forced FP16 on HL2 fixes FX performance problem???

That is lovely and all, but why? Why can't valve release a patch which allows FX users to use ONLY fp16 then they don't have to do anything really and just say "warning this may produce undesireable artifacts"

What are they afraid of that suddenly there will be porn popping up by random combination of artifacts in the textures?

 

damn you ivzk you made me visist nvnews ><

Aynways back to the topic, anyone seen the screenies that were linked in the steam forum? Has anyone compared them to how they look in DX8 mode?
Cause of those i've seen so far it looks like shit in still, does it look better in motion?
http://img62.exs.cx/img62/9071/GlassFP16.jpg
http://img62.exs.cx/img62/6974/GlassFP32.jpg
http://img62.exs.cx/img62/1950/MetalFP16.jpg
http://img62.exs.cx/img62/5243/MetalFP32.jpg

Some say that people wont complain about some small anomalies, but which of the some do people mean then?
The people that visit forums or the ones that buys game-insert game-plays game and don't give a crap as long as it works and looks ok?
Implementing a mixed mode would be good, but for the good will of the company it better be tested and proven to work ok before releasing it buggy to please a bunch of forum copy pasters or the "millions" of fx owners wanting to see some DX9 effects or the few fx owners who wont mind some anomalies.

There are lots of factors to consider here and to blindly think that just change this change that will just take little time nothing to worry about. And it will only take a few days. That's just being moronic

 

Psstt:
A little factoid for you. Back in the early days with Farcry there was a simular issue with banding on FX cards. One fix for that was to change the Device ID to an ATI card. Since Farcry is on the NV TWIMTBP programs me doubts that the device id is doing anything "evil" for NV in HL2.

 

Maybe they will. There are many things with HL2 that are not done yet...give them some time...

 

To make sure that I have the same frame I used an external player and modify the shader there:

 

Ok why not just leave the glass shader running at 32 bit and all the other shaders in the world at 16 bit? Also there shouldn't be that much banding, you might right that might be a ps 1.4 shader for the glass.

 

Why not just leave full precision for texture lookups and partial for everything else if that's why this banding is occuring? Btw, isn't FP24 an absolut minimum precision for texture lookups in DX9 specs?

 

This was out from a joint article from Nvidia and ATI

http://developer.nvidia.com/docs/IO/8230/D3DTutorial1_Shaders.ppt

when using texture lookups its not the texture it self that you are calculating precision with, its the vertex normals that you extrapolate of that texture. later in the article for a 512x512 texture you only need 10 bits to read and filter 2 times.

 

how come there`s no 1 from valve here to comment on this? there some many ppl here who work for graphics related company`s and no 1 from valve?

 

No, I don't believe "no 1 from valve" has ever posted up here.

 

i didnt get the joke

 

One would have to go a step further and keep the operations used to generate the texture coordinate at full precision. An FP16 value in an FP32 register still has the precision error of the former, after all.

Sorta. The spec requires that hardware be able to perform texture lookup with at least FP24 precision. However, if a developer wishes, he could still use partial precision (FP16 minimum) to sample a texture.

 

Thanks Demirug!

 

Sorry, it wasn't really meant as a joke so much as an exasperated comment in regards to how people feel that game developers owe them some kind of response for posting up at forums they happen to belong to.

They don't, and they don't like being "called out" so to speak...takes all the fun out of hanging at a place in a hurry.

Besides, I've really never seen anyone from Valve post here before.

 

To address a couple of concerns, I started programming on the C64 in assembler (not my favorite, but good for understanding how computers really work), moved on to Lattice (SAS) C & C++ on the Amiga (you needed a working knowledge of C to understand how the Amiga OS worked), moved to visual studio 6 on pc then visual studio.net

Digitalwanderer, not all of HL2's shaders are SM2.0+ shaders. We don't need to touch the DX8.x shaders.

 

Sheesh, not disputing your point, but does anyone buy hardware from tigerdirect? Especially when the likes of Canada Computers are generally cheaper:

Gigabyte GeForce 6800GT 256MB DDR3 w/ DVI, TV-out - $559
XFX GeForce 6800 Ultra / 256MB DDR3 / AGP 8X / Dual DVI / TV Out / Video Card - $709.99
Sapphire Radeonâ„¢ X800 PRO 256MB GDDR3 w/ TV-Out - $565

(they don't sell the Leadtek or Visiontek cards you quoted)

 

Thx for the shots Demirug!

Not much difference in the shaders used in that scene... I sure couldn't tell which was fp16 and which was fp32.

 

Well there you go. Never heard of Canada Computers on the net. Thanks for the link and I will have to check them out in the future. Still, even those prices are quite high and way out of the range of most peoples video card budget.

 

Hy guys, i am new here but i decided to drop in to give you some solid numbers on DX FP16 vs. 32 vs DX8.1 on FX line of cards.

I learned about the issue yesterday and i up till now i've been doing some heavy benchmarking.

I will write an article about the issue when i finish with the "manual" work.


So..."my" numbers:

Tested on:
Athlon 64 3200+ Socket 754 Clawhammer core
512 MB DDR400 memory running on CAS 2-3-3-6
Abit KV8-MAX3 + Hyperion 4.53

Video Cards:
Radeon 9800 Pro @ Catalyst 4.11 + CC
GF FX5900 @ ForceWare 66.93

HL2 Steam version (latest patches and so on...) + 3DVis. app v2,26


1024x768 (all high except "reflect world", vsync off, trilinear filtering)

Radeon 9800 Pro - 73,7 fps
FX5900 @ DX8.1 - 62,44 fps
FX5900 @ DX9 forced FP16 - 31,45 fps
FX5900 @ DX9 full p. aka FP32 - 19,46 fps


1280x960 (same as above)

Radeon 9800 Pro - 52 fps
FX5900 @ DX8.1 - 45,18 fps
FX5900 @ DX9 forced FP16 - 23,31 fps
FX5900 @ DX9 full p. aka FP32 - 14,4 fps


1600x1200 (same as above)

Radeon 9800 Pro - 33,65 fps
FX5900 @ DX8.1 - 29,49 fps
FX5900 @ DX9 forced FP16 - 15,08 fps
FX5900 @ DX9 full p. aka FP32 - 9,35 fps


Seems to me that FP16 really brings considerable performance bonuses on FX cards. Visual degradation surely exists but i failed to noticed it during the playtesting. Difference between DX8.1 and DX9? Clearly visible.

Steam users with relatively adequate harware for FP16 + additional optimizations DX9 pass? About 70.000 ~ 5,5 %

That's not much, but after so many delays i expected mixed shading path for unfortunate NV3x users. For me that is sloppiness with a touch of malice (probably ATI funded).

On the other hand, i can't say i regard FX5900 FP16 numbers competitive with those of Radeon 9800 Pro. Maybe ForceWare 67.02 introduces some kind of benefit. Anyone tried it (i am to tired to do that now)?

P.S.

I used a custom demo recorded at the beginning of the water hazard level. Warehouse with winch and stuff...

 

This will no doubt amuse those attributing this to sabotage on ATI's behalf.

There's a line along the bottom of the Copyright notice just after the Valve guy splash screen: "this product includes code licensed from Nvidia".