Godot performance and release

Target: Godot 4.7.2.stable, macOS Apple Silicon.

Expert level here means every claim about speed comes with a before and after number from the same harness, every build is verified as a file before anyone calls it shipped, and the agent knows which numbers Godot reports wrongly. The stance: measure, classify the bound, fix the biggest cost, re-measure, compare pictures, then ship builds you have opened, run and listed. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-godot-expert (channels, the review loop, toolkit, 4.7 traps).

Stance (the expert delta)

1. Classify before fixing. Frame time is the slower of CPU and GPU, so a fix on the other side gains nothing (general optimization doc). Live: cutting draw calls from 2,501 to 2 dropped CPU from 2.7 to 0.35 ms and left the GPU at 6.2 ms; only the shadow budget moved the GPU (to 1.7 ms). Run a half-resolution pass: when GPU time does not fall, resolution is not the problem [added]. First log RenderingServer.get_video_adapter_name(): a hybrid laptop may run on its integrated GPU; --gpu-index N picks another [added].
2. Nodes cost before pixels do. Per-node _process, thousands of nodes and redundant physics objects are the usual Godot CPU cost (Firebelley 1PnG3r1rKcs: 115 ms to about 8 ms with RenderingServer items; DwMoEdAhtYQ, O5a4AKemkiQ). Replace repeats with MultiMesh and a shader, and centralize ticking. Live: 2,000 MeshInstance3D created in 2.3 ms, 0.6 ms through RenderingServer.instance_create2.
3. Spread periodic work. Random phase per poller, force_raycast_update() on a disabled RayCast (Dan Does Dev, s2C2RO_WMh0 [00:05:34]). Live: 600 agents, worst tick 1.93 to 0.64 ms, spike frames 62 to 16.
4. Hand-sized occluders, judged by objects drawn. Box occluders per big wall beat a whole-scene bake (Zenva, VBiBZBVxu1s: about 300 to 80 FPS baked versus 350 with boxes). Occlusion is a CPU raster: judge by objects drawn from several camera spots.
5. Hitches are loading, pipelines and runtime bakes. Load AND instantiate off the main thread, ship a warm pipeline cache (Shader Baker export option, 4.5), and log the PIPELINE_COMPILATIONS_* deltas per frame so each hitch is classed as load or compile. Navmesh bakes and ReflectionProbes on Always at area boundaries are the other suspects [added]. Live worst frame: synchronous 453 ms cold, threaded load plus instantiate 230 ms cold, 9.6 ms warm.
6. Test on the real target, long enough. Export With Debug does not slow the game; throttling shows only on long runs (Ian Bolton, Arm, WrjaUNAXYqk [00:05:20, 00:09:35]). Soak a phone 30 minutes in the heaviest area and gate on minutes 20 to 30 [added duration]. Clean build for every perf report: a stale cache hid a perf bug for weeks (Claire Blackshaw, grdqHJOL5F4 [00:16:15]).
7. Exports fail on environment and credentials, not on Godot. Templates per exact version, JDK and SDK paths, ETC2/ASTC import, bundle ids, signing identities, server headers. Credentials go in GODOT_* env vars, never in export_presets.cfg (export docs; Nihamkin _n9l15CANag [00:25:38]).
8. Web is Compatibility only and single-threaded by default. Threads need COOP same-origin plus COEP require-corp over HTTPS (web doc). Live: the threaded build without headers does not start.
9. A written plan must run without this toolkit. Next to every gd_perf call, give the engine code or CLI it wraps: the logger below, godot --headless --export-release, the env vars, bundletool, xcodebuild. Round-1 blind grading docked plans a reader could not run.

Establish first

Input	Why it changes the plan	Default if unknown
Target hardware and frame budget	16.67 ms at 60 fps, 33.33 at 30, 11.11 at 90 (VR), 8.33 at 120	60 fps on the lowest named device
Renderer per platform	auto-instancing, occlusion, features differ; web forces Compatibility	Forward+ desktop, Mobile on phones, Compatibility on web
Where it is slow	steady low fps versus hitches when loading or turning	profile both
Platforms and stores	signing, ids, size limits, AAB versus APK	ask; each store needs account steps
Download budget	wasm, pck, APK sizes	the bytes the host really serves
Who holds credentials	keystores, Apple team, Steam builder	the agent never invents or stores secrets

Workflow

1. Baseline on a clean project copy. Clone with gd_env.base_project, gd_perf.install(P), rename application/config/name (Base3D clones otherwise share one user:// and shader cache). Read project.godot back: Jolt for 3D, the renderer you mean. gd_perf.scene_audit on the heavy scene. GATE: audit JSON saved, physics engine and renderer confirmed.
2. Measure windowed, per variable. gd_perf.profile_variant(P, scene, size=(1280, 720), driver="vulkan"): per-frame CSV, visible and shadow draw calls, scene _process time from sentinel nodes; it compares settings, toggles or rendering_method variants after gd_perf.install(P), while a single verdict on an unchanged scene needs only the lead's gd_run.profile_scene(driver="vulkan"). Never use Performance.TIME_PROCESS per frame: it holds one value for about a second. Metal reports GPU 0 on macOS; use Vulkan. Windowed frame time is capped at the refresh; judge gpu_ms and cpu_ms. Deeper: --gpu-profile, the Visual Profiler, the script Profiler (read Self, not Inclusive; it does not cover C#), RenderDoc or the vendor tool for one frame. GATE: two runs within 10 % [added].
3. Classify. gd_perf.classify(full, half_res). GATE: a written verdict (CPU scripts, CPU render, GPU fill, GPU geometry or shadows) backed by numbers.
4. Fix one thing, re-measure, look. gd_perf.apply_fix(P, scene, out, fix) writes a new scene: share_materials, multimesh, shadow_budget, box_occluders, visibility_range, strip_process (lab only: drops _process scripts once a shader does their work). gd_perf.compare(before, after) applies a 5 % and 0.1 ms noise floor. Capture before and after, contact sheet, open it. Resolution scale below an agreed floor (0.67) needs sign-off [added]. GATE: metric improved beyond noise, nothing regressed, mean luma difference under 0.02 or the change approved.
5. Hitch pass. perf_bench.gd:hitch with a fresh private HOME for a cold shader cache. GATE: worst frame under 2x the budget on warm cache, pipeline compilations off gameplay frames.
6. Ship builds. Presets through gd_run.ensure_preset plus gd_perf.preset_fields; exports under gd_perf.export_env(HOME=private_home). Plain CLI: godot --headless --path . --export-release "<preset>" <out>. Verify each artifact: verify_macos, verify_apk, verify_xcode_project, pck_listing, size_report, web_smoke. GATE: per platform below.
7. Run what you exported. gd_perf.run_exported(exe, ["--headless", "--", "--perf-log", "--perf-quit"]) with perf_logger.gd; web through headless Chrome. GATE: SHIP_BOOT shows the expected tags and renderer, PERF_SUMMARY p95 within budget, no errors.
8. Release audit and handover. gd_perf.release_audit(P); write signing, notarization and upload steps for the human holding the accounts (notarize_plan, steam_vdf). GATE: no high flag; account steps listed as "not run" with the reason.

Phones. Mobile renderer, or Compatibility for the widest low-end reach (Ray Hayes, KWhVVMpihsc [00:06:03]); fallback_to_opengl3 is on by default and changes the look, so capture the fallback too [00:05:12]. Low tier: scaling_3d_scale 0.67 to 0.8, Engine.max_fps = 30, physics tick 30 with interpolation on (O5a4AKemkiQ [00:03:43]). Apply the safe area on Android too: 4.7.2 APKs target SDK 36 (measured), so Android 15+ draws edge to edge [added]. Android build: --install-android-build-template only together with an export command (with --quit it installs nothing, measured), keystore in GODOT_ANDROID_KEYSTORE_RELEASE_{PATH,USER,PASSWORD}, size with bundletool build-apks --connected-device then get-size total [not run]. Thermal soak: log adb shell dumpsys thermalservice every 10 s; Streamline region markers and Performance Advisor in CI (WrjaUNAXYqk [00:14:43]); Performance.add_custom_monitor for on-device game metrics. Wire a crash reporter and test the next OS beta before launch (8zqjiRuhQLE [00:03:40, 00:09:11]). iOS: export writes an Xcode project, then xcodebuild archive and -exportArchive; create the App Store Connect record before the first upload, raise the build number each time, no simulator for Vulkan renderers, sudo xcode-select -s /Applications/Xcode.app/Contents/Developer when export says "xcodebuild requires Xcode" (x_ZFAV2id3I [00:14:41, 00:17:04, 00:40:34]). IAP: store plugins and accounts, never faked.

Web. Single-threaded unless threads were measured to matter. itch.io serves files as uploaded (no on-the-fly compression), so budget the bytes it sends. Background tabs pause _process and drop networked games; HTTPRequest has no chunked responses and one progress step per frame; a stale service worker can show another project (web doc). Once isolated, cross-origin fetches need CORS or CORP [added]. C# projects cannot export to web (version deltas 16). Leaderboard read-only in the browser.

Engine-only frame logger (no toolkit, run 2026-10-02)

extends Node   # autoload "Bench"; run the game with: -- --bench
var ai_ms := 0.0
var _vp: RID
var _rows := PackedStringArray(["frame_ms,render_cpu_ms,gpu_ms,draw_calls,pipelines_total"])
func _ready() -> void:
	if not "--bench" in OS.get_cmdline_user_args():
		set_process(false)
		return
	_vp = get_viewport().get_viewport_rid()
	RenderingServer.viewport_set_measure_render_time(_vp, true)
	Performance.add_custom_monitor("game/ai_ms", _get_ai_ms)   # a method, never a lambda
	print("BENCH adapter=", RenderingServer.get_video_adapter_name())
func _get_ai_ms() -> float:
	return ai_ms
func _process(delta: float) -> void:
	var pipes := Performance.get_monitor(Performance.PIPELINE_COMPILATIONS_SURFACE) \
		+ Performance.get_monitor(Performance.PIPELINE_COMPILATIONS_SPECIALIZATION)
	_rows.append("%.2f,%.3f,%.3f,%d,%d" % [delta * 1000.0,
		RenderingServer.viewport_get_measured_render_time_cpu(_vp),
		RenderingServer.viewport_get_measured_render_time_gpu(_vp),
		Performance.get_monitor(Performance.RENDER_TOTAL_DRAW_CALLS_IN_FRAME), pipes])
func _exit_tree() -> void:
	if is_processing():
		Performance.remove_custom_monitor("game/ai_ms")
		FileAccess.open("user://bench.csv", FileAccess.WRITE).store_string("\n".join(_rows))

Pipeline counters are cumulative: a hitch row whose total rose is a compile. GPU time reads 0 for the first frames.

Numbers

Value	Relative to	Source
hitch = frame over 2x the median	per run	toolkit [added]
2,501 to 33 draw calls	2,500 cubes, 8 shared materials, Forward+	live, auto-instancing
2,501 to 2,501	same on Mobile	live: no auto-instancing
400 transparent cubes: 1 draw, 2 alternating materials: 117 (opaque: 2)	Forward+	live: depth sorting breaks batches
400 Label3D: 800 draws; one shared TextMesh: 1	Forward+	live: outline plus text, never instanced
omni shadow: Cube (default) 19 draws, Dual Paraboloid 2	Forward+, 400 casters	live; Mobile reversed (124 vs 220)
6.6 to 1.7 ms GPU	16 to 4 shadowed omni lights	live
1,260 to 90 objects	one box occluder	live, all three renderers
38.1 MB raw, 9.7 gzip, 7.8 brotli	4.7.2 web export, small 3D scene	live; wasm dominates
26.5 to 28.3 MB	Android arm64 APK, release / debug	live; engine library 24 MB

Quality gates

• Measured: p95 frame or GPU time within budget on the target class; worst frame on area load; compare shows the intended metric moved and none regressed; scene_audit has no high flag.
• Visual: before and after contact sheet (1600 px max) opened; occluder debug capture; web canvas screenshot.
• macOS: codesign --verify --deep --strict OK, Debugging entitlement off, both archs; Gatekeeper accepts only after notarization.
• Android: apksigner verifies, signer SHA-256 equals the keystore, versionCode raised, release not debuggable, planned ABIs, no password in presets.
• iOS: Xcode project with scheme, bundle id and team; unsigned xcodebuild passes once the iOS platform is installed.
• Web: boots in headless Chrome as served; crossOriginIsolated true if threaded; no page errors; served size within budget.
• Packs: base pck holds no excluded content; DLC loads with replace_files=false; demo tag only in the demo.

Common mistakes

Mistake	What it looks like	Fix
process_ms p95 from the monitor CSV	the same value on 120 rows	sentinel timing
GPU numbers on Metal	gpu_ms 0	--rendering-driver vulkan
Headless frame times	"60 fps" with nothing drawn	windowed runs
MultiMesh built in a headless job	saved scene has no instances	write MultiMesh.buffer
Shared materials on Mobile	draw calls unchanged	MultiMesh
Mixed transparent materials in one area	draw calls climb with object count	fewer transparent materials; split mixed opaque and alpha surfaces (3D doc)
Label3D for nameplates or damage numbers	2 draws per label	shared TextMesh or one text node (grdqHJOL5F4 [00:09:33])
Lights left Dynamic beside LightmapGI	dynamic light cost on baked areas	Static bake mode, DirectionalLight3D stays Dynamic (3D doc)
Custom monitor as a lambda	stays at its first value; crashed 4.7.2 at exit (signal 11)	bind a method; remove_custom_monitor on exit
Occlusion on, no occluders	objects drawn unchanged	OccluderInstance3D boxes
Threaded web build on itch without headers	"Cross-Origin Isolation missing"	single-threaded, headers, or PWA workaround
Android paths seeded once	"A valid Java SDK path is required"	set paths in a private editor settings file
DLC mounted with replace_files true	old scripts override new ones	load_resource_pack(path, false)
Passwords in export_presets.cfg	secrets in git	GODOT_ANDROID_KEYSTORE_*, GODOT_MACOS_* env vars

Handoffs

scenario-godot-rendering-lighting (shadow budgets, bake modes, quality tiers), scenario-godot-shaders (MultiMesh vertex animation, Shader Baker), scenario-godot-3d-world (LOD, visibility ranges, occluders, Jolt, Large World Coordinates or origin shifting for big maps), scenario-godot-gameplay and scenario-godot-architecture (managers, pooling, staggered timers, threaded loading), scenario-godot-ui (touch layout, safe areas), scenario-godot-multiplayer (web transport, dedicated server, trusted leaderboard writes), scenario-godot-pipeline-automation (CI exports, clean builds), scenario-godot-audio (Sample playback on web). Humans: Apple Developer ID, App Store Connect, Play Console, Steamworks, consoles, IAP products. Audit flags carry the owner (perf_audit.gd).

Godot 4.7 notes

• OS.get_thermal_state does not exist in 4.7.2: log frame times over the soak, and OS thermal status from adb or Instruments.
• Defaults checked: rendering_method.mobile="mobile", rendering_method.web="gl_compatibility", occlusion culling off, import_etc2_astc false in a hand-written project, physics interpolation off, physics 60 Hz, fallback_to_opengl3 true, OmniLight3D.omni_shadow_mode Cube, light_bake_mode Dynamic, web audio Sample.
• Pck format 4 (pck_listing reads it). Jolt is default for new 3D projects since 4.6; a hand-written project.godot needs the setting.
• Android: JDK 17; Godot seeds JDK and SDK paths from JAVA_HOME and ANDROID_HOME only when it creates its editor settings; permissions are not auto-requested since 4.3.
• iOS export needs application/export_project_only without a signing team; gd_run.export reports failure because no .ipa exists.
• Physics interpolation (2D 4.3, 3D 4.4) makes a lower physics tick viable.

References

• references/procedures.md: procedures 1 to 16 with code, live results and what was not run.
• references/expert-notes.md: claims by expert with timestamps, where live tests disagreed.
• references/critique.md, references/gui-paths.md, references/sources.md.
• scripts/gd_perf.py and scripts/agentkit/perf/ (perf_stress, perf_fixes, perf_audit, perf_bench, perf_logger, ship_boot, web_bridge, mobile_touch).