This is an Android app that profiles the Volk library and selects the best implementations for the architecture, similar to volk_profile that comes as part of Volk. It is configured to run in benchmark mode, i.e., it will profile all kernels, even the ones without alternative implementations. Running the application generates two files:
volk_results.jsonwith benchmark resultsvolk_configwith the best implementations for the architecture
Files are stored on External Storage (which might be an SD card or some internal storage). The directory can be accessed when the phone is connected to a PC and is available to all applications on the phone that have the corresponding permission. This way apps that use Volk (like GNU Radio Android apps) can use it to select the implementations that perform best.
Volk comes with a basic evaluation script for the JSON file.
Building the app requires the GNU Radio Android toolchain. Please see this repository for further instructions on how to build the toolchain and apps that use it.
The app is minimalistic. There is only a button to start profiling and a textbox that indicates when it is done. You can use the logcat output to monitor progress.
The following plot shows the speedup of unaligned implementations on a OnePlus 5T smartphone.
If you use this project, we would appreciate a reference to:
-
Bastian Bloessl, Lars Baumgärtner and Matthias Hollick, “Hardware-Accelerated Real-Time Stream Data Processing on Android with GNU Radio,” Proceedings of 14th International Workshop on Wireless Network Testbeds, Experimental evaluation & Characterization (WiNTECH’20), London, UK, September 2020. [DOI, BibTeX, PDF and Details…]
