Generating Baseline Profiles Agent Skill

Generating Baseline Profiles — ship the AOT compilation hint list and prove it moved the needle

Baseline Profiles ship an AOT compilation hint list inside the APK so ART pre-compiles hot Compose code paths on install instead of relying on JIT during the first runs. Cited gains: roughly 30% faster cold startup and 40% smoother first-scroll on the journeys that were profiled. Compose ships unbundled from the platform, so every Compose UI app benefits — there is no version of Android where Compose is already AOT-compiled by the system image.

This skill is the measurement spine for the rest of the performance work. Profiles are generated with BaselineProfileRule from androidx.benchmark:benchmark-macro-junit4 and measured with MacrobenchmarkRule from the same artifact. Generation and measurement are two distinct @Test files in a separate :baselineprofile module that AGP 8.2+ scaffolds via New Module → Baseline Profile Generator.

When to use this skill

• Cold startup is slow and the developer wants AOT-compiled hot paths on first launch.
• First-scroll on a LazyColumn / LazyVerticalGrid is janky on real devices even after stability fixes.
• Preparing a release build and the developer needs a perf baseline before shipping.
• Verifying that a stability or strong-skipping fix actually moved cold-startup or frame-timing numbers — Macrobenchmark is the ground truth.
• The user mentions "baseline profile", "macrobenchmark", "TTFD", "time-to-fully-drawn", "StartupTimingMetric", "FrameTimingMetric", "CompilationMode", aosp_cf_x86_64_phone-userdebug, or "ReportDrawn".

When NOT to use this skill

• Still in early prototyping with no release-ready user journeys yet — Baseline Profiles encode hot paths, and there are no hot paths to encode while screens are still churning.
• The developer wants per-composable recomposition counts — that is ../../recomposition/debugging-recompositions/SKILL.md (debug Layout Inspector) or ../tracing-recompositions-at-runtime/SKILL.md (release @TraceRecomposition).
• The developer wants a CI gate against stability regressions — that is ../../stability/enforcing-stability-in-ci/SKILL.md. Baseline Profiles measure runtime; stabilityCheck measures compile-time skippability.
• Numbers are being collected from a debug build — debug builds run interpreted with Live Literals and produce non-representative numbers. See ../testing-compose-in-release-mode/SKILL.md.

Prerequisites

• AGP 8.2+ so the New Module → Baseline Profile Generator template is available (it scaffolds the androidx.baselineprofile Gradle plugin and a :baselineprofile module).
• A physical low-end test device or a Cuttlefish emulator (aosp_cf_x86_64_phone-userdebug). High-end pixel devices mask perf wins; emulators with default API images are not representative.
• A release variant configured (isMinifyEnabled = true, isShrinkResources = true, proguard-android-optimize.txt). Debug builds are not measurable. See ../../build/configuring-r8-for-compose/SKILL.md for R8 setup if missing.
• <profileable android:shell="true"/> in the app AndroidManifest.xml under <application> so the Macrobenchmark process can attach simpleperf. The android: namespace prefix is required — without it, manifest merger fails.
• At least one scroll journey identified beyond startup — a feed list, a paginated grid, opening a detail screen. Profiling startup alone leaves first-scroll uncompiled.
• Familiarity with ../../stability/diagnosing-compose-stability/SKILL.md and ../../recomposition/debugging-recompositions/SKILL.md if startup is dominated by avoidable recomposition rather than initial composition.

Workflow

1. Add the Baseline Profile Generator module

In Android Studio: File → New → New Module → Baseline Profile Generator. Pick the target application module. AGP scaffolds:

• A new :baselineprofile module with the androidx.baselineprofile Gradle plugin applied.
• A BaselineProfileGenerator.kt skeleton with a BaselineProfileRule @Test.
• A StartupBenchmarks.kt skeleton with a MacrobenchmarkRule @Test.
• The androidx.baselineprofile plugin applied in the app module too, so ./gradlew :app:generateBaselineProfile is wired up.

If editing manually instead of using the template, add to :baselineprofile/build.gradle.kts:

plugins {
    id("com.android.test")
    id("org.jetbrains.kotlin.android")
    id("androidx.baselineprofile")
}
android { targetProjectPath = ":app" }
dependencies {
    implementation("androidx.benchmark:benchmark-macro-junit4:1.3.0")
    implementation("androidx.test.ext:junit:1.2.1")
    implementation("androidx.test.uiautomator:uiautomator:2.3.0")
    implementation("androidx.test.espresso:espresso-core:3.6.1")
}

And to :app/build.gradle.kts:

plugins { id("androidx.baselineprofile") }
dependencies { baselineProfile(project(":baselineprofile")) }

See references/macrobenchmark-harness.md for the full Gradle + manifest setup including <profileable>.

2. Write the generator with both startup AND scroll

The generator is a @Test using BaselineProfileRule.collect. MUST cover startup plus at least one scroll — startup-only profiles leave first-scroll cold.

@RunWith(AndroidJUnit4::class)
class BaselineProfileGenerator {
    @get:Rule val rule = BaselineProfileRule()

    @Test
    fun startupAndScroll() = rule.collect(packageName = "com.example") {
        startActivityAndWait()
        val feed = device.findObject(By.res("feed"))
        feed.setGestureMargin(device.displayWidth / 5)
        feed.fling(Direction.DOWN)
        feed.fling(Direction.DOWN)
        device.findObject(By.res("feed_item_0"))?.click()
        device.wait(Until.hasObject(By.res("detail")), 3_000)
        device.pressBack()
    }
}

Use Modifier.testTag("feed") in the Compose source so By.res("feed") resolves. A gesture margin of displayWidth / 5 keeps flings off the system gesture area.

3. Generate the profile

./gradlew :app:generateBaselineProfile

For a specific variant: ./gradlew :app:generateReleaseBaselineProfile.

Output lands at:

app/src/<variant>/generated/baselineProfiles/baseline-prof.txt

(e.g. app/src/release/generated/baselineProfiles/baseline-prof.txt. On older AGP layouts the file may instead land at app/src/main/generated/baselineProfiles/baseline-prof.txt — AGP writes the variant-specific path; merge to main for shipping if all variants share a profile.)

4. Verify the profile shipped in the APK

Build → Analyze APK on the release .apk / .aab and confirm:

assets/dexopt/baseline.prof
assets/dexopt/baseline.profm

If those files are missing, the profile was generated but not packaged — usually because baselineProfile(project(":baselineprofile")) was not added to the app module's dependencies. Without these files in the APK, ART has nothing to AOT-compile and the perf gain is zero.

5. Measure with Macrobenchmark — startup

A separate @Test, in the same :baselineprofile module, using MacrobenchmarkRule. The compilation mode must be CompilationMode.Partial(BaselineProfileMode.Require) so the test fails loudly if the profile is missing rather than silently measuring an unprofiled build.

@RunWith(AndroidJUnit4::class)
class StartupBenchmarks {
    @get:Rule val rule = MacrobenchmarkRule()

    @Test
    fun startupCompilationBaselineProfiles() = rule.measureRepeated(
        packageName = "com.example",
        metrics = listOf(StartupTimingMetric()),
        iterations = 10,
        startupMode = StartupMode.COLD,
        compilationMode = CompilationMode.Partial(BaselineProfileMode.Require),
    ) {
        pressHome()
        startActivityAndWait()
    }
}

For an A/B comparison, add a sibling @Test with compilationMode = CompilationMode.None and compare medians. This is the only way to prove the profile moved the number.

6. Measure with Macrobenchmark — scroll

@Test
fun feedScrollPerformance() = rule.measureRepeated(
    packageName = "com.example",
    metrics = listOf(FrameTimingMetric()),
    iterations = 5,
    compilationMode = CompilationMode.Partial(BaselineProfileMode.Require),
) {
    startActivityAndWait()
    val feed = device.findObject(By.res("feed"))
    feed.setGestureMargin(device.displayWidth / 5)
    feed.fling(Direction.DOWN)
    feed.fling(Direction.DOWN)
}

FrameTimingMetric reports frameDurationCpuMs percentiles (P50 / P90 / P95 / P99) and frameOverrunMs (negative = on time, positive = missed deadline). The headline number is P95 frameOverrunMs — the worst-case missed-deadline frame in the high tail of a scroll.

7. Wire ReportDrawn for accurate TTFD

StartupTimingMetric reports timeToInitialDisplay by default. For timeToFullDisplay (TTFD) — the number that actually matches user perception — the app must call ReportDrawn once the first meaningful screen state is rendered. Without it, TTFD falls back to timeToInitialDisplay and undercounts.

import androidx.activity.compose.ReportDrawn
import androidx.activity.compose.ReportDrawnWhen
import androidx.activity.compose.ReportDrawnAfter

@Composable
fun Feed(state: FeedState) {
    ReportDrawnWhen { state.items.isNotEmpty() }
    LazyColumn { items(state.items, key = { it.id }) { SnackRow(it) } }
}

• ReportDrawn — fire immediately on composition (use when the first frame is the meaningful frame).
• ReportDrawnWhen { predicate } — fire when the predicate first turns true (typical case: state hydrated).
• ReportDrawnAfter { suspend block } — fire after the suspend block completes (typical case: an explicit awaitFirstFrame()).

8. Run on a real device, compare medians, iterate

Run on a connected physical device or aosp_cf_x86_64_phone-userdebug Cuttlefish via ./gradlew :baselineprofile:connectedReleaseAndroidTest. Report medians across iterations, not means — Macrobenchmark logs both, and means are sensitive to a single thermal-throttled run. The IDE displays a result table; the JSON lives at :baselineprofile/build/outputs/connected_android_test_additional_output/.

See references/macrobenchmark-harness.md for parsing the JSON in CI.

Patterns

Pattern: profile startup AND scroll, never startup alone

// WRONG
@Test fun startup() = rule.collect(packageName = "com.example") {
    startActivityAndWait()
}
// WRONG because: only the startup code paths get AOT-compiled. First-scroll, list-item
// composition, and detail-screen entry stay interpreted on first run, which is exactly
// where the user perceives jank. Profile every hot user journey, not just launch.

// RIGHT
@Test fun startupAndScroll() = rule.collect(packageName = "com.example") {
    startActivityAndWait()
    val feed = device.findObject(By.res("feed"))
    feed.setGestureMargin(device.displayWidth / 5)
    feed.fling(Direction.DOWN)
    feed.fling(Direction.DOWN)
}

Pattern: assert the profile is applied with BaselineProfileMode.Require

// WRONG
compilationMode = CompilationMode.Partial(BaselineProfileMode.UseIfAvailable)
// WRONG because: UseIfAvailable silently falls back to no profile if assets/dexopt/baseline.prof
// is missing. The Macrobench then measures an unprofiled build and the developer thinks
// the profile is working when it is not. Use Require to fail loudly.

// RIGHT
compilationMode = CompilationMode.Partial(BaselineProfileMode.Require)

Pattern: TTFD undercount when ReportDrawn is missing

// WRONG
@Composable
fun Feed(state: FeedState) {
    LazyColumn { items(state.items) { SnackRow(it) } }
}
// WRONG because: TTFD is reported when the first frame draws. An empty LazyColumn draws
// a frame too — TTFD then matches timeToInitialDisplay and undercounts the time the user
// actually waited for content. Add ReportDrawnWhen { state.items.isNotEmpty() }.

// RIGHT
@Composable
fun Feed(state: FeedState) {
    ReportDrawnWhen { state.items.isNotEmpty() }
    LazyColumn { items(state.items, key = { it.id }) { SnackRow(it) } }
}

Pattern: never measure Compose perf in a debug build

# WRONG
"./gradlew :app:installDebug && adb shell am start … && record startup with systrace"
# WRONG because: debug builds run Compose interpreted, with Live Literals turning constants
# into getters that defeat compile-time folding. Cold-start numbers are inflated 2–4×;
# scroll FrameTiming numbers are dominated by the interpreter overhead, not by the code.
# Measurement only counts on release + R8 + real device. Cross-link
# ../testing-compose-in-release-mode/SKILL.md.

# RIGHT
"./gradlew :baselineprofile:connectedReleaseAndroidTest" on a physical device, with
release variant minified, baseline profile generated, BaselineProfileMode.Require asserting it.

Pattern: report medians, not means, across enough iterations

# WRONG
"Mean cold startup with profile: 412 ms (5 iterations)."
# WRONG because: a single thermal-throttled iteration drags the mean. Macrobenchmark
# reports min / median / max; the median is the resilient summary. 5 iterations is also
# thin for startup variance — use ≥10 for StartupTimingMetric, ≥5 for FrameTimingMetric.

# RIGHT
"Median cold startup, 10 iterations: BaselineProfile 318 ms vs None 462 ms (–31%).
 P95 frameOverrunMs scrolling 30 items: BaselineProfile –6 ms vs None +9 ms."

Pattern: gesture-margin so flings are not eaten by the system

// WRONG
val feed = device.findObject(By.res("feed"))
feed.fling(Direction.DOWN)
// WRONG because: on gesture-nav devices, a fling that starts inside the system gesture area
// is intercepted as a back-swipe or recents-swipe and the scroll never happens.
// Macrobenchmark then measures an idle screen and reports artificially-good numbers.

// RIGHT
val feed = device.findObject(By.res("feed"))
feed.setGestureMargin(device.displayWidth / 5)
feed.fling(Direction.DOWN)

Mandatory rules

• MUST generate Baseline Profiles for both startup AND at least one scroll journey. Startup-only profiles leave first-scroll cold.
• MUST measure on a real low-end physical device or aosp_cf_x86_64_phone-userdebug Cuttlefish. High-end devices mask the perf delta; default API emulator images are not representative.
• MUST use CompilationMode.Partial(BaselineProfileMode.Require) in the measurement test so a missing or stale profile fails the test loudly. BaselineProfileMode.UseIfAvailable silently measures an unprofiled build.
• MUST use ReportDrawn / ReportDrawnWhen / ReportDrawnAfter from androidx.activity.compose to mark the meaningful first-drawn moment. Without it timeToFullDisplay undercounts.
• MUST NOT measure Compose perf in debug builds — debug runs interpreted with Live Literals, the numbers are not representative. Cross-link ../testing-compose-in-release-mode/SKILL.md.
• MUST verify the profile shipped in the APK: assets/dexopt/baseline.prof must exist after ./gradlew :app:assembleRelease. If absent, the baselineProfile(project(":baselineprofile")) wiring on the app module is missing.
• MUST report medians, not means, across iterations. Means are sensitive to a single thermal-throttled run.
• PREFERRED: ≥10 iterations for StartupTimingMetric, ≥5 for FrameTimingMetric.
• PREFERRED: keep an A/B sibling test pinned to CompilationMode.None so every PR can prove the profile is still moving the number.
• PREFERRED: add Modifier.testTag("feed") (or whatever ID the journey uses) in the Compose source rather than relying on text matchers — text changes with localization, test tags do not.

Verification

• [ ] :baselineprofile module exists with the androidx.baselineprofile plugin applied.
• [ ] :app/build.gradle.kts has baselineProfile(project(":baselineprofile")) in dependencies.
• [ ] ./gradlew :app:generateBaselineProfile produces baseline-prof.txt under app/src/<variant>/generated/baselineProfiles/ (or app/src/main/generated/baselineProfiles/).
• [ ] Build → Analyze APK on the release artifact shows assets/dexopt/baseline.prof and assets/dexopt/baseline.profm.
• [ ] Generator @Test covers cold startup AND at least one scroll fling.
• [ ] Measurement @Test uses CompilationMode.Partial(BaselineProfileMode.Require).
• [ ] ReportDrawn / ReportDrawnWhen / ReportDrawnAfter is invoked from a composable that renders only when the screen is meaningfully complete.
• [ ] Macrobench results — median across ≥10 iterations for startup, ≥5 for scroll — show a measurable delta vs CompilationMode.None.
• [ ] Measurement run is on a physical low-end device or aosp_cf_x86_64_phone-userdebug, not a stock API emulator.
• [ ] Debug-build numbers are not used as evidence anywhere in the report.

References

• Baseline Profiles overview — https://developer.android.com/topic/performance/baselineprofiles/overview
• Baseline Profiles with Compose — https://developer.android.com/develop/ui/compose/performance/baseline-profiles
• Benchmark Baseline Profiles with Macrobenchmark — https://developer.android.com/topic/performance/baselineprofiles/measure-baselineprofile
• Macrobenchmark library — https://developer.android.com/topic/performance/benchmarking/macrobenchmark-overview
• androidx.activity.compose.ReportDrawn reference — https://developer.android.com/reference/kotlin/androidx/activity/compose/package-summary
• Ben Trengrove, "Why you should always test Compose performance in release" — https://medium.com/androiddevelopers/why-should-you-always-test-compose-performance-in-release-4168dd0f2c71
• skydoves, "Improving Android App Performance with Baseline Profiles" (Stream blog) — https://getstream.io/blog/android-baseline-profile/
• Compose performance overview — https://developer.android.com/develop/ui/compose/performance
• references/macrobenchmark-harness.md — full Gradle + manifest harness setup, the metric catalogue (StartupTimingMetric, FrameTimingMetric, TraceSectionMetric, MemoryUsageMetric, PowerMetric), and how to parse the JSON output for CI dashboards.

For confirming that a generated profile actually moved the recomposition count of a specific composable in release, see ../tracing-recompositions-at-runtime/SKILL.md. For why debug numbers are not measurement evidence, see ../testing-compose-in-release-mode/SKILL.md. For the build-time R8 setup that release measurement assumes, see ../../build/configuring-r8-for-compose/SKILL.md. For preventing stability regressions between profile-generation runs, see ../../stability/enforcing-stability-in-ci/SKILL.md.