MobileGL/tools/trace_replay/skills/trace-fixture-authoring/SKILL.md

name, description

name	description
trace-fixture-authoring	Author a deterministic MobileGL trace-replay fixture from a captured apitrace - build the in-tree apitrace fork, capture a reproducible scene, frame-trim with gltrim, generate and verify a golden image, package under the archive-size budget, register the case in trace_cases.json, and validate on Linux and Android. Use when adding or re-trimming a trace_replay regression fixture.

Trace fixture authoring

Variables

export REPO="$PWD"
export WORK="$PWD/.trace-work"
export CASE="case-name"
export WIDTH=854
export HEIGHT=480
export TARGET_FRAME=0
export TARGET_CALL=0

Prerequisites

git clone --recursive <mobilegl-repo-url> MobileGL
cd MobileGL
git lfs install
git lfs pull

Install:

• CMake
• Ninja
• C++ compiler
• Python 3
• Mesa OpenGL/EGL runtime
• Vulkan loader and ICD for DirectVulkan
• Pillow or ImageMagick for alpha cleanup
• Android SDK, Android NDK, JDK, Gradle, and adb for Android replay

Build apitrace

Build the in-tree fork, not an upstream release: it carries the frametrim handlers for DSA and ARB_multi_bind call streams and shadow-based tracking of persistent-mapped buffers, all of which modern Minecraft mod stacks need.

cmake -S "$REPO/3rdparty/apitrace" -B "$WORK/build-apitrace" -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DENABLE_GUI=OFF
cmake --build "$WORK/build-apitrace" --target apitrace glretrace gltrim --parallel

export APITRACE="$(find "$WORK/build-apitrace" -type f -name apitrace -perm -111 | head -n 1)"
export GLRETRACE="$(find "$WORK/build-apitrace" -type f -name glretrace -perm -111 | head -n 1)"
export GLTRIM="$(find "$WORK/build-apitrace" -type f -name gltrim -perm -111 | head -n 1)"
test -n "$APITRACE"
test -n "$GLRETRACE"
test -n "$GLTRIM"

On Windows, set APITRACE and GLRETRACE to the corresponding .exe files and also build the wgltrace target: apitrace trace fails with "failed to find opengl32.dll wrapper" unless wrappers/opengl32.dll was built. gltrim may be built and run on a Linux/WSL checkout instead; traces are portable, and trimming a multi-hundred-MB trace is much faster from a native filesystem copy than through /mnt/c.

Prepare the capture

• Set the target window size to WIDTH x HEIGHT.
• Disable unintended overlays, frame counters, notifications, and launcher UI.
• Fix language, resource packs, mods, shader pack, world seed, time, weather, player position, camera direction, FOV, GUI scale, and render distance.
• Trace the final OpenGL process, not the launcher.
• For Minecraft, document version, mod loader, mods, shader pack, language, world, time, and camera setup.

Minecraft specifics that keep the capture deterministic and small:

• Freeze the world in level.dat: doDaylightCycle, doWeatherCycle, doMobSpawning, randomTickSpeed 0, a fixed DayTime, and the player Rotation that frames the intended subject. The camera snaps to the saved rotation on world join, so composition is edited in the save, not in-game.
• options.txt: pauseOnLostFocus:false, a low maxFps (10 works), a small renderDistance (3), and the capture resolution pinned to WIDTH x HEIGHT (854x480) via overrideWidth/overrideHeight (or --width/--height). These are the main levers on fixture size: a low frame rate keeps the full trace short, a small render distance keeps per-frame geometry down, and the 854x480 resolution keeps every render target the frame references small (a trimmed frame's framebuffer/attachment textures scale with resolution squared). A ~35 s in-world session at 10 fps and 854x480 lands well under the archive budget after repack.
• maxFps has a practical floor: Minecraft ignores values below ~10 and falls back to unlimited/vsync (a maxFps:1 capture rendered ~60 fps and ballooned the trace). 10 is as low as this lever goes, so do not count on a lower frame rate to shrink the frame count further.
• Enter the world non-interactively with --quickPlaySingleplayer <world> so every capture takes the same path from boot to gameplay.
• Keep the game window UNFOCUSED for the whole capture (focus the desktop right after launch, and again before closing). A focused Minecraft window grabs the mouse, and any physical mouse motion rotates the camera - the resulting goldens show a drifted view that is easy to misread as a rendering bug.
• On Windows with JDK 21+, pass -Djdk.net.unixdomain.tmpdir=<short-path-without-spaces>: NIO selectors create AF_UNIX sockets under %TEMP%, which fails on some hosts and kills the game at boot with "Unable to establish loopback connection".

Capture

mkdir -p "$WORK/$CASE"
"$APITRACE" trace --api=gl \
  --output "$WORK/$CASE/full.trace" \
  -- <application-command> <application-args>

For Java:

"$APITRACE" trace --api=gl \
  --output "$WORK/$CASE/full.trace" \
  -- "$JAVA_EXE" <jvm-args> <main-class-or-jar> <game-args>

An @argfile with the full JVM+game command line keeps the invocation reproducible across recaptures.

NEVER put a real credential on the traced command line. apitrace records the traced process's argv into the trace as a process.commandLine property, so anything passed there - --accessToken, session tokens, API keys - is embedded in the trace and ships inside the committed fixture. Minecraft never validates --accessToken for singleplayer, so pass a placeholder (--accessToken 0); --username/--uuid are public and may stay real. Before packaging, grep the UNCOMPRESSED trace for the secret to confirm it is absent:

"$APITRACE" repack "$WORK/$CASE/trace.trace" /tmp/plain.trace   # decompress
grep -ac "<secret-prefix>" /tmp/plain.trace                     # must be 0

If a secret has already been captured, it can be scrubbed in place instead of recapturing: apitrace's snappy container is [length][raw snappy] chunks with no checksum, and a high-entropy secret is stored as literal bytes, so replacing those bytes with an EQUAL-LENGTH filler keeps the container valid and leaves the GL call stream byte-identical. Blank every maximal run of the secret (it splits across chunks), then verify: frame count unchanged, the decompressed trace no longer contains the secret, and the replayed target frame still matches the golden. Treat any already-pushed trace as leaked regardless - rotate the credential, since a force-push does not purge the LFS object from the remote.

Watch for vendor-gated shader paths. Shader packs branch on the GL vendor that Iris injects (MC_GL_VENDOR_NVIDIA / _AMD / ...) and compile a vendor-exclusive path, so capturing on an NVIDIA card can bake NVIDIA-only GLSL into the fixture (iterationRP selects subgroupPartitionNV / GL_NV_shader_subgroup_partitioned instead of the portable subgroupShuffleXor). Iris resolves the #ifdef before glShaderSource, so only the taken branch is in the trace and the fixture cannot replay on the mobile GPUs MobileGL targets. Rather than hunting for a second GPU (the Windows per-app GPU preference does NOT change which OpenGL ICD is loaded), mask the vendor at capture time with apitrace's own config - point GLTRACE_CONF at a file containing:

GL_VENDOR = "NoVIDIA (MobileGL spoof)"
GL_RENDERER = "NoVIDIA (MobileGL spoof)"

The wrapper then returns that from glGetString, so the pack compiles the portable path while still running on the fast driver. Pick a string that does NOT contain the real vendor name as a substring (Iris matches by substring, so "Not NVIDIA ..." would still match) and that is self-describing, so nobody later mistakes the trace for a capture on different hardware. Afterwards, grep the decoded trace to confirm the vendor-exclusive symbols are gone:

"$APITRACE" dump "$WORK/$CASE/full.trace" | grep -c subgroupPartitionNV   # must be 0

Software rasterisers are not a substitute here: llvmpipe exposes no GL_KHR_shader_subgroup at all, and packs that use subgroup ops unguarded cannot run on it in any vendor configuration.

Keep full.trace until both backends are validated.

Persistent-mapped buffers: apps may legally write a GL_MAP_PERSISTENT_BIT mapping and let the GPU read it without an explicit flush (Flywheel's indirect backend writes its compute scatter descriptors this way). Stock apitrace never records those writes, so the trimmed fixture silently loses the content that depends on them - the symptom is geometry that renders live but disappears in replay. The in-tree fork shadow-tracks persistent mappings unconditionally; if a replay of full.trace is already missing content that the live run showed, fix capture (wrapper) first - no amount of trimming will bring the data back, and the case must be recaptured. There is also a replay-side requirement: MobileGL only forwards such never-flushed writes when MOBILEGL_COHERENT_AS_FLUSH=1, so register the case with "coherent_as_flush": true (see "Register the case").

Select target frame

Fixture selection must be frame-based. Do not trim the fixture from a full trace by filtering arbitrary call ranges or single full-trace calls. Pick a rendered frame, then trim with gltrim -f.

"$APITRACE" dump --calls=frame "$WORK/$CASE/full.trace" \
  > "$WORK/$CASE/frames.txt"

A quick way to bound the choice is the total frame count from a benchmark replay:

"$GLRETRACE" -b "$WORK/$CASE/full.trace"   # "Rendered N frames in ..."

Pick a LATE frame (roughly N - 20): early frames still contain loading screens, chunk pop-in, and animation warm-up, while the last few frames may overlap the window-close path. Inspect frames.txt or snapshot the candidate frame to confirm it contains the intended visual state, then set:

export TARGET_FRAME=<chosen-frame-number>

Trim and package

"$GLTRIM" \
  -f "$TARGET_FRAME" \
  --output "$WORK/$CASE/trace.trace" \
  "$WORK/$CASE/full.trace"

Then VERIFY the trim before doing anything else: replay trace.trace, snapshot its final swap, and compare the content against the same frame of full.trace. gltrim bugs fail silently - the classic symptom is a trimmed trace whose static world renders fine while everything driven by less common call patterns (DSA texture binds, glBindBuffersRange multi-bind setup, compute-written buffers) is missing or garbled, often with "invalid buffer name"-style retrace warnings. If content is missing from the trimmed trace but present in the full trace, the fix belongs in 3rdparty/apitrace's frametrim, not in the fixture.

Temporal shaders (auto-exposure / eye adaptation, TAA, temporal reflections - e.g. the iterationRP shader pack) break a single-frame gltrim -f: the target frame reads its predecessors' feedback buffers, which the isolated frame no longer contains, so a mid-sequence frame replays overexposed to white (and any timed name/version overlay the pack draws in its first seconds silently drops). The symptom is a trimmed frame that looks blown-out or washed while the same frame of full.trace renders correctly, and it gets worse the later the frame. When a single-frame trim of such a pack cannot be made to render correctly, keep the temporal history instead of the dependency slice: select an early in-world target frame and trim a PREFIX with apitrace trim --calls=0-<target-swap-call> (it preserves call numbers, so target_call is just that swap call). The prefix replays every frame up to the target, so its temporal buffers are correct. Prefer the earliest frame that already shows the intended subject - fewer lead-in frames means a smaller archive and a faster CI replay. This deviates from the single-frame rule deliberately; note it in the README entry.

Generate golden

Generate frame snapshots from the trimmed trace, then choose the snapshot that matches the selected frame. The target call used by replay registration must come from the trimmed trace, not from a call-filtered full-trace selection.

Generate the golden with the same GL stack the scene was captured on. A headless software renderer (llvmpipe) is fine for vanilla and light packs, but heavy ray-traced shader packs (compute-driven atmosphere LUTs, screen-space tracing - e.g. iterationRP) render as solid black or blown-out white under llvmpipe. Drive the golden from a real GPU instead: on Windows a stock glretrace.exe (an upstream apitrace release works for replay even on an in-tree-fork trace) replays the trace on the discrete GPU and snapshots the target call. Read the resulting PNG back and confirm the subject actually rendered before trusting it as golden.

mkdir -p "$WORK/$CASE/golden"
"$APITRACE" replay --headless \
  --snapshot-prefix "$WORK/$CASE/golden/$CASE." \
  --call-nos \
  "$WORK/$CASE/trace.trace"

For a single-frame trim the target is simply the trimmed trace's final swap; finding it and snapshotting just that call is much faster than dumping every frame:

"$APITRACE" dump "$WORK/$CASE/trace.trace" | grep SwapBuffers | tail -n 1
"$GLRETRACE" -s "$WORK/$CASE/golden/$CASE." -S <final-swap-call> \
  "$WORK/$CASE/trace.trace"

Set TARGET_CALL to the call number in the chosen trimmed-trace snapshot filename:

export TARGET_CALL=<chosen-trimmed-trace-snapshot-call>
GOLDEN_SRC="$WORK/$CASE/golden/$CASE.$(printf '%010d' "$TARGET_CALL").png"

Remove unintended alpha:

python3 - "$GOLDEN_SRC" "$WORK/$CASE/$CASE.$(printf '%010d' "$TARGET_CALL").png" <<'PY'
import sys
from PIL import Image

src, dst = sys.argv[1], sys.argv[2]
img = Image.open(src).convert("RGBA")
bg = Image.new("RGBA", img.size, (0, 0, 0, 255))
bg.alpha_composite(img)
bg.convert("RGB").save(dst)
PY

Or copy directly:

cp "$GOLDEN_SRC" "$WORK/$CASE/$CASE.$(printf '%010d' "$TARGET_CALL").png"

Verify the golden CONTENT against a reference (a screenshot of the live run, or the same scene on a known-good backend), not just that a file exists. The subject must be present, correctly shaped, and framed as intended - a golden captured through a drifted camera or a half-loaded scene will happily pass authoring and then permanently enshrine the wrong image.

Package. Compress the trace itself with repack --brotli first - it shrinks a gzip-resistant trace by an order of magnitude (a ~70 MB single-frame Minecraft trim lands around 7 MB) and glretrace reads it directly:

"$APITRACE" repack --brotli "$WORK/$CASE/trace.trace" "$WORK/$CASE/trace-brotli.trace"
mkdir -p "$WORK/$CASE/archive"
cp "$WORK/$CASE/trace-brotli.trace" "$WORK/$CASE/archive/trace.trace"
tar -czf "$REPO/tools/trace_replay/fixtures/$CASE.tgz" \
  -C "$WORK/$CASE/archive" trace.trace
cp "$WORK/$CASE/$CASE.$(printf '%010d' "$TARGET_CALL").png" \
  "$REPO/tools/trace_replay/fixtures/"

If the case is ever re-trimmed, REDO the repack and the archive: a .tgz whose repack predates the latest trim silently packages the stale trace, and the mismatch only surfaces later as "did not create expected snapshot" when the registered target call no longer exists.

Check the final archive size. The committed fixture archive should be less than 20 MiB, and should preferably be less than 10 MiB. If it is larger, recapture with a shorter run or a lower frame rate / render distance instead of adding call-based filtering.

Some packs have an irreducible size floor: a large static lookup table baked into the pack (iterationRP ships a ~17 MiB half-float atmosphere LUT that the target frame samples) lands in the trace once and does not compress, so every variant - single frame, prefix, or full - sits near the same size regardless of frame count. When the floor alone exceeds the budget, neither a lower frame rate nor fewer frames helps; confirm the fixture is worth the exception and record the measured size in the case's README entry rather than chasing an unreachable target.

du -h "$REPO/tools/trace_replay/fixtures/$CASE.tgz"
tar -tzf "$REPO/tools/trace_replay/fixtures/$CASE.tgz"

Track with Git LFS:

git lfs track "tools/trace_replay/fixtures/*.tgz"
git lfs track "tools/trace_replay/fixtures/*.png"
git add .gitattributes tools/trace_replay/fixtures/$CASE.tgz \
  tools/trace_replay/fixtures/$CASE.$(printf '%010d' "$TARGET_CALL").png
git lfs status

Register the case

Both Linux ctest and the Android CI matrix are generated from the single registry tools/trace_replay/trace_cases.json (via tools/trace_replay/trace_cases.py); there is nothing to edit in CMakeLists.txt or apk.yml. Append one object to cases:

{
  "name": "case-name",
  "trace_archive": "case-name.tgz",
  "golden": "case-name.0000000000.png",
  "target_call": 0,
  "timeout_seconds": 900
}

Values matching the defaults block (854x480, trace.trace, ssim 0.99, zero crop, 900 s timeout) may be omitted. Available per-case keys: name, trace_archive, trace_file, golden, alternate_golden, target_call, width, height, ssim_threshold, crop_x/y/width/height, timeout_seconds, ci, coherent_as_flush. Long single-frame replays of heavy in-world scenes need a raised timeout_seconds (the Create fixtures use 1800). Set "coherent_as_flush": true for Flywheel-style engines that let the GPU read persistent GL_MAP_FLUSH_EXPLICIT_BIT mappings they never flush (both Create fixtures need it); the case then replays with MOBILEGL_COHERENT_AS_FLUSH=1 on every runner.

Update tools/trace_replay/README.md with one fixture sentence and one golden image link.

Validate on Linux

cmake -S "$REPO" -B "$WORK/build-linux" -G Ninja \
  -DCMAKE_BUILD_TYPE=RelWithDebInfo \
  -DMOBILEGL_BUILD_TEST=ON \
  -DMOBILEGL_BUILD_BENCHMARK=OFF \
  -DMOBILEGL_BUILD_TRACE_REPLAY=ON \
  -DMOBILEGL_LOG_ACTIVE_LEVEL=MOBILEGL_LOG_LEVEL_INFO
cmake --build "$WORK/build-linux" --target mobilegl_trace_replay --parallel

Run the registered case:

ctest --test-dir "$WORK/build-linux/tools/trace_replay" -V \
  -R "MobileGLTraceReplay\\.$CASE\\."

Run one backend manually:

cmake \
  -DTRACE_REPLAY_EXE="$WORK/build-linux/tools/trace_replay/mobilegl_trace_replay" \
  -DMOBILEGL_LIBRARY="$WORK/build-linux/libMobileGL.so" \
  -DTRACE_CASE_NAME="$CASE" \
  -DTRACE_ARCHIVE="$REPO/tools/trace_replay/fixtures/$CASE.tgz" \
  -DTRACE_FILE=trace.trace \
  -DTRACE_GOLDEN="$REPO/tools/trace_replay/fixtures/$CASE.$(printf '%010d' "$TARGET_CALL").png" \
  -DTRACE_BACKEND=DirectGLES \
  -DTRACE_TARGET_CALL="$TARGET_CALL" \
  -DTRACE_WIDTH="$WIDTH" \
  -DTRACE_HEIGHT="$HEIGHT" \
  -DTRACE_SSIM_THRESHOLD=0.99 \
  -DTRACE_CROP_X=0 \
  -DTRACE_CROP_Y=0 \
  -DTRACE_CROP_WIDTH=0 \
  -DTRACE_CROP_HEIGHT=0 \
  -DTRACE_OUTPUT_DIR="$WORK/$CASE/linux-DirectGLES" \
  -DTRACE_ARTIFACT_DIR="$WORK/$CASE/linux-artifacts" \
  -P "$REPO/tools/trace_replay/run_trace_case.cmake"

DirectGLES CI env:

export EGL_PLATFORM=surfaceless
export LIBGL_ALWAYS_SOFTWARE=1
export MESA_GL_VERSION_OVERRIDE=3.3
export MESA_GLSL_VERSION_OVERRIDE=330

DirectVulkan check:

vulkaninfo | grep -E 'deviceName|VK_EXT_headless_surface'

Validate on Android

Build one generic trace APK. Both backends use this APK and the same package; select the backend at replay time with --backend.

gradle --no-daemon -p "$REPO/android-plugin" \
  :app:assembleTraceRelease \
  -Pmobilegl.abis=all \
  -Pmobilegl.debuggableRelease=true \
  -Pmobilegl.logLevel=MOBILEGL_LOG_LEVEL_INFO \
  --parallel

TRACE_APK=$(find "$REPO/android-plugin/app/build/outputs/apk/trace/release" \
  -maxdepth 1 -name 'MobileGL-plugin-trace-release-*.apk' -print -quit)
TRACE_PACKAGE=top.mobilegl.plugin.trace

Release APKs are only signed when SIGNING_STORE_PASSWORD, SIGNING_KEY_ALIAS, and SIGNING_KEY_PASSWORD are set and android-plugin/keystore.jks exists - an unsigned build still "succeeds" but installs fail later with INSTALL_PARSE_FAILED_NO_CERTIFICATES. If the device or emulator already has top.mobilegl.plugin.trace from a different keystore, uninstall that one package first or the install fails with INSTALL_FAILED_UPDATE_INCOMPATIBLE.

Match the CI environment (.github/workflows/apk.yml matrix): the emulator boots with --gpu software + MOBILEGL_ESPRYT_USE_ANGLE=1 for DirectGLES and --gpu lavapipe + MOBILEGL_MAGMA_R11G11B10F_FALLBACK=1 for DirectVulkan. The emulator's ANGLE-on-Vulkan GLES stack exercises genuinely different driver semantics than physical devices (e.g. indirect-draw gl_InstanceID handling), so treat AVD-only image mismatches as real signal, not emulator noise.

Known emulator flake: the FIRST DirectVulkan replay after a fresh lavapipe AVD boot segfaults intermittently (~50%), for any trace; subsequent runs in the same boot are stable. Burn a warm-up replay and discard its result before the measured runs.

Result directories keep result.json from previous runs - delete the case's result directory before each run, or an earlier failure/success can masquerade as the current one (identical-to-the-last-digit ssim across "different" runs is the tell).

Run both backends against the same APK:

run_android_retrace() {
  backend="$1"
  shift
  sh "$REPO/android-plugin/trace-replay-ci.sh" \
    --apk-file "$TRACE_APK" \
    --package "$TRACE_PACKAGE" \
    --backend "$backend" \
    --result-root "$WORK/$CASE/android-result-$backend" \
    --fixture-root "$WORK/$CASE/android-fixture" \
    --case "$CASE" \
    --trace-archive "$REPO/tools/trace_replay/fixtures/$CASE.tgz" \
    --trace-file trace.trace \
    --golden "$REPO/tools/trace_replay/fixtures/$CASE.$(printf '%010d' "$TARGET_CALL").png" \
    --target-call "$TARGET_CALL" \
    --width "$WIDTH" \
    --height "$HEIGHT" \
    --ssim-threshold 0.99 \
    --crop-x 0 \
    --crop-y 0 \
    --crop-width 0 \
    --crop-height 0 \
    --timeout-seconds 900 \
    "$@"
}

run_android_retrace DirectGLES --use-pbuffer
run_android_retrace DirectVulkan

Inspect:

• $WORK/$CASE/android-result-DirectGLES/$CASE-DirectGLES/result.json
• $WORK/$CASE/android-result-DirectVulkan/$CASE-DirectVulkan/result.json
• Each backend's *-actual.png, *-diff.png, retrace.log, and logcat.txt

Checklist

• Golden content is verified against a live-run reference (subject present, correct shapes, intended camera framing).
• Golden matches the committed trace and target call.
• Archive contains only trace.trace, and that file is the brotli repack of the CURRENT trim.
• Both Linux backends pass before registration in trace_cases.json.
• Both Android backends pass on a CI-equivalent AVD (software/ANGLE + lavapipe) before relying on APK CI.
• actual.png and case-diff.png are inspected.
• Fixture .tgz and .png files are tracked by Git LFS.
• No build output, extracted trace directory, temporary report, or debug text is staged.