See CMakeLists.txt and cmake/CPM.cmake as examples of setting up a simple project that uses gtensor.
Reproduce bug seen with ROCm 5.2.0, -O2, and per dim kernels off (default). On machine with ROCm installed and supported gpu, run like this:
export ROCM_PATH=/opt/rocm
export HIP_GPU_ARCHITECTURES=gfx803,gfx906,gfx908,gfx90a:xnack+,gfx1030
./run-unravel-reproducer.sh
Tested and reproduced with gfx1030 and gfx90a:xnack+.
For related gtensor code, see: