kicad-developers team mailing list archive

[jp charras <jp.charras@xxxxxxxxxx>]

Le 10/05/2019 à 18:33, Tomasz Wlostowski a écrit :
> Hi,
> 
> I've been recently playing with Victor's huge 32-layer PCB design and
> trying to improve the performance of pcbnew for larger designs. This
> board causes even pretty decent PCs to crash/render glitches due to
> pcbnew's enormous VBO (Vertex Buffer) memory consumption.
> 
> It turns out it's caused by the way KiCad renders filled zones:
> - the inside of a zone is drawn/plotted as a filled polygon with 0-width
> boundary. This one not a problem - we already triangulate the polygons
> and I recently developed a patch for the OpenGL GAL that allows reusing
> vertices of triangulated polys in the VBO/Index buffer to further reduce
> memory footprint.
> - the thick outline is drawn with rounded segments with the width =
> minimum width of the polygon. Since we don't have arcs in polygons, each
> of round features (e.g. vias) surrounded by a zone gets a ton of tiny
> segments in the polygon outline. Each rounded segment in OpenGL is
> composed of 2 triangles, hence 6 vertices (that can't be reused...). For
> Victor's board it means 1 GB (sic!) of the VBO goes for outlines of the
> polygons alone. Disabling the outline drawing makes the renderer work
> smooth again.
> 
> I've been experimenting with some ways to fix this:
> - generating the thick outline strokes using a Geometry Shader (which
> means bumping up GL 3.0+), which means farewell to many Linux/older
> integrated Graphics users.
> - caching a triangulated polygon which is a boolean sum of the filled
> inside and the thick stroked outline. This takes a lot of time (~2
> minutes for Victor's design) to load and still takes quite a bit of VBO
> memory. Another downside is that the polygons are not fully WYSIWYG
> (outline segments have true rounded corners, while the corners of the
> displayed shape would be approximated with line segments).
> - change the way KiCad handles filled zones to calculate the (stroke +
> inside) boolean sum during zone filling process. It means changes to the
> plotting/GAL/3D code, but no changes to the file format. We'll also be
> forced to inform the users that they have to refill the zones if they
> read a file generated by an older KiCad version.
> 
> 
> Which solution would you prefer?
> Cheers,
> Tom

Hi Tom,

I have a *strong preference* for the solution 3:
- It does not require a bumping to GL 3.0. I am not thrilled by this idea.
- It works also in Cairo engine.
- It fixes a issue in connection calculations: a connection (pad, via
track end) is found if the reference position is inside the zone.
Inside the zone is currently inside the main polygon, not really the
full area (polygon+thick outline), so a few connections are missed.
Why I have used polygons with thick outlines:
Because the polygon library used at this time did not a good job to
inflate polygons (read a really bad job).
But this is not the case now with Clipper, that is a really great
polygon library.

I made a very fast test, and just tried these 2 changes:
- do not draw thick outlines
- Use 16 segments by circle to create holes and polygon inflate: this is
enough for this purpose. I am not sure other ECADs are better.
- and inflate the filled polygons by zone min thickness/2
the result is very good:
- filled polygon shapes are very good.
- the number of segments to draw filled polygons does not significantly
change.

Of course, using segments to draw thick outlines give a better look for
acute angles, but do not mix cosmetic criteria and technical criteria.

-- 
Jean-Pierre CHARRAS