Open Bug 557423 Opened 11 years ago Updated 9 years ago
GL quake "too slow to be supported" in Firefox
240.64 KB, image/jpeg
62.19 KB, image/jpeg
42.44 KB, patch
|Details | Diff | Splinter Review|
130.07 KB, text/plain
291.21 KB, text/plain
1019.19 KB, application/x-xz
82.19 KB, application/x-xz
157.75 KB, image/jpeg
126.24 KB, text/plain
76.03 KB, application/x-xz
216.87 KB, text/plain
546.73 KB, text/plain
481.88 KB, text/plain
47.74 KB, text/plain
Oh sweet jeebers, you have to compiler a bunch of Java stuff. I'll...sit on this for a while.
Assignee: general → shaver
I think GWT produces JS as output, not an interpreter running JVM bytecode. But the server side of it is a bunch of Java that I have to set up an environment for, which will take some cycles.
Could we get them to test it? Send them a debug build and have them send back TM spew and/or TraceVis results?
Or just point us to a public url where the thing is set up....
I followed their build instructions and got the Jetty server up and running, but it says my WebGL support is turned off. I think I have the right prefs... which is it?
webgl.enabled_for_all_sites should be set to true. Does the page at https://cvs.khronos.org/svn/repos/registry/trunk/public/webgl/sdk/tests/webgl-conformance-tests.html say "WebGL not enabled" in red near the "run tests" button at the top? (I don't recommend running the tests, since the first one I clicked crashed me. :-P )
@ Robert Sayre especially: quake2-gwt does *not even try* to create a WebGL context before saying "WebGL Support Required". I know that because I set breakpoints in our WebGL initialization code, and they are never reached when loading quake2-gwt, while of course they are reached by every WebGL-using page. So at least this particular error message, "WebGL Support Required", is erroneous. There must be other reasons why quake2-gwt doesn't run in minefield, for example web sockets or something like that. I filed a bug report but didn't get a reply: http://code.google.com/p/quake2-gwt-port/issues/detail?id=26 In any case I'm not happy about this erroneous error message that gives our WebGL implementation a bad name, and the above-quoted comment on our JS performance doesn't hold water either, because: * if they were able to test then why aren't _we_ able to test? * it's not like we're 10 times slower than google, so if it's playable for them then (on good enough hardware) it must be playable for us.
This adds "webgl" as a canvas id that we support.
This patch mostly adds debugging messages in the code. It also adds firebug output to console, I didn't write it in my original post but the method for sending messages to the browser console (for some unknown reason, maybe an oversight) also in some instances sends messages to the server console, meaning some of the server's console messages appear twice and in a little garbled mode- Not that big of an issue for now :)
Great. I see that your patch also has the change making it test "experimental-webgl" first, so one can test stuff without applying my patch to firefox.
Yeah, I guess I wasn't clear enough in comment 13- the attached patch includes everything I've talked about in the previous comment :)
Comment on attachment 452714 [details] [diff] [review] add "webgl" canvas-id No, it can't be called "webgl" until we have WebGL 1.0; all the browsers agreed to support "experimental-webgl" until then (and do so) -- is something not working there? The bug actually seems to be that we throw when we get an invalid context string, instead of returning null. That's an easy fix.
Attachment #452714 - Flags: review-
Well, I did as Benoit suggested and added the getParameter() method and used it instead of getString() - it makes Firefox show the GL info: GL_VENDOR: Mozilla GL_RENDERER: Mozilla GL_VERSION: WebGL 1.0 GL_EXTENSIONS: null But breaks Chrome: GL_VENDOR: null GL_RENDERER: null GL_VERSION: null GL_EXTENSIONS: null Which means the Chrome guys aren't supporting this yet, right? Looking at the code, I can see that GL_EXTENSIONS are supposed to be filled with all my GL extensions but I guess that isn't working in both Chrome (see the comment I wrote before) and Firefox. The code retrieves it with getParameter(GLAdapter.GL_EXTENSIONS), GLAdapter.GL_EXTENSIONS == 0x1F03 - is that the "correct" way? My video cards is Opengl 3.2 compatible and an external extensions viewer (no glxinfo type prog in windows right?) reports a lot of extensions- so it definitely shouldn't be null :) Any ideas?
(In reply to comment #17) > Well, I did as Benoit suggested and added the getParameter() method and used it > instead of getString() - it makes Firefox show the GL info: > GL_VENDOR: Mozilla > GL_RENDERER: Mozilla > GL_VERSION: WebGL 1.0 > GL_EXTENSIONS: null > > But breaks Chrome: > GL_VENDOR: null > GL_RENDERER: null > GL_VERSION: null > GL_EXTENSIONS: null > > Which means the Chrome guys aren't supporting this yet, right? > Looking at the code, I can see that GL_EXTENSIONS are supposed to be filled > with all my GL extensions but I guess that isn't working in both Chrome (see > the comment I wrote before) and Firefox. > The code retrieves it with getParameter(GLAdapter.GL_EXTENSIONS), > GLAdapter.GL_EXTENSIONS == 0x1F03 - is that the "correct" way? Nope, it should be null -- there are no WebGL extensions defined yet, so none are listed. Though, perhaps, an empty array is more correct here rather than null.
You were right, that wasn't the problem here. The problem was that gwt-quake expects the returned GL_VERSION to be only numbers and hits an exception when trying to convert it to float (btw- the same problem exists in Chrome and the fix posted on the project website by users doesn't work for Chrome either). I've fixed it to cut everything and now (after a few other fixes as well) it goes a little further- it runs the game's inner loop (you can see in the messages it responds to browser window resizes) but it still doesn't render properly- I figured it's because it doesn't load the game's config files properly (like I wrote earlier) but for some odd reason it now loads the game's config files! :) I'd post the exception it throws but since it's gwt it won't mean much to anyone (method names translated by gwt are random letters- to make it compress better) - to get to the bottom of this someone will have to debug the hell out of this thing :) Like I said- for the next couple of weeks I'll have absolutely no free time for this so I'm leaving this patch for prosperity :)
Attachment #452725 - Attachment is obsolete: true
note: to cut everything = to cut everything other than the version numbers
It really shouldn't care what the version number is -- I don't think there's anything useful with that information that it can do currently anyway. What's the exception that's generated?
oh, I thought you were getting an xpcom exception. Still, TypeError is interesting there -- "yqb is null" is the actual error. So, if you want to do some spelunking, take a look at what's on that line and see how yqb is being used. Don't they have some way to debug this goop?
There actually is- gwt can be passed an argument to build the methods with meaningful names - http://java.dzone.com/news/understanding-gwt-compiler. This project uses the Maven build system and I couldn't gwt to work by passing these argument or by setting it in it's config files (http://mojo.codehaus.org/gwt-maven-plugin/compile-mojo.html) - I probably missed something obvious... Oh, according to what I read enabling these modes also outputs to Firebug's console so the change I made (so that Firebug output works) wouldn't be needed anymore :)
OK, so- I've flipped on DETAILED mode which gave me the exact place this failed (I forgot to set some variable, silly me :)) and fixed the part that was failing. It now gets into the main menu of the game, when I tried to start a new game it failed on trying to create/load a save game- I've disabled that part but the game still won't load. Looking back, I can see that after the graphics initialization it throws a glGetEror() = 0x500. As I understand it that means GL_INVALID_ENUM which shouldn't cause the game to not work... Anyway, I'm done with messing with this for a while, a few pointers to anyone brave enough to touch this in the mean time: * Looks like previous versions of Chrome did output the GL Extensions and by parsing that the game knew which rendering options to load (it is after all originally based on OpenGL), what I've done is to set only the flags that seem necessary for the game to run (all of this hackery is done in /src/jake2/render/gl/Main.java, line:1277) - it is very possible that rendering doesn't work because of these flags not set correctly :) * The comment I made above about a save game not created/loaded is probably due to some deeper problem with saving / loading files Good luck :)
(In reply to comment #25) > Looking back, I can see that after the graphics initialization it throws a > glGetEror() = 0x500. > As I understand it that means GL_INVALID_ENUM which shouldn't cause the game to > not work... This rings a bell: I saw code in quake2-gwt throwing an exception on every GL error. So the next step is to disable that. Many thanks for your work - seeing the main game menu is a huge progress already. > Anyway, I'm done with messing with this for a while, a few pointers to anyone > brave enough to touch this in the mean time: > * Looks like previous versions of Chrome did output the GL Extensions and by > parsing that the game knew which rendering options to load (it is after all > originally based on OpenGL), what I've done is to set only the flags that seem > necessary for the game to run (all of this hackery is done in > /src/jake2/render/gl/Main.java, line:1277) - it is very possible that rendering > doesn't work because of these flags not set correctly :) > * The comment I made above about a save game not created/loaded is probably due > to some deeper problem with saving / loading files > > Good luck :)
Attachment #452887 - Attachment is obsolete: true
That's alright, I'm pretty ignorant at JS as well :) That message appears when the Firebug console isn't enabled... Do you have Firebug installed? if not you need to install it from http://getfirebug.com/downloads (latest alpha versions), open it and enable console...
OK, thanks, once this was fixed, I got the game actually loading !!!! It just took veeeeery long. Should retry with a release build, instead of debug --- that's going to make a huge difference here. Here is a screenshot! Anyone happy about this, should thank Barak!
Attachment #453225 - Attachment description: quake2 in minefield → screenshot: quake2 scene in minefield
Just tried with a release build of Minefield, and it's already playable!
Benoit, Barak: awesome work!
Great to hear :) What did you change to make it get past the main menu? Did you change back from DETAILED mode (back to OBF)? it probably also has a lot to do with the current slowness of the game...
The only difference between what you told me and what I did, is that I commented out all the console... lines in the generated HTML, as I had not yet read your reply about this requiring the firebug console to be opened. Also, note that with a debug build of firefox, it takes a good minute or two from the time you launch a game in the main menu, to the time you actually start playing. With a release build of firefox, this goes much faster.
I see, well- I'm using a nightly build from the mozilla-central trunk builds (http://ftp.mozilla.org/pub/mozilla.org/firefox/tinderbox-builds/mozilla-central-win32/)- any idea if those are built in release or debug mode? I tried to wait a few minutes but the game still won't load, I think the difference is that I'm running the browser in a Windows box and the server is running in a Linux VM (since quake-gwt server won't compile/run on Windows)- while you're obviously running both on the same box- I'm not sure if my configuration is even supported in quake-gwt... The exceptions I get are about load/save file permission problems on the Linux box- so it sounds plausible that the "server/client on different environments" issue is causing this. I've tried connecting via the multiplayer option but the default config tries to connect to the localhost, and when I manually try to connect to the VM's ip address the server won't accept my connection... oh well, at least it works on your end :) If anyone else cares to try this, I'd love to hear more success/fail stories :) And like I said in my previous comment- changing the build options back from DETAILED mode to OBF mode (in the diff, see the second change in pom.xml) means you won't have to install Firebug to get this working...
Indeed, I am running both client and server on the same machine. With a colleague at Mozilla, we tried letting him use my server from his machine. After unblocking port 8080 on my machine, he could access my server from his on the local network. He then tried 2 machines. With his windows one, all he got was "WebGL support required". But with his Mac, he could actually play!
Ah, thanks for the DETAILED/OBF idea, I had overlooked this!
I tried the OBF config, by removing the corresponding line from your diff, that didn't work, even on a clean build directory. Still the same error about 'console' being undefined. I then tried with firebug console open: still the same error! So I redid my trick about removing manually the 'console' thing from the generated HTML, namely this file (only 1 file matches this pattern): war/gwtquake/*.cache.html and that fixed again the game.
The benchmark is running! However, it seems to be running many times and I was not patient enough to wait for it to complete all its runs. Here's the result from the 1st run, which was the best out of 4: 632 frames, 71 seconds, 8.9 FPS This is largely independent of the size of the windows; fwiw, it was 1324x768.
Attachment #452714 - Attachment is obsolete: true
That's weird :) could you check that the file /maven-build/client/pom.xml has <style>OBF</style> instead of <style>DETAILED</style> - just to make sure that the diff was applied correctly... Firebug seems to have a "work/no-work" relationship with the latest builds- seems to work fine, close and open Firefox and for some reason it stops working... what usually fixes that for me is a reset it's settings/uninstall/reinstall loop. Could you try and hold off on setting the style back and try to get it work with the Firebug console on a windows machine? I'd love to see if that box has the same problem I'm having (I get a NoFileFoundException, trying to open a RandomAccessFile)
That's weird = about console, not about the benchmark :)
HOWTO play Quake2 in Firefox: Prerequisites: you need a Linux or Mac box, and some packages installed: Lame, Ogg Vorbis, Java. These instructions are adapted from http://code.google.com/p/quake2-gwt-port/wiki/BuildingAndRunning Check out the Quake2 code: $ hg clone https://quake2-gwt-port.googlecode.com/hg/ quake2-gwt-port $ cd quake2-gwt-port Now apply Barak's patch, the version keeping OBF: go to https://bug557423.bugzilla.mozilla.org/attachment.cgi?id=453365 and save as diff-keeping OBF, and apply it: $ patch -p1 < /path/to/diff-keeping-OBF Now build quake2 and install the resources (will download them for you). $ ./build-dedicated-server $ ./install-resources $ cp -r maven-build/server/target/gwtquake/war/gwtquake war Note, the 'cp' line above is really needed for now (explained on above-linked install instructions). Now it where I have to remove the console.log(...) calls. I use this command: $ find war/gwtquake/ -name *.cache.html | xargs sed -i 's/console\.log([^)]*);*//g' Now you should have a working quake2, just launch the server: $ ./run-dedicated-server Lauch a RELEASE BUILD of Minefield and navigate to: http://localhost:8080/GwtQuake.html
(In reply to comment #41) > That's weird :) > could you check that the file /maven-build/client/pom.xml has > <style>OBF</style> instead of <style>DETAILED</style> - just to make sure that > the diff was applied correctly... Yes, I checked. In OBF, there are only 3 console.log() calls to remove, but I still have to remove them. > Firebug seems to have a "work/no-work" relationship with the latest builds- > seems to work fine, close and open Firefox and for some reason it stops > working... what usually fixes that for me is a reset it's > settings/uninstall/reinstall loop. > Could you try and hold off on setting the style back and try to get it work > with the Firebug console on a windows machine? Sorry, I don't have a windows setup at the moment. But the quake2-gwt wiki says windows is not supported anyway.
Windows isn't supported as a server, should work as a client... I don't really understand the reason behind my problem but I'm just ecstatic this works on anything :)
Here is the output of 'perf' (the new linux kernel profiler). Quick summary: we're spending most of our time in the JS engine, very little time in WebGL. That can be explained by the fact that Quake2, released in 1997, does a lot of stuff in software, it doesn't take full advantage of GL. I hope that this inspires our JS gurus to make us run faster! There seems to be room for improvement: I get only 8.9 FPS on my big Core i7 @ 1.6 GHz; I think to have read somewhere that Chromium gets much better framerates. *** How this was generated *** On a release build of minefield: $ perf record -f -- dist/bin/firefox -P test-release -no-remote Then go to quake2-gwt, run the benchmark, let it complete two runs, quit firefox, then do: $ perf report --sort symbol | c++filt > ~/quake2-perf *** More analysis of the results *** The symbols accounting for more that 1% of execution time are: 33.45% [.] js_Interpret 5.88% [.] JSScopeProperty::trace(JSTracer*) 4.77% [.] js_TraceObject 4.46% [.] js_CallGCMarker(JSTracer*, void*, unsigned int) 4.25% [.] 0x000038200802bf 3.39% [.] 0x0000383bc6ba20 3.01% [.] JSScope::searchTable(long, bool) 2.15% [.] gfxUtils::PremultiplyImageSurface(gfxImageSurface*, gfxImageSurface*) 2.09% [.] js_LookupPropertyWithFlags 1.55% [.] js_GetPropertyHelper 1.43% [.] JS_TraceChildren 1.30% [.] js::SweepScopeProperties(JSContext*) 1.27% [.] JSScope::changeTable(JSContext*, int) 1.22% [.] js_NativeGet This is mostly JS stuff. Where are the WebGL symbols? Here are the top 10: $ grep WebGL ~/quake2-perf | head -10 0.15% [.] mozilla::WebGLContext::QueryInterface(nsID const&, void**) 0.09% [.] mozilla::WebGLContext::ValidateBuffers(unsigned int) 0.08% [.] mozilla::WebGLContext::AddRef() 0.06% [.] nsICanvasRenderingContextWebGL_Uniform1i(JSContext*, unsigned int, long*) 0.04% [.] int xpc_qsUnwrapThis<nsICanvasRenderingContextWebGL>(JSContext*, JSObject*, JSObject*, nsICanvasRenderingContextWebGL**, nsISupports**, long*, XPCLazyCallContext*) 0.04% [.] mozilla::WebGLUniformLocation::QueryInterface(nsID const&, void**) 0.04% [.] nsICanvasRenderingContextWebGL_DrawArrays(JSContext*, unsigned int, long*) 0.03% [.] unsigned int xpc_qsUnwrapArg<nsIWebGLUniformLocation>(JSContext*, long, nsIWebGLUniformLocation**, nsISupports**, long*) 0.03% [.] mozilla::WebGLContext::Uniform1i(nsIWebGLUniformLocation*, int) 0.03% [.] mozilla::WebGLContext::Release() Not only this is very little, but much of that is actually XPCOM stuff! In conclusion, to make Quake2 go faster, we need to look at the JS side of things, not the WebGL side. As you guys had apparently guessed in the very first comments on this bug :)
This version is more detailed, it has been generated by this command: perf report --sort comm,dso,symbol | c++filt > ~/quake2-perf-moreinfo
Here's the raw output of 'perf', compressed. Uncompress it: $ bunzip2 perf.data.bz2 Then you can generate reports for yourself with 'perf report'.
Amongst others we GC ourselves to death here. We are working on that part. Great work everyone. Fun to watch this bug make progress.
We should try this with the fatval branch.
Can we please get bugs blocking this one filed on at least the following issues: 1) 15% or more is spent in GC, based on comment 47. 2) We're somewhat failing to trace here (30+% in js_Interpret!).
bz, if you guys want to try the fatval branch, I wouldn't be surprised if the GC stuff is doubles.
Does browser run on fatval branch? If so, are there nightlies? If not, can you point Benoit at the right thing to pull? I still think it's worth the separate bugs to keep issues straight.
(In reply to comment #54) > Does browser run on fatval branch? If so, are there nightlies? If not, can > you point Benoit at the right thing to pull? > > I still think it's worth the separate bugs to keep issues straight. The fatval browser works. There are no nightlies yet. But I think you want to wait until we get the traceable-native quickstubs re-enabled there before trying this one out. I'll post here when it is ready.
This was recorded with the -g option, so it has call graph info. You need xz to uncompress it (it was really big). You can then do: $ perf report | c++filt | wc -l and you should see a nice graph in ASCII art.
Attachment #453397 - Attachment is obsolete: true
Uncompress this with xz. The resulting file is a plain text call-graph.
(In reply to comment #52) > Can we please get bugs blocking this one filed on at least the following > issues: > > 1) 15% or more is spent in GC, based on comment 47. fatvals may help here, as Andreas said. Compartments could conceivably mitigate this a bit too. > 2) We're somewhat failing to trace here (30+% in js_Interpret!). Feel free to find out why; maybe it's just a simple missing feature. Otherwise, JM hopefully will help with this. I'm going to be trying to get the 'moo' browser in a testable state very soon, so we can see how it does there.
Oh, one more thing from the profile: it seems like some property access functions are in the top group. Some of this may be fixed by bug 558451. But there may be other kinds of property accesses that we are not caching, but should. We should figure out what's going on there so we know whether current scope and PIC work will cover it, or if we need more.
In case people don't know, "fatvals" refers to a branch currently being developed in which the representation of basic JS values has changed. Most significantly for this bug, doubles are represented directly, rather than being represented by pointers to values on the heap. This make them faster to access and also means much less garbage collection is required.
OK, fatvals is ready to test this. You can check out and build from: http://hg.mozilla.org/users/lwagner_mozilla.com/fatval/ on any *32-bit* OS (don't worry, 64-bit support is getting done even as we speak). Or, we could supply someone with an opt build. Just let me know what platform.
My platform is linux x86-64, please let me know when it is supported by fatvals.
OS: Windows 7 → All
Hardware: x86 → All
Hi Everyone, Well- I've snooped a little more into the code (to try and understand why it doesn't work on my setup - Windows browser, linux server on a different machine) and have a good understanding of how the code is supposed to work but have no idea why it doesn't: It seems to get stuck at loading certain files (mainly map files). To load these files the game uses XMLHttpRequest and has an async callback function which is supposed to run whenever XMLHttpRequest's internal status changes (http://msdn.microsoft.com/en-us/library/dd576252%28v=VS.85%29.aspx is how the async callback is set up, http://msdn.microsoft.com/en-us/library/ms534361%28v=VS.85%29.aspx are possible states it checks) - inside it checks the current state and by that determines if the file has finished loading or not (Sorry if this stuff is obvious to you- all new to me :)). Using debug messages I can see that the callback is only called once (after it's been setup - readyState = 1) but never called again, even if there is some kind of a connection issue it should be called again at the end of the call (after it's command has been set and the send command was called - should change to readyState = 2). Trying to load a large map file only shows that the browser sent the command to the server (via Firebug console- GET http://10.0.0.66:8080/baseq2/maps/demo1.bsp 200 OK 13ms) but it never changes past that and eventually times out (internal game timer). At a small demo files (GET http://10.0.0.66:8080/baseq2/demos/q2demo1.dm2 200 OK 99ms - total size 160kb for example) I can see via the Firebug net column that the file has finished downloading but the callback function is never called! The function that's responsible for setting all of this up (and has the callback function) can be seen here: http://www.google.com/codesearch/p?hl=en#bwYYd3m9KuA/src/jake2/gwt/client/GwtResourceLoaderImpl.java&q=loadResourceasync%20package:http://quake2-gwt-port\.googlecode\.com&sa=N&cd=3&ct=rc&l=68 Who should I talk to about this? anyone with any insight on this? Thanks :)
Hooray it works on my setup as well :) Going over XMLHTTPREQUEST bugs I ran into Bug 343769 - (XMLHttpRequest onreadystatechange not firing after readyState = 1) which basically means the behavior I described is exactly what happens when Firebug is running. So- I removed any firebug related stuff in the code (OBF + Benoit's script) and it worked :) Boy do I feel dumb for not following Benoit's exact instructions, oh well- no harm done... so this is tested to work on Linux and Windows (with a different machine as the server of course, since windows is not supported as a server) :)
Hm, indeed the WebGL implementation has improved a lot recently, and fatvals seems not to have merged in mozilla-central changes in the past few weeks, so fatvals doesn't have these WebGL fixes. You could try doing the merge yourself, e.g. "hg merge ../mozilla-central", but I have no idea if this is likely to give a lot of conflicts, you need fatvals people here :)
Like I wrote in the previous comment- changed it so that we no longer query any of the GL variables (GL_VENDOR, GL_RENDERER etc...) and instead these variables are preset. This also makes the latest Chrome run (although it outputs tons of error messages about one of the GL functions being obsolete) but it freezes a couple of seconds into the game... boo hoo :)
The fatvals branch will be merged into the TraceMonkey repository (hg.mozilla.org/tracemonkey) soon, perhaps by this weekend? TraceMonkey is kept pretty well up-to-date w.r.t. mozilla-central. So perhaps it's best to wait until that merge happens.
Thanks to Stefan Haustein the Firefox fixes were merged to mainline and latest Minefield nightly works out of the box :) (http://code.google.com/p/quake2-gwt-port/issues/detail?id=26)
Thanks a lot Barak and Stefan, for handling this, this is beautiful!
As I understand it fatvals was recently merged into Tracemonkey so... anyone up for profiling this with a Tracemonkey build? :)
Does it support linux x86-64 now?
Well, the TraceMonkey nightly builds directory in http://ftp.mozilla.org/pub/mozilla.org/firefox/nightly/latest-tracemonkey/ has a linux x86_64 package so I guess so... Can any of the JS guys confirm this assumption?
Yes (it couldn't have landed without x64 support).
OK, indeed the tracemonkey branch runs (very well) here on linux x86-64. I just had to apply the patch from bug 578215 which fixes an important canvas-related crash. I have run and profiled quake2's benchmark with the tracemonkey branch. Summary (by "before" I mean the results reported above in this bug report, i.e. mozilla-central from a few weeks ago): * In-game performance is noticeably improved, and the quake2 benchmark now reports 9.5 FPS, up from 8.9 FPS before. * What these numbers don't say, is that the benchmark *loading* speed is much improved. Runs an order of magnitude faster now. * From what I understand, profiler results (see below) seem to show that the crazy GC'ing is indeed fixed now --- presumably thanks to fat values. * But, profiler results still show 37% of the time spent in js::Interpret, and, as far as I understand, only 4% of the time spent in JIT'ed code. I am attaching below the profiler results, let me just paste here the top of the list of most expensive functions: # Overhead Symbol # ........ ...... # 37.58% [.] js::Interpret(JSContext*) 3.85% [.] 0x00003825818769 3.74% [.] JSScope::searchTable(long, bool) 3.73% [.] JSScopeProperty::trace(JSTracer*) 3.21% [.] js_TraceObject(JSTracer*, JSObject*) 2.82% [.] js::Mark(JSTracer*, void*, unsigned int) 2.61% [.] js_LookupPropertyWithFlags(JSContext*, JSObject*, long, unsigned int, JSObject**, JSProperty**) 2.53% [.] gfxUtils::PremultiplyImageSurface(gfxImageSurface*, gfxImageSurface*) 1.62% [.] js_GetPropertyHelper(JSContext*, JSObject*, long, unsigned int, js::Value*) 1.48% [.] JSScope::changeTable(JSContext*, int) 1.44% [.] 0x000038218448f0 1.38% [.] js_NativeGet(JSContext*, JSObject*, JSObject*, JSScopeProperty*, unsigned int, js::Value*) 1.15% [.] js_SetPropertyHelper(JSContext*, JSObject*, long, unsigned int, js::Value*) 0.95% [.] JSScope::addProperty(JSContext*, long, int (*)(JSContext*, JSObject*, long, js::Value*), int (*)(JSContext*, JSObject*, long, js::Value*), unsigned int, unsigned int, unsigned int, int) 0.90% [k] _nv006601rm 0.86% [.] js::SweepScopeProperties(JSContext*) 0.84% [k] audit_syscall_entry 0.80% [.] JS_TraceChildren 0.50% [.] JSScope::addPropertyHelper(JSContext*, long, int (*)(JSContext*, JSObject*, long, js::Value*), int (*)(JSContext*, JSObject*, long, js::Value*), unsigned int, unsigned int, unsigned int, int, JSScopeProperty**)
Profiler results for quake2 benchmark in a build of the tracemonkey branch from today. This is a simple table of most expensive functions, sorted by decreasing own time.
This file is XZ-compressed because it was big. It's an ASCII-art call graph of today's profiler results with the tracemonkey branch. I'm not uploading the raw perf.data file this time, because it's quite big and I'm not sure if anybody is using it. Just tell me if you would like to have it.
Hey Everyone, I use the Firefox nightly builds on a daily basis so every now and then I run the quake2-gwt-port and see if it's improved (it mostly has :)). Anyway, when I ran it today I saw that it no longer works- the error console shows errors regarding texImage2D and after some reading I realized that Firefox was changed to no longer accept the deprecated form of texImage2D (Bug 584815). The attached patch makes mostly trivial changes to the function's signature/call in the quake2-gwt-port code. I've opened a bug on the quake2-gwt-port project (http://code.google.com/p/quake2-gwt-port/issues/detail?id=30) to take back the patch but in the mean time this patch is needed to make it run with the latest nightlies.
(In reply to comment #78) > Created attachment 466158 [details] [diff] [review] > Patch to address texImage2D changes > > Hey Everyone, > I use the Firefox nightly builds on a daily basis so every now and then I run > the quake2-gwt-port and see if it's improved (it mostly has :)). > Anyway, when I ran it today I saw that it no longer works- the error console > shows errors regarding texImage2D and after some reading I realized that > Firefox was changed to no longer accept the deprecated form of texImage2D (Bug > 584815). Yep. Chromium is also no longer accepting the old form. > > The attached patch makes mostly trivial changes to the function's > signature/call in the quake2-gwt-port code. > I've opened a bug on the quake2-gwt-port project > (http://code.google.com/p/quake2-gwt-port/issues/detail?id=30) to take back the > patch but in the mean time this patch is needed to make it run with the latest > nightlies. Awesome! So you have done everything that was useful to do, I don't have anything to add :-)
Hey again, Firefox stopped working a few days ago, after a little investigation I found that the problem was with the fragment shader validation (due to http://bugzilla.mozilla.org/show_bug.cgi?id=589011) and required a pretty straight forward fix. As before, I've opened a bug on the quake2-gwt-port project (http://code.google.com/p/quake2-gwt-port/issues/detail?id=31) to take back the patch but in the mean time this is needed to run on Firefox... :)
Thanks once again Barak! Notice that another work-around is to go to about:config and set webgl.shader_validator to false.
(In reply to comment #81) > Thanks once again Barak! Sure thing :) > Notice that another work-around is to go to about:config and set > webgl.shader_validator to false. Sure, though solving it with the patch is better since shader validation will always be on by default, right?
Right! I was just mentioning this as a work-around for testing.
I tested today the tracemonkey repo, changeset e2e1ea2a39ce. My limited understanding is that this is _after_ a big jaegermonkey merge. The results are not very good: it's sluggish and tends to crash while trying to load the game's files. The only useful info that I have is that when it crashes, there are 512 threads (!!!). Most of them are @pthread_cond_wait. Is there something in Jaegermonkey that would explain why so many threads are being created? Do you need a good backtrace (requires me to make a debug build) or a profiler run?
Is there a reason your opt builds don't have debugging symbols? That's the simplest way to make sure you always get useful stacks... Are you willing to check whether it's the merge itself that caused the performance drop? Things to maybe test: 1) Disable the method jit in preferences. How does that affect things? 2) Check builds from right before and after the merge. 3) Check things from right before and after recursion-tracing was disabled.
(In reply to comment #85) > Is there a reason your opt builds don't have debugging symbols? That's the > simplest way to make sure you always get useful stacks... Right, I didn't know how to get debugging symbols, now I've found it, thanks. Rebuilt with that option; now I sometimes (not consistently) get a different crash: Program received signal SIGILL, Illegal instruction. (gdb) bt #0 0x00007fff853fa9bc in ?? () #1 0x0300000000000000 in ?? () #2 0x0000000000000000 in ?? () My only guess is a JIT assembler bug. When this crash happens, there are roughly 500 other threads. > > Are you willing to check whether it's the merge itself that caused the > performance drop? Things to maybe test: > > 1) Disable the method jit in preferences. How does that affect things? Disabling method jit, and also disabling all four jit options doesn't make any visible difference: still same speed, still hundreds of threads created. I couldn't reproduce any crash with method jit disabled. But that doesn't mean much as the crash was hard to reproduce anyway. > 2) Check builds from right before and after the merge. I tried the revision just before the Jaegermonkey merge. That is fc11a310d9a2 "Merge mozilla-central to tracemonkey." Unfortunately it doesn't build here (linux x86-64). Going before it is nontrivial as it is a merge, so I decided I don't have time, sorry.
Hey Everyone, First of all- the above patches are no longer needed as they were merged back to the quake2-gwt source. Regarding the recent TM builds: To test the issue I downloaded an opt TM build from the TM tryserver builds and more specifically the one built from http://hg.mozilla.org/tracemonkey/rev/319b1a4e0883 (a Win32 build). I've had the same problem as Benoit- the browser crashes after about a minute into the game and there seems to be a consistent pause every second or so (GC problem?) of gameplay which makes it seem very slow, looking at the crash log I see around 200 threads when it crashed. Just as Benoit noted he browser doesn't seem to crash once I turn off methodjit but I have no idea how to check the number of threads. Other than looking at the crash log, how can I see the number of threads in windows? (Process Explorer to the rescue maybe?) Crash reports: http://crash-stats.mozilla.com/report/index/bp-8df09db8-5eaa-4a53-8791-850752100901 http://crash-stats.mozilla.com/report/index/bp-aa13c280-cd1a-4ba7-9211-690ed2100901 > 3) Check things from right before and after recursion-tracing was disabled. Is this still suspected? if so, around which date/cset did this change happen?
Oh and I forgot to mention that it still crashes even when I disable tracing and only enable methodjit...
(In reply to comment #87) > the above patches are no longer needed Then, for each of them, can you click 'Details' and check 'obsolete' ? Thanks. > but I have no idea how to check the number of threads. > Other than looking at the crash log, how can I see the number of threads in > windows? (Process Explorer to the rescue maybe?) I don't know much about windows, will let others reply. On my side, this is something that I see when running Firefox in a debugger.
(In reply to comment #89) > I don't know much about windows, will let others reply. On my side, this is > something that I see when running Firefox in a debugger. Well, Process Explorer does let me see the number of threads for a given process. There are 250+ threads even with tracing only (shoots up when the game starts). I have the latest nightly (which TM was merged to just before JM was merged to TM as I understand it) and this 250+ threads behaviour happens there as well.
In a debugger, can someone look at what all those threads are doing, please?
Here's a Windbg log of the crash... interesting stuff :)
That's unfortunately not useful -- you don't actually have symbols for any of the gecko files, so it's just picking nearest exported symbols. (*** ERROR: Symbol file could not be found. Defaulted to export symbols for C:\Users\Barak\Downloads\firefox\TM After merge\01092010\firefox-4.0b5pre.en-US.win32\firefox\xul.dll -) If these are TM nightlies, do we know that TM nightlies push their symbols to the symbol server?
This is a Windbg log of the latest nightly, it doesn't crash but I did a break in the middle of gameplay so you could see all the background threads...
Here are backtraces for the first dozen threads or so. There are hundreds more threads omitted there. The active thread is thread 1.
Hmm. Those last several threads with the ThreadFunc thing going on bother me. What about breakpointing in thread creation and seeing who does it?
In particular, nsThread::Init.
OK, here's a backtrace of all hundreds of threads when the crash occurs. Thread 1 is the active thread (grrr, it is the only thread not to have debug info... does that mean that it's JITted code?)
Attachment #471225 - Attachment is obsolete: true
OK, so I have let quake2 run until it starts getting crazy, interrupted it, set a breakpoint on nsThread::Init, and got some traces, see attachment.
> does that mean that it's JITted code?) Seems likely, yes. Or just a busted stack. All those extra threads seem to be for media elements (<audio> or <video>). iirc we currently use 3 threads per such element... and it sounds like this page has dozens to hundreds of them?
Indeed, this page is loading a large number of <audio> elements. Is it really reasonnable to let a web page coerce us into spawning hundreds of concurrent threads? Naive question here --- I really don't know if it's OK to have hundreds of threads. Perhaps the crash and sluggishness are unrelated.
No, it's not ok for us to fail to pool and share threads. Bug on file on that? /be
(In reply to comment #102) > No, it's not ok for us to fail to pool and share threads. Bug on file on that? OK, filed bug 592833 blocking this one.
(In reply to comment #101) > Indeed, this page is loading a large number of <audio> elements. > > Is it really reasonnable to let a web page coerce us into spawning hundreds of > concurrent threads? Naive question here --- I really don't know if it's OK to > have hundreds of threads. Perhaps the crash and sluggishness are unrelated. If you disable sound does Quake2 performance improve significantly? We've been trying to rewrite the audio subsystem to remove one of the threads per media element, but that's not going to make FF4 unfortunately. We probably could do something clever to pool and reduce all the media element threads, it would be a major piece of work, and need significant testing though.
(In reply to comment #105) > Anyway, if having hundreds of threads is really the issue here, then removing > one thread per element will not fix this issue, as it will still leave 2 > threads per element... Well, if we reduce the thread count by 1 per element, we're 30% of the way to a solution. ;) But ultimately yes, we do need to pool the media element threads.
I couldn't see any easy way to disable sound either so I reverted stefan's patch to add ogg support for Firefox, since by default quake2-gwt used mp3 which Firefox doesn't support (http://code.google.com/p/quake2-gwt-port/source/detail?r=8a4cb2e062e8e08cc44c173a3725392aa001c53f has the details, if you don't feel like editing it yourself I can attach a diff). Doing this does seem (at least to my eyes) to make the game alot better (speed wise) but it still crashes when methodjit is enabled... Here's the crash report for this (unlike the previous reports this one seems to have a stack trace- has something changed in the last few hours?): http://crash-stats.mozilla.com/report/index/f74d8bd7-a903-4362-89fb-4aa562100901
(In reply to comment #107) > Doing this does seem (at least to my eyes) to make the game alot better (speed > wise) but it still crashes when methodjit is enabled... Well, the browser works without crashing just long enough to run the built-in benchmark, here are my machine's results while enabling/disabling audio: Audio enabled: 6.4-7.1 fps Audio disabled: 10.5 fps
(In reply to comment #105) > > Yes, I guess that if the large number of threads is really what causes > problems Hang on... although hundreds of audio threads isn't good, the program was working well before the JM merge, right? So either the methodjit or some other changes that were merged in along with the methodjit is the cause of the slowdown. When the methodjit is disabled is it still worse than it was before the JM merge?
(In reply to comment #111) > (In reply to comment #105) > > > > Yes, I guess that if the large number of threads is really what causes > > problems > > Hang on... although hundreds of audio threads isn't good, the program was > working well before the JM merge, right? See comment 107. Between today and the previous time that we tried Quake2-GWT, there have also been changes in Quake2-GWT, adding sound support for Firefox. So the previous time, we just were not doing any sound.
(In reply to comment #112) > See comment 107. Between today and the previous time that we tried Quake2-GWT, > there have also been changes in Quake2-GWT, adding sound support for Firefox. > So the previous time, we just were not doing any sound. Hold on... sound support for Firefox was added about two months ago (see the date of the revision I linked to) - I'm not sure if the audio threads issue is a new regression or not (I've never checked out the number of threads before), I'll try and find a regression range (if one exists) tomorrow. > Hang on... although hundreds of audio threads isn't good, the program was > working well before the JM merge, right? As far as performance before the JM merge and after it- subjectively speaking the "after JM merge" builds seem to have pauses every second or so while the "pre JM merge" ones doesn't seem to have that. I've tried to use the built-in benchmark as my guide but it seems to fluctuate too much to be a decent tool for measuring these changes (at least on my system).
I am working on some patches to add debugging prefs that disable the ICs (inline caches) used by the methodjit. I'll let you know when they are available in TM nightlies. Seeing if the JM-related problems still occur with those pref'd off will give us some clues about what's going on.
(In reply to comment #114) > I am working on some patches to add debugging prefs that disable the ICs > (inline caches) used by the methodjit. I'll let you know when they are > available in TM nightlies. Seeing if the JM-related problems still occur with > those pref'd off will give us some clues about what's going on. Actually, that plan looks like it would be more annoying and crufty than I expected. I'm just going to make a few special builds and do some testing myself.
(In reply to comment #113) > (In reply to comment #112) > > See comment 107. Between today and the previous time that we tried Quake2-GWT, > > there have also been changes in Quake2-GWT, adding sound support for Firefox. > > So the previous time, we just were not doing any sound. > > Hold on... sound support for Firefox was added about two months ago (see the > date of the revision I linked to) Until yesterday, I was using a revision of Quake2-GWT dating back to June. Basically, I hadn't updated since I wrote my HOWTO (comment 44). Also, yesterday was the first time that I could hear any sound.
(In reply to comment #110) > (In reply to comment #107) > > Doing this does seem (at least to my eyes) to make the game alot better (speed > > wise) but it still crashes when methodjit is enabled... > > Well, the browser works without crashing just long enough to run the built-in > benchmark, here are my machine's results while enabling/disabling audio: > Audio enabled: 6.4-7.1 fps > Audio disabled: 10.5 fps You could do the following to improve sound perf: 1. Convert the (short) sounds into WAV files. I assume they're mostly short SFX files? The Wave format isn't compressed, so is much less compute intensive to playback. The wave backend also requires fewer threads per media element FWIW. 2. Set the <audio> element's 'preload' attribute to 'none'. This means no network traffic will occur until the first time the sound is loaded. If you're creating 100 audio elements at startup to preload all the sounds (including the ones you're not going to use), this will also put a strain on your HTTP connection limit. Incidentally, WAV files require fewer connections than Ogg to load even if you've not set preload='none'. The above measures will probably help much more than us refactoring the media decoders to have fewer threads. I discussed this today with the other guys in Auckland, reducing the thread count won't reduce the amount of work we need to do to play back or preload multiple media files concurrently. We're pretty careful to put the threads to sleep when we're not using them. We'd just be swapping the overhead of thread context switches for event dispatching overhead (or however we implemented it). Reducing the thread count would reduce memory overhead (particularly on Linux), so is still worth doing however.
You need to log in before you can comment on or make changes to this bug.