mirror of
https://github.com/vitorpamplona/amethyst.git
synced 2026-09-14 00:55:08 +00:00
Vitor PamplonaandClaude Opus 5
4c606c2571
Follow-up to the hand-authored amethyst/src/main/baseline-prof.txt, which was a
stopgap: whole-package wildcards that compile methods which never run. This adds
the machinery to produce a real profile from a recorded journey.
./gradlew :amethyst:generateBaselineProfile
New :baselineprofile module (com.android.test + androidx.baselineprofile) with a
BaselineProfileRule journey, and :amethyst applies the plugin and consumes it.
Notes on the wiring, since three things needed working out:
- com.android.test must be applied WITHOUT a version. AGP is already on the
buildscript classpath, so `alias(libs.plugins.androidTest)` fails with
"already on the classpath with an unknown version".
- targetSdk belongs in defaultConfig, not testOptions, for a test module.
- :amethyst has a `channel` dimension, so the test module needs
missingDimensionStrategy("channel", "play") or the dependency is ambiguous.
The generator ran on a connected device: Macrobenchmark 1.5.0-alpha07 supports
non-rooted generation on API 33+, so no root or AOSP emulator was needed. It
produced 31,497 rules (3.1 MB).
WHAT THE GENERATED PROFILE DOES AND DOES NOT COVER — it captures cold-start on a
LOGGED-OUT app, not the ingest burst. The generator installs the release
applicationId (com.vitorpamplona.amethyst), which is a fresh install with no
account, so the journey recorded a login screen. Measured on the output: zero
rules for justConsume and only 33 of 31,497 rules marked hot. Capturing ingest
needs a journey against a logged-in app, which needs a seeded test identity —
a design decision (a key in the repo is not acceptable), so it is left open.
Both profiles are therefore kept, because they cover different things: the
hand-authored one covers ingest (measured: nterp 42.9% -> 4.2% of ingest worker
CPU, ~1.6x more ingest per core-second), the generated one covers startup and
class loading.
Verified they both reach the shipping artifact by inspecting assets/dexopt/
baseline.prof across variants:
11,425 bytes neither profile
16,341 bytes hand-authored only (benchmark build type)
25,541 bytes hand-authored + generated (RELEASE variant)
The generated profile only lands in `release`; the custom `benchmark` build type
gets just the hand-authored file, because the plugin wires generated profiles
into release variants. That is why the earlier ingest measurements — taken on
the benchmark build type — could not see it. The runtime effect of the generated
half is NOT measured here: the release APK is unsigned and could not be
installed on the test device.
The 3.1 MB generated file is committed on purpose (saveInSrc), so release builds
do not need a device attached at build time.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>