Open Software Stack for Compression-Aware Adaptive Edge Offloading

Edge offloading of semantic segmentation with dynamically selectable compression and local/remote processing. Includes remote server discovery.

This repository was used to produce results for the paper Jakub Žádník, Robin Bijl, Jan Solanti, Erno Joensuu, Markku Mäkitalo, Pekka Jääskeläinen, "Open Software Stack for Compression-Aware Adaptive Edge Offloading", in Proc. WCNC 2025.

Cloning the Repository

clone the repo with its submodules with the following command:

git clone --recursive-submodules [email protected]:cs/cpc/aisa-demo.git

or initialize submodules after cloning with:

git submodule update --init --recursive

Relevant Versions

Android constantly updates packages and libraries that can break with newer versions. Below is a list of relevent versions currently used:

What	Version	where to check
Android Studio	Electric Eel (2022.1.1)	file > settings > appearance & behavior > system settings > updates
JDK	17	file > project structure > SDK Location > gradle settings > gradle JDK
Minimum SDK version	24	android/app/build.gradle > minSdk
Targeted SDK version	33	android/app/build.gradle > targetSdk
Gradle	8.0	file > project structure > project > gradle version
Android Gradle plugin version	8.0.2	file > project structure > project > android gradle plugin version
CMake	3.22.1	android/app/src/main/cpp/CMakeLists.txt > cmake_minimum_required
NDK	23.1.7779620	(android layout in project tab) app/cpp/inlcudes/NDK
C++	17	android/app/build.gradle > cppFlags '-std=c++<version>'

Installation and Requirements

The smartphone application can be found in the android/ directory. Open and install it with Android Studio. You should be able to run the semantic segmentation locally. For remote inference, you need to build the server separately (see below) and add its IP address into the application.

The project requires installation of libraries inside the external directory. The Tracy profiler in external/tracy is not required unless you plan to profile the application.

pocl

This project depends on a custom fork of an OpeCL implementation pocl, included in the external/pocl directory. To build the server, use the attached build_pocl_remote.sh script (make sure to modify the paths as necessary). To run the server, use run_pocl_remote.sh (again adapting paths as necessary). When the server is running, it is possible to offload computation to it by adding its IP address to the smartphone application.

Check that phone exposes the OpenCL library

Some phones provide a OpenCL library that can be used in C. This library needs to be whitelisted by the vendor. This can be checked like so:

1. adb into the phone
2. run:

   cat /vendor/etc/public.libraries.txt  
   

   and check that libOpenCL.so is there

Android Studio

1. install Android Studio
2. install the android sdk from Tools > SDK Manager > SDK Platforms
3. install the android Native Development Kit (NDK) from Tools > SDK Manager > SDK Tools
4. install CMake from Tools > SDK Manager > SDK Tools

OpenCV

follow the instructions in the scripts/build_opencv.bash script

Environment variables

In order to build and run OpenCV and BiK, it is recommended to have the following variables set:

CUDA_VERSION='11.7'
CUDADIR="/usr/local/cuda-11.7"
CUDA_PATH="/usr/local/cuda-11.7"
CUDA_INSTALL_PATH="/usr/local/cuda-11.7"
LD_LIBRARY_PATH=/usr/local/cuda-11.7/lib64:/lib/x86_64-linux-gnu

Of these LD_LIBRARY_PATH is the most important.

Development

While developing, there are a number of logging and debug variables that can be set. For poclImageProcessor.ccp, "PRINT_PROFILE_TIME" can be defined. On the java side, the DevelopmentVariables.java contains variables that make the program more verbose

Common trouble shootingshooting

Can not find platform -1001 error

This is can happen after installing cuda-11-7. The ICD loader grabs the libopencl.so library that cuda-11.7 provides instead of the one from ocl-icd-libopencl1. the easiest way to fix this is by moving the libOpenCL.so* files in /usr/local/cuda-11/lib64 into a new directory so that they won't be found.

clinfo can not find any platforms when using OCL_ICD_VENDORS

This is the same issue as the previous troubleshooting case.

CL Error -1 (CL_DEVICE_NOT_FOUND) while setting POCL_DEVICES="basic"

Add POCL_BUILDING=1 as a parameter as well.

OpenCV error while loading model

If you get an error like this:

[ INFO:[email protected]] global onnx_importer.cpp:595 populateNet DNN/ONNX: loading ONNX v7 model produced by 'pytorch':2.0.0. Number of nodes = 297, inputs = 1, outputs = 2 OpenCV(4.7.0) Error: Assertion failed (blob_0.size == blob_1.size) in parseBias, file ./modules/dnn/src/onnx/onnx_importer.cpp, line 1067

make sure you built OpenCV and PoCL while having the LD_LIBRARY_PATH environment variable set to contain:

/usr/local/cuda-11.7/lib64:/lib/x86_64-linux-gnu

in that order

Development notes

• Use CPP files with extern "C" as the JNIEnv is different between C and CPP and the editer will have an issue with it otherwise.

License

Apache 2.0 license