Supermodel Forum

One for Harry
I changed the normal conversion code (as well as the code for fixed shading) to this

Code: Select all

#define BYTE_TO_FLOAT(B)   ((2.0f * (B) + 1.0f) * (1.0F/255.0f))

Code: Select all

         // Per vertex normals
         if (ph.SmoothShading()) {
            p.v[j].normal[0] = BYTE_TO_FLOAT((INT8)(ix & 0xFF));
            p.v[j].normal[1] = BYTE_TO_FLOAT((INT8)(iy & 0xFF));
            p.v[j].normal[2] = BYTE_TO_FLOAT((INT8)(iz & 0xFF));
         }

from

Code: Select all

p.v[j].normal[0] = (INT8)(ix & 0xFF) / 128.f;

You can see for positive numbers this will effect the length of the normal.
Before for 127 / 128 = 0.99
Now
((127 * 2) + 1 ) / 255 = 1.0

It's only a 1% change, but didn't know if this would effect specular at all?
Maybe the bias you added might not be needed anymore?

Interesting... I'm gonna try that now and let you know as soon as possible.

Did that effect specular at all?

Sorry, I was testing when I noticed I've some problem in my private build with transparent polygons. So I'm currently struggling with some unknown bug.

Going from r626 to current at some point broke something and now light math doesn't give me the exact results as before, the bad thing is I don't know what's happened, and happens in some very corner case.

In particular (left: Model 3, right: Supermodel)

See that highlight? Before it worked perfectly, now has disappeared! It's been like... 4 hours that I'm trying to fix that.

Shader debugging is hard!
Not like you can step through it with a debugger

Damn!

I hope to try tomorrow the type conversion thing, until i fix this madness I cannot trust any more shader math, let alone specular highlight, it seems to be something that blows up at a particular light/normal angle, but who knows?

I hope to not have hit another driver bug, my old Radeon HD4850 is considered legacy from 2014...

I'm using a gtx 960
a 2014 card shouldn't be too old at all .. I mean we are only using gl 2.1

One way of debugging shaders if you are really struggling is to dump the value out as the output colour

Ian wrote:One way of debugging shaders if you are really struggling is to dump the value out as the output colour

Yeah, infact is what I'm doing to see what's happening...

Finally, problem solved!

It was again an issue with the polygon header hash computation and the caching mechanism. Like happened recently with a fog issue and VON 2, this time it was necessary to insert also the bits relating to shininess and specular coefficient into the mix.

Many of the weirdness present in LeMans24 are now fixed. The few remaining very little quirks (single polys) are, maybe, related to proper quad poly rendering. For example the following case is now solved and back to the same behavior as the original hw:

(Left = Model 3, Right = Supermodel old incorrect rendering)


I'll post soon a patch.


P.S.
Hey Ian, since it's clear that polygon header attributes remain constant at (obviusly) polygon level but not at mesh level, either we expand the hash to be much more that 64 bits (to include as many attributes as possible) or we disable the caching mechanism altogether but, I think, with a huge performance hit. What do you think?