Re: [Mesa3d-users] Software rendering problem

[Brian P. <br...@va...>]

Rob Hynes wrote:
> 
> Hi,
> 
> For a while I've been having problems with software rendering of triangle
> strips arranged into rings.  The problem is that if the grid is too fine,
> or the object too distant, or the window too small then parts or all of
> the strip are invisible.  It appears as either a hole in an object where
> the grid is finest, or as a speckled pattern. Background objects can be
> seen through the gaps.  It seems to depend on the size of the triangles in
> relation to the pixel size - changing the window size will change the
> effect even if no changes are made to the object definition.  The problem
> only occurs with software rendering, in 16 or 32 bit colour depth, eg with
> the make linux target.  Using the Glide driver on a Voodoo3 card (make
> linux-glide) I get the problem in windowed mode, but it goes away in
> fullscreen mode with no modification to the code.  I have tried enabling
> and disabling depth testing, back face culling, double buffering and
> dithering and none of these make any difference.  A program that produces
> the problem is attached below.  It should produce a white circle with a
> small hole in the middle.  With software rendering a ring of black
> speckles also appears around the central hole.  The size of the ring
> depends on the window resolution.  There may be some crucial setting I've
> missed or maybe this could be a software rendering bug?  Any suggestions
> would be welcome.

Rendering very tiny triangles is actually a tricky problem because of
accuracy problems in the arithmetic.  Other people have had problems
with this in the past.  Usually, tweaking the code in s_tritemp.h which
computes the triangle area and thresholds it has minimized the problem,
if not eliminated it.

I've been tinkering with your test program and Mesa's triangle rasterizer
for a couple hours now.  It looks like the only solution will involve
snapping the triangle x/y vertices to some integral subpixel position
and adjusting the threshold for processing tiny triangles.

I've tried to avoid doing this in the past because of the performance hit.
However, I think it can be done without too much of a penalty.  When I've
got something for you to test I'll let you know.

-Brian