README_OACC.md

Building configuring for the OpenACC backend

CMake Basics

Alpaka requires cmake version >= 3.15.

In the root of the alpaka dir, run:

mkdir build
cd build

Configuring Using CMake

In the build directory, invoke cmake to configure. Use the options below to enable only the OpenACC backend.

cmake .. \
  -DALPAKA_ACC_ANY_BT_OACC_ENABLE=on \
  -DBUILD_TESTING=on \
  -Dalpaka_BUILD_EXAMPLES=on \

All other backends are disabled for faster compilation/testing and reduced environment requirements. The cmake package OpenACC is used to detect the required OpenACC flags for the compiler. Additional flags can be added, e.g:

• gcc, target x86:

    -DCMAKE_CXX_FLAGS="-foffload=disable"

  As of gcc 9.2 no test will compile if the nvptx backend is enabled. If cmake fails to set the -fopenacc flag, it can be set manually.
• pgi/nvhpc, target tesla (set $CC, $CXX and $CUDA_HOME to appropriate values for your system to use pgi):

    -DCMAKE_CXX_FLAGS="-ta=tesla -Minfo"

Limitations

• No separabel compilation. OpenACC requires functions for which device code should be generated for a not-inlined call in a target region to be marked with pragmas. This cannot be wrapped by macros like ALPAKA_FN_DEVICE because they appear between template parameter list and function name. alpaka-group#1126 (comment)

Test targets

helloWorld

make helloWorld
./examples/helloWorld/helloWorld

The output should end with something like

[z:3, y:7, x:15][linear:511] Hello World

Numbers can vary when teams are executed in parallel: 512 teams, with one worker each are started in a 3d grid. Each worker reports its grid coordinates and linear index.

compiler	compile status	target	run status
GCC 10	ok	x86	ok
NVHPC 20.7	ok	tesla	ok

vectorAdd

make vectorAdd
./examples/vectorAdd/vectorAdd

The output should end with

Execution results correct!

compiler	compile status	target	run status
GCC 10(dev)	ok	x86	ok
NVHPC 20.7	ok	tesla	ok

Building and Running all tests

make
ctest