Graphics performance just (r4474) increased by 66%

[tomreyn]

Thanks to the work of Softcoder, Will, JohnDH and Silnarm (listed in reverse chronological order ), r4474 renders faster by a magnitude of two thirds (+66%) compared to 3.7.1.

We will yet need to verify that this happens without loosing anything important regarding unit, structure and tileset animations, but during initial tests I could not observe any such issues.

So this seems to be a major achievement, especially for us users plagued with slow graphics performance (due to slow graphics hardware).

Quoting my updated table from this thread, running the "benchmark" scenario using...

Code: [Select]

./megaglest --load-scenario=benchmark --fullscreen=true
...and unmodified INIs on my atibox (specs below):

Game version	UpdateFPS	RenderFPS
3.7.1	40	30
r4409	39	29
r4414	39	30
r4430	39	30
r4470	40	30
r4474	40	49

Sadly this doesn't make a difference for intelbox:

Game version	UpdateFPS	RenderFPS
3.5.0	39	18
3.5.2	40	17
3.6.0.3	39	18
3.7.1	40	18
r4362	39	19
r4470	39	19
r4474	40	19

My configuration files on these systems are the same with a single exception, based on --show-ini-settings output: intelbox uses the selectbuf selection mode, while atibox uses color picking.
Switching intelbox to color picking makes no difference to selectbuf on this system running r4474.
Switching atibox to selectbuf (which breaks unit selection there) makes quite the difference on this system running r4474: UpdateFPS 40, RenderFPS: 29.

So apparently whether the recent improvements are noticeable to you is very much a matter of which hardware and driver you are using, as well as your choice of the selection method.

[MuwuM]

r4470	40	30
r4470	40	49

???????????????????

Why has r4470 two value such different

[filux]

In my opinion last row is r.4474 as says the topic's name

[tomreyn]

Right, I had a typo there, that's fixed now. Thanks for pointing it out.

[Pizza90]

I can confirm the improvements (full screen), not so drastic though:

3.7.1   render: 40   update 22
r4472  render: 39   update 24
r4474  render 40:   update 27

+5 fps!

my pc specs:

Processor: AMD Athlon X2 Dual-Core QL-64 2.10 GHz
Video Card: ATI 3200
Installed memory (RAM): 4gb (3,75gb usable)
Operative system: Windows 7 professional N, 64bit Service Pack 1

[MuwuM]

render: 170 fps vs. 230 fps ... + 60 for me

Win 7 64 bit - 16 GB RAM - Nvidea Geforce GTX 570  - intel i7-2700k (4[8 virtual] x 3.5 GHz)

[tomreyn]

Please also state which selection mode you're using, and be sure to list the revision / version you are referring / comparing to.

[will]

Glad it worked out!

Has anyone run all the models through g3dhack (there's a command-line mode so it can be automated as I recall?) and seen how that helps too?

Mr War's models have always been super-inefficient as-exported and g3dhack has always worked wonders.

[tomreyn]

I assume there is no non-C# rewrite of it, is there?
I guess one could run it in a VM, or try to make it work with Mono. But I really like Stereo and high fidelity.

[will]

(I always used to run it with mono.)

[jammyjamjamman]

Sadly, my hardware doesn't seem to reflect these large performance improvements.

Processor: Intel-i5-3210m

My video settings:

Code: [Select]

Resolution: 1600x900

Filter: billinear

Unit Selection type: SelectBuffer (nvidia)

Shadows: projected

Shadow texture size: 512

3d Textures: enabled

Max Lights: 8

Unit Particles: enabled

Tileset Particles: enabled

Animated Tileset Objects: ∞

Texture Compression: disabled

Rain Effect (Menu/ Game): enabled/ enabled

Enable Video Playback: enabled

Results from using benchmark and taking an average from 3 games:

Version	Nvidia-gt630m (using primus)	Nvidia-gt630m	Intel-HD-4000	Intel-HD-4000
	Updatefps	Renderfps	Updatefps	Renderfps
3.7.1 (rev3948)	40	76	40	65
3.8.0-beta1 (rev4518)	40	82	40	68

This shows only a 7.9% increase in fps for the Nvidia-gt630m card and a 4.6% increase in fps for the Intel-HD-4000.

[tomreyn]

Well, at least some improvement. ;-)
My enjoyment was premature, stemming from the fact that some models were crippled which allowed for the (then) large FPS improvements. Once this was corrected, the results were much closer to what the used to be - but there's still an improvement for some graphics cards, as you were able to reproduce, too.

Will: I don't like Mono, but I like it better than Windows. Could you explain how you made g3dhack work with it, reliably?

[will]

Tomreyn, I think the performance improvements are still possible.

I believe the bug when you saw the speedup was from joining selectable meshes with unselectable ones.  And the fix was to stop doing this, meaning reverting to the previous slow behaviour.

If so, then its easy to fix in a way that maintains the performance improvement instead.  Still join the meshes, but put the unselectable vertices at the end; and instead of a boolean selectable flag, have a count of selectable vertices.

I won't be around nearly often enough to keep reminding titi and softcoder, but hopefully you can be on the case if you want to get back the performance improvements Tomreyn

https://forum.megaglest.org/index.php?topic=8560.msg88133#msg88133