openmpllvmoffloading

Using OpenMP target offloading in llvm-8.0.0

When trying to use openmp target offloading with llvm I get the following error

$ cat offload.cpp
#include <omp.h>
int main() {
#pragma omp target teams distribute parallel for
    for(int i=0; i<100; i++);
     return 0;
}
$ clang++ -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda --cuda-path=$CUDA_TOOLKIT_ROOT_DIR offload.cpp -o offload
$ ./offload
Libomptarget fatal error 1: default offloading policy must switched to mandatory or disabled
$

.

I have llvm-8.0.0 installed in my system at LLVM_PATH. I downloaded openmp-8.0.0 source from the llvm download page. To build openmp I used the following command:

$ mkdir build && cd build
$ cmake -DCMAKE_BUILD_TYPE=Debug -DCMAKE_INSTALL_PREFIX=$LLVM_PATH -DCMAKE_C_COMPILER=$LLVM_PATH/bin/clang -DCMAKE_CXX_COMPILER= $LLVM_PATH/bin/clang++ -DLIBOMPTARGET_NVPTX_COMPUTE_CAPABILITIES=35,60,70 ..
$ make && make install

OpenMP got built without giving any error. But when trying to use target offloading with OpenMP I get the above mentioned fatal error. I tried this on 3 different machines, with same result.

I am able to build and run CUDA examples on my system with nvcc.

Regards,

Alok

Solution

  • I found the answer. To make sure we compile for offloading we need to use the flags:

    -Xopenmp-target -march=sm_XX

    where XX is the compute capability. So for me

    clang++ -Xopenmp-target -march=sm_35 -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda --cuda-path=$CUDA_TOOLKIT_ROOT_DIR offload.cpp -o offload

    worked