Intel ITANIUM ARCHITECTURE - SOFTWARE DEVELOPERS VOLUME 3 REV 2.3, Руководство пользователя, Страница: 776 / 1898

2:528 Volume 2, Part 2: MP Coherence and Synchronization
2.4.2 Simple Barrier Synchronization
A barrier is a common synchronization primitive used to hold a set of processes at a
particular point in the program (the barrier) until all processors reach the location.
Once all processes arrive at the barrier, they may all continue to execute.
Figure 2-5
shows a sense-reversing barrier synchronization based on the
fetchadd
instruction
from Hennessy and Patterson [HP96].
This type of barrier prevents a process that races ahead to the next instance of the
barrier from trapping other (slow) processors that are in the process of leaving the
barrier.
The barrier code begins by atomically updating the number of processors that are
waiting at the barrier,
count
, using a
fetchadd
instruction. For the last processor that
reaches the barrier, the
fetchadd
instruction returns the same value as the
total
shared variable, which is one less than the number of processors that wait at the
barrier. Other processors each get a unique value on the interval [0,
total
) based on
the order in which they arrive at the barrier.
All processors except the last processor wait in the
wait_on_others
loop for the signal
that all have arrived at the barrier. The last processor to arrive at the barrier provides
this signal.
The signal to leave the barrier is deduced from the value of the
release
shared variable
and the
local_sense
local variable. Upon entering the barrier, each processor
complements the value in its private
local_sense
variable. Once in the barrier, all
processors always have the same value in their
local_sense
variables. This variable
Figure 2-5. Sense-reversing Barrier Synchronization Code
// The total shared variable is one less than the number of processors
// that wait at the barrier.
// The release shared variable indicates if the processor must wait at
// the barrier (initially, this variable is 0).
// local_sense is a per-processor local variable that indicates the
// "sense" of the barrier (initially, this variable is 0).
sr_barrier:
fetchadd8.acq r1 = [count], 1 // update counter
ld8 r2 = [total] // get number of procs - 1
ld8 r3 = [local_sense] ;; // get local “sense” variable
xor r3 = 1, r3 // local_sense =! local_sense
cmp.eq p1, p2 = r1, r2;; // p1 => last proc to arrive
st8 [local_sense] = r3 // save new value of local_sense
(p1) st8 [count] = r0 // last resets count to 0
(p1) st8.rel [release] = r3 ;; // last allows other to leave
wait_on_others:
(p2)
ld8 r1 = [release] ;; // p2 => more procs to come
(p2) cmp.ne.and p0, p2 = r1, r3 // have all arrived yet?
(p2) br.cond.sptk wait_on_others ;; // nope, continue waiting
// This mf prevents memory operations that follow the barrier code
// from moving ahead of memory operations that precede the barrier
// code
mf ;;