Move the device timer conversion in cvk_device instead of commands

[rjodinchr]

No description provided.

[kpet]

What's the motivation for this change? Performance improvements?

[kpet]

Suggested change

	if (dev > m_sync_dev) {
	if (dev <= m_sync_dev) {

Isn't your condition backwards? You want a new pair of aligned timestamps only when you've wrapped around and the timestamp to convert to the host's time base is smaller than the device timestamp used as a reference.

[rjodinchr]

I don't think it is backwards.
The idea is to get new sync points when the time we want is bigger than what we have converted so far. Because they might have been a slight divergence since we have the last sync point.

[kpet]

So you'll end up requesting a new pair of timestamps for pretty much each command since the device timestamp will almost always be greater than the last one obtained when getting a host/device pair. The only time where you're not requesting a host/device pair is when the device timestamp wraps around. At which point I don't see much motivation for the change.

[rjodinchr]

What's the motivation for this change? Performance improvements?

From what I remember, this is needed to fix issues introduced by the new pattern from using timeline semaphore.
But as without timeline semaphore this is just a different implementation that works as well, I thought it would be better to submit it before in order to reduce the size of the timeline semaphore PR.

[rjodinchr]

I have rerun the CTS without this patch (but with the timeline semaphore) to remind me why this patch is needed with timeline semaphore.

In the implementation I have, there is a race condition between the application thread and the executor thread. It can lead to the thread app looking for an event profiling info, because the event is complete, but the executor has not computed the profiling info.
But this event might be related to a cvk_command inside of a cvk_command_batch. As the sync points are taken care by the batch command, it was making things complicated. And this is the way I dealt with it, moving the sync point to the device instead of the batch command.