VOGONS


First post, by 386SX

User metadata
Rank l33t
Rank
l33t

Hi,
I am trying a Voodoo3 3000 AGP with latest drivers on WinME and I didn't know they included this "Geometry assist" feature to "help" geometry and lightning thank to the Pentium 3 or K7 architectures.
I was curious to understand more about this, if it actually is the same D3D software T&L of the Dx7/8 or it differs in some way (and if it's similar to the Kyro2 latest drivers option). Running 3DMark2000 with this features it "see" an hardware T&L cap and from the usual Athlon cap it improved the score by 150 points (!) (~3200 with T&L hw and ~3050 without).
(By the way I didn't remember the V3 with the Athlon flying so much, it surpassed the Savage 2000 by 500 points in the same bench)
Thank

Reply 1 of 5, by Scali

User metadata
Rank l33t
Rank
l33t

Sounds like it is indeed a software T&L pipeline, but optimized for the specific video card, as opposed to D3D's built-in generic pipeline.
D3D also uses SSE and 3DNow! optimizations where possible, but it doesn't know anything about the underlying hardware.
If you perform hardware-specific optimizations, you can output the T&L data in a way that it can be used directly by the rasterizer, rather than having to convert it from D3D's generic output to the hardware-specific format in the driver.

http://scalibq.wordpress.com/just-keeping-it- … ro-programming/

Reply 2 of 5, by 386SX

User metadata
Rank l33t
Rank
l33t
Scali wrote:

Sounds like it is indeed a software T&L pipeline, but optimized for the specific video card, as opposed to D3D's built-in generic pipeline.
D3D also uses SSE and 3DNow! optimizations where possible, but it doesn't know anything about the underlying hardware.
If you perform hardware-specific optimizations, you can output the T&L data in a way that it can be used directly by the rasterizer, rather than having to convert it from D3D's generic output to the hardware-specific format in the driver.

Thanks for the answer, as usual. 😀
It's really interesting considering at that time more T&L games were released and probably supported by this feature improving a bit the frame rate. Maybe in the same way they did with the Kyro2 similar option? (but never understood if the STG4800 of the Kyro2 SE had something more in the hardware over the sw feature to do the same thing).

It's also interesting in 3dmark to test all the caps and see that (at least with the Savage 2000) there was good difference even from 3dnow cap to Athlon cap...

Reply 3 of 5, by Scali

User metadata
Rank l33t
Rank
l33t
386SX wrote:

It's really interesting considering at that time more T&L games were released and probably supported by this feature improving a bit the frame rate.

There were even some games that wouldn't work at all if they didn't detect 'hardware' T&L, so it also gave better compatibility.

386SX wrote:

Maybe in the same way they did with the Kyro2 similar option? (but never understood if the STG4800 of the Kyro2 SE had something more in the hardware over the sw feature to do the same thing).

I believe the Kyro II SE was basically the same chip, with faster memory, and the 'enhanced T&L' feature was a driver-thing, which could also be used on regular Kyro II, so I think it was completely software.
Intel did a similar thing on their iGPUs at the time: they reported hardware T&L, but used a software pipeline. Later various vendors even reported vertex shader hardware, while executing them on the CPU.

One interesting feature of enhanced T&L is 'implicit guard band clipping'. Sounds like they don't perform clipping during T&L at all, because they know it will be done for tile-based rendering anyway. So that is another way to boost performance.

http://scalibq.wordpress.com/just-keeping-it- … ro-programming/

Reply 4 of 5, by 386SX

User metadata
Rank l33t
Rank
l33t
Scali wrote:
There were even some games that wouldn't work at all if they didn't detect 'hardware' T&L, so it also gave better compatibility. […]
Show full quote
386SX wrote:

It's really interesting considering at that time more T&L games were released and probably supported by this feature improving a bit the frame rate.

There were even some games that wouldn't work at all if they didn't detect 'hardware' T&L, so it also gave better compatibility.

386SX wrote:

Maybe in the same way they did with the Kyro2 similar option? (but never understood if the STG4800 of the Kyro2 SE had something more in the hardware over the sw feature to do the same thing).

I believe the Kyro II SE was basically the same chip, with faster memory, and the 'enhanced T&L' feature was a driver-thing, which could also be used on regular Kyro II, so I think it was completely software.
Intel did a similar thing on their iGPUs at the time: they reported hardware T&L, but used a software pipeline. Later various vendors even reported vertex shader hardware, while executing them on the CPU.

One interesting feature of enhanced T&L is 'implicit guard band clipping'. Sounds like they don't perform clipping during T&L at all, because they know it will be done for tile-based rendering anyway. So that is another way to boost performance.

I didn't know this. If we consider the Athlon optimizations and 3dnow and this sw T&L solutions.. well.. sounds like for that time alternative not-T&L solutions did everything they could to be competitive.

Reply 5 of 5, by Scali

User metadata
Rank l33t
Rank
l33t

Should be no surprise. Display drivers are among the most optimized software in the world. There is such a strong competition in this field.

http://scalibq.wordpress.com/just-keeping-it- … ro-programming/