If you compare this to the code in the link above, you'll see that an if statement is missing.
This causes 2 issues:

1. Some threads may access data that is beyond the end of the input array.
2. The output may differ because in some iteration, given two threads A and B (where B_id = A_id + offset), B may update its value, and then A will read the new value.
   This may also explain why you saw an error when using more than 32 threads, or when the global size was larger than 4096.