently).
- The FPS overlay itself must be ON (it is what produces the FCLFPS lines).
Pinning on the two MGPipe campaign devices (2026-09-07)
bench.sh's pin_freqs() writes /proc/ppm/policy/* and /proc/gpufreq/gpufreq_opp_freq; neither
path exists on 35d0befa (SM8750) nor on 3B159D009VZ00000 (MT6993, which dropped both legacy
interfaces for /proc/gpufreqv2/). For those two devices run bench.sh --no-pin and pin with
tools/device_bench/pin_device.sh <serial> pin|unpin|check (pure adb + su; exit 0 PINNED, 1 DRIFT,
2 UNPINNED, so check && measure cannot measure unpinned). The profiles under devices/ carry
PROFILE_VERIFIED=1 for the nodes and pins named in the file, not for bench.sh's ability to drive
them; the verification evidence is docs/Disaggregated/devices/pin-verification-2026-09-07.md.