Benchmark
Latest Update | Stable Release | Release Candidate | Beta Release | Alpha Release |
---|---|---|---|---|
October 2, 2024 | 1.3.2 | - | - | 1.4.0-alpha02 |
Declaring dependencies
To add a dependency on Benchmark, you must add the Google Maven repository to your project. Read Google's Maven repository for more information.
Macrobenchmark
To use Macrobenchmark
in your project, add the following dependencies to your build.gradle
file for
your
macrobenchmark module:
Groovy
dependencies { androidTestImplementation "androidx.benchmark:benchmark-macro-junit4:1.3.2" }
Kotlin
dependencies { androidTestImplementation("androidx.benchmark:benchmark-macro-junit4:1.3.2") }
Microbenchmark
To use Microbenchmark
in your project, add the following dependencies to your build.gradle
file for
your
microbenchmark module:
Groovy
dependencies { androidTestImplementation "androidx.benchmark:benchmark-junit4:1.3.2" } android { ... defaultConfig { ... testInstrumentationRunner "androidx.benchmark.junit4.AndroidBenchmarkRunner" } }
Kotlin
dependencies { androidTestImplementation("androidx.benchmark:benchmark-junit4:1.3.2") } android { ... defaultConfig { ... testInstrumentationRunner = "androidx.benchmark.junit4.AndroidBenchmarkRunner" } }
The Microbenchmark library also provides a Gradle plugin to use with your microbenchmark module.
This plugin sets build configuration defaults for the module, sets up
benchmark output copy to the host,
and provides the
./gradlew lockClocks
task.
To use the plugin, include the following line in the `plugins` block in your top-level
build.gradle
file:
Groovy
plugins { id 'androidx.benchmark' version '1.3.2' apply false }
Kotlin
plugins { id("androidx.benchmark") version "1.3.2" apply false }
Then apply the plugin to your benchmark module's build.gradle
file
Groovy
plugins { id 'androidx.benchmark' }
Kotlin
plugins { id("androidx.benchmark") }
Feedback
Your feedback helps make Jetpack better. Let us know if you discover new issues or have ideas for improving this library. Please take a look at the existing issues in this library before you create a new one. You can add your vote to an existing issue by clicking the star button.
See the Issue Tracker documentation for more information.
Version 1.4
Version 1.4.0-alpha02
October 2, 2024
androidx.benchmark:benchmark-*:1.4.0-alpha02
is released. Version 1.4.0-alpha02 contains these commits.
API Changes
- Moved Gradle tasks
lockClocks
andunlockClocks
to be on benchmark projects, instead of available at the top level. This change was necessary as there is unfortunately no way to register these as top level actions without breaking project isolation. (I02b8f, b/363325823)
Bug Fixes
BaselineProfileRule
now collects profiles for multi-process apps by signaling each running process at the end of the block to dump profiles. If a profile based compilation never successfully finds a process to broadcast to, the compilation will fail, as it's unexpected to have profile data within. Additionally, added an instrumentation argument to control dump wait duration:androidx.benchmark.saveProfileWaitMillis
(I0f519, b/366231469)- From Benchmark
1.3.2
: Fixed Firebase Test Lab (FTL) being unable to pull Baseline Profile or Macrobenchmark result files from the Baseline Profile Gradle Plugin. (I2f678, b/285187547)
To use FTL apply the plugin to the baseline profile module in the plugin block, with:
plugins {
...
id("com.google.firebase.testlab")
}
and then configure firebase test lab with:
firebaseTestLab {
// Credentials for FTL service
serviceAccountCredentials.set(file("credentials.json"))
// Creates one or more managed devices to run the tests on.
managedDevices {
"ftlDeviceShiba34" {
device = "shiba"
apiLevel = 34
}
}
// Ensures the baseline profile is pulled from the device.
// Note that this will be automated as well later with aosp/3272935.
testOptions {
results {
directoriesToPull.addAll("/storage/emulated/0/Android/media/${android.namespace}")
}
}
}
Also the created FTL device needs to be added to the baseline profile extension:
baselineProfile {
managedDevices += "ftlDeviceShiba34"
useConnectedDevices = false
}
Version 1.4.0-alpha01
September 18, 2024
androidx.benchmark:benchmark-*:1.4.0-alpha01
is released. Version 1.4.0-alpha01 contains these commits.
New Feature - App Startup Insights
- Initial version of app startup insights can be enabled in Macrobenchmark. (09fae38)
To enable in a startup benchmark:
@Test
fun startup {
macrobenchmarkRule.measureRepeated(
…
packageName = "com.example.my.application.id"
metrics = listOf(StartupTimingMetric()),
iterations = 5,
startupMode = StartupMode.COLD,
compilationMode = CompilationMode.None(),
experimentalConfig = ExperimentalConfig(startupInsightsConfig = StartupInsightsConfig(isEnabled = true))
) {
scope.startActivityAndWait(...)
}
}
Then running your startup benchmark will analyze the trace to look for common problems, and print them after metrics to Studio test output in the benchmark tab, e.g.:
StartupBenchmark_startup[startup=COLD,compilationMode=None]
├── Metrics
│ ├── timeToFullDisplayMs min 1,147.2, median 1,208.8, max 1,307.4
│ └── timeToInitialDisplayMs min 1,147.2, median 1,208.8, max 1,307.4
├── App Startup Insights
│ ├── App in debuggable mode (expected: false)
│ │ └── seen in iterations: 0(true) 1(true) 2(true) 3(true) 4(true) 5(true) 6(true) 7(true) 8(true) 9(true)
│ ├── Potential CPU contention with another process (expected: < 100000000ns)
│ │ └── seen in iterations: 4(105022546ns)
│ └── Main Thread - Binder transactions blocked (expected: false)
│ └── seen in iterations: 7(true)
└── Traces
└── Iteration 0 1 2 3 4 5 6 7 8 9
This feature is still a work-in-progress, with improvements to documentation and extensibility to follow, but feedback is welcome.
New Features
- Added gradle property
androidx.baselineprofile.suppressWarnings
to suppress all baseline profile warnings. (314153a) - Microbench metrics are now displayed in Perfetto traces as counters. (3214854)
- Add experimental scripts for disabling jit (requires root / runtime restart), and resetting device perf/test state. These are not currently published as gradle tasks. (7c3732b)
- Added benchmark argument to skip tests when running on emulator. When
automaticGenerationDuring
build is enabled, benchmarks will also trigger baseline profile generation. This will fail, if emulators are used. With the new argumentskipBenchmarksOnEmulator
we can instead skip the test. (0c2ddcd) - Change perf event enable logic to run on API 23+ (2550048)
API Changes
- Existing experimental
PerfettoConfig
argument toMacrobenchmarkRule.measureRepeated()
moved to the newExperimentalConfig
object.
Bug Fixes
- Increase
lockClocks.sh
retry count (99e9dac) - Don't create
nonMinified
and benchmark build types if existing. Due to a bug, even ifnonMinified
and benchmark build types existed, they were going to be recreated. (e75f0a5) - Ignore non-terminating slices from
TraceSectionMetric
results. (a927d20) - Improved emulator check to consider
sdk_
prefix. (1587de8) - Treat non-running packages as cleared in
FrameTimingGfxInfoMetric
. (35cc79c) - Fix
androidx.benchmark.cpuEventCounter
producing corrupt values for non-Instruction events. (06edd59) - Fix
resumeTiming/runWithTimingDisabled
to respect metric priority order, and significantly reduce impact of lower priority metric pause/resume on higher priority metric results. For example, if using cpu perf counters viacpuEventCounter.enable
instrumentation argument, timeNs is no longer significantly reduced when pause/resume occur. (5de0968)
Version 1.3
Version 1.3.2
October 2, 2024
androidx.benchmark:benchmark-*:1.3.2
is released. Version 1.3.2 contains these commits.
Bug Fixes
- Fixed Firebase Test Lab (FTL) being unable to pull Baseline Profile or Macrobenchmark result files from the Baseline Profile Gradle Plugin. (I2f678, b/285187547)
To use FTL apply the plugin to the baseline profile module in the plugin block, with:
plugins {
...
id("com.google.firebase.testlab")
}
and then configure firebase test lab with:
firebaseTestLab {
// Credentials for FTL service
serviceAccountCredentials.set(file("credentials.json"))
// Creates one or more managed devices to run the tests on.
managedDevices {
"ftlDeviceShiba34" {
device = "shiba"
apiLevel = 34
}
}
// Ensures the baseline profile is pulled from the device.
// Note that this will be automated as well later with aosp/3272935.
testOptions {
results {
directoriesToPull.addAll("/storage/emulated/0/Android/media/${android.namespace}")
}
}
}
Also the created FTL device needs to be added to the baseline profile extension:
baselineProfile {
managedDevices += "ftlDeviceShiba34"
useConnectedDevices = false
}
Version 1.3.1
September 18, 2024
androidx.benchmark:benchmark-*:1.3.1
is released. Version 1.3.1 contains these commits.
Bug Fixes
- Added gradle property
androidx.baselineprofile.suppressWarnings
to suppress all baseline profile warnings (I7c36e, b/349646646) - Fixed Baseline Profile Gradle Plugin to use pre-existing
nonMinified…
andbenchmark…
if created by the app instead of creating wrappers. (Ia8934, b/361370179) - Fixed
java.lang.AssertionError: ERRORS (not suppressed): EMULATOR
whenautomaticGenerationDuringBuild
is enabled on emulators. New argument is used to instead skip the test. (If3f51, b/355515798) - Microbenchmark minification - keep subclasses of
org.junit.runner.notification.RunListener
in benchmark library proguard (Ic8ed5, b/354264743) - Fix
TraceSectionMetric
to Ignore non-terminating slices. Previously these were considered to have -1 duration, e.g. during summation or finding minimum duration. (If74b7) - Fixed an issue in
FrameTimingGfxInfoMetric
where starting the metric would crash if the process wasn't already running. (I6e412)
Version 1.3.0
August 21, 2024
androidx.benchmark:benchmark-*:1.3.0
is released. Version 1.3.0 contains these commits.
Microbenchmark changes since 1.2.0
- Method tracing is on by default in microbenchmarks when running on most devices
- Method tracing runs as a separate phase, after measurements - this enables accurate measurements and method traces to both be output from a single benchmark run
- Method tracing on some Android OS and ART versions will affect later measurement phases - on these versions, method tracing is off by default and a warning is printed to Studio output
- Main thread benchmarks and ANRs
- Added
measureRepeatedOnMainThread
for UI thread benchmarks (e.g. those that interact with Compose/View UIs) to avoid ANRs when running for many seconds. - Method traces are skipped if expected to overrun the ANR avoidance deadline. Set
androidx.benchmark.profiling.skipWhenDurationRisksAnr
to false to disable this behavior (not recommended for CI runs, as ANRs can cause problem in long CI runs).
- Added
- Minification
- Embedded proguard rules to improve microbenchmarking with minification enabled
- Minification/R8 in a library module requires AGP 8.3, and can be enabled via
android.buildTypes.release.androidTest.enableMinification
in yourbuild.gradle
- Experimental
BlackHole.consume()
API added to prevent dead code elimination (If6812, b/286091643)
- Metrics
- Experimental cpu event counter feature (metrics from
perf_event_open
, which requires root on most versions of the platform), access viaInstrumentationArgument
androidx.benchmark.cpuEventCounter.enable
(can be set totrue
), andandroidx.benchmark.cpuEventCounter.events
can be set e.g. to (Instructions,CpuCycles
). This should be supported on some userdebug emulators, but support has not been tested across all available emulators
- Experimental cpu event counter feature (metrics from
MACRObenchmark changes since 1.2.0
- Method tracing overhaul for macrobenchmarks.
- Now method traces are scoped to the duration of the
measureBlock
, and can capture multiple sessions if the process starts multiple times. - Previously, method tracing would only work for
StartupMode.COLD
benchmarks, and capture nothing formeasureBlocks
that didn't restart the target process - Fixed method traces flush in macrobenchmark, so that method traces should be fully captured and valid, even on slower devices. (I6349a, b/329904950)
- Now method traces are scoped to the duration of the
- Correctly dump ART profile during individual
warmUp
iterations when process is killed soCompilationMode.Partial(warmup=N)
measurements are more accurate. (I17923) - Drop Shader broadcast failure message
- Added debugging suggestions to drop shader broadcast failure message
- Add two instrumentation arguments for overriding shader dropping behavior to workaround crashes when benchmarking apps without
ProfileInstaller
1.3:androidx.benchmark.dropShaders.enable=true/false
: can be used to skip all shader dropping (including that done inStartupMode.Cold
launches), esp when benchmarking apps that don't yet use profileinstaller 1.3androidx.benchmark.dropShaders.throwOnFailure=true/false
: can be used to tolerate failures when trying to drop shaders, for example when benchmarking apps without profileinstaller 1.3 (I4f573)
- Added experimental
MacrobenchmarkRule#measureRepeated
variant which takes a customPerfettoConfig
for fully customized Perfetto trace recording. Note that incorrectly configured configs may cause built in Metric classes to fail. (Idfd3d, b/309841164, b/304038384) - Cancel background dexopt jobs before running a Macrobenchmark to reduce interference. (I989ed)
- Macrobenchmark now waits for 1 second for the target application to flush an ART profile (previously it waited for 500 ms). (I85a50, b/316082056)
- TraceSectionMetric overhaul
- Note:
TraceSectionMetric
changes below can affect outputs in CI usage, and may create discontinuities, or break parsing - Sum is now the default, as most usage of this metric is for repeated events, and first would discard data in these cases
- Changed to be more customizable, with more available modes
- Mode names are now embedded in metric output name (in Studio and JSON)
- Now supports slices created using
Trace.{begin|end}AsyncSection
.
- Note:
- Metrics
- Power - Added
PowerMetric.deviceSupportsHighPrecisionTracking
,PowerMetric.deviceBatteryHasMinimumCharge()
andPowerMetric.deviceSupportsPowerEnergy()
- Renamed
Metric.getResult
togetMeasurements
to match return type - Added log.w / exception labels to all startup detection failures. This does not change current behavior (so some errors throw, and others silently fail to detect the startup), just makes it more understandable. Generally the ones that
Log.w()
and fail to report startup metrics are those where non-frame events are missing, exceptions are thrown when startup is detected except for frame timing information (from UI/RT slices). (Id240f, b/329145809) - Added
frameCount
measurement toFrameTimingMetric
to aid in discovery of scenarios where measurements change because the number of frames produced changed (new animations added, invalidation issues fixed). (I1e5aa) - Clarified that
frameOverrunMs
is the preferred metric for tracking when available in docs, and why. (I18749, b/329478323) - Fixes issue where unterminated frames at the beginning and end of the trace could be paired together, which would incorrectly report as a single extremely long frame. (I39353, b/322232828)
- Improve
FrameTimingMetric
error when frames aren't produced, and always output link to trace when failing metric parsing to assist in diagnosing problem. (I956b9) - Fixed crash in
FrameTimingMetric
failing to parse frame id, especially on certain OEM devices. (Ia24bc, b/303823815, b/306235276) - Relaxed strictness of checks in
FrameMetrics
, and added more detail to error messages. (Iadede)
- Power - Added
Baseline Profile capture / Gradle plugin changes since 1.2.0
- Increased max recommended version of AGP to 9.0.0-alpha01.
- Ensure
mergeArtProfile
andmergeStartupProfile
tasks always wait for baseline profile generation. (I623d6, b/343086054) - Generating a baseline profile successfully will output a summary of what changed (I824c8, b/269484510)
- Added DSL to disable warnings (Ic4deb, b/331237001)
- Fix to ensure benchmarks use generated baseline profiles when
automaticGenerationDuringBuild
is off (Ic144f, b/333024280) - Fix
BaselineProfile
gradle plugin property overrides to enable baseline profile generation and benchmarking when customizing anonMinified
or benchmark build type. (Ib8f05, b/324837887) - Fix for including library baseline profiles in AAR prior to AGP 8.3.0-alpha15. (I1d2af, b/313992099)
- Fixed baseline and startup profile output url at the end of generation task. (I802e5, b/313976958)
Other significant changes since 1.2.0
- Trace capture
- Reduced EXITCODE 2 error when starting perfetto from an error to logged warning
- Enable AIDL tracing by default in benchmarks(requires API 28) (Ia0af2, b/341852305)
- Enable porter tag tracing by default in benchmarks. This captures, for example, wakelock tracepoints. (Icfe44, b/286551983)
- Increased trace capture start timeout to avoid crashes when starting tracing on slower devices (I98841, b/329145808)
- Added public API
PerfettoTraceProcessor.Session.queryMetrics
APIs with JSON, textproto, and proto binary (undecoded) variants. These allow you to query metrics built intoTraceProcessor
(I54d7f, b/304038382) - Enable blocking start on Perfetto trace record to reduce risk of missing data at beginning of trace. Only supported on API 33+. (Ie6e41, b/310760059)
- JSON output
- Added additional information in benchmark context in JSON output:
context.artMainlineVersion
- integer version of Art mainline module (if present on device,-1
otherwise)context.build.id
- Equals android.os.Build.IDcontext.build.version.codename
- Equals android.os.Build.VERSION.CODENAMEcontext.build.version.abbreviatedCodename
- corresponds to first letter of pre-release codename (including on release builds) (Ie5020)
- Added
profilerOutput
list to JSON output for easier tooling around profiling traces (e.g. Perfetto, Method traces) (I05ddd, b/332604449) - Added a warning when Android Test Orchestrator is used in benchmark modules, as this will cause per-module output JSON files to be repeatedly overwritten. (Ia1af6, b/286899049)
- Throw when filenames are longer than 200 chars to avoid unclear crashes when writing or post-processing files. (I4a5ab)
- Added additional information in benchmark context in JSON output:
Version 1.3.0-rc01
August 7, 2024
androidx.benchmark:benchmark-*:1.3.0-rc01
is released. Version 1.3.0-rc01 contains these commits.
Bug Fixes
- Fix
androidx.benchmark.cpuEventCounter
producing corrupt values for non-Instruction events (I7386a, b/286306579) - Fix
resumeTiming
/runWithTimingDisabled
to respect metric priority order, and significantly reduce impact of lower priority metric pause/resume on higher priority metric results. For example, if using cpu perf counters viacpuEventCounter.enable
instrumentation argument, timeNs is no longer significantly reduced when pause/resume occur. (I39c2e, b/286306579, b/307445225) - Reduced chance of stack sampling causing
measureRepeatedOnMainThread
from hitting main thread hard timeout by moving stack sampling conversion off main thread. (I487a8, b/342237318) - Removed manual outlining of access to new platform APIs since this happens automatically via API modeling when using R8 with AGP 7.3 or later (e.g. R8 version 3.3) and for all builds when using AGP 8.1 or later (e.g. D8 version 8.1). Clients who are not using AGP are advised to update to D8 version 8.1 or later. See this article for more details. (I9496c, b/345472586)
- Added agp version check to send package name as instr arg. Previous to AGP 8.4.0 the target app package name cannot be send to the instrumentation app via instrumentation arguments. (0c72a3f)
Version 1.3.0-beta02
July 10, 2024
androidx.benchmark:benchmark-*:1.3.0-beta02
is released. Version 1.3.0-beta02 contains these commits.
Bug Fixes
- Gracefully handle EXITCODE
2
when starting Perfetto to log a warning, but proceed.
Version 1.3.0-beta01
June 12, 2024
androidx.benchmark:benchmark-*:1.3.0-beta01
is released. Version 1.3.0-beta01 contains these commits.
API Changes
- Renamed
MethodTracing.affectsMeasurementOnThisDevice
toAFFECTS_MEASUREMENT_ON_THIS_DEVICE
for consistency. (I1bdfa) - Added experimental
BlackHole.consume()
api to prevent dead code elimination in microbenchmarks. (If6812, b/286091643) - Microbenchmark will now correctly throw to prevent method tracing from interfering with measurements. This occurs on certain devices when method tracing is forced on (via instrumentation args or
MicrobenchmarkConfig
), and if a measurement is attempted after a method trace. Affected devices are running API 26-30 or certain ART mainline module versions affected by this interference, and can be detected at runtime viaProfilerConfig.MethodTracing.affectsMeasurementOnThisDevice
. (Iafb92, b/303660864)
Bug Fixes
- Bumped max agp version recommended to 9.0.0-alpha01. (I5bbb0)
- Added compilation mode to benchmark context (If5612, b/325512900)
- Enable AIDL tracing by default (requires API 28) (Ia0af2, b/341852305)
- Added additional information in benchmark context in JSON output:
context.artMainlineVersion
- integer version of Art mainline module (if present on device, -1 otherwise)context.build.id
- Equalsandroid.os.Build.ID
context.build.version.codename
- Equalsandroid.os.Build.VERSION.CODENAME
context.build.version.abbreviatedCodename
- corresponds to first letter of pre-release codename (even on release builds) (Ie5020)
- Fixes
StackSampling
to respectandroidx.benchmark.profiling.sampleDurationSeconds
(Ib1d53) - Change macro->common dependency to be
api()
, so it's easier to use e.g.PerfettoTrace
andPerfettoConfig
. (Icdae3, b/341851833) - Ensure
mergeArtProfile
andmergeStartupProfile
tasks always wait for baseline profile generation. (I623d6, b/343086054) - Consider variant enable state when deciding whether variant should be enabled. (I5d19e, b/343249144)
- Increased default start timeout for perfetto trace processor. (I87e8c, b/329145808)
Version 1.3.0-alpha05
May 14, 2024
androidx.benchmark:benchmark-*:1.3.0-alpha05
is released. Version 1.3.0-alpha05 contains these commits.
Bug Fixes
- Throw clearer exception when macrobench metric returns zero values for all iterations (Iab58f, b/314931695)
- Additional workaround rules added to microbench proguard rules, including support for listener rules and other observed warnings / errors. (I14d8f, b/329126308, b/339085669)
- Method tracing runs as a separate phase during a Macrobenchmark, and it no longer affects measurements. (If9a50, b/285912360, b/336588271)
- Added extra debugging suggestions to drop shader broadcast failure message. (I5efa6, b/325502725)
Version 1.3.0-alpha04
May 1, 2024
androidx.benchmark:benchmark-*:1.3.0-alpha04
is released. Version 1.3.0-alpha04 contains these commits.
API Changes
- Added experimental
MacrobenchmarkRule#measureRepeated
variant which takes a customPerfettoConfig
for fully customized Perfetto trace recording. Note that incorrectly configured configs may cause built in Metric classes to fail. (Idfd3d, b/309841164, b/304038384) - Rename
PowerMetric.deviceSupportsPowerEnergy
toPowerMetric.deviceSupportsHighPrecisionTracking
for clarity (I5b82f) - Added
PowerMetric.deviceBatteryHasMinimumCharge()
andPowerMetric.deviceSupportsPowerEnergy()
to enable changing or skipping benchmarks based on device power measurement capability. (I6a591, b/322121218)
Bug Fixes
- Added comparison with previous baseline profile (I824c8, b/269484510)
- Added DSL to disable warnings (Ic4deb, b/331237001)
- Changed exception to info log when benchmark variants are disabled (I8a517, b/332772491)
- Make it simpler to capture method traces for a Macrobenchmark is scoped to the duration of the actual
measureBlock()
. Previously, it started at target process launch and only supported cold starts (Iee85a, b/300651094) - Avoid crashing when perfetto trace processor is slow to start (I98841, b/329145808)
Version 1.3.0-alpha03
April 17, 2024
androidx.benchmark:benchmark-*:1.3.0-alpha03
is released. Version 1.3.0-alpha03 contains these commits.
New Features
- Adds public API
PerfettoTraceProcessor.Session.queryMetrics
APIs with JSON, textproto, and proto binary (undecoded) variants. These allow you to query metrics built into TraceProcessor (I54d7f, b/304038382) - Added
profilerOutput
to JSON output for easier tooling around profiling traces (e.g. perfetto, method traces). (I05ddd, b/332604449) - Added power tag to benchmark Perfetto Config. This captures, for example, wakelock tracepoints. (Icfe44, b/286551983)
- Added inst argument
androidx.benchmark.profiling.skipWhenDurationRisksAnr
, can be set to false to avoid skipping method traces when expected duration may cause an ANR - strongly recommended to avoid in CI runs. - Added experimental inst argument
androidx.benchmark.profiling.perfCompare.enable
, set this to true to run comparison timing between measurement and profiling phases. Useful in e.g. evaluating overhead of method tracing. (I61fb4, b/329146942)
API Changes
- Changed
TraceSectionMetric.Mode
to sealed class to enable future expansion without breaking exhaustive when statements (I71f7b) - Added
TraceSectionMetric.Mode.Average
and.Count
, and reordered args so the more common argument (mode) was earlier in the arg list, reducing need for specifying parameter names. (Ibf0b0, b/315830077, b/322167531) - Renamed
Metric.getResult
togetMeasurements
to match return type (I42595)
Bug Fixes
- Fix to ensure benchmarks use generated baseline profiles when
automaticGenerationDuringBuild
is off (Ic144f, b/333024280) - Fix
BaselineProfile
gradle plugin property overrides to enable baseline profile generation and benchmarking when customizing anonMinified
or benchmark build type. (Ib8f05, b/324837887) - Fixed method traces flush in macrobenchmark, so that method traces should be fully captured and valid, even on slower devices. (I6349a, b/329904950)
- Enable blocking start on Perfetto trace record to reduce risk of missing data at beginning of trace. Only supported on API 33+. (Ie6e41, b/310760059)
- Added a warning when Android Test Orchestrator is used in benchmark modules, as this will cause per-module output JSON files to be repeatedly overwritten. (Ia1af6, b/286899049)
- Force ',' (comma) thousands separators for consistency in Studio output, ignoring device locale (I3e921, b/313496656)
TraceSectionMetric
now supports slices created usingTrace.{begin|end}AsyncSection
. (I91b32, b/300434906)- Added log.w / exception labels to all startup detection failures. This does not change current behavior (so some errors throw, and others silently fail to detect the startup), just makes it more understandable. Generally the ones that
Log.w()
and fail to report startup metrics are those where non-frame events are missing, exceptions are thrown when startup is detected except for frame timing information (from UI/RT slices). (Id240f, b/329145809) - Cancel background dexopt jobs before running a Macrobenchmark to reduce interference. (I989ed)
- Added
frameCount
measurement toFrameTimingMetric
to aid in discovery of scenarios where measurements change because the number of frames produced changed (new animations added, invalidation issues fixed). (I1e5aa) - Clarified that
frameOverrunMs
is the preferred metric for tracking when available in docs, and why. (I18749, b/329478323)
Version 1.3.0-alpha02
March 20, 2024
androidx.benchmark:benchmark-*:1.3.0-alpha02
is released. Version 1.3.0-alpha02 contains these commits.
New Features
Experimental R8 support in microbench via embedded proguard rules. Note that this support is experimental, and requires AGP 8.3 for minification of library module tests. Use the following to enable R8 minification/optimization in your benchmark module's
build.gradle
, which should lead to a significant performance increase, depending on workload. (I738a3, b/184378053)android { buildTypes.release.androidTest.enableMinification = true }
Bug Fixes
- Fixes method tracing warning to be on separate line from microbench output. (I0455c, b/328308833)
Version 1.3.0-alpha01
February 21, 2024
androidx.benchmark:benchmark-*:1.3.0-alpha01
is released. Version 1.3.0-alpha01 contains these commits.
API Changes
- Renamed
MicrobenchmarkConfig
boolean parameters to avoid unnecessary word 'should' (Ia8f00, b/303387299) - Added
BenchmarkRule.measureRepeatedOnMainThread
so main thread benchmarks (e.g. ones touching Views or Compose UIs) can avoid triggering ANRs, especially during large suites in CI. (I5c86d) - Added
FrameTimingGfxInfoMetric
, an experimental alternate implementation ofFrameTimingMetric
with measurements coming directly from the platform, rather than extracted from the Perfetto trace. (I457cb, b/322232828) - Add the ability to dump an ART profile during individual
warmUp
iterations. (I17923) - Several changes to
TraceSectionMetric
API:- Add
Mode.Min
,Mode.Max
- Add label argument to override section name as metric label
- Added mode name to output to clarify metric meaning
- Changed default to sum, as most usage of this metric is for repeated events Be aware of this changes in CI usage, as it may create discontinuities or break parsing. (Ic1e82, b/301892382, b/301955938)
- Add
Bug Fixes
- Improved error message in baseline profile gradle plugin when specified managed device does not exist (Idea2b, b/313803289)
- Fix for including library baseline profiles in AAR prior to AGP 8.3.0-alpha15 (I1d2af, b/313992099)
- Fixed baseline and startup profile output url at the end of generation task (I802e5, b/313976958)
- Adjusted data source timeouts to attempt to fix
java.lang.IllegalStateException: Failed to stop [ProcessPid(processName=perfetto, pid=...)]
(I8dc7d, b/323601788) - Add two instrumentation arguments for overriding shader dropping behavior to workaround crashes when benchmarking apps without
ProfileInstaller
1.3:androidx.benchmark.dropShaders.enable=true/false
: can be used to skip all shader dropping (including that done inStartupMode.Cold
launches), esp when benchmarking apps that don't yet use profileinstaller 1.3androidx.benchmark.dropShaders.throwOnFailure=true/false
: can be used to tolerate failures when trying to drop shaders, for example when benchmarking apps without profileinstaller 1.3 (I4f573)
- Skip method tracing on UI thread when expected to take longer than a few seconds, and cleanup method traces when throwing. (I6e768)
- Throw when filenames are longer than 200 chars to avoid unclear crashes when writing or post-processing files. (I4a5ab)
- Fixes issue where unterminated frames at the beginning and end of the trace could be paired together, which would incorrectly report as a single extremely long frame. (I39353, b/322232828)
- Use
--skip verification
on API 30+ when reinstalling a package on API 30-33 to clear ART profiles on user builds. This helps bypass Play Protect warnings that cause failures on some class of devices. (Ic9e36) - Use
am force-stop
to kill apps when not a system app like System UI or Launcher. (I5e028) - Macrobenchmark now waits for
1 second
for the target application to flush an ART profile (previously it waited for500 ms
). (I85a50, b/316082056) - Improve
FrameTimingMetric
error when frames aren't produced, and always output link to trace when failing metric parsing to assist in diagnosing problem. (I956b9) - Fixed crash in
FrameTimingMetric
failing to parse frame id, especially on certain OEM devices. (Ia24bc, b/303823815, b/306235276) - Relaxed strictness of checks in
FrameMetrics
, and added more detail to error messages. (Iadede)
Version 1.2
Version 1.2.4
April 17, 2024
androidx.benchmark:benchmark-*:1.2.4
is released. Version 1.2.4 contains these commits.
Bug Fixes
- Fixes baseline profile srcset not being set up in benchmark variants. Also fixes
automaticGenerationDuringBuild
in libraries causing a circular dependency. (I28ab7, b/333024280) - Use
am force-stop
to kill apps when not a system app like System UI or Launcher. This fixesStartupMode.COLD
benchmarks crashing from "Package $package must not be running prior to cold start!" due to process kill not fully succeeding. (I5e028)
Version 1.2.3
January 24, 2024
androidx.benchmark:benchmark-*:1.2.3
is released. Version 1.2.3 contains these commits.
Bug Fixes
- Removed exception from Baseline Profile Gradle Plugin when AGP version is 8.3.0 or higher.
- Fix for including library baseline profiles in AAR prior to AGP 8.3.0-alpha15.
Version 1.2.2
December 1, 2023
androidx.benchmark:benchmark-*:1.2.2
is released. Version 1.2.2 contains these commits.
Baseline Profiles
- Execution logs will show the baseline profile output file path as a local file URI (aosp/2843918, aosp/2853665, b/313976958)
Version 1.2.1
November 15, 2023
androidx.benchmark:benchmark-*:1.2.1
is released. Version 1.2.1 contains these commits.
New Features
- Improved error message when user disables test variants (b/307478189)
- Added properties to support AS test run integration (b/309805233), (b/309116324)
Version 1.2.0
October 18, 2023
androidx.benchmark:benchmark-*:1.2.0
is released. Version 1.2.0 contains these commits.
Important changes since 1.1.0
Baseline Profiles
- New Baseline Profile Gradle Plugin automates capturing and including baseline profiles in your test and build workflow.
BaselineProfileRule.collect
now stable, a streamlined and simplified version of the previous experimentalBaselineProfileRule.collectBaselineProfile
API- Just specify
packageName
, and drive your app
- Just specify
- For libraries generating baseline profiles, you can now filter the rules generated either in code (
BaselineProfileRule.collect
argument), or even more simply in the gradle plugin - Fixes
- Fixed baseline profile collection on Android U+ (Id1392, b/277645214)
Macrobenchmark
- Compilation
- Macrobenchmark now correctly fully resets compilation state for each compile - this requires reinstalling the APK prior to Android 14, so benchmarking on Android 14+ is strongly recommended if you want to persist state (like user login) in what's being measured.
- You can also work around this by controlling app compilation separately, and skipping compilation with
CompilationMode.Ignore()
or instrumentation argument
Instrumentation Arguments
- Support for
androidx.benchmark.dryRunMode.enable
instrumentation argument, (already available in microbenchmark) for quicker validation runs (e.g. when creating the benchmark, or in presubmit) - Support for
androidx.benchmark.profiling.mode=StackSampling
andMethodTracing
. - Added
androidx.benchmark.enabledRules
to allow runtime filtering baseline profile vs macrobenchmark rule tests - Added
androidx.benchmark.perfettoSdkTracing.enable
argument to enable tracing with tracing-perfetto, e.g. Compose recomposition tracing. Note that when used withStartupMode.COLD
, timing will be significantly affected as the tracing library is loaded and enabled during app startup.
- Support for
Requirements
- Macrobenchmark now requires
ProfileInstaller
1.3.0 or greater in the target app, to enable profile capture / reset, and shader cache clearing.
- Macrobenchmark now requires
New Experimental Metric APIs
- Added experimental
TraceSectionMetric
, which allows for extracting simple timing fromtrace("") {}
blocks in your app, or TraceMetric for leveraging the full query capability of PerfettoTraceProcessor
. - Added experimental
PowerMetric
to capture power usage information - Added experimental
MemoryCountersMetric
to count page faults - Added experimental
PerfettoTraceProcessor
API, which is used internally to extract metrics from System traces (aka Perfetto traces)
- Added experimental
Fixes
- Fixed crashes when installing or extracting profiles from an app installed from multiple APKs (e.g. from app bundle).
- Fixed
FrameTimingMetric
ignoring frames with inconsistent frame IDs (generally, frames during ripples on API 31+) (I747d2, b/279088460) - Fixed parsing errors on traces > 64MB (Ief831, b/269949822)
- Clarified errors when device (especially emulator) OS image not correctly configured for tracing, or compilation
- Skip battery level check for devices without battery (micro and macro)
- Improved file output, with more clear errors for invalid output directories, and safer defaults
- Improved stability of
StartupMode.COLD
by consistently dropping the shader cache (also exposed viaMacrobenchmarkScope.dropShaderCache
) - Fixed leanback fallback for
startActivityAndWait
.
Microbenchmark
- Features
- Profiling was moved to a separate phase, after other metrics, so one test run can display both accurate timing and profiling results.
- Experimental APIs
- Added experimental
MicrobenchmarkConfig
API for defining custom metrics and configuring tracing and profiling. Can be used to capture method traces, or capture tracepoints (but be aware of tracing overhead). - Added experimental APIs for controlling
BenchmarkState
separately fromBenchmarkRule
, without JUnit - Added experimental
PerfettoTrace
record to enable capturing Perfetto traces, with custom configuration, separate from benchmark APIs.
- Added experimental
- Fixes
- Workaround missing leading whitespaces in Android Studio benchmark output.
- Fix issue where warnings could fail to print in Android Studio benchmark output.
- Fixed
SampledProfiling
crash on Android 13 (API 33) and higher. - Massively improved performance of
dryRunMode
by skippingIsolationActivity
and Perfetto tracing (Up to 10x faster dry run mode on older OS versions).
Version 1.2.0-rc02
October 6, 2023
androidx.benchmark:benchmark-*:1.2.0-rc02
is released. Version 1.2.0-rc02 contains these commits.
Bug Fixes
- Fix Benchmark file output to no longer break
BaselineProfile
Plugin file copying. Files were generated and copied off device, but had been renamed such that the gradle plugin wouldn't see them. (I8dbcc, b/303034735, b/296453339) - Clarified
tracing-perfetto
loading error messages when injecting from macrobenchmark module into target application.
Version 1.2.0-rc01
September 20, 2023
androidx.benchmark:benchmark-*:1.2.0-rc01
is released. Version 1.2.0-rc01 contains these commits.
Bug Fixes
- An exception (with remedy instructions) is now thrown when Perfetto SDK tracing fails to initialize in a Benchmark. (I6c878, b/286228781)
- Fix OOM crash when converting ART method trace -> perfetto format. (I106bd, b/296905344)
- (Macrobenchmark) Clarified method tracing label when linked in Studio test output, and fixed method tracing filenames to be unique on device/host, so they won't be overwritten when more than one benchmark is run. (I08e65, b/285912360)
- Ensures that the device is awake when capturing a baseline profile. (I503fc)
Version 1.2.0-beta05
August 30, 2023
androidx.benchmark:benchmark-*:1.2.0-beta05
is released. Version 1.2.0-beta05 contains these commits.
New Features
- The Baseline Profile Gradle Plugin now supports Android Gradle Plugin 8.3. (aosp/2715214)
Version 1.2.0-beta04
August 23, 2023
androidx.benchmark:benchmark-*:1.2.0-beta04
is released. Version 1.2.0-beta04 contains these commits.
New Features
- The Baseline Profiles Gradle plugin now supports Android Gradle Plugin 8.3. (aosp/2715214)
Bug Fixes
- Fix failures in writing / moving and pulling files (especially those from parameterized tests) by sanitizing output file names further, avoiding '=' and ':' in output file names. (I759d8)
Version 1.2.0-beta03
August 9, 2023
androidx.benchmark:benchmark-*:1.2.0-beta03
is released. Version 1.2.0-beta03 contains these commits.
API Changes
- Added argument to filter
TraceSectionMetric
to only the target package, on by default (Ia219b, b/292208786)
Bug Fixes
- Renamed
fullTracing.enable
instrumentation argument toperfettoSdkTracing.enable
for consistency with artifact name, and other references.fullTracing.enable
will continue to work as a fallback. (I7cc00) - Benchmark library internal tracepoints (including microbenchmark loop/phase tracing) will now show up in Studio system trace viewer, and nest under the correct process in Perfetto. (I6b2e7, b/293510459)
- Removed macrobenchmark NOT-PROFILEABLE error on API 31+, and skip profileable check on eng/userdebug rooted devices. (I2abac, b/291722507)
- When using Dex Layout Optimizations, startup profile rules are also now considered as baseline profile rules. (aosp/2684246, b/293889189)
Version 1.2.0-beta02
July 26, 2023
androidx.benchmark:benchmark-*:1.2.0-beta02
is released. Version 1.2.0-beta02 contains these commits.
API Changes
- Added experimental APIs for microbench custom metrics and configuration (e.g. profiler, and tracing). (I86101, b/291820856)
Bug Fixes
- Report error in macrobench when OS is misconfigured for tracing, as was recently fixed in API 26/28 ARM64 emulators. (I0a328, b/282191686)
- Added detail to compilation reset failure to suggest updating emulator, as some emulators have recently fixed this issue. (I8c815, b/282191686)
- Make
androidx.test.uiautomator:uiautomator:2.2.0
anapi
instead of animplementation
dependency. (I1981e)
Version 1.2.0-beta01
July 18, 2023
androidx.benchmark:benchmark-*:1.2.0-beta01
is released. Version 1.2.0-beta01 contains these commits.
Bug Fixes
- Fix warnings being sometimes suppressed in Benchmark output in Studio, and workaround leading whitespaces from Benchmark output not showing up in Studio (Ia61d0, b/227205461, b/286306579, b/285912360)
- Fixed comment for
FrameTimingMetric
. The submetric is namedframeDurationCpuMs
. (Ib097f, b/288830934).
Version 1.2.0-alpha16
June 21, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha16
is released. Version 1.2.0-alpha16 contains these commits.
API Changes
BaselineProfileRule.collectBaselineProfile()
API has been renamed toBaselineProfileRule.collect()
. (I4b665)
Bug Fixes
- Macrobenchmark support for
androidx.benchmark.profiling.mode = MethodTracing
. (I7ad37, b/285912360) - Microbenchmark profiling moved to a separate phase, so it occurs in sequence after measurement, instead of replacing it.
MethodTracing
trace sections are also now included in the captured Perfetto trace, if present. (I9f657, b/285014599) - Add count measurement to
TraceSectionMetric
withMode.Sum
. (Ic121a, b/264398606)
Version 1.2.0-alpha15
June 7, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha15
is released. Version 1.2.0-alpha15 contains these commits.
New Features
- Added experimental
MemoryUsageMetric
for tracking memory usage of a target application. (I56453, b/133147125, b/281749311) - Add support for fully custom Perfetto configs with
PerfettoTrace.record
(If9d75, b/280460183) - Added property to skip baseline profile generation. Usage:
./gradlew assemble -Pandroidx.baselineprofile.skipgeneration
. (I37fda, b/283447020)
API Changes
- The
collectBaselineProfile
API always generates stable baseline profiles. ThecollectStableBaselineProfile
API has been removed andcollectBaselineProfile
should be used instead. (I17262, b/281078707) - Changed
BaselineProfileRule
'sfilterPredicate
arg to non-null, with a equivalent default value so that the default filter behavior is more clear in docs. (I3816e)
Bug Fixes
- Disable
IsolationActivity
and Perfetto tracing indryRunMode
to significantly improve performance, as these were majority of runtime. (Ie4f7d) - Support for call stack sampling in Macrobenchmarks using instrumentation test arguments
androidx.benchmark.profiling.mode=StackSampling
andandroidx.benchmark.profiling.sampleFrequency
. (I1d13b, b/282188489) - Fixes crash when dropping shaders on Android U (API 34), as well as on emulators. (I031ca, b/274314544)
Version 1.2.0-alpha14
May 3, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha14
is released. Version 1.2.0-alpha14 contains these commits.
Bug Fixes
- Fix
FrameTimingMetric
ignoring frames with inconsistent frame IDs. This would cause some animations on recent platform versions (API 31+) to ignore many frames whileRenderThread
was animating (e.g. during a ripple). (I747d2, b/279088460) - Fixed trace processor parsing for traces larger than 64Mb. (Ief831, b/269949822)
- Fixed baseline profile generation on Android U failing because of the different output of
pm dump-profiles
command. (Id1392, b/277645214) - Fix GPU clock locking script to compare strings correctly (I53e54, b/213935715)
Version 1.2.0-alpha13
April 5, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha13
is released. Version 1.2.0-alpha13 contains these commits.
API Changes
- Added profile type parameter when generating baseline profiles to support upcoming startup profile feature (Ie20d7, b/275093123)
- Added new experimental
TraceMetric
API for defining fully custom metrics based on content of a Perfetto trace. (I4ce31, b/219851406) - Add an experimental metric to determine the number of page faults during a benchmark. (I48db0)
Version 1.2.0-alpha12
March 22, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha12
is released. Version 1.2.0-alpha12 contains these commits.
New Features
- The new baseline profile gradle plugin is released in alpha version, making it easier to generate a baseline profile and simplifying the developer workflow.
API Changes
- Removed Perfetto tracing support on API 21 and 22, which includes both Microbenchmarks and the experimental
PerfettoTrace
APIs. Prior to this version,UiAutomation
connections were unreliable on some devices. (I78e8c) - Added public experimental API for
PerfettoTraceProcessor
to enable parsing trace content. This is a step toward fully custom metrics based on Perfetto trace data. (I2659e, b/219851406)
Version 1.2.0-alpha11
March 8, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha11
is released. Version 1.2.0-alpha11 contains these commits.
Bug Fixes
- Fixed crashes in
MacrobenchmarkRule
andBaselineProfileRule
when reinstalling or extracting profiles from an app bundle with multiple APKs. (I0d8c8, b/270587281)
Version 1.2.0-alpha10
February 22, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha10
is released. Version 1.2.0-alpha10 contains these commits.
New Features
- On Android 14+, Macrobenchmark no longer reinstalls target applications to reset compilation state, thanks to a new platform feature. Previously it was necessary to have a rooted device, or to deal with all application state (e.g. user login) being removed before each benchmark runs. (I9b08c, b/249143766)
Bug Fixes
- Fix
DryRunMode
to no longer crash with empty profile, due to compilation skipping. Instead, it runs a single iteration and extracts the profile to ensure something is captured. (I2f05d, b/266403227) - Fix
PowerMetric
crash when checking for powerstats presence on old API levels. (5faaf9, b/268253898)
Version 1.2.0-alpha09
January 11, 2023
androidx.benchmark:benchmark-*:1.2.0-alpha09
is released. Version 1.2.0-alpha09 contains these commits.
Bug Fixes
- Enabled passing
None
toandroidx.benchmark.enabledRules
instrumentation arg to disable all benchmarks / baseline profile generation. (I3d7fd, b/258671856) - Fix
PerfettoTrace
capture in app modules (i.e. non-self-instrumenting test APKs) (I12cfc) - Fixed baseline profile adb pull argument order in Studio output (I958d1, b/261781624)
- Arm emulator api 33 is now correctly recognized as such when trying to run a macrobenchmark and will correctly print the warning. (69133b,b/262209591)
- Skip battery level check on devices without battery in Macrobenchmark (fe4114, b/232448937)
Version 1.2.0-alpha08
December 7, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha08
is released. Version 1.2.0-alpha08 contains these commits.
API Changes
- Added experimental new APIs
PerfettoTrace.record {}
andPerfettoTraceRule
to capture Perfetto traces (also known as System Traces) as part of a test, to inspect test behavior and performance. (I3ba16) BaselineProfileRule
now accepts a filter predicate instead of a list of package prefixes. This gives the test full control on filtering. (I93240)- Add an experimental API
BaselineProfileRule.collectStableBaselineProfile
which waits until a baseline profile is stable for N iterations. (I923f3) - Add the ability to specify an output file name prefix when generating baseline profiles using
BaselineProfileRule
. (I7b59f, b/260318655)
Bug Fixes
- Improve safety of file output writing, which should prevent output files from silently not being written / appended, especially on API 21/22. (If8c44, b/227510293)
- Fix
simpleperf
trace output to create and place the file correctly. This should also more generally fix issues where a file is unsuccessfully pulled by gradle. (I12a1c, b/259424099) - Improve profileinstaller error message printed when profileinstaller is too old. This now tells you to update profileinstaller version (1.2.1) for measuring baseline profiles on API 31 through 33, instead of saying it's not supported. (Ia517f, b/253519888)
- Fix several shell command failures onerror message Print needed API <=23, including failed perfetto capture binary setup and trace capture failures (Ib6b87, b/258863685)
- Automatically sort generated profile rules to minimize the number of changes as they change over time (when checking-in profile rules into source control). (Ie2509)
- Fixed crash on unrooted builds below Android 13 (API 33) with message
Expected no stderr from echo 3 > /proc/sys/vm/drop_caches
(I6c245, b/259508183)
Known Issues
- MacrobenchmarkScope.dropShaderCache()
may crash due to a missing broadcast registry in profileinstaller manifest, which has not yet been released. (I5c728, b/258619948) To workaround the issue in profileinstaller:1.3.0-alpha02
, add the following to your application's (not your benchmark's) AndroidManifest.xml:
<!-- workaround bug in profileinstaller 1.3.0-alpha02, remove when updating to alpha03+ -->
<receiver
android:name="androidx.profileinstaller.ProfileInstallReceiver"
android:permission="android.permission.DUMP"
android:exported="true">
<intent-filter>
<action android:name="androidx.profileinstaller.action.BENCHMARK_OPERATION" />
</intent-filter>
</receiver>
Version 1.2.0-alpha07
November 9, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha07
is released. Version 1.2.0-alpha07 contains these commits.
API Changes
- Adds
PowerMetric
API for measuring energy and power in Macrobenchmarks. (Ife601, b/220183779) - Fixed
MacrobenchmarkScope.dropShaderCache()
to actually drop the shader cache. This removes roughly 20ms of noise fromStartupMode.COLD
benchmarks, as shaders are now consistently cleared each iteration. Previously,Partial
compilation using warmup iterations would report incorrectly fast numbers, as shader caching was more likely to happen during warmup. This fix requires either a rooted device, or usingprofileinstaller:1.3.0-alpha02
in the target app. ForProfileInstaller
library’s API changes, please refer to ProfileInstaller 1.30-alpha02 page. (Ia5171, b/231455742) - Added
TraceSectionMode("label", Mode.Sum)
, allowing measurement of total time spent on multiple trace sections with the same label. For instance,TraceSectionMetric("inflate", Mode.Sum)
will report a metricinflateMs
for the total time in a macrobenchmark spent on inflation. Also removed API 29 requirement, asTraceSectionMetric
works together withandroidx.tracing.Trace
back to lower API levels, with the use offorceEnableAppTracing
within the target app. (Id7b68, b/231455742)
Bug Fixes
- Improved safety of all internal shell commands by validating all output/errors. (I5984d, b/255402908, b/253094958)
- Specify device in baseline profile
adb pull
command, so the pull command can be simply copied if multiple devices are connected (up to one emulator) (I6ac6c, b/223359380) - Add error if macrobenchmark test apk isn't set up as self-instrumenting. This error prevents macrobenchmarking from within the target app's process. In process, macrobench wouldn't be able to compile/kill/cold start the app, or control its own permissions (I4279b)
- Fixed an issue in
measureRepeated()
whereStartupMode.COLD
wouldn't kill the target process aftersetupBlock
. NowsetupBlock
interacting with the app will not leave the app process running, and an invalid cold start measurement. (I8ebb7)
Version 1.2.0-alpha06
October 24, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha06
is released. Version 1.2.0-alpha06 contains these commits.
API Changes
BaselineProfileRule
no longer requires root on Android 13 (API 33), and is no longer experimental. (Ie0a7d, b/250083467, b/253094958)- This change also fixes how profiles from an app are flushed to disk on unrooted devices, but requires updating the target app's profileinstaller dependency.
- To use
BaselineProfileRule
orCompilationMode.Partial(warmupIterations)
on an unrooted device, you must also update your target app to useandroidx.profileinstaller.profileinstaller:1.3.0-alpha01
. This enables flushing the profile to disk correctly, so that it can be compiled/extracted.
Bug Fixes
- Fixes
SampledProfiling
crash on API 33+. (I40743, b/236109374)
Version 1.2.0-alpha05
October 5, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha05
is released. Version 1.2.0-alpha05 contains these commits.
Bug Fixes
- Fix frame breakdown in Studio system trace viewer for benchmark captured traces (I3f3ae, b/239677443)
- Correct
FrameTimingMetric
to listFrameOverrun
as requiring API 31 instead of 29 (I716dd, b/220702554) - Set iteration in
BaselineProfileRule
, and throw clearly if target package not installed (was already done for MacrobenchmarkRule). (Ic09a3, b/227991471)
Version 1.2.0-alpha04
September 21, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha04
is released. Version 1.2.0-alpha04 contains these commits.
New Features
Add support for
dryRunMode.enable
instrumentation argument to macrobenchmark (already available in micro) for faster local development, and validating app automation (e.g. in presubmit). This overrides iterations to 1, skips compilation, suppresses all configuration errors, and disables measurement .json file output. (Ib51b4, b/175149857)On Gradle command line:
./gradlew macrobenchmark:cC -P android.testInstrumentationRunnerArguments.androidx.benchmark.dryRunMode.enable=true
In build.gradle:
android { defaultConfig { testInstrumentationRunnerArgument 'androidx.benchmark.dryRunMode.enable', 'true' } }
Bug Fixes
- Fixed
StartupTimingMetric
to no longer require measured Activities to be launched throughMacrobenchmarkScope.startActivityAndWait()
. This means the metric can pick up launches from e.g. notifications,Context.startActivity()
, in-app Activity based navigation, or shell commands. (Ia2de6, b/245414235) - Fix bug where
startActivityAndWait
would timeout trying to wait for launch completion on emulators by reducing strictness of frame detection. (Ibe2c6, b/244594339, b/228946895)
Version 1.2.0-alpha03
September 7, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha03
is released. Version 1.2.0-alpha03 contains these commits.
New Features
- Added experimental APIs for using
BenchmarkState
independently, separate fromBenchmarkRule
/JUnit4
. (Id478f, b/228489614)
Bug Fixes
- Added Leanback fallback for
startActivityAndWait
. (01ed77, b/242899915)
Version 1.2.0-alpha02
August 24, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha02
is released. Version 1.2.0-alpha02 contains these commits.
API Changes
- Default to
am force stop
forMacrobenchmarkScope.killProcess()
, even when rooted, except during Baseline Profile generation. This can be overridden with an optional boolean argument. (02cce9, b/241214097)
Bug Fixes
- Support baseline profile generation for System apps. (I900b8, b/241214097)
- Support checking for ODPM power metrics on unrooted devices. (a38c78, b/229623230)
Version 1.2.0-alpha01
July 27, 2022
androidx.benchmark:benchmark-*:1.2.0-alpha01
is released. Version 1.2.0-alpha01 contains these commits.
New Features
- New tracing-perfetto-common component allowing tooling to enable Perfetto SDK tracing in an app that exposes it (I2cc7f)
Added
androidx.benchmark.enabledRules
instrumentation argument to enable filtering macrobenchmark runs to just benchmarks, or just baseline profile generation. Pass in 'Macrobenchmark', or 'BaselineProfile' to just run one type of test, e.g. when just generatingBaselineProfiles
on an emulator. Comma-separated list also Supported. (I756b7, b/230371561)E.g. in Your macrobenchmark's build.gradle:
android { defaultConfig { testInstrumentationRunnerArgument 'androidx.benchmark.enabledRules', 'BaselineProfile' } }
Or from the Gradle command line:
./gradlew macrobenchmark:cC -P android.testInstrumentationRunnerArguments.androidx.benchmark.enabledRules=BaselineProfile
API Changes
- Added new
PowerMetric
for measuring energy and power tasks in benchmarks. (I9f39b, b/220183779) - Added a new compilation mode
CompilationMode.Ignore
to skip profile reset and compilation. (Ibbcf8, b/230453509) - Added a new parameter to
BaselineProfileRule#collectBaselineProfile
to filter output file by package names (If7338, b/220146561) - Enables developer to discharge device to measure power drain. (I6a6cb)
- Added the ability to clear shader cache in
MacrobenchmarkScope
. (I32122) - Enables developer to configure display of metric type and detail desired subsystem categories. (I810c9)
- Previously an
UnsupportedOperationException
was thrown in the benchmark if run on an unsupported device. Now UOE only occurs if the metric is used on the unsupported device (ie:PowerMetric.configure
). (I5cf20, b/227229375) - Added
TotalPowerMetric
andTotalEnergyMetric
for measuring total power and energy in each system category in macrobenchmarks. (I3b26b, b/224557371)
Bug Fixes
- Fixed an issue where compiled methods were not correctly being reset between each macrobenchmark on unrooted builds. This unfortunately requires reinstalling the apk each iteration, which will clear application data for each macrobenchmark. (I31c74, b/230665435)
- Fix trace recording crash on API 21/22 (If7fd6, b/227509388, b/227510293, b/227512788)
- Overhaul activity launch completion detection to fix 'Unable to read any metrics' exception in startup macrobenchmarks. (Ia517c)
Version 1.1.1
Version 1.1.1
November 9, 2022
androidx.benchmark:benchmark-*:1.1.1
is released. Version 1.1.1 contains these commits.
Bug Fixes
- Fixes
android.system.ErrnoException: open failed: EACCES
which would occur on some Android11 (API 30)+ devices. This is a cherry-pick of a fix from1.2.0-alpha01
. (aosp/2072249)
Version 1.1.0
Version 1.1.0
June 15, 2022
androidx.benchmark:benchmark-*:1.1.0
is released. Version 1.1.0 contains these commits.
- This version is identical to
androidx.benchmark:benchmark-*:1.1.0-rc03
.
Important changes since 1.0.0
Support for Jetpack Macrobenchmarks, which allows you to measure whole-app interactions like startup and scrolling, provides the ability to capture traces & measure trace sections.
Support for Baseline Profiles
CompilationMode.Partial
to measure the effectiveness of Baseline Profiles.@BaselineProfileRule
to automatically generate Baseline profiles for a given critical user journey.
Support for Allocation metrics & profiling during Microbenchmark runs.
Version 1.1.0-rc03
June 1, 2022
androidx.benchmark:benchmark-*:1.1.0-rc03
is released. Version 1.1.0-rc03 contains these commits.
Bug Fixes
Avoid reinstalling the target package on every benchmark iteration. ( aosp/2093027, b/231976084)
Remove the
300ms
delay frompressHome()
. (aosp/2086030, b/231322975)Improve Macrobenchmark iteration speed by optimizing Shell commands used under the hood. (aosp/2086023, b/231323582)
Support for Managed Gradle Devices when generating Baseline Profiles with Macrobenchmarks. (aosp/2062228, b/228926421)
Version 1.1.0-rc02
May 11, 2022
androidx.benchmark:benchmark-*:1.1.0-rc02
is released. Version 1.1.0-rc02 contains these commits.
- Note that this release includes a behavior change, as apps are now fully reinstalled in between each benchmark to ensure accurate measurements.
Bug Fixes/Behavior Changes
Fixed an issue where app compilation was not correctly reset between macrobenchmarks, and not reset at all on unrooted builds. This fixes many cases where running multiple tests would result in
CompilationMode
having little to no effect on measurements. To workaround this problem, the target app is now fully reinstalling each test method, which will clear application data between each macrobenchmark. (I31c74, b/230665435)As this prevents apps from setting up state before tests, it is now possible to skip compilation / reinstallation to enable working around this. You can for example fully compile the target with a shell command
cmd package compile -f -m speed <package>
, and then bypass macrobenchmark's compilation step.E.g. in Your macrobenchmark's build.gradle:
android { defaultConfig { testInstrumentationRunnerArgument 'androidx.benchmark.compilation.enabled, 'false' } }
Or from the Gradle command line:
./gradlew macrobenchmark:cC -P android.testInstrumentationRunnerArguments.androidx.benchmark.compilation.enabled=false
Made it possible to share a module between macrobenchmarks and baseline profile generating tests by adding
androidx.benchmark.enabledRules
instrumentation argument. Pass in 'Macrobenchmark', or 'BaselineProfile' to just run one type of test, e.g. when generatingBaselineProfiles
on an emulator. (I756b7, b/230371561)E.g. in Your macrobenchmark's build.gradle:
android { defaultConfig { testInstrumentationRunnerArgument 'androidx.benchmark.enabledRules', 'BaselineProfile' } }
Or from the Gradle command line:
./gradlew macrobenchmark:cC -P android.testInstrumentationRunnerArguments.androidx.benchmark.enabledRules=BaselineProfile
Version 1.1.0-rc01
April 20, 2022
androidx.benchmark:benchmark-*:1.1.0-rc01
is released. Version 1.1.0-rc01 contains these commits.
Bug Fixes
- Baseline profile output links in Android Studio now use a unique file name. This way the output always reflects the latest results of using a
BaselineProfileRule
. ( aosp/2057008, b/228203086 )
Version 1.1.0-beta06
April 6, 2022
androidx.benchmark:benchmark-*:1.1.0-beta06
is released. Version 1.1.0-beta06 contains these commits.
Bug Fixes
- Fix trace recording crash on API 21/22 (If7fd6, b/227509388)
- Overhaul activity launch completion detection to fix 'Unable to read any metrics' exception in startup macrobenchmarks. (Ia517c)
- Fix startup metrics for Macrobenchmarks when
CompilationMode.None()
is used. Before this change,CompilationMode.Partial()
would appear to be slower thanCompilation.None()
. (611ac9).
Version 1.1.0-beta05
March 23, 2022
androidx.benchmark:benchmark-*:1.1.0-beta05
is released. Version 1.1.0-beta05 contains these commits.
Bug Fixes
- Kill package after skipping profile installation when using
CompilationMode.None
. (aosp/1991373) - Fixed an issue where Macrobenchmarks is unable to collect startup metrics when using
StartupMode.COLD
. (aosp/2012227 b/218668335)
Version 1.1.0-beta04
February 23, 2022
androidx.benchmark:benchmark-*:1.1.0-beta04
is released. Version 1.1.0-beta04 contains these commits.
Bug Fixes
Fix missing metrics on Android 10, and
NoSuchElementException
caused by process names not being captured correctly in traces. (Ib4c17, b/218668335)Use
PowerManager
for thermal throttling detection on Q (API 29) and higher. This significantly reduces frequency of false positives in thermal throttling detection (benchmark retry after 90 second cooldown), and speeds up benchmarks significantly on user builds. It also provides throttle detection even when clocks are locked (if they're locked too high for the device's physical environment). (I9c027, b/217497678, b/131755853)Filter simpleperf sampled profiling to
measureRepeated
thread only to simplify inspection (Ic3e12, b/217501939)Support metrics from named UI subprocesses in multi-process apps (Ice6c0, b/215988434)
Filter Baseline Profile rules to target Android 9 (SDK 28). aosp/1980331 b/216508418
Skip Profile Installation when using
Compilation.None()
. Additionally, report warnings when the app is using an older version ofandroidx.profileinstaller
and Android Gradle Plugin. aosp/1977029
Version 1.1.0-beta03
February 9, 2022
androidx.benchmark:benchmark-*:1.1.0-beta03
is released. Version 1.1.0-beta03 contains these commits.
API Changes
- Added
AudioUnderrunMetric
into macrobenchmark library under experimental flag to allow detection of audio underruns (Ib5972) BaselineProfileRule
no longer accepts asetup
block as this functioned the same as theprofileBlock
. (Ic7dfe, b/215536447)For e.g.
@Test fun collectBaselineProfile() { baselineRule.collectBaselineProfile( packageName = PACKAGE_NAME, setupBlock = { startActivityAndWait() }, profileBlock = { // ... } ) }
@Test fun collectBaselineProfile() { baselineRule.collectBaselineProfile( packageName = PACKAGE_NAME, profileBlock = { startActivityAndWait() // ... } ) }
Bug Fixes
- Fixed issue where microbench profiler traces would fail to be updated in subsequent runs when linked in Studio output (I5ae4d, b/214917025)
- Prevent compilation shell commands on API 23 (Ice380)
- Renamed
FrameCpuTime
->FrameDurationCpu
,FrameUiTime
->FrameDurationUi
to clarify these are durations, not timestamps, and to match prefixes. (I0eba3, b/216337830)
Version 1.1.0-beta02
January 26, 2022
androidx.benchmark:benchmark-*:1.1.0-beta02
is released. Version 1.1.0-beta02 contains these commits.
Bug Fixes
- Microbenchmark Stack Sampling / Method Tracing Profile results are now linked in Studio output, similar to other profiling outputs, and do not suppress the allocation metric. (Idcb65, b/214440748, b/214253245)
- BaselineProfileRule now prints the
adb pull
command in logcat and Studio output for pulling generated BaselineProfile text file. (f08811)
Version 1.1.0-beta01
January 12, 2022
androidx.benchmark:benchmark-*:1.1.0-beta01
is released. Version 1.1.0-beta01 contains these commits.
Bug Fixes
- Fixes profiler argument enable being ignored. (I37373, b/210619998)
- Removed deprecated
CompliationModes
(I98186, b/213467659) - Switched baseline profile arg of
CompilationMode.Partial
to enum for clarity. (Id67ea)
Version 1.1.0-alpha13
December 15, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha13
is released. Version 1.1.0-alpha13 contains these commits.
API Changes
- Add low-overhead System Tracing to microbench output on Android Q (API 29+). Note that this does not currently capture custom tracing (via
android.os.Trace
orandroidx.tracing
Jetpack APIs) to avoid affecting results. This tracing should be useful in diagnosing instability, especially from sources outside the benchmark. (I298be, b/205636583, b/145598917) - Clarify
CompilationModes
into three classes - Full, None, Partial. Previously they were inconsistently named after compilation arguments (which we now treat as implementation details) and features. This makes the tradeoffs, potential combinations, and behavior across platform versions more clear. (I3d7bf, b/207132597) - Setup and measure are now always in pairs, in order. You can now query the package name and iteration (though the iteration may be
null
in certain warmup scenarios). (Id3b68, b/208357448, b/208369635)
Bug Fixes
- Fixed
CompilationMode.Speed
incorrectly treated asNone
(I01137)
Version 1.1.0-alpha12
November 17, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha12
is released. Version 1.1.0-alpha12 contains these commits.
New Features
- Add experimental TraceSectionMetric for custom trace-based timing measurements. (I99db1, b/204572664)
Bug Fixes
- Wake device each iteration, to ensure UI can be tested - requires lockscreen is disabled. (Ibfa28, b/180963442)
- Fixes multiple crashes in StackSampling profiling mode on emulators and non-rooted devices (Icdbda, b/202719335)
- Removed 0.5 second sleep at the end of each iteration - if you see missing metrics with this change, please file a bug. (Iff6aa)
- Reduce chances of dropped data, and lower memory overhead from tracing (Id2544, b/199324831, b/204448861)
- Reduce trace size by ~40% by switching to compact sched storage format. (Id5fb6, b/199324831)
- Updated implementations of startup metrics to always end at end of renderthread. This will be more contistent across platform versions, and more closely map to in-app measurements. (Ic6b55)
Version 1.1.0-alpha11
November 3, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha11
is released. Version 1.1.0-alpha11 contains these commits.
API Changes
- Macrobenchmark now has a
minSdkVersion
of23
. (If2655) - Adds a new experimental
BaselineProfileRule
which is capable of generating baseline profiles for app's critical user journey. Detailed documentation to follow. (Ibbefa, b/203692160) - Removes measureRepeated interface variant, which was added for java callers, as it caused ambiguity in completing/resolving the method. Java callers will again need to return Unit.Instance from measureRepeated. If this is an inconvenience, please file a bug, we can revisit this in a future version. (Ifb23e, b/204331495)
Version 1.1.0-alpha10
October 27, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha10
is released. Version 1.1.0-alpha10 contains these commits.
API Changes
- Backport StartupTimingMetric to work back to API 23. This new implementation also better handles reportFullyDrawn() to wait until corresponding content has been rendered. (If3ac9, b/183129298)
- Added JvmOverloads to multiple MacrobenchmarkScope methods for Java callers. (I644fe, b/184546459)
- Provide alternative MacrobenchmarkRule.measureRepeated function that uses a
Consumer<MacrobenchmarkScope>
for idiomatic usage in Java language. (If74ab, b/184546459)
Bug Fixes
- Fix for traces not starting early enough, and missing metric data. This is expected to fix "Unable to read any metrics during benchmark" exceptions that were caused by the library itself. (I6dfcb, b/193827052, b/200302931)
- FrameNegativeSlack has been renamed to FrameOverrun to clarify its meaning - how much the frame went over its time budget. (I6c2aa, b/203008701)
Version 1.1.0-alpha09
October 13, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha09
is released. Version 1.1.0-alpha09 contains these commits.
Bug Fixes
- Support dropping Kernel page cache without root on API 31/S+, which will increase accuracy of StartupMode.COLD launches. (Iecfdb, b/200160030)
Version 1.1.0-alpha08
September 29, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha08
is released. Version 1.1.0-alpha08 contains these commits.
API Changes
- Enable scrolling macrobenchmarks to run back to API 23 (If39c2, b/183129298)
- Add new type of sampled metric to UI and JSON output, focused on percentiles of multiple samples per iteration. (I56247, b/199940612)
- Switch to floating point metrics throughout the benchmark libraries (truncated in the Studio UI). (I69249, b/197008210)
Version 1.1.0-alpha07
September 1, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha07
is released. Version 1.1.0-alpha07 contains these commits.
API Changes
- Raised min API to 21 to reflect the intended lowest API level to be supported in the future. Current min API supported continues to be conveyed via RequiredApi(), and is currently 29 (I440d6, b/183129298)
Bug Fixes
- Fixes
ProfileInstaller
to make it easier for apps using baseline profiles to run MacroBenchmarks usingCompilationMode.BaselineProfile
. (I42657, b/196074999) NOTE: requires also updating toandroidx.profileinstaller:profileinstaller:1.1.0-alpha04
or greater. StartupMode.COLD
+CompilationMode.None
benchmarks are now more stable. (I770cd, b/196074999)
Version 1.1.0-alpha06
August 18, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha06
is released. Version 1.1.0-alpha06 contains these commits.
API Changes
- Added
androidx.benchmark.iterations
instrumentation argument to allow manual overriding of iteration count when testing/profiling locally. (6188be, b/194137879)
Bug Fixes
- Switched to Simpleperf as default sampling profiler on API 29+. (Ic4b34, b/158303822)
Known Issues
CompilationMode.BaselineProfile
is a work in progress. Avoid using it to determine how good a profile is for now.
Version 1.1.0-alpha05
August 4, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha05
is released. Version 1.1.0-alpha05 contains these commits.
1.1.0-alpha04
was cancelled before release due to a sporatic crash. b/193827052
API Changes
- Switched startActivityAndWait to invoke launch via
am start
, which reduces the time of each measurement iteration by approximately 5 seconds, at the cost of no longer supporting intent parcelables. (I5a6f5, b/192009149
Bug Fixes
- Reduce aggressiveness of thermal throttle detection, and recompute baseline if throttles are detected frequently. (I7327b)
- Fixes FrameTimingMetric to work on Android S beta (Ib60cc, b/193260119)
- Use an
EmptyActivity
to bring the target app out of a force-stopped state to better supportCompilationMode.BaselineProfile
. (Id7cac, b/192084204) - Changed trace file extension to
.perfetto-trace
to match platform standard. (I4c236, b/174663039) - StartupTimingMetric now outputs the "fullyDrawnMs" metric to measure time until your application has completed rendering. To define this metric for your app, call Activity.reportFullyDrawn when your initial content is ready, such as when your initial list items are loaded from DB or network. (reportFullyDrawn method available without build version checks on ComponentActivity). Note that your test must run long enough to capture the metric (startActivityAndWait doesn't wait for reportFullyDrawn). (If1141, b/179176560)
- Reduce cost of appending Ui metadata to traces by 50+ ms (Ic8390, b/193923003)
- Drastically increased polling frequency when stopping tracing, which can reduce e.g. startup benchmark runtime by 30+% (Idfbc1, b/193723768)
Version 1.1.0-alpha03
June 16, 2021
androidx.benchmark:benchmark-*:1.1.0-alpha03
is released. Version 1.1.0-alpha03 contains these commits.
New Features
- Added a new
CompilationMode.BaselineProfile
to support profiles installed using the Jetpack ProfileInstaller library. (aosp/1720930)
Bug Fixes
The sample Gradle code for suppressing benchmark errors has been updated to use a non-deprecated API with a syntax that also supports .gradle.kts users.
E.g.,
testInstrumentationRunnerArguments["androidx.benchmark.suppressErrors"] = "EMULATOR,LOW-BATTERY"
Version 1.1.0-alpha02
May 18, 2021
Benchmark version 1.1.0-alpha02 brings a big component to benchmarking - Macrobenchmark. In addition to benchmark allowing you to measure CPU loops, macrobenchmark allows you to measure whole-app interactions like startup and scrolling, and capture traces. For more information see the library documentation.
androidx.benchmark:benchmark-*:1.1.0-alpha02
is released. Version 1.1.0-alpha02 contains these commits.
New Features
Macrobenchmark artifacts added (androidx.benchmark:benchmark-macro-junit4
and androidx.benchmark:benchmark-macro
)
- Capture startup, scrolling/animation performance metrics from your app, locally or in CI
- Capture and inspect traces from within Android Studio
Bug Fixes
- Workaround shell permissions issue with output directory on Android 12 (Note - may require updating Android Gradle Plugin to 7.0.0 canary and Android Studio to Arctic Fox (2020.3.1), to continue capturing output files on affected devices). (Icb039)
- Support configuration caching in BenchmarkPlugin (6be1c1, b/159804788)
- Simplified file output - on by default, in a directory that doesn't require
requestLegacyExternalStorage=true
(8b5a4d, b/172376362) - Fixes library printing logcat warnings about not finding JIT thread on platform versions where it is not present. (I9cc63, b/161847393)
- Fix for reading device max frequency. (I55c7a)
Version 1.1.0-alpha01
June 10, 2020
androidx.benchmark:benchmark-common:1.1.0-alpha01
, androidx.benchmark:benchmark-gradle-plugin:1.1.0-alpha01
, and androidx.benchmark:benchmark-junit4:1.1.0-alpha01
are released. Version 1.1.0-alpha01 contains these commits.
New Features of 1.1
- Allocation Metric - Benchmarks now run an additional phase after warmup and timing, capturing allocation counts. Allocations can cause performance problems on older versions of the platform (140ns in O became 8ns in M - measured on Nexus5X, with locked clocks). This metric is displayed in Android Studio console output, as well as in the
- Profiling support - You can now capture profiling data for a benchmark run, to inspect why your code may be running slowly. Benchmark supports capturing either method tracing, or method sampling from ART. These files can be inspected with the Profiler inside Android Studio using File > Open.
- The Benchmark Gradle plugin now provides defaults for simpler setup:
testBuildType
is set to release by default, to avoid using dependencies with code coverage built-in. The release buildType is also configured as the default buildType, which allows Android Studio to automatically select the correct build variant when opening a project for the first time. (b/138808399)signingConfig.debug
is used as the default signing config (b/153583269)
** Bug Fixes **
- Significantly reduced the warmup transition overhead, where the first measurement for each benchmark was artificially higher than others. This issue was more pronounced in very small benchmarks (1 microsecond or less). (b/142058671)
- Fixed
InstrumentationResultParser
error printed for each benchmark when running from command line. (I64988, b/154248456)
Known Issues
- Command line, gradle invocations of Benchmark do not print out results directly. You can work around this by either running through Studio, or parsing the JSON output file for results.
- Benchmark reporting fails to pull the report from devices that have an app installed with an applicationId ending with either “android” or “download” (case insensitive). Users hitting this issue should upgrade the Android Gradle Plugin to 4.2-alpha01 or later.
Version 1.0.0
Benchmark Version 1.0.0
November 20, 2019
androidx.benchmark:benchmark-common:1.0.0
, androidx.benchmark:benchmark-gradle-plugin:1.0.0
, and androidx.benchmark:benchmark-junit4:1.0.0
are released with no changes from 1.0.0-rc01. Version 1.0.0 contains these commits.
Major features of 1.0.0
The Benchmark library allows you to write performance benchmarks of app code and get results quickly.
It prevents build and runtime configuration issues and stabilizes device performance to ensure that measurements are accurate and consistent. Run the benchmarks directly in Android Studio, or in Continuous Integration to observe code performance over time, and to prevent regressions.
Major features include:
- Clock stabilization
- Automatic thread prioritization
- Support for UI performance testing, such as in the RecyclerView Sample
- JIT-aware warmup and looping
- JSON benchmark output for post-processing
Version 1.0.0-rc01
October 23, 2019
androidx.benchmark:benchmark-common:1.0.0-rc01
, androidx.benchmark:benchmark-gradle-plugin:1.0.0-rc01
, and androidx.benchmark:benchmark-junit4:1.0.0-rc01
are released. Version 1.0.0-rc01 contains these commits.
New features
- Added systrace tracing to benchmarks
Bug fixes
- Fixed metric instability issue where JIT wouldn't finish before warm up due to deprioritization (b/140773023)
- Unified JSON output directory across Android Gradle Plugin 3.5 and 3.6
Version 1.0.0-beta01
October 9, 2019
androidx.benchmark:benchmark-common:1.0.0-beta01
, androidx.benchmark:benchmark-gradle-plugin:1.0.0-beta01
, and androidx.benchmark:benchmark-junit4:1.0.0-beta01
are released. Version 1.0.0-beta01 contains these commits.
New features
- Run garbage collection before each warmup to reduce memory pressure from one benchmark to leak to the next (b/140895105)
Bug fixes
- Added
androidx.annotation:android-experimental-lint
dependency, so that Java code will correctly produce lint errors when experimental API is not used, similar to what is provided by the Kotlin experimental annotation for Kotlin callers. - Now correctly detects usage of
additionalTestOutputDir
instrumentation argument for output in Android Gradle Plugin 3.6, to know when AGP will handle data copy. - Fix undetected clock frequency in JSON to correctly print
-1
(b/141945670).
Version 1.0.0-alpha06
September 18, 2019
androidx.benchmark:benchmark-common:1.0.0-alpha06
, androidx.benchmark:benchmark-gradle-plugin:1.0.0-alpha06
, and androidx.benchmark:benchmark-junit4:1.0.0-alpha06
are released. Version 1.0.0-alpha06 contains these commits.
New features
- Added a check for incorrectly using the old package for the test runner, which now provides a more-helpful error message
API changes
- The experimental annotation
ExperimentalAnnotationReport
is now correctly public. Usage of the experimental BenchmarkState#report API now requires this annotation
Version 1.0.0-alpha05
September 5, 2019
androidx.benchmark:benchmark-common:1.0.0-alpha05
, androidx.benchmark:benchmark-gradle-plugin:1.0.0-alpha05
, and androidx.benchmark:benchmark-junit4:1.0.0-alpha05
are released. The commits included in this version can be found here.
API changes
BenchmarkState.reportData
API is now marked experimental
Bug fixes
- Fix for the clock-locking script, which would fail on devices that were either missing the
cut
orexpr
shell utilities. - Fixed an issue with
./gradlew lockClocks
task that would hang on devices that were rooted with an older version of the su utility, which did not support the-c
flag.
Version 1.0.0-alpha04
August 7, 2019
androidx.benchmark:benchmark-common:1.0.0-alpha04
, androidx.benchmark:benchmark-gradle-plugin:1.0.0-alpha04
, and androidx.benchmark:benchmark-junit4:1.0.0-alpha04
are released. The commits included in this version can be found here.
New documentation has also been added for how to use the Benchmark library without Gradle, both for usage with different build systems (such as Bazel or Buck), and when running in CI. For more information, see Build benchmarks without Gradle and Run benchmarks in Continuous Integration.
New features
- Gradle plugin
- Now automatically disables test coverage, and sets the
AndroidBenchmarkRunner
by default (b/138374050) - Added support for new AGP-based data copy, when running benchmarks and when using AGP 3.6+
- Now automatically disables test coverage, and sets the
- JSON format additions
- Output total benchmark test run time (b/133147694)
@Parameterized
benchmarks that use a name string (for example@Parameters(name = "size={0},depth={1}")
) now output parameter names and values per benchmark in the JSON output (b/132578772)
- Dry Run mode (b/138785848)
- Added a "dry run" mode for running each benchmark loop only once, to check for errors/crashes without capturing measurements. This can be useful e.g. for, for example, quickly running benchmarks in presubmit to check that they're not broken.
API changes
- Module structure has changed, splitting the library (b/138451391)
benchmark:benchmark-junit4
contains classes with JUnit dependency:AndroidBenchmarkRunner
, andBenchmarkRule
, both of which have moved into theandroidx.benchmark.junit4
packagebenchmark:benchmark-common
contains the rest of the logic, including the BenchmarkState API- This split will allow the library to support benchmarking without JUnit4 APIs in the future
- Configuration warnings are now treated as errors, and will crash the test (b/137653596)
- This is done to further encourage accurate measurements, especially in CI
- These errors can be reduced back to warnings with an instrumentation argument. For example:
-e androidx.benchmark.suppressErrors "DEBUGGABLE,LOW_BATTERY"
Bug fixes
- Errors when writing to external storage on Q devices provide more-descriptive messages, with suggestions of how to resolve the issue
- Screens are automatically turned on during benchmark runs, instead of failing when the screen is off
External contributions
- Thanks to Sergey Zakharov for contributing JSON output improvements and the fix for screen off issues!
Version 1.0.0-alpha03
July 2, 2019
androidx.benchmark:benchmark:1.0.0-alpha03
and androidx.benchmark:benchmark-gradle-plugin:1.0.0-alpha03
are released. The commits included in this version can be found here.
New features
- Expose sleep duration due to thermal throttling per benchmark in the full JSON report
Bug fixes
- The Gradle plugin should no longer be required to be applied after Android plugins and the Android block
- Adds support for benchmark reports on Android 10 devices using scoped storage
Version 1.0.0-alpha02
June 6, 2019
androidx.benchmark:1.0.0-alpha02
and
androidx.benchmark:benchmark-gradle-plugin:1.0.0-alpha02
are released. The
commits included in this version can be found here.
Note that we are treating the JSON schema as an API. We plan to follow the same stability constraints as other APIs: stable (with very rare exceptions) once in beta, and fixed in final release, with only additions in minor releases and changes/removals in major releases.
API changes
Overhauled JSON schema. Further changes to the JSON schema are likely to be limited to additions:
- Reorganized the result object structure to support additional metric groups in the future (b/132713021)
- Added test run context information, such as device and build info and whether clocks are locked, to the top-level object (b/132711920)
- Time metric names now have ‘ns’ in their name (b/132714527)
- Additional stats added per reported metric (maximum, median, minimum), and removed simplified 'nanos' summary stat (b/132713851)
Removed XML output (b/132714414)
Thermal throttle detection removed from
BenchmarkState.reportData
API (b/132887006)
Bug fixes
- Fixed
./gradlew lockClocks
not sticking on some recent OS devices (b/133424037) - Throttling detection disabled for emulator (b/132880807)
Version 1.0.0-alpha01
May 7, 2019
androidx.benchmark:benchmark:1.0.0-alpha01
is released. The commits included
in this version are available
here.