OpenGL Immediate Mode performance question

Started by
9 comments, last by Bacterius 12 years ago
I've been playing about with OpenGL for a while now, but have never bothered to use anything other than immediate mode until now. I constructed the quick test code below to test the performance of immediate mode, but found that drawing 1000 textured quads to the screen yields a framerate of only ~30fps according to fraps.

What is going on? I remember reading on here once that Quake 2 use immediate mode and rendered something like 10000 triangles per frame!

My computer is a couple of years old but has an i5 and a 4890. Why isn't my code below running at a higher frame rate?



void display()
{
glClear(GL_COLOR_BUFFER_BIT);

glLoadIdentity();

glBegin(GL_QUADS);
for(int i=0; i<1000; i++)
{

glTexCoord2f(0.0f, 0.0f); glVertex2f(0.0f, 0.0f);
glTexCoord2f(1.0f, 0.0f); glVertex2f(screenWidth, 0.0f);
glTexCoord2f(1.0f, 1.0f); glVertex2f(screenWidth, screenHeight);
glTexCoord2f(0.0f, 1.0f); glVertex2f(0.0f, screenHeight);

}
glEnd();
glFlush();
}


Many thanks
Advertisement
I'm not exactly sure how many quads the software I work on for a living renders (in immediate mode), but I'm sure is more than your 30k/sec, so you should be able to go a lot faster than that, too! smile.png

You did compile this with a "release" configuration, didn't you? "Debug" really is not good for immediate mode...

Plus, your quads are rather large, try rendering smaller ones. I'm pretty sure quake 2 couldn't render 10k full-screen quads, because this is really bad for the fill rate.

And finally: Have you actually TRIED to render more than 1000 Quads per frame? After all, it may have something to do with glFlush() waiting for vertical sync, and you'd get away with rendering another 9000 quads "for free"...

Regards,
Tilmann
Are you in release or debug mode (debug might slow it down or something)?

Did quake use display lists if they even existed at that time >_>?

o3o

If the name of the variables implies that you are drawing 1000 full screen quads per frame, then I'm not at all surprised about the low frame rate. While, as your example, Quake 2 can draw ten times the number of quads, you'll also find that they are much less than one tenth the size of your full screen quads.
Try drawing them all in one draw call, instead of one thousand draw calls. Draw calls are slow - Games render 2,500 triangle meshes by drawing them in one call, and by drawing them instanced (where the graphics card draws the mesh at many different places and with many different variables in one draw call).

And that's how some RTS's can get hundreds or thousands of soldiers on-screen. So, something to look into. :)
Thanks all for your advice, I never new about debug running slower!

Thanks all for your advice, I never new about debug running slower!


Just out of curiosity, what performance did you get with a release build ?

Also, what resolution are you rendering at ? as the quads cover the full screen fillrate will be an issue at high resolutions even on a modern GPU. (at 30fps and 1000 fullscreen quads per frame you're basically redrawing the full screen 30000 times per second, at 800x600 that would be 14.4 billion pixels per second.
The radeon HD 4890 has a fillrate of around 13.6 billion pixels per second(according to techreport) so just below 30 fps is pretty reasonable for what you are doing if that is the resolution you're using.
[size="1"]I don't suffer from insanity, I'm enjoying every minute of it.
The voices in my head may not be real, but they have some good ideas!
It's not immediate mode that's killing you, it's overdraw. Quake may have used immediate mode, but it only had 3x/4x/5x or thereabouts overdraw max for any given pixel; you've got 1000x overdraw for the entire screen.

You also seem to be using a single-buffered context, which is not going to be the best option on modern hardware. You may get a slight improvement by going double-buffered, but it will be slight as it's not your primary bottleneck.

Direct3D has need of instancing, but we do not. We have plenty of glVertexAttrib calls.


[quote name='altich88' timestamp='1333481040' post='4927967']
Thanks all for your advice, I never new about debug running slower!


Just out of curiosity, what performance did you get with a release build ?

Also, what resolution are you rendering at ? as the quads cover the full screen fillrate will be an issue at high resolutions even on a modern GPU. (at 30fps and 1000 fullscreen quads per frame you're basically redrawing the full screen 30000 times per second, at 800x600 that would be 14.4 billion pixels per second.
The radeon HD 4890 has a fillrate of around 13.6 billion pixels per second(according to techreport) so just below 30 fps is pretty reasonable for what you are doing if that is the resolution you're using.
[/quote]

It was 800x600, *applauds the maths*. As for performance it doubled to around 60fps in release mode. Thanks for the info on overdraw etc, I've never tried that test before so was a bit worried something was wrong with my GPU but if it seems in the normal range that's cool.

I never new about debug running slower!

A debug build typically does not use a lot of the optimizations your compiler is able to offer, and also does extra work to help you with trouble-shooting any problems that might occur. In some situations there can be some value to turning on certain optimizations or trying to make your debug build run faster, but as a general rule you should not concern yourself with debug performance.

- Jason Astle-Adams

This topic is closed to new replies.

Advertisement