979026 - [Messages] SMS app Checkerboards while scrolling

Status: RESOLVED WORKSFORME

(Firefox OS Graveyard :: Gaia::SMS, defect, P1)

Description

[Mason Chang [Inactive] [:mchang]]

Profile - https://people.mozilla.org/~bgirard/cleopatra/#report=28c3f9f7d854d63d2ee93e743ecf2cd5405cb4ff. We see that we also spend a bunch of time in pgralloc, having ~20-40ms grallocs.

This is with patches from bug 977680, bug 979018, and bug 972731. It's much harder to checkerboard a full screen anymore, but it still happens here and there.

Comment 1

[Benoit Girard (:BenWa)]

The latest development of tiling on our graphics branch improve gralloc/buffer rotation bounds testcase a great deal.

Comment 3

[Mason Chang [Inactive] [:mchang]]

Just tested on master, here's a profile up - https://people.mozilla.org/~bgirard/cleopatra/#report=0b06da00a84de76614defd4930ae351431fd7a78. The recent improvements to tiling have helped a lot.

We no longer checkerboard the SMS app with a reference workload light. Closing as WFM. The only time left spent while scrolling is in BrowserElementPanning.js - bugs 976605 and bug 970125 should fix this.

Comment 4

[Milan Sreckovic [:milan] (needinfo for best results)]

We have to be careful here.  Bug 986687 was tested on QRD, with the latest tiling and other graphics improvements.  If we're tagging it as a duplicate of this bug, than this probably isn't a work for me.  Or the other one isn't a duplicate.

Comment 5

[Mason Chang [Inactive] [:mchang]]

I tested on today's master with a hamachi device and scrolled as fast as I could. If you do a casual scroll as if I was actually reading my SMS, I couldn't checkerboard.

Comment 6

[Jason Smith [:jsmith]]

(In reply to Mason Chang [:mchang] from comment #3)
> Just tested on master, here's a profile up -
> https://people.mozilla.org/~bgirard/cleopatra/
> #report=0b06da00a84de76614defd4930ae351431fd7a78. The recent improvements to
> tiling have helped a lot.
> 
> We no longer checkerboard the SMS app with a reference workload light.
> Closing as WFM. The only time left spent while scrolling is in
> BrowserElementPanning.js - bugs 976605 and bug 970125 should fix this.

Per QC's testing in the dupe, this isn't working right now. You need test this with the heavy workload to reproduce this problem right now.

Comment 7

[Mason Chang [Inactive] [:mchang]]

Attachment: SMS Layer Tree

Comment 8

[Mason Chang [Inactive] [:mchang]]

Attachment: Heavy Ref Workload - fast scroll

Comment 9

[Mason Chang [Inactive] [:mchang]]

Attachment: Heavy Ref Workload - Normal Scroll

Comment 10

[Milan Sreckovic [:milan] (needinfo for best results)]

Thanks Mason for recording this.  I can see this is Buri/Hamachi - probably the phone you have access to, so it's great.  I don't know how QC was testing, but I know that in the past we ran into checkerboarding quickest when scrolling fast in one direction, then reversing the direction of the scroll, also going fast, and repeating. Basically, something to mess up the heuristic :)

Comment 11

[Mason Chang [Inactive] [:mchang]]

We see 1 checkerboard with the normal scroll and 1 slight blip with the really fast scroll. But it's not very perceivable. 

Profile w/ a checkerboard - https://people.mozilla.org/~bgirard/cleopatra/#report=ec7afe69a5b5e30a20d8c6f6eb99d57d8cdeefb5

In range 8441 - 9206, we see a fairly long rasterize - 200 ms + a 100ms buildDisplayList.

Comment 12

[Mason Chang [Inactive] [:mchang]]

Attachment: Heavy Ref Workload - Stop/Start scrolling

Comment 13

[Mike Lee [:mlee]]

Per Mason's videos in comment 8, comment 9, and comment 12 checkerboarding during SMS scrolling is no longer an issue. Closing as WFM.

Comment 14

[Mason Chang [Inactive] [:mchang]]

Attachment: Display List Dump

In case we have to come back to this.

Comment 15

[Inder]

Tapas/Vikram - can you guys please take a look again and see if the issue we raised in bug 986687 is fixed now?

Comment 16

[Mandyam Vikram]

We'll verify it in a build early next week.

Comment 17

[Inder]

Thanks Vikram. Keeping the ni as a reminder.

Comment 18

[Mandyam Vikram]

On the latest build (04/01), the checkerboarding is not observed. This issue can be considered as resolved.