[Mesa-dev] [PATCH-RFC] draw: Rewrite primitive decomposer

[Zack Rusin]

On Wednesday 04 August 2010 13:06:33 Chia-I Wu wrote:
> On Wed, Aug 4, 2010 at 11:01 PM, Zack Rusin <zackr at vmware.com> wrote:
> > On Wednesday 04 August 2010 04:17:15 Chia-I Wu wrote:
> >> I've updated the geometry shader so that given a triangle with
> >> adjacency, 3 triangles will be emitted:
> >> 
> >>   - the triangle formed by the first edge and the first adjacent vertex
> >>   - the triangle formed by the second edge and the second adjacent
> >> vertex - the triangle formed by the third edge and the third adjacent
> >> vertex
> >> 
> >> and the color is decided by the respective adjacent vertices.
> >> 
> >> The vertex array is updated to consist of 8 vertices (2 triangles),
> >> which layout just like the diagram in the D3D doc, except the diagram
> >> has 14 vertices (5 triangles).  The color of the adjacent vertices for
> >> the first and second triangles are red and green respectively.
> >> 
> >> With this setup, you can find the results of the running the demo with
> >> master and draw-decompose-2 branches, and with nVidia's driver on
> >> Windows.
> >> 
> >> This says nothing about D3D triangle strip with adjacency, but it now
> >> seems to follow wat nVidia's OpenGL driver does.
> > 
> > Nice job! That looks pretty good, I ported the example to D3D to see what
> > it does. It looks like at least the right vertices are being used. On my
> > NVIDIA linux setup it looks like D3D flips the verts in the second
> > triangle (as we expected).
> > Compare tristripadj8.png to what the GL  version of my example does with
> > glDrawArrays(..., 0, 8) and tristripadj10.png to glDrawArrays(..., 0,
> > 10). Attached is the D3D example that I used (just change the Vertex
> > data to the data from my example to see it since right now it uses the
> > data from yours and that looks the same on d3d windows as it does on
> > gl). It accepts "-reference" flag which will use Windows reference
> > rasterizer.
> 
> I've uploaded the results, due to the size, to
> 
>   http://people.freedesktop.org/~olv/tmp/tristripadj/
> 
> tristripadj6.png, tristripadj8.png, and tristripadj10.png are the results
> of running the original tri-adj.c with 6, 8, and 10 vertices.  The
> vertices of ever other triangles are flipped.
> 
> tristripadj-d3d-order.patch forces draw_decompose_tmp.h to use D3D
> decomposition rule.  With it applied, the results are exactly the same as
> the D3D's.
> 
> With GS, the order of the vertices of a primitive becomes significant.  We
> can no longer permute the vertices to support flatshade_first (at least,
> not before GS runs).


Awesome stuff. Thanks for looking into this. I'd commit that patch and we'll 
abstract it under some settable "use d3d or gl rules" flag for draw - we'll 
need it for clipping anyway.

z