HIPAcc

Getting Started

Requirements

HIPAcc is a source-to-source compiler based on Clang/LLVM. To build HIPAcc from source, either the development package of Clang/LLVM and libc++ is required, or the corresponding components from the Clang/LLVM have to be installed manually:

HIPAcc works only with the release version of Clang/LLVM listed in dependencies.sh, which corresponds to the current Clang/LLVM 3.8 release version.

Windows
While HIPAcc is mainly developed under GNU Linux and macOS, it compiles and runs under Windows as well. Use msys2 to compile HIPAcc on Windows. Make sure that development packages are installed. The source files generated by HIPAcc can then be compiled in Visual Studio.

Installation

The build system of HIPAcc uses CMake (2.8.12 or later) and requires installation to an absolute path:

          cd hipacc
          mkdir build && cd build
          cmake ../ -DCMAKE_INSTALL_PREFIX=`pwd`/release
          make && make install
          
Note that Xcode and the command line tools are required to build HIPAcc on macOS (use xcode-select --install to install command line tools).

A list of CMake options that are used by HIPAcc to find headers and libraries for its various back ends (CUDA, OpenCL, and Renderscript) can be found in the INSTALL file.

First Run

After building and installing HIPAcc successfully, you are ready to use HIPAcc. To verify that everything is working, generate and execute a program using HIPAcc. Go to your installation directory and call make cuda. This translates the source file specified by the TEST_CASE environment variable into a version that runs on the GPU using CUDA and executes it. Similarly, the same program can be translated to OpenCL for different device types (opencl-acc, opencl-cpu, opencl-gpu) as well as standard C++ code (cpu). For execution on Android, Renderscript (renderscript) and Filterscript (filterscript) code can be generated.

          cd hipacc/build/release
          TEST_CASE=tests/opencv_gaussian_8uc1 make cuda
          ./main_cuda
          

For execution on Android devices, the generated binary is configured to store data in /data/local/tmp. The same directory can be used for the binary itself:

          adb shell mkdir -p /data/local/tmp
          adb push main_renderscript /data/local/tmp
          adb shell /data/local/tmp/main_renderscript