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

This is completely unrelated to Humus's tweak. AF is not affected.

When Doom 3 is allowed to set the AF level (and control panel is set to application preference) the lines separating mip levels are clearly visible akin to the kind of thing you would see with bilinear filtering instead of the blending between levels you would get with trilinear.

When you force AF through the driver control panel, the mip levels are blended properly.


This is the case with or without the tweak, but it has left me curious. There was a thread about this at Rage3d before the tweak ever came out.

 

Re: HUMUS

Oh, allow me. 8)



Although I think work too.

Of course, I think is appropriate too. 8)

 

Re: HUMUS

 

OK, I checked mipmap lines with app.pref/forced and Humus/no-Humus. With Humus' tweak I can now play with force 8xAF with no mipmap lines and good fps. Using 4.9beta.

 

Chris Lux (from 3dcenter) checked a frame in Doom and it looks like that it use the same specular falloff LUT for all objects.

After analyse this texture I was able to find a identical math replacment for it. The shader have to change in the following way:

Code:

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

New:

# perform a dependent table read for the specular falloff
SUB_SAT R1, specular, 0.75;
MUL     R1, R1, R1;
MUL_SAT R1, R1, 16.0;

In the case of a good shadercompiler in the driver this should better:

Code:

Old:
# 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:
# 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];

 

for what are those specular texture maps then good for when you simply can do a calculation ?
i assume your guess about "more accurate" is only true if the specular texture maps are generated automatically always with exactly that calculation of POW 16 and that only in NON ULTRA modes because of the compression.

 

Re: Answers...

Tried this change, framerate in timedemo dropped from ~52fps to ~45fps (6800GT stock @ 1600x1200, 8xAF), so no improvement there.

 

i think the difference comes from the hardware differences - because NV rethought their hardware and relayouted it. ATI didnt.

besides ... that hack seems to be great (long thread already), still not sure for what games are using specular texture maps then and why directx is offering that feature?

 

Maybe i've done something wrong but it looks kinda red now. :?

 

Did some test runs, AthlonXP 1800+, stock speed 9600Pro, 768MB PC266, Cat4.7
800x600 High Quality noVSync
timedemo demo1, three runs, first discarded
Edit: noAF was not true, DoomConfig.cfg sets to x8 on hq Edit2: now also tested 0xAF in DoomConfig.cfg

Code:

noAA 0xAF via DoomConfig.cfg
TEX (normal)      MUL_SAT (humus)        Gain
   30,1               30,5                ~1.5%
   30,0               30,6                         

noAA 8xAF via DoomConfig.cfg
TEX (normal)      MUL_SAT (humus)        Gain
   27,8               28,7                ~3%
   27,7               28,6                         

noAA 8xAF performance via CP
TEX (normal)      MUL_SAT (humus)        Gain
   25,3               29,6               ~17%
   25,3               29,5                         

noAA 8xAF quality via CP
TEX (normal)      MUL_SAT (humus)        Gain
   23,9               28,3               ~18.5%
   23,9               28,4                         

 

There is no connection between so called "specular map" and the tex lookup discussed here.
If you look closely in the shader sode you can see that the specular component is still there.
What's changed is the way the shader calculates the light contribution of the specular component.
Instead of table lookup it's now done with math. That's it.

 

Ok where do I plug that into the interaction file? Do i just copy and paste that and insert it in the first *changes by humus* section? And which one do you use?

 

Re: Answers...

Maybe you've just broken NVIDIA's shader replacements for DOOM 3? Remember what Carmack said about "any little change causes a serious drop in performance when you fall off the fast path in NVIDIA's driver" (paraphrase obviously).

-FUDie

 

ok - then that "hack" seems to be a really good improvement/optimization for your ATI cards and i am honestly happy with you (having a NV though).

 

Did you use Version 1 or 2? In the case you use the second please try the first. It is possible that the second version did not work correct with every driver.

 

Re: Answers...

Maybe someone with an NV30/NV40 should take the default shader and reorder its' instructions. For instance, try moving the "# perform the diffuse bump mapping..." section above the "# load the specular half angle first" section. The order of these two sections will not in any way change the final result of the shader and if performance drops then this *may* indicate shader detection occuring (it may also indicate a less than stellar shader compiler in NV's driver). The same should be tried with ATI hardware. Actually I will try it on my R300 and report back.

EDIT:
Tried my suggested instruction reorder and incurred no performance change. This was done to the default interaction.vfp shader.

 

Ok, I did some checking here and it appears to be that, yes, AF done by the game only applies to the textures it feels needs them while AF applied from the CP applies it across the board.

The effect the humus tweak is having is that it is allowing the ATI cards to operate more effeciently based on their hardware designs, thus allowing you to force CP based AF on the game. The side effect of this is you can now run AF at a higher level than the in game AF allows and get roughly equal performance, thus boosting image presentation.

Now, on a different note, fanboys, please check yourselves at the door and go someplace else. It is obvious looking over the code that the design choice Carmack made was:

To implement the shader method he felt would work most effeciently across the broadest range of cards possible. So instead of having upteen million card specific optimizations, he chose the single best one that would work across as many cards as possible.

That is simple coding effeciency and removes the burden of having to deal with next-gen card issues(ie you remove yourself from the trap of needing to add in new optimizations for every single card release).

It just so happens, as we are all aware in the first place, Nvidia designed their current generation of cards to ensure they got optimal performance off this game. There is nothing wrong with that, its a solid card no matter what. But, since Nvidia altered their card design based on the method Carmack used, the side effect is current generation Nvidia cards are already "optimized" in Doom 3.

If ways are found to optimize nvidia's performance any further, I'll be all for it. This has nothing to do with one card performing better than another and competition. It has everything to do with being a technical discussion of the doom 3 shader code, figuring out how it works, and finding ways individual users can boost their performance by optimizing the code for their hardware.

Please don't turn this into an ATI vs Nvidia fanboy war, it is anything but.

 

I have used the original shader but it is the same place that humus changed.

 

first version works