I tried cobalt browser on our platform (arm-v8 linux), and I found the key response is a little bit slow. When pressing up/down key on the youtube home page, it takes 200ms-400ms from InjectKeyboardEvent to DoLayoutAndProduceRenderTree, and DoLayoutAndProduceRenderTree takes 100ms-150ms then starts to render. I saw movie rows start drawing after 500ms. Any suggestion for this?
I tried to change javascript_engine from mozjs to javascriptcore, but the performance was similar. (I found javascriptcore not enable JIT by default?) Also we run cobalt PC version, and the key response is similar, delayed some time then start scrolling.
Does Cobalt has any performance measurement for developers to check this? how to enable it?
Thanks a lot.
Non-gold builds of Cobalt have a "debug console" HUD that can be used to display live-updating debug values, including performance measurements. qa is the fastest build type that still has the debug console.
The debug console HUD can be enabled with the --debug_console=hud command-line flag, or toggled at runtime by pressing Ctrl-O.
The HUD is an overlay that displays a bunch of "CVals" (console values). These can be shown or hidden in the debug console itself (type help). In particular, you probably want to enable Event.Duration.MainWebModule which will be updated every time you hit a key. Take a look at src/cobalt/browser/web_module_stat_tracker.cc for a description of the event-timing CVals.
There is also a build target layout_benchmarks that tests the non-Javascript portions of a full layout. This is not quite the same as an incremental layout, and clearly Javascript (InjectKeyboardEvent) is a majority of your key-handling cost, but it may be a decent proxy for overall performance, and is reasonably comparable between platforms.
JavaScriptCore will soon be deprecated in favor of SpiderMonkey (a.k.a. mozjs), and removed from the Cobalt tree, so it's not an avenue of exploration with much future.
Note that Linux X11 Cobalt is not necessarily as fast as device platforms, as the X11 implementation is not particularly optimized. MesaGL is software-rendered, for example. Video composition is done fairly crudely, and YUV conversion is also done on the CPU.
Related
I am creating a three.js powered website, and on some browsers (I am looking at you firefox), if other tabs are also running webGL, my performance stutters.
I would like to know if there is a way to find out (in the browser) if other tabs are consuming webgl services so that I can alert the user.
I appreciate any and all comments!
That would be a security violation, so no, you can't do that.
Update:
I'll just add, you could include stats.js (more likely, you'll need to do something very similar to establish to what stats.js is doing to get an idea of what the average frame rate is like and look for dips in that performance), and if that is regularly dropping then alert the user. That said, you are likely to get the calculation wrong, and there are always many variables you can't control which can affect performance. Most of those can be resolved with a browser restart, particularly Firefox doesn't seem to be releasing its GPU memory across page reloads. When that memory is full the performance drops badly.
In any case, any warnings you give out should probably not be intrusive for the user in any way.
Also note that properly written WebGL programs (using requestAnimationFrame as intended) should to my knowledge not be consuming much in compute resources when the tab is in the background, though this may also vary per browser. But a tab in the background will still consume the memory (GPU memory, and JavaScript code and objects).
Can someone explain me why TWebBrowser control is working so slow on all XE editions of Delphi including XE5 and possibly XE6? To test this you need to create a new Delphi project and put TWebBrowser control in it. On form show event, navigate to this website:
http://ie.microsoft.com/testdrive/Performance/setImmediateSorting/Default.html
Please test this on Windows 7 or later. When navigation is complete, run setImmediate test and watch the results. It will take huge amount of time to complete the test. It will take about a minute to finish this.
When you open true Internet Explorer browser and do the same thing - test will be completed instantly (~200 miliseconds).
Some additional wierd informations:
When you recreate this procedure on old versions of Delphi (Delphi 7 to be precise) the web-control works as fast as it should be working and test is completed instantly. But the HTML5 speed test will still works slow (alternative test on this page).
Another weird thing is, the same slow behavior can be seen on C++ Builder but not in Visual Studio products. Is Microsoft deliberately slowing down the TWebBorwser in Embarcadero products?? I can't belive this.
I was trying to overcome this problem with diffrent methods such as:
Trying different feature options in registry such as:
FEATURE_GPU_RENDERING,
FEATURE_BROWSER_EMULATION (11001),
FEATURE_ALIGNED_TIMERS (undocumented option),
FEATURE_ALLOW_HIGHFREQ_TIMER (undocumented option),
Setting timerBeginPeriod(1) - no effect.
Please, if someone have any clue how to fix this issue - share this information with me.
UPDATE1
I made standalone test app if anyone cares. It can be downloaded here: http://mp.org.pl/download/ietest.zip It contains source and exe app with htm file. HTM file contains some js procedure that works 10 times faster in standalone IE than in TWebBrowser control. It uses setImmediate as a test (the same procedure used in test described above). But it can be easier for testing this way.
I can also see the behavior described (in your original post and in the comments). I have a few thoughts, but not necessarily an answer.
One should expect some difference in performance between the WebBrowser control and IE, in part because your Delphi app will need to build in support for certain features/APIs that IE supports out of the box.
For example, the WebBrowser control fires notifications related to tabbed browsing (old, but relevant), but it does not intrinsically handle those notifications or update the UI. You have to respond to the notifications and draw the tabs yourself. By default, IE is hardware accelerated and uses certain Windows APIs that may not be directly supported by Delphi's VCL (for resource/performance) reasons. (Hardware acceleration could account for some of the performance differences you've noticed.)
(And, for the record, I don't believe a list of differences between IE and the WebBrowser control was ever documented. I certainly don't remember seeing one in the portfolio.)
Also, the default values for various feature controls vary between IE and applications hosting the WebBrowser control. Part of the reason for this stems from the idea that IE needs to highlight performance over compatibility whereas applications generally need to emphasize compatibility over performance. You may wish to review the feature control reference to see if there are other FCKs you need to enable for your app.
Second, your loops are very tight, perhaps too tight. You've got one request piling on earlier requests and you're not really leaving much room for processing, even with setImmediate. (IIRC, we're not really supposed to use anything smaller than 250ms for setInterval without risking performance hits from the sheer number of requests.) The remarks in the setImmdiate ref. page provide some guidance, as does this article on requestAnimationFrame.
One reason why dragging the window appears to improve performance may due to the priority of window drag repaint requests. They may be forcing your loops to hold long enough (or even break) to allow other events to process. Hard to say without with tracing the system with a debugger.
Have you ever had to add application.processMessages() to your Delphi apps in order to allow the system a chance to handle the work you've already assigned? A similar need may be coming into play given the nature of your test.
Performance testing and timing is a tricky thing. You need to make sure the test isn't imposing so much overhead that it interferes with the actual work you're trying to perform.
Finally, there were some questions about the document mode of the page as it's loaded into your project. When I first started messing around with your sample, I couldn't get project4 to load slowtest.html in anything other than IE5 quirks mode (notoriously slow). Here's what eventually started working for me:
<!DOCTYPE html>
<!-- saved from url=(0023)http://www.contoso.com/ -->
<html>
<head>
<meta http-equiv="X-UA-Compatible" content="IE=edge"/>
<script type="text/javascript">
...
(Note, I deleted your initial doctype declaration and rewrite it to resolve a syntax error that was being reported by the F12 tools debugger.)
A few style key points here:
I used a mark of the web to load the page in the Internet zone. I find this makes it easier to load the page in edge mode, as pages in the intranet zone are loaded in compatibility view by default (unless you map the zones differently).
The x-ua-compatible header needs to be one of the first in the head block. It can follow title, but not much else.
Stylistically, elements need to be specified in lower case these days. There's a possibility that not following the current conventions forces the parser to fall back to an earlier rendering that supports the conventions.
Once I was able to control the documentMode at runtime, I found results I expected: older document modes ran more slowly. I also found that using requestAnimationFrame instead of setImmediate led to even better performance, but also surfaced the timing issue almost immediately.
In the end, this may be a case where the test highlights a problem, but not necessarily one you're trying to solve. (Insert Inigo meme here.) I get that you're trying to resolve a bottleneck. Are you sure you've found the correct bottleneck?
You may not be able to replicate the same performance of the native browser, but perhaps you can refactor the code to perform adequately without the extra overhead? Is there anything that might be better handled with a worker or some other implementation technique?
Hope this helps...
So far, I have always tested the performance (i.e. "smoothness") of my iOS user interfaces informally, by testing the user interface myself. This is obviously not a very accurate way to profile the performance, so I wondered whether there were some methods / tools that are designed to do this. Are there?
Use the Instruments tool 'Core Animation' to measure graphics (and thus UI) performance. Mostly in the form of frame rate (which is a formal way of measuring smoothness), but you can also configure it to show overlapping and blended views (which your GPU absolutely hates).
Also, there are some great WWDC sessions available for iOS developers on this topic.
I want my program to be able to launch any Windows game, and while the user is playing it, intermittently display some text or pictures in some part of the game window. The game may be in windowed or full-screen mode. From what I have been able to figure out from online resources, this could be done using a graphics library that supports overlays and using Windows Hooks to keep track of the target application's window. In this context I have some questions.
Will the overlays affect the game's performance?
How will hooking the application affect performance?
Is there any other way one could achieve this? For example, how do you think PIX, the DirectX debugging and analysis tool, work?
Fraps is the archetypal example of doing this sort of thing to a fullscreen DirectX application from a third-party app. It works by hooking some system calls and inserting itself into the call-chain between an app and DirectX. There is some performance hit, but in general its minimal.
This page seems to have some details and sample code on how to hook the app in this way.
If I recall correctly, from other forum discussions (can't find the link at the moment. search for things like "how does fraps work", it's a popular question), Fraps hooks a few things to force the app to load its DLL, then hooks Present() calls and executes a device->Clear() call before calling the real Present(), with a list of small rectangles to set to a different color, which can spell out the FPS number that it displays. This has a minimal performance impact and is widely compatible with whatever rendering the app is doing. Overlaying a bitmap would be more complicated since it wouldn't be as easy to do at Present-time. Perhaps if you could hook EndScene, then you could do more, but you would have to be careful to not change the device state.
PIX has privileged access to the DirectX driver, so I wouldn't expect to be able to use that as a model to emulate.
If an the target app is running in windowed mode, hooking DirectX still work, but you could also just use GDI instead.
Edit: I think this is the link I was originally thinking of.
When writing DirectX applications, obviously it's desirable to support the user suspending the application via Alt-Tab in a way that's fast and error-free. What is the best set of practices for ensuring this? Things that need to be addressed include:
The best methods of detecting when your application has been alt-tabbed out of and when it has been returned to.
What DirectX resources are lost when the user alt-tabs, and the best ways to cope with this.
Major things to do and things to avoid in application architecture for purposes of alt-tab support.
Any significant differences between major DirectX versions as they apply to the above.
Interesting tricks and gotchas are also good to hear about.
I will assume you are using C++ for the purposes of my answers, but if you can afford to use C#, XNA (http://creators.xna.com/) is an excellent game platform that handles all of these issues for you.
1]
This article is helpful for windows events in the window procedure to detect when a window loses or gains focus, you could handle this on your main window: http://www.functionx.com/win32/Lesson05.htm. Also, check out the WM_ACTIVATEAPP message here: http://msdn.microsoft.com/en-us/library/ms632614(VS.85).aspx
2]
The graphics device is lost when the application loses focus from full screen mode. Microsoft offers an article on how to handle this: http://msdn.microsoft.com/en-us/library/bb174717(VS.85).aspx This article also has a lost device tutorial: http://www.codesampler.com/dx9src/dx9src_6.htm
DirectInput can also have a device lost error state, here is a link about that: http://www.toymaker.info/Games/html/directinput.html
DirectSound can also have a device lost error state, this article has code that handles that: http://www.eastcoastgames.com/directx/chapter2.html
3]
I would make sure to never disable Alt-Tab. You probably want minimal CPU load while the application is not active because the user probably Alt-Tabbed because they want to do something else, so you could completely pause the application, or reduce the frames rendered per second. If the application is minimzed, you of course don't need to render anything either. After thinking about a network game, my best solution is that you should still reduce the frames rendered per second as well as the amount of network packets handled, possibly even throwing away many of the packets that come in until the game is re-activated.
4]
Honestly I would just stick to DirectX 9.0c (or DirectX 10 if you want to limit your target operating system to Vista and newer) if at all possible :)
Finally, the DirectX sdk has numerous tutorials and samples: http://www.microsoft.com/downloads/details.aspx?FamilyID=24a541d6-0486-4453-8641-1eee9e21b282&displaylang=en
We solved it by not using a fullscreen DirectX device at all - instead we used a full-screen window with the top-most flag to make it hide the task bar. If you Alt-Tab out of that, you can remove the flag and minimize the window. The texture resources are kept alive by the window.
However, this approach doesn't handle the device lost event happening due to 'lock screen', Ctrl+Alt+Delete, remote desktop connections, user switching or similar. But those don't need to be handled extremely fast or efficiently (at least that was the case in our application)
All serious D3D apps should be able to handle lost devices as this is something that can happen for a variety of reasons.
In DX10 under Vista there is a new "Timeout Detection and Recovery" feature that makes it common in my experience for graphics devices to be reset which would cause a lost device for your app. This seems to be improving as drivers mature but you need to handle it anyway.
In DX8 and 9 (and 10?) if you create your resources (vertex and index buffers and textures mainly) using D3DPOOL_MANAGED they will persist across lost devices and will not need reloading. This is because they are stored in system memory and the DX runtime copies to video memory automatically. However there is a performance cost due to the copying and this is not recommended for rapidly changing vertex data. Of course you would profile first to determine if there is a speed issue :-)