example profiler refCount issue

[fishautumn]

1. taskEventBase.refCount:

nccl/ext-profiler/example/event.h

Line 90 in 2ea4ee9

int refCount; // number of references for this operation

at proxyOp start, it is assigned to 1 rather than increment

nccl/ext-profiler/example/plugin.c

Line 341 in 2ea4ee9

__atomic_store_n(&parent->base.refCount, 1, __ATOMIC_RELAXED);

nccl/ext-profiler/example/plugin.c

Line 357 in 2ea4ee9

__atomic_store_n(&parent->base.refCount, 1, __ATOMIC_RELAXED);

is it a typo?

2. group.refCount:

nccl/ext-profiler/example/event.h

Line 131 in 2ea4ee9

int refCount;

it is initialized to 1 at group start:

nccl/ext-profiler/example/plugin.c

Line 206 in 2ea4ee9

__atomic_store_n(&event->refCount, 1, __ATOMIC_RELAXED);

and is incremented at coll and p2p task start:

nccl/ext-profiler/example/plugin.c

Line 252 in 2ea4ee9

__atomic_fetch_add(&parent->refCount, 1, __ATOMIC_RELAXED);

nccl/ext-profiler/example/plugin.c

Line 285 in 2ea4ee9

__atomic_fetch_add(&parent->refCount, 1, __ATOMIC_RELAXED);

it is only decremented in updateEvent() function (invoked at op stop -> updateEvent for task -> updateEvent for group)

nccl/ext-profiler/example/plugin.c

Line 382 in 2ea4ee9

if (__atomic_fetch_sub(&event->refCount, 1, __ATOMIC_RELAXED) == 1) {

but is not decremented in group stop event:

nccl/ext-profiler/example/plugin.c

Line 440 in 2ea4ee9

if (type == ncclProfileGroup) {

After all tasks are released, the group will have refCount==1, so it cannot get released.

[gcongiu]

This is a known issue and is fixed in the next NCCL release.