build instructions:

1. bash scripts/configure.sh
2. run/profile hip: bash scripts/hip_build_and_run.sh / bash scripts/hip_build_and_profile.sh
3. run/profile cuda: bash scripts/hip_build_and_run.sh / bash scripts/hip_build_and_profile.sh

tests

1. mem_transfers.cpp, for comparing speed of device to host memcpy of small amount of bytes
2. pass_writes.cpp

last test is related to integrator_shade_shadow (intern/cycles/kernel/integrator/shade_shadow.h). it is 3 times slower for hip. I see that it happens because of film_write_direct_light (https://github.com/blender/blender/blob/main/intern/cycles/kernel/film/light_passes.h#L461) this function just writes values to passes with atomicAdd