OK, I think it simply gives up after opencl_mandatory_timeout.
The sleep time is 5 ms (usec = 5000), my opencl_mandatory_timeout is 200, so it gives up after 1 second, if I understand correctly.
I could have seen it here:
Note the 1 second gap between the two.
Raising the loop count (opencl_mandatory_timeout) so it allows up to 100 s wait results in GPU-only processing:
47.088374 [pixelpipe_process] [full] using device 0
...
54.892174 [opencl_profiling] profiling device 0 ('NVIDIA GeForce GTX 1060 6GB'):
54.892212 [opencl_profiling] spent 7.7388 seconds totally in command queue (with 0 events missing)
54.892535 [dev_process_image] pixel pipeline processing took 7.804 secs (5.978 CPU)
54.893041 [pixelpipe_process] [preview] using device 0
...
63.725147 [opencl_profiling] spent 8.7267 seconds totally in command queue (with 0 events missing)
63.725630 [dev_process_preview] pixel pipeline processing took 16.661 secs (13.020 CPU)
63.750731 [histogram] took 0.004 secs (0.004 CPU) scope draw