Hi,

On 27/02/2019 18:14, Ivaylo Dimitrov wrote:
> Hi,
>
> On 27.02.19 г. 0:22 ч., Merlijn Wajer wrote:
>> Hi freemangordon,
>>
>> You're not on IRC atm, so email might work better :p. Spinal figured out
>> how we can work around some of the driver buggyness in powervr.
>
> Thats great.
>
>> Specifically, with older hildon-desktop and clutter 0.8, he doesn't get
>> any of the hangs or slowness that we're seeing. He thinks that this is
>> likely because of one of these patches to clutter 0.8:
>> https://github.com/spinal84/clutter-0.8/tree/blit/debian/patches
>
> Could be, afaik Nokia did lots of work trying to squeeze the last bit of
> performance from powervr

Makes sense.

>
>> We don't want to use old hildon-desktop, since there have been lots of
>> improvements to the maemo leste one. So I'm wondering what the next
>> sensible step would be. Should be try to build "modern" hildon-desktop
>> from maemo leste with clutter 0.8 and see if that also works? (And if it
>> does, see how we can fix clutter 1.0/1.2 to work well with PowerVR?)
>
> There are tools from ImgTek(TI?) that can help a lot to see what is the
> real bottleneck with clutter 1.x. When I ported to clutter 1.x my goal
> was to have h-d working at all, leaving the performance aside. So the
> thing we should do is to capture a couple of apitrace (and pvrtrace)
> logs from both 0.8 and 1.x, compare and then fix h-d or clutter 1.x to
> be in par with 0.8 in terms of performance. It could be that shaders I
> wrote are suboptimal, or... dunno, I was focused on functionality, not
> performance.

This sounds good wrt performance improvements. How would this help with
the hang. Would we see it what opengles call it likely hangs?

>> I went over the hildon-desktop commit logs, and from the log messages
>> there seemed to be only one (not too large) commit that mentions moving
>> to clutter 1.0 instead of 0.8, do you think it'd be a lot of work to
>> make it work with clutter 0.8?
>
> Yes, it will be and to be honest, I have strong opinion against doing
> that. We should fix h-d (or clutter 1.x) and move forward instead of
> going back.

OK. As I indicated above, I feel the same way. But I can't help but
wonder if such an exercise this will bring us closer to understanding
the 'hang' in powervr userspace<->kernel interaction. Since it just
doesn't happen with old clutter and old hildon-desktop. So I am trying
to answer the question: what change in either clutter or hildon-desktop
would cause this?

That is separate from the performance question, I think.

Cheers,
Merlijn