 Andra Hugo
|
46367259cb
refatctoring to use j->task_size = nworkers of ctx
|
11 years ago |
 Nathalie Furmento
|
f0caab0f9a
src/drivers: use a more meaningful name than args for the worker parameter
|
11 years ago |
|
Nathalie Furmento
|
a590112c14
fix fxt code
|
11 years ago |
|
Nathalie Furmento
|
4c43618b9b
merge trunk
|
11 years ago |
 Samuel Thibault
|
b25760aa94
fix warnings
|
11 years ago |
|
Andra Hugo
|
afade57efa
allow ctx to execute parallel codes by reusing starpu workers (keep the workers alive instead of getting them to sleep) -> TODO try to factorize with parallel tasks on combined workers (doing practically the same, duplicate tasks and barrier_wait before & after exec)
|
11 years ago |
|
Nathalie Furmento
|
892665a9b3
src/drivers/driver_common/driver_common.c: when dealing with parallel contexts, also generate trace and stats for non-master workers
|
11 years ago |
|
Samuel Thibault
|
f14e68dc2c
Separate worker and thread state, to see streamed kernel executions
|
11 years ago |
|
Samuel Thibault
|
f3423acafe
Set the worker status within the scheduling mutex section, for coherency with starpu_wakeup_worker()
|
11 years ago |
|
Samuel Thibault
|
778b4c444a
* Add codelet size, footprint and tag id in the paje trace.
|
11 years ago |
|
Samuel Thibault
|
af2b4ca2a8
port r12834 from scheduler_modularization: Also show "pop" scheduling state on CUDA
|
11 years ago |
|
Samuel Thibault
|
3dd3186a6c
Also show scheduling cost at push time
|
11 years ago |
|
Samuel Thibault
|
30d5f814d3
Show 'scheduling' state in paje trace
|
11 years ago |
|
Andra Hugo
|
3c8000976d
fix deadlock: we have to lock the worker's sched_mutex when iterating on its list of contexts but unlock it before the worker is going to sleep
|
11 years ago |
|
Andra Hugo
|
c8def5adac
sharing workers btw ctxs with no sched_policy
|
11 years ago |
|
Andra Hugo
|
3ca3790ef1
* sched_ctxs without sched_policy
|
11 years ago |
|
Andra Hugo
|
5a219c8a8f
* patch Terry: fix book workers (book workers that have already been booked and wake up eventually workers that we don't want anymore in the group)
|
12 years ago |
|
Andra Hugo
|
6fd68c0e07
execute parallel code (openmp for eg) inside contexts ------ temporary solution
|
12 years ago |
|
Samuel Thibault
|
be20ab0bcc
Advise the user to use --enable-blocking-drivers to avoid consuming real CPU for virtual CPU idling
|
12 years ago |
|
Andra Hugo
|
52459c8508
*eager: wake up workers when we add them to a context
|
12 years ago |
|
Nathalie Furmento
|
483f8e2979
merge trunk
|
12 years ago |
|
Andra Hugo
|
646c61a828
improve idle computations in order to better compute max_workers for the resizing lp
|
12 years ago |
|
Nathalie Furmento
|
98b4469b6a
merge trunk up to r11088
|
12 years ago |
 Pierre-André Wacrenier
|
9c421fc50e
unlock worker's sched_mutex before busy waiting
|
12 years ago |
 Thibaud Lambert
|
e2dd19c39c
mic (perfmodels): merge trunk
|
12 years ago |
|
Thibaud Lambert
|
df72333352
mic (perfmodels): Update perfmodels initialization
|
12 years ago |
|
Samuel Thibault
|
984d306f6d
16384 makes workers sleep for almost a whole ms, this is way too much. Keep not-too-intrusive value in the trunk for now
|
12 years ago |
|
Thibaud Lambert
|
cb72812881
mic (perfmodels): merge trunk
|
12 years ago |
|
Pierre-André Wacrenier
|
85bec148ba
define BACKOFF_MAX to 16384 for spinning
|
12 years ago |
|
Pierre-André Wacrenier
|
64c28778a9
add BACKOFF_MAX & BACKOFF_MIN macros for spinning
|
12 years ago |
