Re: Results differs for GPU and CPU devices depending on data.
I recorded the new screencast, now it about integer data. As you can see, there is no dependence from data type. Source files also in attachment (it little differs from float sources.)
Re: Results differs for GPU and CPU devices depending on data.
I manually unrolled main loop and replaced async_work_group_copy to simple copying inside single tread. Results are the same. This unrolled and more commented version of test_prop.cl in attachment.
Re: Results differs for GPU and CPU devices depending on data.
In VS2013 I also saw warnings about calloc/malloc, but after including malloc.h it gone. Which error you see?
In attached video my new experiment - instead last propagation I simple copy data from layer 1 to local buffer and from buffer to layer 2. In straight order it shows from 0 to 6 errors on Devstator, with reverse order it also appended 63 errors - seems like it early accessed global data that is waiting computation and have not yet filled. And the same reason is why it sometimes got bad data from 1st propagation, sometimes not - sometimes it have time to finish, sometimes not. There is a question - how to properly sync computation for guaranteed sharing data between work-items? Global barriers not helps in this case.
