Skip to content

README.md

Latest commit

 

History

History
100 lines (80 loc) · 2.85 KB

README.md

File metadata and controls

100 lines (80 loc) · 2.85 KB

C++/Halide implementation

This version in this directory allows the benchmarking and accelerated execution in multiple targets (desktop and mobile). It contains the four algorithms that are compared in the paper:

  • PDRT 2
  • PDRT 32
  • PS DRT
  • MDD DRT

Setup

Download with Halide 15.0.0 from the releases page. And put extract it in a directory of your choice $HALIDE_DIR.

Host (x86 desktop CPU)

Building

These instructions are for building the executable on the desktop CPU for benchmarking and checking results.

mkdir build
cd build
export HALIDE_DIR=<YOUR_HALIDE_INSTALLATION_PATH>
cmake -DHalide_DIR=$HALIDE_DIR/lib/cmake/Halide ..
make barcode_segmentation_host

Running

cd host
./barcode_segmentation_host

This will output the benchmarked times of the four algorithms and save the outputs in outputs.

Building a dynamic library for Python

For convenience, the algorithms can be compiled into a dynamic library that can be called from python. For this run:

mkdir build
cd build
export HALIDE_DIR=<YOUR_HALIDE_INSTALLATION_PATH>
cmake -DHalide_DIR=$HALIDE_DIR/lib/cmake/Halide ..
make barcode_segmentation_lib

Android (CPU)

These instructions are for building the executable on an Android CPU for benchmarking and checking results.

Setup

You'll need ADB available in your path. An Android toolchain is also needed. You can use the one that comes within the Android NDK that can be installed by the SDK manager of Android Studio. It is usually located in ~/Android/Sdk/ndk/<ndk_version>. I'll refer to this path as <NDK_PATH>.

Build libpng

git clone https://github.com/glennrp/libpng
cd libpng
mkdir build 
cd build
export NDK_PATH=<YOUR_NDK_INSTALLATION_PATH>
cmake \
   -DCMAKE_TOOLCHAIN_FILE=$NDK_PATH/build/cmake/android.toolchain.cmake \
   -DANDROID_ABI=arm64-v8a\
   ..
make

Build libjpeg

git clone https://github.com/libjpeg-turbo/libjpeg-turbo
cd libjpeg-turbo
mkdir build 
cd build
export NDK_PATH=<YOUR_NDK_INSTALLATION_PATH>
cmake \
   -DCMAKE_TOOLCHAIN_FILE=$NDK_PATH/build/cmake/android.toolchain.cmake \
   -DANDROID_ABI=arm64-v8a\
   ..
make

Building

# Omit these three lines if already configured previously for host
mkdir build
cd build
export HALIDE_DIR=<YOUR_HALIDE_INSTALLATION_PATH>
cmake -DHalide_DIR=$HALIDE_DIR/lib/cmake/Halide ..

export NDK_PATH=<YOUR_NDK_INSTALLATION_PATH>
export LIBPNG=<YOUR_LIBPNG_INSTALLATION_PATH>
export LIBJPEG=<YOUR_LIBJPEG_INSTALLATION_PATH>

cd arm64-android
CPP=$NDK_PATH/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android21-clang++ \
    LIB_PNG_LOCATION=$LIBPNG \
    LIB_JPEG_LOCATION=$LIBJPEG \
    make build_and_run_android_executable

This will output the benchmarked times of the four algorithms and save the output in outputs.