On 12/28/2012 11:04 AM, Mark Zhang wrote: > On 12/28/2012 04:50 PM, Arto Merilainen wrote: >> >> In my opinion asking tegra_stream_begin() to put a bad fence into the >> stream is a case we should never be. assert() kills the application >> immediately (in debug builds) and usually this helps the programmer for >> 1) finding bugs 2) not doing bad code. >> > > Yep, I agree. But in release builds, assert does nothing. So this > checking doesn't make sense and also a wrong fence will be pushed into > command buffer silently. And we always use release version in real > products, so we can't count on this "assert".
The only pro of using assert is low (=no :-) ) overhead in release builds. > >> "Silencing" is not a good solution especially in this case: >> tegra_stream_flush() returns an invalid fence when flushing fails. If >> the application chains submits (i.e. do a blit and then do another using >> the output of the first blit) it is crucial to be sure the first submit >> has been performed before starting the second one. >> > > Yes. So I suggest doing fence checking at the beginning of the > "tegra_stream_begin", if invalid fence found, return an error. > This sounds reasonable. - Arto
