/*
* Copyright 2021 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package androidx.benchmark.macro
import android.content.pm.ApplicationInfo
import android.content.pm.ApplicationInfo.FLAG_DEBUGGABLE
import android.content.pm.ApplicationInfo.FLAG_SYSTEM
import android.content.pm.ApplicationInfo.FLAG_UPDATED_SYSTEM_APP
import android.content.pm.PackageManager
import android.os.Build
import android.util.Log
import androidx.annotation.RestrictTo
import androidx.benchmark.Arguments
import androidx.benchmark.ConfigurationError
import androidx.benchmark.DeviceInfo
import androidx.benchmark.InstrumentationResults
import androidx.benchmark.Profiler
import androidx.benchmark.ResultWriter
import androidx.benchmark.Shell
import androidx.benchmark.checkAndGetSuppressionState
import androidx.benchmark.conditionalError
import androidx.benchmark.inMemoryTrace
import androidx.benchmark.json.BenchmarkData
import androidx.benchmark.perfetto.PerfettoCapture.PerfettoSdkConfig
import androidx.benchmark.perfetto.PerfettoCapture.PerfettoSdkConfig.InitialProcessState
import androidx.benchmark.perfetto.PerfettoCaptureWrapper
import androidx.benchmark.perfetto.PerfettoConfig
import androidx.benchmark.perfetto.PerfettoTrace
import androidx.benchmark.perfetto.PerfettoTraceProcessor
import androidx.benchmark.perfetto.UiState
import androidx.benchmark.perfetto.appendUiState
import androidx.test.platform.app.InstrumentationRegistry
import androidx.tracing.trace
import java.io.File
/**
* Get package ApplicationInfo, throw if not found.
*/
@RestrictTo(RestrictTo.Scope.LIBRARY_GROUP)
@Suppress("DEPRECATION")
fun getInstalledPackageInfo(packageName: String): ApplicationInfo {
val pm = InstrumentationRegistry.getInstrumentation().context.packageManager
try {
return pm.getApplicationInfo(packageName, 0)
} catch (notFoundException: PackageManager.NameNotFoundException) {
throw AssertionError(
"Unable to find target package $packageName, is it installed?",
notFoundException
)
}
}
/**
* @return `true` if the [ApplicationInfo] instance is referring to a system app.
*/
@RestrictTo(RestrictTo.Scope.LIBRARY_GROUP)
fun ApplicationInfo.isSystemApp(): Boolean {
return flags and (FLAG_SYSTEM or FLAG_UPDATED_SYSTEM_APP) > 0
}
internal fun checkErrors(packageName: String): ConfigurationError.SuppressionState? {
Arguments.throwIfError()
val applicationInfo = getInstalledPackageInfo(packageName)
val instrumentation = InstrumentationRegistry.getInstrumentation()
val errors = DeviceInfo.errors +
// TODO: Merge this debuggable check / definition with Errors.kt in benchmark-common
listOfNotNull(
conditionalError(
hasError = applicationInfo.flags.and(FLAG_DEBUGGABLE) != 0,
id = "DEBUGGABLE",
summary = "Benchmark Target is Debuggable",
message = """
Target package $packageName is running with debuggable=true in its manifest,
which drastically reduces runtime performance in order to support debugging
features. Run benchmarks with debuggable=false. Debuggable affects execution
speed in ways that mean benchmark improvements might not carry over to a
real user's experience (or even regress release performance).
""".trimIndent()
),
conditionalError(
// Profileable is currently only needed on API 29+30, since app trace tag no longer
// requires profileable on API 31, and macrobench doesn't currently offer other
// means of profiling (like simpleperf) that need the flag.
hasError = DeviceInfo.profileableEnforced &&
Build.VERSION.SDK_INT in 29..30 &&
applicationInfo.isNotProfileableByShell(),
id = "NOT-PROFILEABLE",
summary = "Benchmark Target is NOT profileable",
message = """
Target package $packageName is running without <profileable shell=true>.
Profileable is required on Android 10 & 11 to enable macrobenchmark to capture
detailed trace information from the target process, such as System tracing
sections defined in the app, or libraries.
To make the target profileable, add the following in your target app's
main AndroidManifest.xml, within the application tag:
<!--suppress AndroidElementNotAllowed -->
<profileable android:shell="true"/>
""".trimIndent()
),
conditionalError(
hasError = instrumentation.targetContext.packageName !=
instrumentation.context.packageName,
id = "NOT-SELF-INSTRUMENTING",
summary = "Benchmark manifest is instrumenting separate process",
message = """
Macrobenchmark instrumentation target in manifest
${instrumentation.targetContext.packageName} does not match macrobenchmark
package ${instrumentation.context.packageName}. While macrobenchmarks 'target' a
separate app they measure, they can not declare it as their instrumentation
targetPackage in their manifest. Doing so would cause the macrobenchmark test
app to be loaded into the target application process, which would prevent
macrobenchmark from killing, compiling, or launching the target process.
Ensure your macrobenchmark test apk's manifest matches the manifest package, and
instrumentation target package, also called 'self-instrumenting':
<manifest
package="com.mymacrobenchpackage" ...>
<instrumentation
android:name="androidx.benchmark.junit4.AndroidBenchmarkRunner"
android:targetPackage="mymacrobenchpackage"/>
In gradle library modules, this is the default behavior. In gradle test modules,
specify the experimental self-instrumenting property:
android {
targetProjectPath = ":app"
// Enable the benchmark to run separately from the app process
experimentalProperties["android.experimental.self-instrumenting"] = true
}
""".trimIndent()
),
conditionalError(
hasError = DeviceInfo.misconfiguredForTracing,
id = "DEVICE-TRACING-MISCONFIGURED",
summary = "This ${DeviceInfo.typeLabel}'s OS is misconfigured for tracing",
message = """
This ${DeviceInfo.typeLabel}'s OS image has not correctly mounted the tracing
file system, which prevents macrobenchmarking, and Perfetto/atrace trace capture
in general. You can try a different device, or experiment with an emulator
(though that will not give timing measurements representative of real device
experience).
This error may not be suppressed.
""".trimIndent()
),
conditionalError(
hasError = Arguments.macrobenchMethodTracingEnabled(),
id = "METHOD-TRACING-ENABLED",
summary = "Method tracing is enabled during a Macrobenchmark",
message = """
The Macrobenchmark run for $packageName has method tracing enabled.
This causes the VM to run more slowly than usual, so the metrics from the
trace files should only be considered in relative terms
(e.g. was run #1 faster than run #2). Also, these metrics cannot be compared
with benchmark runs that don't have method tracing enabled.
""".trimIndent()
),
).sortedBy { it.id }
// These error ids are really warnings. In that, we don't need developers to have to
// explicitly suppress them using test instrumentation arguments.
// TODO: Introduce a better way to surface warnings.
val alwaysSuppressed = setOf("METHOD-TRACING-ENABLED")
val neverSuppressed = setOf("DEVICE-TRACING-MISCONFIGURED")
return errors.checkAndGetSuppressionState(
Arguments.suppressedErrors + alwaysSuppressed - neverSuppressed
)
}
/**
* macrobenchmark test entrypoint, which doesn't depend on JUnit.
*
* This function is a building block for public testing APIs
*/
private fun macrobenchmark(
uniqueName: String,
className: String,
testName: String,
packageName: String,
metrics: List<Metric>,
compilationMode: CompilationMode,
iterations: Int,
launchWithClearTask: Boolean,
startupModeMetricHint: StartupMode?,
perfettoSdkConfig: PerfettoSdkConfig?,
setupBlock: MacrobenchmarkScope.() -> Unit,
measureBlock: MacrobenchmarkScope.() -> Unit
) {
require(iterations > 0) {
"Require iterations > 0 (iterations = $iterations)"
}
require(metrics.isNotEmpty()) {
"Empty list of metrics passed to metrics param, must pass at least one Metric"
}
val suppressionState = checkErrors(packageName)
var warningMessage = suppressionState?.warningMessage ?: ""
// skip benchmark if not supported by vm settings
compilationMode.assumeSupportedWithVmSettings()
val startTime = System.nanoTime()
// Ensure method tracing is explicitly enabled and that we are not running in dry run mode.
val launchWithMethodTracing = Arguments.macrobenchMethodTracingEnabled()
val applicationInfo = getInstalledPackageInfo(packageName)
val scope = MacrobenchmarkScope(
packageName,
launchWithClearTask = launchWithClearTask
)
// Capture if the app being benchmarked is a system app.
scope.isSystemApp = applicationInfo.isSystemApp()
scope.launchWithMethodTracing = launchWithMethodTracing
// Ensure the device is awake
scope.device.wakeUp()
// Stop Background Dexopt during a Macrobenchmark to improve stability.
if (Build.VERSION.SDK_INT >= 33) {
scope.cancelBackgroundDexopt()
}
// Always kill the process at beginning of test
scope.killProcess()
inMemoryTrace("compile $packageName") {
compilationMode.resetAndCompile(scope) {
setupBlock(scope)
measureBlock(scope)
}
}
// package name for macrobench process, so it's captured as well
val macrobenchPackageName = InstrumentationRegistry.getInstrumentation().context.packageName
// Perfetto collector is separate from metrics, so we can control file
// output, and give it different (test-wide) lifecycle
val perfettoCollector = PerfettoCaptureWrapper()
val tracePaths = mutableListOf<String>()
val methodTracingResultFiles = mutableListOf<Profiler.ResultFile>()
try {
metrics.forEach {
it.configure(packageName)
}
val measurements = PerfettoTraceProcessor.runServer {
val runIterations = if (Arguments.dryRunMode) 1 else iterations
List(runIterations) { iteration ->
// Wake the device to ensure it stays awake with large iteration count
inMemoryTrace("wake device") {
scope.device.wakeUp()
}
scope.iteration = iteration
inMemoryTrace("setupBlock") {
setupBlock(scope)
}
val iterString = iteration.toString().padStart(3, '0')
val fileLabel = "${uniqueName}_iter$iterString"
val tracePath = perfettoCollector.record(
fileLabel = fileLabel,
config = PerfettoConfig.Benchmark(
/**
* Prior to API 24, every package name was joined into a single setprop
* which can overflow, and disable *ALL* app level tracing.
*
* For safety here, we only trace the macrobench package on newer platforms,
* and use reflection in the macrobench test process to trace important
* sections
*
* @see androidx.benchmark.macro.perfetto.ForceTracing
*/
appTagPackages = if (Build.VERSION.SDK_INT >= 24) {
listOf(packageName, macrobenchPackageName)
} else {
listOf(packageName)
},
useStackSamplingConfig = true
),
perfettoSdkConfig = perfettoSdkConfig,
inMemoryTracingLabel = "Macrobenchmark"
) {
try {
trace("start metrics") {
metrics.forEach {
it.start()
}
}
trace("measureBlock") {
measureBlock(scope)
}
} finally {
trace("stop metrics") {
metrics.forEach {
it.stop()
}
if (launchWithMethodTracing && scope.isMethodTracing) {
val (label, tracePath) = scope.stopMethodTracing(fileLabel)
val resultFile = Profiler.ResultFile.ofPerfettoTrace(
label = label,
absolutePath = tracePath
)
methodTracingResultFiles += resultFile
scope.isMethodTracing = false
}
}
}
}!!
tracePaths.add(tracePath)
val measurementList = loadTrace(PerfettoTrace(tracePath)) {
// Extracts the metrics using the perfetto trace processor
inMemoryTrace("extract metrics") {
metrics
// capture list of Measurements
.map {
it.getMeasurements(
Metric.CaptureInfo(
targetPackageName = packageName,
testPackageName = macrobenchPackageName,
startupMode = startupModeMetricHint,
apiLevel = Build.VERSION.SDK_INT
),
this
)
}
// merge together
.reduce { sum, element -> sum.merge(element) }
}
}
// append UI state to trace, so tools opening trace will highlight relevant part in UI
val uiState = UiState(
highlightPackage = packageName
)
File(tracePath).apply {
appendUiState(uiState)
}
Log.d(TAG, "Iteration $iteration captured $uiState")
// report just the metrics
measurementList
}.mergeMultiIterResults()
}
require(measurements.isNotEmpty()) {
"""
Unable to read any metrics during benchmark (metric list: $metrics).
Check that you're performing the operations to be measured. For example, if
using StartupTimingMetric, are you starting an activity for the specified package
in the measure block?
""".trimIndent()
}
InstrumentationResults.instrumentationReport {
if (launchWithMethodTracing && methodTracingResultFiles.size < iterations) {
warningMessage += "\nNOTE: Method traces cannot be captured during iterations" +
" that start while the target process is already running (including HOT/WARM" +
" launches)."
}
reportSummaryToIde(
warningMessage = warningMessage,
testName = uniqueName,
measurements = measurements,
iterationTracePaths = tracePaths,
profilerResults = methodTracingResultFiles
)
warningMessage = "" // warning only printed once
measurements.singleMetrics.forEach {
it.putInBundle(bundle, suppressionState?.prefix ?: "")
}
measurements.sampledMetrics.forEach {
it.putPercentilesInBundle(bundle, suppressionState?.prefix ?: "")
}
}
val warmupIterations = when (compilationMode) {
is CompilationMode.Partial -> compilationMode.warmupIterations
else -> 0
}
val mergedProfilerOutputs = (tracePaths.mapIndexed { index, it ->
Profiler.ResultFile.ofPerfettoTrace(
label = "Trace Iteration $index",
absolutePath = it
)
} + methodTracingResultFiles).map {
BenchmarkData.TestResult.ProfilerOutput(it)
}
ResultWriter.appendTestResult(
BenchmarkData.TestResult(
className = className,
name = testName,
totalRunTimeNs = System.nanoTime() - startTime,
metrics = measurements.singleMetrics + measurements.sampledMetrics,
repeatIterations = iterations,
thermalThrottleSleepSeconds = 0,
warmupIterations = warmupIterations,
profilerOutputs = mergedProfilerOutputs
)
)
} finally {
scope.killProcess()
}
}
/**
* Run a macrobenchmark with the specified StartupMode
*/
@RestrictTo(RestrictTo.Scope.LIBRARY_GROUP)
fun macrobenchmarkWithStartupMode(
uniqueName: String,
className: String,
testName: String,
packageName: String,
metrics: List<Metric>,
compilationMode: CompilationMode,
iterations: Int,
startupMode: StartupMode?,
setupBlock: MacrobenchmarkScope.() -> Unit,
measureBlock: MacrobenchmarkScope.() -> Unit
) {
val perfettoSdkConfig =
if (Arguments.perfettoSdkTracingEnable && Build.VERSION.SDK_INT >= Build.VERSION_CODES.R) {
PerfettoSdkConfig(
packageName,
when (startupMode) {
null -> InitialProcessState.Unknown
StartupMode.COLD -> InitialProcessState.NotAlive
StartupMode.HOT, StartupMode.WARM -> InitialProcessState.Alive
}
)
} else null
macrobenchmark(
uniqueName = uniqueName,
className = className,
testName = testName,
packageName = packageName,
metrics = metrics,
compilationMode = compilationMode,
iterations = iterations,
startupModeMetricHint = startupMode,
perfettoSdkConfig = perfettoSdkConfig,
setupBlock = {
if (startupMode == StartupMode.COLD) {
// Run setup before killing process
setupBlock(this)
// Shader caches are stored in the code cache directory. Make sure that
// they are cleared every iteration. Must be done before kill, since on user builds
// this broadcasts to the target app
dropShaderCache()
// Kill - code below must not wake process!
killProcess()
// Ensure app's pages are not cached in memory for a true _cold_ start.
dropKernelPageCache()
// validate process is not running just before returning
check(!Shell.isPackageAlive(packageName)) {
"Package $packageName must not be running prior to cold start!"
}
} else {
if (iteration == 0 && startupMode != null) {
try {
iteration = null // override to null for warmup
// warmup process by running the measure block once unmeasured
setupBlock(this)
measureBlock()
} finally {
iteration = 0 // resume counting
}
}
setupBlock(this)
}
},
// Don't reuse activities by default in COLD / WARM
launchWithClearTask = startupMode == StartupMode.COLD || startupMode == StartupMode.WARM,
measureBlock = measureBlock
)
}