--- Quote Start ---
If you are using num_compute_units to replicate the kernel, and your accesses are random in a way that two different work-items from different work-groups might try to read/write the same memory location, then yes, this is certainly possible. Without num_compue_units, this shouldn't happen unless there is some race condition in the kernel itself (i.e. incorrect code which would also give incorrect results on CPU/GPU).
--- Quote End ---
I am not using num_compute_units and I don't know what it is so I guess that is not a problem. So running that kernel sequentially or in parallel should not be a problem I guess. I'll write here If I find something else. Thanks for the help.
Honored Contributor
