Strange problem in PS2 with Volari

[Marc]

Hi

I've got some problem to understand what's going on with my Volari :

As you can see the quality is not correct, here is the correct result on ATI in FP24 and on NVIDIA with PP forced (det 44.03) :

I try too look the precision in mandelbrot. In OpenGL quality is perfect, exactly like ATI FP24, and less than NVIDIA FP32.

But in Direct3D (Spotlight is D3D too) i've got some banding problem as you can see here :

=>

You can check here the quality of ATI FP24
http://www.hardware.fr/medias/photos_news/00/06/IMG0006351_1.jpg
And check here the quality on NVIDIA with PP forced (det 44.03) :
http://www.hardware.fr/medias/photos_news/00/06/IMG0006350_1.jpg

When i zoom more, the precision seem to be the same as ATI, but there is this banding problem. Pixel Shader precision test utility from Digit Life did not work but i can confirm thanks to shaders provided by Tridam that exponent is up to 7 bits and mantissa is up to 16 bits, so it's FP24.

If someone got an idea of the reason for this problem in PS2 ....

 

[demalion]

Difference noted between D3D and OpenGL mandelbrot shader:

D3D

...
dcl_2d	s0

def		c0, 0.0, 1.0, 8.0, 0.0
...
mov		r1.z, c0.x

dp3_sat	r0, r1, r1
mul		r0.x, r0.x, c0.z

exp		r0.x, -r0.x
sub		r0, c0.y, r0

texld	r0, r0, s0

mov		oC0, r0

OpenGL

DP3_SAT	tmp, curr1, curr1;
#RSQ		curr0.x, tmp.x;
#MUL		tmp, tmp, curr0.x;

TEX  output, tmp, texture[0], 2D;

  
END

Hopefully I didn't miss a difference in the "MAD" sequences aside from these sections...

Why does the OpenGL implementation not require those steps, and is there something particular about them that is likely to be the problem? (Why were they put in and then remarked out anyways...maybe it says in the original Mandelbrot thread?) The teapot shading problem looks both offset towards "bright" and having a repeating discontinuity, cycling over and under what it should be to a marked degree as it progresses...what instructions does it use for its precision dependent output?

 

[Reverend]

This is just OTOH but I think there is a persistent filtering problem that is related to their design wrt looking out for their multi-chip possibilities.

 

[Humus]

Why does the OpenGL implementation not require those steps, and is there something particular about them that is likely to be the problem?

It's just that when I wrote the D3D version I decided to do the final step a little differently. No particular reason why. Consequently, the output is slightly different too.

 

[joe_deng]

Hi,
The PS 2.0 precision of Volari is 24 bits floating point(S7.16) as MS DX9.0 specs suggest.

The quality of your dumped image is so strange, especial for banding problem.

Can you provide your testing program to me ?
I will give you feedback ASAP.

Thanks.

Joe Deng
3D Software Manager of XGI Technology Inc.

E-mail: joe_deng@xgitech.com

 

[KimB]

Btw, that's almost certainly FX12 forced. I don't think FP16 would show any issues with that shot.

 

[Marc]

Hi,
The PS 2.0 precision of Volari is 24 bits floating point(S7.16) as MS DX9.0 specs suggest.

The quality of your dumped image is so strange, especial for banding problem.

Can you provide your testing program to me ?
I will give you feedback ASAP.

Thanks.

Joe Deng
3D Software Manager of XGI Technology Inc.

E-mail: joe_deng@xgitech.com

Hi and thanks for your reply,

You can download Humus Mandelbrot D3D program here :
http://esprit.campus.luth.se/~humus/3D/MandelbrotSet.zip

This banding problem seems to be not limited to this program since i see it in shadermark too. What's strange is that S3 seems to have the same problem on his DeltaChrome from what i've seen on Tech Report screenshot.

 

[X2]

Maybe the exp is done at a lower precision?