Game.EndDraw take 77% of draw time

1

Hi,

I found something strange with MonoGame behavior.

I have two computers for my test:
Desktop: i7 5820k + GTX 1080
Laptop: i5 2410M + GT540M

the two computers have windows 10 and Monogame 3.5.1.1679 installed.

I have performance issue on the laptop, I think that was because of the difference of hardware perf.

but after some test I found something strange.
My project (Windows/OpenGl) act differently on this two computer.

The draw time of 3d models are about two time slower on the laptop as expected i guess.

But the strange thing is this is not the cause of my bad fps.

I measured the execution time of Microsoft.Xna.Framework.Game.DoDraw(GameTime) with Resharper, here my result:

On the desktop platform :
99.90% MyGame.Draw(GameTime)
0.09% Microsoft.Xna.Framework.Game.EndDraw

On the Laptop:
22.31% MyGame.Draw(GameTime)
77.69% Microsoft.Xna.Framework.Game.EndDraw

With the laptop 72% of the time used by Game.DoDraw are used by Game.EndDraw, I don’t know if this is normal…

I try to search a reason for this in MonoGame’s source without success.

I don’t know how to get out of this do you have an idea ?

Thanks for reading.

2

Isn’t it possible that EndDraw is waiting for vsync?

3

I have test with Vsync on and off the results are similar.

On the laptop I have like 10fps at best, VSync can’t be slower.

4

You might have some hardware setting forcing vsync that causes this. All EndDraw does is wait for vsync and then swap the buffers.

5

Draw(GameTime) sends a drawing command to the GPU driver. The GPU executes those commands asynchronously from a queue, Draw doesn’t wait for the drawing to finish.

EndDraw() wait for the GPU command queue to finish and then swaps the backbuffers to present the new frame.

GT540M takes a bit longer to draw your scene, that’s why you see some execution time shift into EndDraw(). The CPU doesn’t do anything during that time, it waits for the GPU.

That, or you forced vsync from the GPU panel.

6

Thank you Jjagg and nkast that’s very helpful!

but… if the scene run at 10fps on a screen at 60hz, VSync should not affect framerate, right ?

7

Nope, that’s not correct because it has to wait for vblank anyway, it doesnt affect much though if the (potential) framerate is a small fraction of refresh.

8

Disable vsync and you probably get more fps

9

Ok thank for the precision!

I have made more tests. with a blank scene I got more than 60fps so VSync isn’t the cause of slowdown.

I also notice that Texture2D.SetData is disproportionately slower on this laptop.

I use it to create 1 by 1 Texture2D to be able to draw rectangle for UI.

Texture2D texture = new Texture2D(graphicsDevice, 1, 1)
texture.SetData(new { color });
Even if the pixels was cached in memory this code was so slow that the application was completely unusable.
More than 1 second to set only one pixel color.

I replace it by this code:

RenderTarget2D pixel = new RenderTarget2D(graphicsDevice, 1, 1);
graphicsDevice.SetRenderTarget(pixel);
graphicsDevice.Clear(color);
graphicsDevice.SetRenderTarget(null);
With this code all seams to works good.

It is possible that something are missing on those laptop and cause this performance issues ?

10

Setdata is incredibly slow on any platform. Do not use it per frame, only to initialize stuff if you must.

Thats by design and not monogame specific. Changing textures on graphics memory is very slow if done from CPU code

11

No, he won’t (at least by much)… at 10fps in the worst case a frame will be ready to display just a cycle after the vblank so it’ll have to wait 1/60s (-1ht) and that won’t happen a lot of times unless there’s a huge std dt deviation, so if he’s getting 10fps with vsync on, by disabling it he may get say 11fps if he disabled buffering.

That profiling result doesn’t seem correct, but if you’re having half the framerate of a desktop GTX1080 in a GT540M, I’d say that’s somewhat expected in terms of raw performance. I mean, I’d expect the former to be at (very) least 2x as fast as the latter, especially because these mobile cards are nothing special really.

12

In terms of raw performance it’s 250 gflops vs 8000 gflops, the 1080 is around 32x faster in that metric.

Vsync below 30 fps is working very poorly in my experience and the monogame timestep implementation can have horrible results if targets can not be hit.

I’d still recommend trying to see how the laptop version performs without vsync

13

Yes, very well said. I’d also recommend checking it without vsync just in case, because you’re right in that regard that sometimes vsync can cause unexpected behaviours, for example I’ve seen recently a strange case in which framerate in windowed mode would always stick to a fraction of the refresh. That being said I really doubt it’s the case and if the test is GPU-bound, at least partially, then that could very well explain the difference. Settings like frame buffering size and in nvidia cards the pre-rendered frames can also cause a difference in benchmarks by releasing the CPU spinning while accepting frames by overriding vblank signaling even when it’s not buffered yet. These nvidia cards have some pretty cool tricks that could render benchmarks useless.

14

On nvidia drivers the vsync is set by default and overrides any app’s option about that. Look in the nvidia control panel to disable it by default.

But is seems the algorithm is not the best one :wink:

15

Thank you for your answer.

VSync is OFF, I have got more than 60fps with an empty scene.

I know my engine isn’t optimized yet, I use multiple material per mesh that means I need to change the effect parameters multiple time for each object. Maybe the 540M is just too slow to handle this amount of Effect.Apply…

16

How many objects/materials have you each frame ?

I have a nvidia620 and in cases with very numerous objects, and numerous materials, it drops FPS to 10 or below.
On a 870M it runs fine at 120fps. Maybe use a “lower” profile when on the 540: in my engine, when coding with the 620, i use lower specs (LOD, number of details etc aggressively lowered) to not wonder about that.
But it helps to see what happens when a user has a low end computer and test to reproduce their problems, or better, improve algorithms and optimize code.

17

My test scene have 266 models.

image.png1280×720 1.53 MB

A computer have in average 12 models, each models have 3 to 10 materials (5 in average).
Only motherboard and certain graphic card have more than 6 materials
The most common: Drive, Ram and Fan have only 2 or 3 materials

All textures are atlased in the same file and all normal map as well.
…

On the GTX1080 I can run this scene at 66fps with a 2160p resolution
On the GT540M I have 8 to 10 fps in 720p

18

Did you do a profiling with a tool to narrow the bottleneck?
It could be the way you do culling, sort, etc

19

Yes that is what I try to do right now.

if I scope the profiling on Game.Draw i got this:

image.png840×587 227 KB

This profiling has been done on my desktop computer.

Effect127 is the effect used on all my models.

WorldModel is a class I use for containing Model,position,rotation, hitbox, etc… for each object.

3.1% WorldModel.EffectParam(…) apply the material, world, view ect to effect.

1.3% WorldModel.DrawMesh(…) is a slightly modified version of Microsoft.Xna.Framework.Graphics.ModelMesh.Draw()

5.7% WorldModel.Refresh(…) is basically this in pseudo code:

foreach( mesh in this Model) {
GetTheMaterial
ApplyMaterialAndMatrix // with WorldModel.EffectParam(…)
DrawMesh // with WorldModel.DrawMesh(…)
}

5.4% Game_Bank.GetTexture return one of the two texture showed in my previous post with coordinate of the needed area.

20

Some ideas coming…
Do u set effect parameters even when they havent not changed?
Do u cache them or call them with parameter[thestring]?
Have you tried sorting the meshes by material to lower the calls?
Maybe MRT could help by drawing the 6 materials in one for loop. And then combining the results.
Or 2 loops of 3 if the hardware does not support up to 8