The barrier runs at about 0.42 us per call


this results agrees with one post from Daniel B (brokensh) from IBM
(note the time about PPE-SPE mailbox barrier is essencitally the same as the lantecy test in lantency/ with multiple spes
6700 cycels agrees well with the ~2.2us latency with 8 spes)

http://www.ibm.com/developerworks/forums/thread.jspa?threadID=153049&start=15&tstart=0

brokensh   

Posts: 223 
Registered: Nov 08, 2005 05:01:44 PM  
  Re: Barrier 
Posted: Apr 12, 2007 11:00:21 AM    in response to: vishwasbc's post         Reply  
 
Vishwas,

I don't see the code that you were referring to. You are achieving considerably different results than expected or I observe.

For PPE-SPE mailbox barrier synchronization (case 1) using direct problem state) in which 8 SPEs are participating in the barrier, I measure ~6700 cycles. This number improves to ~4690 if I assume that no possible inbound mailbox overflow will occur and not check the mailbox status before writing. 

I have an modified implementation of case 3 (SPE-SPE signotify) that achieves an average of ~1320 cycles. For my (inferior) implementation, I configure all the SPEs signotify registers in OR mode and each broadcasts its assigned bit to all the other SPE's signal notification registers and then waits for the all the bits of its signotify register to be set.

Dan B.

I would really like to get to the bottom of this discrepancy of results.