Automatic Feedback-Directed Optimization (12 or higher)

Introduced in Android 12, the Android build system supports optimizing native Android modules that have blueprint build rules using Automatic Feedback-Directed Optimization (AutoFDO). AutoFDO is a sampling-based optimization technique. AutoFDO captures the runtime behavior of system binaries, enabling compilers to make better performance optimizations while also reducing binary size. AutoFDO supports collecting data from AArch32 and AArch64 architectures, although the profiles can be used across architectures.

AutoFDO is the successor for instrumentation-based Profile-Guided Optimization (PGO).

Compared to other instrumentation based profiles, AutoFDO has these additional benefits:

  • Unobtrusive data collection: AutoFDO profiles can be collected from development or user devices without any modification to build rules.

  • Real-world usage representation: AutoFDO represents real-world usage behavior assuming profiles are collected from user devices, while instrumentation PGO is representative only of the synthetic collection workload. It isn't generally straightforward to create a collection workload that perfectly matches real world usage.

AOSP ships with AutoFDO profiles for most performance-critical projects. The profiles were collected from phone and tablet devices, and are representative of their general usage patterns. The profiles are located under toolchain/pgo-profiles/sampling. AFDO is enabled by default on AOSP.

Enable AutoFDO for a blueprint build rule

To enable AutoFDO for blueprint build rules, append afdo: true to the shared library or binary rule.

Collect profiles

The prebundled profile provided with AOSP doesn't support these specific scenarios:

  • The inclusion of additional AutoFDO projects
  • The presence of locally modified code
  • Unique usage patterns associated with your system

If you have one of these scenarios, you must gather profiles directly from development or user devices.

For detailed instruction on how to prepare ARM devices for collecting AutoFDO profiles, refer to Collect ETM data for AutoFDO.

For information on how to enable automatic background profile collection, processing, and uploading, see Profcollect.

Analyze AutoFDO profiles

Android uses standard LLVM AutoFDO profiles. The AFDO profiles can be read using the LLVM's llvm-profdata tool. The script (toolchain/pgo-profiles/scripts/ automatically generates the most frequently executed functions according to the AutoFDO profile.