glmark2 on anland: the P15 present-wait stack, before and after

Measured 2026-10-10 on the anland Plasma desktop of the Lenovo TB321FU (Adreno 750). Single runs, one per arm and configuration: windowed light scenes move about +-15 % run to run (see the per-scene tables), so only differences well beyond that mean anything. Method and traps: glmark2-handoff.md.

Arms

arm	commit	notes
before	`a759b4b7` (the `origin/feat/disaggregated` of that day)	server and container client built from the same commit
after	`eb08b0b1` (branch `p15stack`)	present-wait levers, hot-path TLS cut, small-store drain, share-group dirty-bit fix. The pushed tip `bbd555fd` is `eb08b0b1` plus one test-only line in `MG_Test/Pipe/TrackerTest.cpp`, so the measured runtime code is the pushed code
after+x2	`eb08b0b1`	Magma only, `MOBILEGL_MAGMA_EXTRA_SWAPCHAIN_IMAGES=2` in the server process
stock	the original anland app and its container	the stock kgsl driver stack, same glmark2 binaries

Server: NDK RelWithDebInfo for android-30 (native TLS segment present in both libs). Client: Release cross build in WSL. The server and client of an arm are a matched pair: a client of one commit against the other commit's server is refused (Refuse{WireFingerprint} for every session, black desktop), so the arms were switched as pairs (libs swapped, container restarted, app reopened).

Device state and validity

CPU pinned to 1.80 / 2.25 / 2.25 / 2.25 GHz (policy0/2/5/7), GPU 680 MHz (kgsl thermal_pwrlevel 4 is that pin, no cap below it). The msm_performance requests were released, and thermal-engine, the perf HAL, the vendor perf service, performance and hyperschedule were stopped for the whole session (all restarted afterwards).
Sampled every 2 s during every run (CPU cur/max per policy, GPU cur/max/pwrlevel, zone and battery temperatures): every sample of every run shows the pinned frequencies exactly (min = max = pin, GPU 680 MHz, no cap). Peak CPU zone 72-88 C, GPU zone 62-75 C, battery 36.5-40.7 C. Battery 93-94 % throughout (USB attached).
Before each run the clocks fell to their lowest OPP and the CPU/GPU zones had to be below 46 C (at least 45 s). The "battery <= 33 C between reps" cool-down was not applied (the battery sat at 35-40 C and the job was told to start without waiting); the frequency samples above are what validate the runs.
Screen probes every 15 s (two screenshots 1 s apart): no black frame in any scored MobileGL run, nor in the scored stock runs. Two more stock fullscreen runs (1365, 1322) showed the known 19838-byte black desktop mid-run and were discarded; the scored stock fullscreen run (1325) had none.
glmark2 reported Surface Size 800x600 windowed and 2560x1600 fullscreen on every arm (tablet in landscape, auto-rotate untouched). The MobileGL app's extra_keys_mode was set to with_keyboard for the session so its fullscreen size equals stock's (the default shows the keys bar: 2560x1412); restored afterwards.
One run per arm and configuration.

Scores (glmark2-es2-wayland, `-b :duration=5`)

	stock	Espryt before	Espryt after	Magma before	Magma after	Magma after+x2
es2 window 800x600	1038	1900	1839 (-3 %)	1278	1289 (+1 %)	1343 (+5 %)
ratio to stock	1.00	1.83	1.77	1.23	1.24	1.29
es2 fullscreen 2560x1600	1325	521	518 (-1 %)	428	423 (-1 %)	426 (0 %)
ratio to stock	1.00	0.39	0.39	0.32	0.32	0.32

Reading: no before/after difference is outside run-to-run noise. Fullscreen runs are steady (scene by scene within about 5 %) and the stack changes nothing on either backend. Windowed, the whole-score deltas are within +-5 % and the per-scene swings flip sign between arms (Magma "build vbo=true" is -35 % after but the same scene is -6 % in the +x2 arm, linear texture +47 % and cel shading +47 % after), which is noise in 5 s samples, not an effect.

Per-scene FPS

windowed 800x600 (FPS, one run each; * = after differs from before by more than 10 %)

scene	stock	Espryt before	Espryt after	Magma before	Magma after	Magma after+x2
build vbo=false	1185	1340	1359	1660	1676	1447*
build vbo=true	1237	2688	2472	1782	1150*	1677
texture nearest	1153	2184	2331	1377	1628*	1512
texture linear	1173	2405	2226	1168	1712*	1678*
texture mipmap	1391	2065	2330*	1641	1367*	1504
shading gouraud	1335	1912	2359*	844	1152*	1323*
shading blinn-phong-inf	1210	2711	2254*	1667	1463*	1583
shading phong	1087	2411	2394	1322	1185*	1648*
shading cel	1229	2572	2436	1100	1613*	1530*
bump high-poly	1520	2369	2128*	1738	1632	1600
bump normals	1023	2375	2424	1696	1835	1645
bump height	1310	2451	2277	1590	1485	1695
effect2d kernel=0,1,0;1,-4,1;0,1,0;	1168	2033	2147	1562	1512	1350*
effect2d kernel=1,1,1,1,1;1,1,1,1,1;1,1,1,1,1;	1250	2147	2208	1739	1113*	1501*
pulsar light=false:quads=5:texture=false	1131	2196	2327	1616	1761	1717
desktop blur-radius=5:effect=blur:passes=1:sep	881	947	620*	746	804	764
desktop effect=shadow:windows=4	921	748	580*	691	702	944*
buffer columns=200:interleave=false:update-dis	394	592	548	733	793	786
buffer columns=200:interleave=false:update-dis	395	868	556*	821	821	794
buffer columns=200:interleave=true:update-disp	432	400	564*	602	792*	829*
ideas speed=duration	755	522	668*	686	681	634
jellyfish	1230	2097	2079	1651	1597	1590
terrain	402	479	477	423	428	429
shadow	1049	1829	1736	1434	1468	1351
refract	734	1024	1025	653	820*	797*
conditionals f-steps=0:v-steps=0	1160	2388	2157	1497	1525	1468
conditionals f-steps=5:v-steps=0	1109	2475	2110*	1207	1610*	1626*
conditionals f-steps=0:v-steps=5	1030	2308	2326	1770	1530*	1582*
function f-complexity=low:f-steps=5	1091	2321	1949*	1440	1283*	1384
function f-complexity=medium:f-steps=5	1075	2704	2414*	1519	1486	1605
loop f-loop=false:f-steps=5:v-steps=5	1018	2252	2454	1232	1179	1467*
loop f-steps=5:f-uniform=false:v-steps=5	1032	2335	2505	1562	1331*	1379*
loop f-steps=5:f-uniform=true:v-steps=5	1188	2601	2312*	1040	1440*	1518*
score	1038	1900	1839	1278	1289	1343

fullscreen 2560x1600 (FPS, one run each; * = after differs from before by more than 10 %)

scene	stock	Espryt before	Espryt after	Magma before	Magma after	Magma after+x2
build vbo=false	1658	613	621	476	473	468
build vbo=true	1601	718	700	543	554	530
texture nearest	1817	613	552	493	525	527
texture linear	1743	646	646	474	484	489
texture mipmap	1838	649	571*	541	495	509
shading gouraud	1687	555	602	501	504	495
shading blinn-phong-inf	1606	501	530	491	428*	490
shading phong	1753	592	576	496	473	475
shading cel	1675	577	575	476	460	495
bump high-poly	1629	525	574	514	500	510
bump normals	1628	606	579	496	493	533
bump height	1582	616	713*	508	538	539
effect2d kernel=0,1,0;1,-4,1;0,1,0;	1083	491	534	420	421	416
effect2d kernel=1,1,1,1,1;1,1,1,1,1;1,1,1,1,1;	544	399	402	359	365	359
pulsar light=false:quads=5:texture=false	1498	564	504*	383	368	383
desktop blur-radius=5:effect=blur:passes=1:sep	494	349	437*	334	338	318
desktop effect=shadow:windows=4	865	583	459*	379	372	382
buffer columns=200:interleave=false:update-dis	341	497	410*	365	357	357
buffer columns=200:interleave=false:update-dis	416	502	511	407	401	417
buffer columns=200:interleave=true:update-disp	373	507	453*	411	412	396
ideas speed=duration	827	440	431	313	316	315
jellyfish	1462	476	538*	294	289	282
terrain	104	143	144	127	127	127
shadow	628	569	583	453	444	437
refract	217	206	207	208	208	207
conditionals f-steps=0:v-steps=0	1545	525	547	460	455	442
conditionals f-steps=5:v-steps=0	2024	546	548	474	463	459
conditionals f-steps=0:v-steps=5	1558	530	533	458	461	478
function f-complexity=low:f-steps=5	1850	531	508	466	447	454
function f-complexity=medium:f-steps=5	1965	550	547	466	468	466
loop f-loop=false:f-steps=5:v-steps=5	2005	535	542	451	457	458
loop f-steps=5:f-uniform=false:v-steps=5	1849	545	547	468	455	447
loop f-steps=5:f-uniform=true:v-steps=5	1894	549	518	465	457	460
score	1325	521	518	428	423	426

Engagement of the present-wait levers

Magma presents through a VkSwapchain. The server logs Set minImageCount = 5 (surface min 5, max 64, extra 0) and Swapchain created, extent = 1600x2560, swapchain imageCount = 6 for the compositor's on-screen session (the display surface, pre-rotated), and surface min 11 swapchains (800x600, 2560x1600, 664x438, 1x1, ...) for the clients' window surfaces. The clients' frames themselves go to shared images.
MOBILEGL_MAGMA_EXTRA_SWAPCHAIN_IMAGES=2 applies: Set minImageCount = 7 (surface min 5, max 64, extra 2) for the display swapchain, 13 (surface min 11 ... extra 2) for the others. How the server gets an environment variable: the app's server process only takes what its own process environment holds (the app sets its list with Os.setenv; the server never reads debug.mobilegl.env or the config files), so the knob was injected with the Android wrap property setprop "wrap.com.anland.consumer.mobilegl:mobilegl" "NAME=value" before the app starts the process (empty string clears it). The +x2 arm scored +5 % windowed and 0 % fullscreen: within noise.
The Magma "off-screen work ahead of the acquire" lever never logged. The line a frame's off-screen work goes to the queue ahead of its window pass appeared 0 times in all three Magma "after" arms (glmark2 runs, the compositor, a browser WebGL page, window hide/show). The pre-acquire flush with pending off-screen work did not happen on anland: the compositor's frames draw straight into the display image with no off-screen pass in front of it, and the clients present to shared images, not to a swapchain image.
Espryt does render to framebuffer 0 of an EGL window surface: the server logs this process owns a display (ServerDisplay installed); a ServerOwned window surface is served on its window. The once-per-frame flush runs when a frame binds framebuffer 0 after a user framebuffer; it has no log line or counter, so its firing was not observed directly (no tracing, by request). No score effect either way.
Net: on anland these levers cannot show up in glmark2. The wait they target (the display release gating the first swapchain / window-buffer use) sits on the compositor's presentation to the Android surface, which glmark2 only reaches through KWin's composition of its window.

Functional check of the after build

Espryt after: cold start (kwin +4.0 s, plasmashell +5.1 s, desktop on screen +6.6 s), swipe-away and reopen (0.2 s), a browser WebGL page at 114-120 fps, idle visible / hidden / locked at 319 % / 28 % / 29 % phone CPU, lock and unlock, hide and show: no fatal signal, no crash, desktop and page correct in the screenshots.
Magma after: cold start, reopen, the WebGL page (150+ fps), idle visible and hidden, and every glmark2 run were fine. After waking the locked screen with the animated browser page open, the Magma server's memory ran away: system swap fell from 11 GB free to under 1 GB within seconds, the server logged VkBufferObject::Create failed: vmaCreateBuffer returned -1 and Magma wire decline [UniformBufferBinding], the low-memory killer then killed it with signal 9, and the app's restart aborted with DirectVulkan: the initial surface could not be created while the screen was off. The same sequence on the before build (a759b4b7) failed the same way (one run each), so the stack did not cause it. It is worth chasing separately before trusting a long Magma desktop session with animated clients across lock/wake. The glmark2 measurements ran without the browser page and without lock cycles and never lost the server.

13 KiB Raw Blame History