cancel
Turn on suggestions
Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

BarsMonster
Journeyman III
‎01-05-2009 07:40 PM

Branching penalty

Did anyone noticed huge branching penalty?

For 500-vliw long core, having 1 branch (success chance ~1/400'000'000, so everyone goes on same branch most of the time) reduces speed by around 10-20%. Is there anyone else noticed that?

That is quite unusual after CUDA where branches are almost free as long as all threads follow the same code path.

0 Likes
2 Replies
gaurav_garg
Adept I
‎01-06-2009 03:17 AM

That's very strange. I haven't done any experiment on this. But, documentation says Branch Granularity is same as wavefront size (64 threads).

Could you post a sample code showing this behavior?

0 Likes
udeepta
Staff
Staff
‎01-13-2009 06:47 AM

Can you post a code showing example of the behavior?

See the Stream Computing User Guide, section 1.3. Do the examples help explain the behavior?

0 Likes