forced FP16 on HL2 fixes FX performance problem???

I think I'd like one of Brent's confusing graphs to see on this one actually, I like how it shows ya the low/high dips so you can get a better feel for what the number means.

 

There's a difference between playing a game, and a recording of someone playing a game. If you did that in something like UT2K3/4, the playthrough would be full physics and AI. With the recording, it would be just positional information recorded and then everything moved to the right places at the right time for playback. It's a world of difference.

 

And doesn't just taking the players input take some time/effort on the PCs part?

(I've always kind of wondered about that, I keep picturing all them "get" statements in the code....I'm an old skool basic whore at heart. )

 

it does but its almost nothing , if input is slow, well there goes the whole gaming experience

Good example, try using windows input calls instead of dx input calls for a game.

Was making an Ogl demo once, and the frame rates with windows input call, were about 80% slower, but the dx input calls there was no frame rate change at all.

 

Not much, I'd imagine, given its Shadermark and Rightmark FP16/32 numbers (IIRC). Almost certainly not as much as a 5900.

 

I like so called "real world" testing. I will try forcing FP16 on GF6800 this weekend, now i don't have time to do it.

 

IIRC NV40 was getting even more performance when using FP16 with complex shaders than NV35. So it's hard to say, better do some benchmarking

According to our latest syntetics test FP16 on NV40 is the only way of getting comparable with R420 PS2.0 performance:





In fact in case of NV43 it the way to outperform RV410.

 

Wow, DesgustatoR, it seems those benchs were not optimized with NV40's register limitations in mind.

Perhaps this is one of the reasons for the performance descrepencies in HL2. I think this calls for a mixed mode (FP16 w/FP32 when required) patch.

 

They definately were This is D3D RightMark, btw. Taken from here.

As for HL2 + forced pp...

Screenshot

I'm not getting any shadows at all. What am i doing wrong?

 

In the case of NV40 register limitations are a smaller problem. But there are still instructions that run faster with FP16. If the shader use many of this instruction it will get a speedup.

 

It depends on the game. Some record random number seeds.

 

I don't think it's a texture lookup that's causing this. It's a specular reflection on a flat surface. One of the vectors used in the calculation has small changes across the wall/window relative to it's length, so the reduced mantissa of FP16 is hurting.

Anyway, FP16 DX9 is still half the performance of DX8.1, right? Even if Valve did use _pp, I doubt they (or any FX user) would think DX9 was worth a 50% performance hit for the marginal improvement in image quality. Why is Valve being berated for this?

I'm just trying to understand what all the fuss is about here.

 

If you are not an FX user yourself, kindly don't try and tell us what we might or might not find worthwhile!

 

Unfortunatly for geforce fx owners forcing fp 16 on everything as this hack does means you aren't running dx 9 anyway .

So i still don't understand what this is all about .

 

What does that have to do with his comment in terms of quoting him. In any event it still looks better. It should be his choice.

 

You are forcing partial precision (which happens to be FP16 on nVidia and S3 products), and partial precision most definitely is and always has been a part of the DirectX9.0x specification.

 

Well, obviously Valve didn't find it worthwhile for GeforceFXs to play the game slower than a Geforce 4.

EDIT: Actually, does the SDK allow one to edit the shaders? I'm guessing it does. . . If that's the case and you feel so strongly about it -- and you think it'd be so easy --, make a mod. Convert the shaders to primarily use partial precision, with full precision only certain instructions that need it. It's clear that Valve won't bother to do anything about it.

 

Actually I believe that Valve's actions (and if Valve fails to release a patch to remedy the situation in a timely fashion) will provide nVidia with a perfect justification to include a user switchable optimization in future drivers that allows users to force partial, full or default precsion.

 

They can add that anytime they like, but I doubt it'll pass Microsoft's WHQL certification.

 

What a sad, sad thread. C'mon people.

Geforce FX is a bad joke on the industry. NV screwed the hell out of a product cycle. They hyped up NV30 (a POS unless you are playing DX8 games) to a ridiculous level, their Nvidiots bought it hook, line, and sinker, and now are demanding Valve put hundreds of man hours into trying to get a tiny quality boost with a guaranteed significant performance loss. And, all this when the primary difference between DX8 path and DX9 path is only the water, which isn't all that great anyway, even in full DX9 mode.

You know, I wouldn't be surprised if NV requested Valve just use DX8. If they had had the game default to DX9 FP16 for HL2, you'd have people raving about how slow the game is and how it should be optimized or some nonsense. It would look bad for NV and Valve.

Toss out the trash (GFFX) and go buy a 6800, 6600, cheap 9700, 9800, whatever. Hell my 9600 in my notebook runs the game amazingly in full FP32 at 1280x800 high detail.