[Mesa-dev] [PATCH 3/8] mesa: Add interpolation override bitfields.
brianp at vmware.com
Tue Oct 25 10:21:07 PDT 2011
On 10/24/2011 05:37 PM, Paul Berry wrote:
> On 24 October 2011 15:58, Brian Paul <brianp at vmware.com
> <mailto:brianp at vmware.com>> wrote:
> On 10/24/2011 03:38 PM, Paul Berry wrote:
> This patch adds the bitfields InterpOverridesFlat,
> InterpOverridesSmooth, and InterpOverridesNoperspective to
> gl_fragment_program. These bitfields keep track of which fragment
> shader inputs are overridden with the GLSL "flat", "smooth", and
> "noperspective" interpolation qualifiers.
> The names of those fields seems a little confusing to me. For
> example, "InterpOverridesFlat" sounds like a field that overrides
> flat shading with something else.
> How about just "InterpFlat", "InterpSmooth", etc?
> The meaning I was trying to convey with "overrides" is that bits are
> only set in these bitfields if the interpolation mode is overridden in
> GLSL. For fragment shader inputs that don't have their interpolation
> mode overridden, all of the bitfields have a zero. (I had to do this
> in order to avoid adding a bunch of code to the handling of fixed
> function fragment shading and ARB fragment programs).
> Or, how about an enum that indicates the interpolation mode for
> each fragment shader input? Something like this:
> enum interp_mode
> INTERP_DEFAULT, /* I _think_ we need a default mode, right? */
> struct gl_fragment_program
> enum interp_mode InterpMode[FRAG_ATTRIB_MAX];
> This would also prevent a non-sensical state where a particular
> bit is accidentally set in more than one of the masks.
> Yeah, I prefer this too, and that's essentially what I did in patch 2
> of the series. To be honest I find all the bitfield stuff rather
> ugly, and I would rather drop this patch entirely, but a lot of the
> i965 code still relies on things being bitfields. Is it the only
> driver that does?
Which bitfields are you referring to specifically? I think the ones
in gl_program are used in the gallium state tracker, but I'd have to
> If so, I suppose I might could be convinced to move
> all the bitfield handling of interpolation into i965 where it won't
> contaminate the other implementations.
I don't know what works best for i965 but using an enumerated type for
the interpolation modes seems cleaner for core Mesa.
> Finally, maybe use "CONSTANT", "LINEAR" and "PERSPECTIVE" instead
> of "FLAT", "NO_PERSPECTIVE" and "SMOOTH". That's what we have in
> Hmm, I'm ambivalent about this one. The terms "linear" and
> "perspective" are certainly clearer than "noperspective" and "smooth",
> but "constant" seems downright misleading for flatshaded attributes,
> since they are only constant across a single primitive. I also think
> there's a lot to be said for being consistent with GLSL terminology,
> which is to call these things "flat", "noperspective", and "smooth".
It's not a huge deal to me either.
> I also admit I'm a little surprised to hear that Gallium has already
> dealt with this question. In my greps, I couldn't find any Gallium
> code that was referring to the ir_variable::interpolation field, so I
> assumed i965 was the first driver to try to implement interpolation
> qualifiers. Did I miss something? If I did, then I probably broke
> gallium when I renamed things in patch 1 of the series.
The gallium state tracker doesn't yet handle interpolation qualifiers
for Mesa's fragment programs, but we emit them when we create TGSI
fragment program input register declarations (frag pos is linear,
texcoords are perspective, etc). For implementing glShadeModel(),
however, we still rely on a rasterization state flag rather than emit
different fragment programs with different input qualifiers. That's
something I'd like to fix someday.
More information about the mesa-dev