How to boost your Doom3 performance by 40% on ATI hardware

now is that the fixed one or the original one ?

 

Much too high. The sweet spot for me seems to be 1152x864, 2AA, 16AF forced by CP, LOD -2, Demiurgs latest tweak installed. Gets me around 30fps and looks great.

1024x768, 4AA, 8AF forced, tweak installed, is slower. 4AA seems to kill performance on 128MB cards for anything above 800x600.

TFT panel users with 1280x1024 native may get away with 0AA/8AF, 2AA again is too much.

 

http://img44.exs.cx/img44/2203/cptweak1.jpg vs http://img44.exs.cx/img44/157/original1.jpg. Just look at the floor to the right. It gets much worse if you set LOD to a negative value in DoomConfig.cfg, as many people do to get sharper textures.

 

uh, im still on page 23, sooo much to read....

i think its time to make an faq as people keep asking the same questions over and over and dont just read the thread to find their answers are already there.

again, big thx to john for letting us have a look into the code!

 

Re: On the topic of getting good performance with AF

I appreciate that a lot.

I've been busy with personal/beta testing/online gaming clan issues.

 

Can someone please link me to Demirug's latest tweak?

 

And can someone tell me how to enable trillinear filtering on Doom 3?

 

It's the Fixed one(or final version??)

US

 

The .x makes it do the math on only one componenet...

Hey guys, do you want more speed?

replace these three lines in Demirugs code:

Code:

SUB_SAT R1.x, specular.x, 0.75; 
MUL     R1.x, R1.x, R1.x; 
MUL_SAT R1.x, R1.x, 16.0;

With these two lines:

Code:

MAD_SAT R1.x, specular, 4.0, -3.0;
MUL     R1.x, R1.x, R1.x;

It seems that everybodybody just ignores my sugestion that I've been sugesting for a long time (but I had the two instruction switched the wrong way around at first) But I do see a small speed increase (r9600xt) with this modification, so if anybody else would test it, I would be very happy!

Note: The speed difference is very small, but consistent. The output is identical (checked by subtracting two images)

 

I believe Demirugs latest tweak is this:

Old:

Code:

# perform a dependent table read for the specular falloff 
TEX   R1, specular, texture[6], 2D; 

# modulate by the constant specular factor 
MUL   R1, R1, program.env[1]; 

New:

Code:

# perform a dependent table read for the specular falloff 

SUB_SAT R1.x, specular.x, 0.75; 
MUL     R1.x, R1.x, R1.x; 
MUL_SAT R1.x, R1.x, 16.0; 

# modulate by the constant specular factor 
MUL   R1, R1.x, program.env[1];

-----------------------------

Neeyik.. please add this to your new topic you posted for people who don't know what settings to use.

For best performance:

9600 - 9800 Series cards
AF 8 should be applied
2x AA
Application in the Control panel

X800 Series
AF 16 should be applied
4X AA
Application in the Control panel

Correct me if i'm wrong

US

 

I registered just to post my results.

P4 1.6a @ 2.133ghz
Abit IT7 MAX (i845D)
512mb DDR 266
Radeon 9800 Pro 256mb

Before Humus:
In game anisotropy 8x
first run: 30.8
second run: 36.1

Control panel forced 8x
first run: 28.0
second run: 32.8

After Humus:
In game anisotropy 8x
first run: 31.2
second run: 37.5

Control panel forced 8x
first run: 31.2
second run: 37.5

It is my theory that anyone reporting the large increases while using a 9700 or 9800 card had forced anisotropy in the control panel. I've seen a visible difference with the forced anisotropy. This would go along with the bilinear/trilinear theory. So... thanks for the fix. I now get proper anisotropy without a performance loss.

Edit: This was all done with demo1. I'm using the 4.7 Omega drivers with the 4.9beta opengl driver. And yes, the after Humus stuff has the exact same numbers. That's how it turned out.

Second Edit:
Control panel forced 16x
first run: 30.6
second run: 37.0

Holy crap man. Talk about nice. Getting the 256mb version of the 128mb has given me free 16x anisotropy in almost every game that I own (with the exception of the Battlefield games, where I need to scale it down a bit).

 

Re: On the topic of getting good performance with AF

I guess those are good reasons to be less active here

 

Thowllly, I've just tested your shader-suggestion. Unfortunately I couldn't detect any change between your shader and Demirug's on my 9800Pro, demo1 with 16x CP-Perf-AF is still unchanged 40,2fps und also my overdraw-test (had that demo still recorded) didn't change at all.

Maybe there is an improvement without AF or on low AF-levels, but 16x CP-AF is unchanged.


In addition I've made 2 comparison shots to compare AF-IQ between 8x ingame, 8x CP performance and 8x CP quality.




While there *is* a difference between quality and performance, it is much less noticable than the texture-aliasing that occurs with ingame-AF. The most noticable differences are in the background, as shown in the second shot.

But to be honest, without making screenshots every few seconds and comparing them in photoshop I'd never know if I'm just running performance- or quality-AF, but I'll always know if I'm running ingame-AF.

I couldn't verify if ingame-AF tinkers with the LOD-bias, but at least it doesn't change the corresponding cvar (image_lodbias).


btw someone concluded that because ingame-AF only affects specific textures it is so much faster. I can't agree with that one, here are a few of my results (all with Demirug's tweak) to prove my point:

8x AF ingame: 36.4fps
8x AF CP-Qual: 36.1fps
8x AF CP-Perf: 37.7fps

So if ingame-AF is so much faster because it only affects what needs to be affected, why comes that quality-AF which should affect everything is just 0.8% slower and performance-AF which also can't decide which layer to filter is actually 3.5% faster?

-BlueSteel

 

Ah, thank you for testing. I tested on a 9600xt with 8x af (in game). The difference between my and Demirugs code was very small, but it was there every time I ran the test. I might just be that the compiler optimises both into the same code, and the fps difference I saw was just random fluctuations. Anyway, thanks for taking the time to test it, I was starting to wonder if I had turned invisible or something

 

results

I've got a crazy tweaked DoomConfig.cfg file and have been using Humus' modification. I decided to do some timedemo work to see if it was making any difference on my 9500Pro. Humus's trick gave me 1fps or so improvement with no AF. THe DoomConfig.cfg tweaks gave me less than one fps improvement. I believe 0.2fps or so. Then I tried forcing different levels of AF up to 8x through the control panel. What I saw was that 8xAF came with no performance degradation, whether set to quality or performance. Quality was less than 1fps slower. But when I actually went to play the game, quality AF felt significantly slower than performance. So, I'm not really confident of how usefull the timedemo is for measuring the impact of Humus' modification or any other tweaks. Has anyone tried a program like fraps to do some analysis?

 

Timedemo is great for measuring performance, just the included Demo1 is not exactly perfect if you just want to test filtering and stuff (too much smoke etc).

I recorded my own demo where I just walked around in an area I saw fit for my purpose. To record your own demos use "recordDemo [demoname]" and "stoprecording", to benchmark use "timedemo [demoname]".

I've bound recording and stop to keys for easier access, just type "bind [key] [command]" to do this. Note that if you don't specify a name for the demo you'll always record and overwrite "demo000".

So if you want to test stuff just record a demo where you think you'll see a difference with filtering and test again. Doesn't have to be as long as demo1, just be sure to make it at least 10 seconds to get usable results.

-BlueSteel

 

thanks

I'll try that. I just found it odd because during the timdemo there didn't seem to be a significant drop for using quality over performance AF, but when I jumped back into the game, it was noticably slower.

I'll see if I can record a number of demos and test them all tonight with and without Humus' mod. Depends on the time I get home tonight. I was looking into it, because I gave my friend the modified shader and he said it was running worse on his PC with his 9700pro than with the original shader.

 

Also registered to post results.
Running 1024x768 HQ, 2xAA (In Doom Advanced options), set to 2xAA, 8xAF in ATi (Omega 4.7) cpanel, with atioglxx.dll from CAT4.9 in Doom 3 folder.

Without interaction.vfp mods and without CAT 4.9, 36.9 FPS
With interaction.vfp mods and with CAT 4.9 49.1 FPS

33% boost. I pretty much get free 2xAA, since without 2x I get about 57 FPS.

Now this is what I call a boost for free.

Thanks Hummus, Beyond3D, and everyone who contributed to the modification.

 

can somebody update the main file so it has every tweak done in the last 30 pages?

 

I guess you meant -3.0 there.