For educational purposes I need to compare the performance of WebGL with OpenGL. I have two equivalent programs written in WebGL and OpenGL, now I need to take the frame rate of them and compare them.
In Javascript I use requestAnimationFrame to animate, and I noticed that it causes the frame rate to be always at 60 FPS, and it goes down only if I switch tab or window. On the other hand if I always call the render function recursively, the window freezes for obvious reasons.
This is how I am taking the FPS:
var stats = new Stats();
stats.domElement.style.position = 'absolute';
stats.domElement.style.left = '450px';
stats.domElement.style.top = '750px';
document.body.appendChild( stats.domElement );
setInterval( function () {
stats.begin();
stats.end();
}, 1000 / 60 );
var render= function() {
requestAnimationFrame(render);
renderer.render(scene,camera);
}
render();
Now the problem if having always the scene at 60 FPS is that I cannot actually compare it with the frame rate of OpenGL, since OpenGL redraws the scene only when it is somehow modified (for example if I rotate the object) and glutPostRedisplay() gets called.
So I guess if there is a way in WebGL to redraw the scene only when it is necessary, for example when the object is rotated or if some attributes in the shaders are changed.
You can't compare framerates directly across GPUs in WebGL by pushing frames. Rather you need to figure out how much work you can get done within a single frame.
So, basically pick some target framerate and then keep doing more and more work until you go over your target. When you've hit your target that's how much work you can do. You can compare that to some other machine or GPU using the same technique.
Some people will suggest using glFinish to check timing. Unfortunately that doesn't actually work because it stalls the graphics pipeline and that stalling itself is not something that normally happens in a real app. It would be like timing how fast a car can go from point A to point B but instead of starting long before A and ending long after B you slam on the brakes before you get to B and measure the time when you get to B. That time includes all the time it took to slow down which is different on every GPU and different between WebGL and OpenGL and even different for each browser. You have no way of knowing how much of the time spent is time spent slowing down and how much of it was spent doing the thing you actually wanted to measure.
So instead, you need to go full speed the entire time. Just like a car you'd accelerate to top speed before you got to point A and keep going top speed until after you pass B. The same way they time cars on qualifying laps.
You don't normally stall a GPU by slamming on the breaks (glFinish) so adding the stopping time to your timing measurements is irrelevant and doesn't give you useful info. Using glFinish you'd be timing drawing + stopping. If one GPU draws in 1 second and stops in 2 and another GPU draws in 2 seconds and stops in 1, your timing will say 3 seconds for both GPUs. But if you ran them without stopping one GPU would draw 3 things a second, the other GPU would only draw 1.5 things a second. One GPU is clearly faster but using glFinish you'd never know that.
Instead you run full speed by drawing as much as possible and then measure how much you were able to get done and maintain full speed.
Here's one example:
http://webglsamples.org/lots-o-objects/lots-o-objects-draw-elements.html
It basically draws each frame. If the frame rate was 60fps it draws 10 more objects the next frame. If the frame rate was less than 60fps it draws less.
Because browser timing is not perfect you might be to choose a slightly lower target like 57fps to find how fast it can go.
On top of that, WebGL and OpenGL really just talk to the GPU and the GPU does the real work. The work done by the GPU will take the exact same amount of time regardless of if WebGL asks the GPU to do it or OpenGL. The only difference is in the overhead of setting up the GPU. That means you really don't want to draw anything heavy. Ideally you'd draw almost nothing. Make your canvas 1x1 pixel, draw a single triangle, and check the timing (as in how many single triangles can you draw one triangle at a time in WebGL vs OpenGL at 60fps).
It gets even worse though. A real app will switch shaders, switch buffers, switch textures, update attributes and uniforms often. So, what are you timing? How many times you can call gl.drawBuffers at 60fps? How many times you can call gl.enable or gl.vertexAttribPointer or gl.uniform4fv at 60fps? Some combination? What's a reasonable combination? 10% calls to gl.verterAttribPointer + 5% calls to gl.bindBuffer + 10% calls to gl.uniform. The timing of those calls are the only things different between WebGL and OpenGL since ultimately they're talking to the same GPU and that GPU will run the same speed regardless.
You actually do not want to use framerate to compare these things because as you just mentioned you are artificially capped to 60 FPS due to VSYNC.
The number of frames presented will be capped by the swap buffer operation when VSYNC is employed and you want to factor that mess out of your performance measurement. What you should do is start a timer at the beginning of your frame, then at the end of the frame (just prior to your buffer swap) issue glFinish (...) and end the timer. Compare the number of milliseconds to draw (or whatever resolution your timer measures) instead of the number of frames drawn.
The correct solution is to use the ANGLE_timer_query extension when available.
Quoting from the specification:
OpenGL implementations have historically provided little to no useful
timing information. Applications can get some idea of timing by
reading timers on the CPU, but these timers are not synchronized with
the graphics rendering pipeline. Reading a CPU timer does not
guarantee the completion of a potentially large amount of graphics
work accumulated before the timer is read, and will thus produce
wildly inaccurate results. glFinish() can be used to determine when
previous rendering commands have been completed, but will idle the
graphics pipeline and adversely affect application performance.
This extension provides a query mechanism that can be used to
determine the amount of time it takes to fully complete a set of GL
commands, and without stalling the rendering pipeline. It uses the
query object mechanisms first introduced in the occlusion query
extension, which allow time intervals to be polled asynchronously by
the application.
(emphasis mine)
Related
I have an animation loop run by requestAnimationFrame.
On my computer, the time between frames obtained this way are more or less 16.667ms corresponding to 60fps. However, https://developer.mozilla.org/en-US/docs/Web/API/window/requestAnimationFrame states that it doesn't necessarily run at a fixed speed and may try to match the device screen refresh rate which could be 90 or 120 or more these days.
Now what I want to do is to measure the time between frames and simplify the animation on low performance devices if the hardware can't keep up, in pseudocode:
if (elapsed_frame_time > (1000 / magical_function_to_get_the_target_fps_of_requestAnimationFrame()) {
animation.particles *= 0.9;
}
I suppose I could just measure it over a few empty requestAnimationFrame ticks to get the approximate nominal time but is there anything in the standard library that would just tell me what the interval or fps should be?
No there is nothing that let us know the "optimal" frequency of requestAnimationFrame (rAF).
Even measuring it by checking a few rounds is not an exact science. You may very well be measuring while an other app is eating all the available process of the device, your user may be scrolling the page which may have dramatic influence on rAF, you will definitely have outliers and it will be near impossible to tell between a 59.9Hz monitor and a 60Hz one for instance.
So if you really have to go this route
Be sure to use an average FPS, based on a big enough sample and to ignore the first call (because browsers actually execute the callback directly from a non-animated document).
Check the callback's timestamp, since this will represent the time the monitor sent its V-Sync signal and shouldn't be influenced by other scripts on the page.
Probably wait at least a few dropped frames before intervening.
Remember that some users may have dual-monitor systems with various refresh-rates.
Expect some edge cases and browser bugs.
I've got a WebGL scene, and a set of parameters I can use to balance render quality versus speed. I'd like to display the scene to the user in as high as I can make the quality, as long as the frame rate doesn't drop below some threshold due to this. To achieve this, I have to somehow measure “current” frame rate in response to changes in quality.
But the scene is static as long as the user doesn't interact with it (like rotating camera using the mouse). I don't want to have a loop re-rendering the same scene all the time even if nothing changes. I want to stop rendering if the scene stops moving. Which means I can't simply average the time between successive frames, since I can't distinguish between the renderer being slow and the user just moving his mouse more slowly.
I thought about rendering the scene a number of times at start up, and judge the frame rate from this. But the complexity of the scene might change over time, due to the portion of the scene visible from the current camera position, or due to user interaction outside the canvas. So I have to adapt the quality as the scene changes complexity. Running a calibration loop after every mouse release would perhaps be an option.
I also thought about using the finish call instead of the flush call to accurately measure render time. But while I wait for GL to finish rendering, my application will essentially be unresponsive, in particular won't be able to queue mouse events. Since I envision the rendering to ideally take up all the time between two frames at the target threshold frame rate, that would probably be rather bad. I might get away with using finish instead of flush only on some occasions, like after a mouse release.
What's the best way to achieve a desired frame rate in a WebGL application, or to measure render time on a regular basis?
Why can't you use the average render time?
Just because you render less often, does not mean you cannot average the render times. It will just take a bit longer to get an accurate average.
Once the user starts moving their mouse you'll get an influx of data, and that should quickly give you an average render rate anyway.
I've seen a couple of questions asking about this, but they're all over three years old and usually end by saying theres not much of a way around it yet, so im wondering if anything's changed.
I'm currently working on a game that draws onto a canvas using an interval that happens 60 times a second. It works great on my iphone and PC, which has a faily decent graphics card, but I'm now trying it on a Thinkcentre with intel i3 graphics, and I notice some huge screen tearing:
http://s21.postimg.org/h6c42hic7/tear.jpg - it's a little harder to notice as a still.
I was just wondering if there's any way to reduce that, or to easily enable vertical sync. If there isnt, is there somethingthat I could do in my windows 8 app port of the game?
Are you using requestAnimationFrame (RAF)? RAF will v-sync but setTimeout/setInterval will not.
http://msdn.microsoft.com/library/windows/apps/hh920765
Also, since 30fps is adequate for your users to see smooth motion, how about splitting your 60fps into 2 alternating parts:
"calculate/update" during one frame (no drawing)
and then do all the drawing in the next frame.
And, get to know Chrome's Timeline tool. This great little tool lets you analyze your code to discover where your code is taking the most time. Then refactor that part of your code for high performance.
[ Addition: More useful details about requestAnimationFrame ]
Canvas does not paint directly to the display screen. Instead, canvas "renders" to a temporary offscreen buffer. “Rendering” means the process of executing canvas commands to draw on the offscreen buffer. This offscreen buffer will be quickly drawn to the actual display screen when the next screen refresh occurs.
Tearing occurs when the offscreen rendering process is only partially complete when the offscreen buffer is drawn on the actual display screen during refresh.
setInterval does not attempt to coordinate rendering with screen refresh. So, using setInterval to control animation frames will occasionally produce tearing .
requestAnimationFrame (RAF) attempts to fix tearing by generating frames only between screen refreshes (a process called vertical synching). The typical display refreshes about 60 times per second (that’s every 16 milliseconds).
With requestAnimationFrame (RAF):
If the current frame is not fully rendered before the next refresh,
RAF will delay the painting of the current frame until the next screen refresh.
This delay reduces tearing.
So for you, RAF will likely help your tearing problem, but it also introduces another problem.
You must decide how to handle your physics processing:
Keep it in a separate process—like setInterval.
Move it into requestAnimationFrame.
Move it into web-workers (the work is done on a background thread separate from the UI thread).
Keep physics in a separate setInterval.
This is a bit like riding 2 trains with 1 leg on each—very difficult! You must be sure that all aspects of the physics are always in a valid state because you never know when RAF will read the physics to do rendering. You will probably have to create a “buffer” of your physics variables so they always are in a valid state.
Move physics into RAF:
If you can both calculate physics and render within the 16ms between refreshes, this solution is ideal. If not, your frame may be delayed until the next refresh cycle. This results in 30fps which is not terrible since the eye still perceives lucid motion at 30fps. Worst case is that the delay sometimes occurs and sometimes not—then your animation may appear jerky. So the key here is to spread the calculations as evenly as possible between refresh cycles.
Move physics into web workers
Javascript is single-threaded. Both the UI and calculations must run on this single thread. But you can use web workers which run physics on a separate thread. This frees up the UI thread to concentrate on rendering and painting. But you must coordinate the background physics with the foreground UI.
Good luck with your game :)
I'm writing a game in javascript canvas, and I experience lag sometimes. I think this is because I draw pretty much to the canvas.
My background for example isn't static, it moves from the right to the left, so each time, I have to clear the entire canvas. And I'm not clearing with just a color, I'm clearing with an image that moves every cycle of the gameloop.
I think this is an expensive operation and I was thinking if there is something to make this operation less expensive, like for example using double buffering for clearing the background (no idea if you can use double buffering just for clearing the background?).
Or should I use double buffering in the entire game?
(I don't think so because I've read that the browser already does that for you)
Since you experience your lag 'sometimes', and it's not a low frame rate issue, i would rather move my eyes towards the evil garbage collector : whenever it triggers, the application will freeze for up to a few milliseconds, and you'll get a frame miss.
To watch for this, you can use google profiling tools, in TimeLine / Memory / press the record button : you can see that a GC occurred when there's a sudden drop in the memory used.
Beware when using this tool, that it slows down a bit the app, and it creates garbage on its own (!)...
Also, since any function call for example creates a bit of garbage, so you can't have a full flat memory line.
Below i show the diagram of a simple game i made before i optimized memory use,
there's up to 5 GC per second :
And here the diagram of the very same game after memory optimisation : there's something like 1 GC per second. Because of the limitations i mentioned above, it is in fact the best we can get, and the game suffers no frame drop and feels more responsive.
To avoid creating garbage
- never create object, arrays or functions.
- never grow or reduce an array size.
- watch out for hidden object creation : Function.bind or Array.splice are two examples.
You can also :
- Pool (recycle) your objects
- use a particle engine to handle objects that are numerous / short lived.
I made a pooling lib (here) and a particle engine (here) you might use if you're interested.
But maybe first thing to do is to seek where you create objects, and have them created only once. Since js is single-threaded, you can in fact use static objects for quite a few things without any risk.
Just one small expl :
function complicatedComputation(parameters) {
// ...
var x=parameters.x, y = parameters.y, ... ...
}
// you can call creating an object each time :
var res = complicatedComputation ( { x : this.x, y : this.y, ... ... } );
//... or for instance define once a parameter object :
complicatedComputation.parameters = { x :0, y:0, ... ... };
// then when you want to call :
var params = complicatedComputation.parameters;
params.x = this.x ; params.y = this.y; ... ...
var res = complicatedComputation(params);
It has the drawback that previous calls parameters remains, so you don't get undefined if you don't
set them, but previous value, so you might have to change bit your function.
But on the other hand, if you call several times the function with similar parameters, it comes very handy.
Happy memory hunting !
Double buffering is a technique to reduce flickering: You draw to one buffer while another buffer is displayed, and then swap them out in a single operation, so the user doesn't see any of the drawing in a partial state.
However, it does not really help for performance problems at all, as there is still the same amount of drawing. I would not try to use double buffering if you don't have a problem with flickering, as it requires more memory and flickering may be implicitly prevented by the system by similar or other means.
If you think drawing the background is too expensive, there are several things that you could look into:
Do you scale the background image down while drawing? If so, create a downscaled version once and use this for clearing the background, reducing the drawing costs per iteration.
Remember the "dirty" areas and just draw the portions of the background that were obscured. This will add some management overhead but reduce the number of pixels that need to be touched significantly
Make the background the background image of the canvas DOM element and just clear the canvas to transparency in each iteration. If the canvas is very large, you could make this faster by remembering the "dirty" areas and just clearing them.
Are you sure painting the background is the main cause of lag? Do you still have lag when you only repaint the background and do not do much else?
Currently, I am rendering WebGL content using requestAnimationFrame which runs at (ideally) 60 FPS. I'm also concurrently scheduling an "update" process, which handles AI, physics, and so on using setTimeout. I use the latter because I only really need to update objects roughly 30 times per second, and it's not really part of the draw sequence; it seemed like a good idea to save the remaining CPU for actual render passes, since most of my animations are fairly hardware intensive.
My question is one of best practices. setTimeout and setInterval are not particularly kind to battery life and CPU consumption, especially when the browser is not in focus. On the other hand, using requestAnimationFrame (or tying the updates directly into the existing render phase) will potentially enforce far more updates every second than are strictly necessary, and may stop updating altogether when the browser is not in focus or at other times the browser deems unnecessary for "animation".
What is the best course of action for updating, but not rendering content?
setTimeout and setInterval are not particularly kind to battery life and CPU consumption
Let's be honest: Neither is requestAnimationFrame. The difference is that RAF automatically turns off when you leave the tab. That behavior can be emulated with setTimeout if you use the Page Visibility API, though, so in reality the power consumption problems between the two are about on par if used intelligently.
Beyond that, though, setTimeout\Interval is perfectly appropriate for use in your case. The only thing that you may want to be aware of is that you'll be hard pressed to get it perfectly in sync with the render loop. You'll have cases where you may draw one too many times before your animation update hits, which can lead to minor stuttering. If you're rendering at 60hz and updating at 30hz it shouldn't be a big issue, but you'll want to be aware of it.
If staying perfectly in sync with the render loop is important to you, you could simply have a if(framecount % 2) { updateLogic(); } at the top of your RAF callback, which effectively limits your updates to 30hz (every other frame) and it's always in sync with the draw.