1. on Cori GPU, load necessary modules
source ./load_cgpu_modules.sh 

2. run bootstrap scrip
./bootstrap.sh

3. Create a build directory, for example, 
mkdir build-cgpu

3. run configure script in the build directory 
cd build-cgpu
sh config-command

Example config-command for single-GPU omp offload: 

../configure \
    --enable-comms=none \
    --enable-simd=GEN \
    --enable-gen-simd-width=16 \
    CXX=clang++ \
    LDFLAGS="-L${CUDA_ROOT}/lib64 -lcudart" \
    CXXFLAGS="-Wno-unknown-cuda-version -I${CUDA_ROOT}/include -fopenmp -std=c++14 -fopenmp-cuda-mode  -O3 -g -fopenmp-targets=nvptx64-nvidia-cuda -Wformat -DOMPTARGET -DOMPTARGET_MANAGED"

4. compile 
make -j8