Render Order Benchmark

[Dave Baumann]

I was just wondering if anyone knows where to find the 'Rendering Order' benchmark utility that someone (I think it might have been pcchen - although if it wasn't I appologise to the coder!) posted on the old pcstats forum?

I have some more tests to run on the Readon 8500 and I wouldn;t mind using it.

Cheers. :)

[pcchen]

Do you need random rendering order? If not, my ztestdx (http://smart.iis.sinica.edu.tw/~pcchen/files/ztestdx2.rar) could be useful :)

The new version has a new stencil test function which can test whether the 3D chip can do fast stencil test. Its usage is

ztest [-s] [width] [height] [bpp] [texture_bpp] [texture_width] [texture_height]

[Dave Baumann]

Do you need random rendering order?

I was thinking about the one that does back-to-front, front-to-back and random order tests.

Cheers.

[Humus]

I guess you're talking about my GL_EXT_reme benchmark, right?

It's on my site:
http://esprit.campus.luth.se/~humus/?page=3D

[Dave Baumann]

Can't remember if thats exatly the one I was thinking about, but it serves the purpose. Thanks - look out for the use of it in the Radeon 8500 Review/Investigation Part 2! (that is, if you don't mind) :)

[Humus]

I'm just happy if my stuff gets used.
Looking forward to read it :)

[Dave Baumann]

Doh! Crap - I just realised I need it to be DX; with OpenGL we can't be sure if its actually using Hier-Z rejection or just plain early Z rejection.

[Humus]

Uhm, I don't really understand, it is and should be transparent to the application in both API's. The only way to know is by looking at the results, if overdraw * width * height * framerate > teorethical pixel fillrate then Hierarchical Z is definitely being used. Otherwise, the pure performance increase would show this clearly (I measured a 6.15x performance increase between front-to-back and back-to-front last time I checked).

[Humus]

Checked again, performance was 4 times as high as teorethical fillrate in the front-to-back test.

[Dave Baumann]

Uhm, I don't really understand, it is and should be transparent to the application in both API's.

There are no specific registry strings for ‘Hierarchical-Z’ in OpenGL for 8500, only ‘HyperZ’ – AFAIK this only enables one facet of HyperZ (Z-Compression), or at least the differences between Villaemark OpenGL and DX tests would indicate that not all elements are present in OpenGL.

The performance difference could just be attributable to straight early Z rejection.

[Humus]

There are registry keys for HierarchicalZ in OpenGL too. I just tested it now just to see that it works too:

800x600x32 (1024x768x32 turned Vsync on for some reason :-?)

HierarchicalZ on:

Overdraw/HSR:
-------------
Overdraw factor 3, back to front: 460.61 fps
Overdraw factor 3, front to back: 1309.37 fps
Overdraw factor 3, random order: 737.59 fps

Overdraw factor 8, back to front: 168.31 fps
Overdraw factor 8, front to back: 1012.59 fps
Overdraw factor 8, random order: 456.17 fps

HierarchicalZ off:

Overdraw/HSR:
-------------
Overdraw factor 3, back to front: 467.22 fps
Overdraw factor 3, front to back: 582.41 fps
Overdraw factor 3, random order: 530.34 fps

Overdraw factor 8, back to front: 169.35 fps
Overdraw factor 8, front to back: 226.57 fps
Overdraw factor 8, random order: 208.98 fps

These are the relevant OpenGL regkeys (all available in UniTuner)

disableHyperZ
enableHierarchicalZ
enableZCompression
enableFastZMaskClear
ZCompForAllConfigs