mozilla :: #gfx

14 Jul 2017
00:04pulsebotCheck-in: https://hg.mozilla.org/projects/graphics/rev/3f606ae08477 - Matt Woodrow - Set the correct reference frame for text overflow markers when the inner frame is transformed
00:04pulsebotCheck-in: https://hg.mozilla.org/projects/graphics/rev/dfd3958c4c3a - Matt Woodrow - Explictly mark the begin/end of display list building so we can match the behaviour of recreating the builder each frame
02:17Lenzakjgilbert: ping
02:17jgilbertLenzak, pong
02:18Lenzakhttps://bugzilla.mozilla.org/show_bug.cgi?id=1377527#c3
02:18LenzakI have pushed the gecko branch with ANGLE3118 onto my github
02:18firebotBug 1377527 NEW, nobody@mozilla.org Failures in conformance2/textures/canvas_sub_rectangle/tex-2d-r11f_g11f_b10f-rgb-float.html after AN
02:18jgilbertcool
02:20LenzakThe ANGLE format_utils refactor added some format validation and changed their return type.
02:20LenzakSome of them have similar name but different behavior
02:22jgilbertok, I will look
02:28LenzakGreat! Thanks :)
02:55gwGankro: around?
02:56Gankropong gw
02:56gwGankro: in the case of pushA, pushB, transparent text, popB, popA, what order are you expecting those shadows to paint in?
02:57Gankrogw: based on the servo impl, A should be under B, *I think*
02:58Gankrogw: but you can pick whichever and I can change gecko/servo to match, I don't have a strong opinion on what's better
02:58gwGankro: that does slightly complicate the implementation, but it's also intuitively correct to me, so I don't have a strong opinion on which is better either
03:00gwGankro: also, with the current changes, I pop and draw all shadows that existed on the stack, followed by any non-transparent text runs that were pushed during the shadow stack. that sound right to you?
03:00Gankrogw: I think so, but I'm not sure if there's a subtlety I'm missing
03:01gwGankro: ok. I'll tidy this up into a PR shortly. Let me know if anything doesn't seem quite right with ordering, or if you can still repro any artifacts and I'll follow up on Monday morning.
03:03Gankrogw: in case the A/B thing becomes a serious problem, the burden on clients is completely trivial here -- like the difference between .iter() and .iter().rev()
03:03GankroAnyway, have a good weekend!
03:03gwGankro: sounds good, thanks. You too!
14:23aosmondkvark: bug 1352092
14:23firebothttps://bugzil.la/1352092 ASSIGNED, hshih@mozilla.com Crash in mozalloc_abort | abort | webrender::texture_cache::TextureCache::insert
14:25kvarkaosmond: https://github.com/servo/webrender/issues/1405#issuecomment-309694244
14:25kvarknical ^
15:13mchangmconley: ping
15:13mconleymchang: pong
15:14mchangmconley: heyo! If I want to let a telemetry probe expire, do I just delete it from histograms.json?
15:14mconleymchang: it's already past expiry, so it's not collecting information anymore. If you don't need it anymore, we should file a bug to remove the infrastructure for the probe and the entry in Histograms.json.
15:15mchangmconley: ok thanks
15:15mconleymchang: np
15:40milanlsalzman: https://dvander.github.io/moz-gfx-telemetry/www/#view=windows-features, though it will move
15:45milanBas put together a way for us to collect more information around driver resets (hopefully) - https://wiki.mozilla.org/Platform/GFX/Device_Reset_Debugging - if you're in a bugzilla conversation with somebody that is getting the resets and you need more information from them, it's worth a try. And if you find gaps in the description, please edit the page...
17:38mconleyBenWa: hey yo - https://air.mozilla.org/the-joy-of-profiling-episode-6/ will be done transcoding in an hour or so
17:38BenWamconley: Nice! I'll take a look
17:38mconley:D
17:38mconleyBenWa: I'll also cc you on the bug I file.
17:40milandvander: so, we can't currently use the downloadable blocklist to deal with advanced layers - we do have the run time check, but do you want a patch that adds the above?
17:40BenWamconley: if stylo provides useful new constructs, like lazy flushing, and/or different performance behavior around style flushes Mozilla should provide guidance for web developers.
17:41mconleyBenWa: I agree
17:42mconleyWe'd probably need to talk to a Stylo engineer for information on that, I guess. :/
17:44mconleyFiled bug 1381071
17:44firebothttps://bugzil.la/1381071 NEW, nobody@mozilla.org Huge hang on Facebook, spending most of my time in Stylo style recalculations
17:56BenWamconley: jrmuizel: BTW if you want laymen's deminified stack that doesn't need private info this tool is very good: https://jsfiddle.net/uwusgnve/11/ and paste in a stack with this format: https://pastebin.mozilla.org/9027101
17:57BenWamstange: ^^ this works pretty well. Would be nice if the profiler could copy a stack in that full format and/or build in something similar to do
17:57mconleyholy hell
17:57BenWaThis is easier to use than some of the FB internal tools ATM so I use this as a first step before checking our internal source maps
17:58mconleythanks!
17:58BenWanp
18:03mstangeBenWa: which format do these " at [...]" things come from?
18:03mstangeBenWa: is this how the Chrome devtools spit out the call stack?
18:03BenWaI think so but I'm not sure
18:03mstangeok
18:04Gankrokats: the crasher seems to be intermittent :/ not convinced it has anything to do with my patch
18:05milansvargas: dev-tech-gfx@lists.mozilla.org
18:05katsGankro: hmm. it might be a latent thing that's made more frequent somehow
18:05dvandermilan: yeah that would be great, re: downloadable blocklists
18:05Gankrokats: you can see we replayed the crashing tasks a bunch of times, only got one repro: https://treeherder.mozilla.org/#/jobs?repo=try&revision=d912db7fa8c5079fca2ecea932cbe271281c914d
18:09Gankrokats: also, any idea why there's *so many* intermittents on every build? Is this common?
18:09katsGankro: also what's kind of interesting is that it doesn't look like the process actually crashes
18:09katslike, the harness complains that it found a crash dump
18:09katsbut the tests just keep running
18:10Gankrojrmuizel: was saying it looked like we were getting an explicit "please crash" command
18:10katsand then when it analyzes the dump later there's a crash reason of DUMP_REQUESTED
18:11katsi think maybe we started sending crash reports without crashing for some things
18:11katsand maybe the test harness was never updated to deal with that
18:11Gankromystor: are you nsBrowserGlue.js?
18:12mystorGankro: Am I a js file? no
18:12katsGankro: if by "so many" you mean the ~40 that are on your all-platforms try push, yeah that's normal
18:13Gankrokats: 40 crashes every build seems like a lot <.<
18:13katsGankro: they&#39;re not all crashes
18:13Gankroerr, test failures
18:13mystorGankro: What do you want to know about this file
18:13Gankromystor: I am getting a spooky intermittent: https://treeherder.mozilla.org/logviewer.html#?job_id=114144902&repo=try&lineNumber=1806-1807
18:14mystorThat looks like windows are all being destroyed before we transmit telemetry or something?
18:17katsthe dump is actually getting generated during startup
18:18Gankromystor: yeah since it seeemed to be related to window creation/destruction it seemed like your bag
18:20mystorGankro: Unfortunately I don&#39;t think I&#39;ve ever looked at that file before :-/
18:20Gankrodang
18:36katsGankro: so i think the crash is a shutdown crash from a content process that spun up earlier
18:37katslooking at the .extra file attached to the log i see ipc_channel_error=ShutDownKill
18:37Gankrokats: which is normal shutdown?
18:38katsi think the process hung on shutdown
18:38katsand so the parent killed it
18:38katsalso the log shows the stack trace for the wrong dmp file i think
18:38katswell it shows the stack from the -browser.dmp file
18:38katswhich is the parent process
18:38katsthe more interesting one would be the content process
18:39katslet me try to symbolicate the other dmp file
18:48katsdoh, the log already has the stack traces from the other dmp as well
18:48katssadly nothing interesting there
18:49katsso yeah i dunno
18:49katsGankro: maybe just land and hope for the best?
18:49katsif it gets really bad the sheriffs will probably back it out
18:50Gankrokats: sgtm
18:50Gankropatch rebases cleanly so you can just hit merge
18:52katsdone
18:52Gankro
19:07mconleyBenWa: the video has finished encoding: https://air.mozilla.org/the-joy-of-profiling-episode-6/ - Facebook stuff starts around 8:30
19:24gw280jrmuizel: recently I&#39;ve noticed my macbook flips GPUs constantly when browsing :(
19:25gw280in the last two minutes it&#39;s changed every 5 seconds or so
19:25jrmuizelgw280: webgl!
19:25jrmuizelgw280: you can ask jgilbert to fix it
19:25gw280jrmuizel: there&#39;s a known issue?
19:26jrmuizelgw280: sort of
19:26gw280I don&#39;t have any webgl stuff running I think?
19:26gw280it&#39;s very annoying
19:26jrmuizelgw280: webgl is everywhere
19:27gw280time to disable webgl
19:27gw280ok, GPU seems to have stopped changing for now...
19:30mconleyjrmuizel / mstange: that was fast - emilio has a working test case in the bug now. :) Eyyyy, makin&#39; stuff better!
19:30mconley(for the Facebook hang)
19:30mstangemconley: wow!
19:30jrmuizelmconley: where do I go to find more information?
19:31mconleyjrmuizel: bug 1381071
19:31firebothttps://bugzil.la/1381071 NEW, nobody@mozilla.org Huge hang on Facebook, spending most of my time in Stylo style recalculations
19:31mconleytalk about an alley-oop
19:31emiliomconley: well, the problem is pretty clear, and I knew we were going to need to do this eventually... I just didn&#39;t know it was going to blow up so awfully :)
19:31emilio!summon bholley
19:32emilioerr
19:32mconleyemilio: great stuff!
19:32mstangegw280: https://bugzilla.mozilla.org/show_bug.cgi?id=1339215#c8
19:32emiliomconley: anyway, thanks for the report!
19:32firebotBug 1339215 NEW, nobody@mozilla.org WebGL context creation is expensive and can cause the compositor to jank
19:32mconleyemilio: np - thanks for your fast action on the bug!
19:32mconleyBenWa: ^-- thank you too!
19:34emiliomconley: btw, did I do something wrong? Or the test-case is also pretty slow without stylo enabled?
19:34emilioBenWa: thank you, indeed :)
19:35* emilio re-checks prefs
19:35mconleyemilio: hmmm... you seem to be right
19:35mconleyI&#39;ll update the bug
20:11emiliomconley: there you got a patch... that fixes the issue (at least I can load the test-case in a few seconds in my debug build), but it still needs some love and it&#39;s late here in europe
20:12emilioBenWa: mconley: thanks a lot again for the report and the help!
20:13mconleyemilio: cool, good stuff. :)
20:36BenWaemilio: thanks for the quick turn around
21:06Gankrojrmuizel: https://github.com/rust-lang/rfcs/issues/1230
21:53jrmuizeldvander: when&#39;s the last time you updated the gpu-db for the gfx-telemetry?
21:53dvanderjrmuizel: the last time you told me too, probably :)
21:53jrmuizeldvander: you should do it again
21:54dvanderjrmuizel: sure thing, its due for a small overhaul, ill add that in
21:55dvanderjrmuizel: do you need it right now?
21:55jrmuizeldvander: nope, no rush at all
21:56jrmuizeldvander: just let me know when it&#39;s been updated
21:56dvanderk
15 Jul 2017
No messages
   
Last message: 9 days and 19 hours ago