BRL-CAD Boundary-Representation Primitive
-----------------------------------------
-- Introduction --
This document describes the new Boundary-representation (BREP)
primitive that has been implemented in BRL-CAD for use with external
geometry import and improved visualisation.
The primitive is not yet complete, and is currently quite buggy, so
this document also serves to describe the current state of
development, and the knowledge a developer needs to continue working
on this primitive.
-- BREP Description --
A boundary-representation is a method of representing solid geometry
by describing its topology and corresponding geometry. In other words,
the vertexes, edges, and faces as well as the points, curves, and
surfaces belonging to those topological elements.
For example, a cube has 8 vertexes each mapping to 1 point in space, 6
faces each mapping to 1 surface, and 8 edges each shared by two faces
and sharing its two vertexes with two other faces and owning 1 real
curve.
Actually creating/using a BREP requires keeping track of a lot of details,
including everything mentioned above, plus curve / face orientations
(a CCW edge/vertex ordering is usually used for determining front/back
designations.)
BRL-CAD has/had an existing BREP structure: the NMG (non-manifold
geometry) primitive/library.
[[describe reasons for not using the existing library]]
One can use many types of surface and curve geometry, but generally a
small subset are used by any particular implementation: e.g. having
converted a few pieces of Pro/E through IGES to the new BRL-CAD BREP
primitive, I've seen the following curve and surface types used:
* Line
* Arc
* NURBS curve
* Surface of revolution
* NURBS surface
The simple IGES converter used to import and test new geometry handles
those types already. Other limitations of the converter are described
later in the document.
-- Primitive Implementation --
openNURBS was chosen to represent the geometry within BRL-CAD. This
turned out to be a good and bad choice at the same time:
* contains a lot of solid functionality
* missing a lot of useful and important functionality
In other words, the openNURBS API provided methods of functions for
several pieces of functionality that had implementations removed when
McNeil and Associates released openNURBS as essentially public domain
code. This slowed development somewhat and meant the functionality had
to be reimplemented by hand.
The API is relatively straightforward C++, and it is used to evaluate
geometry and represent/store BREPs in the BRL-CAD geometry file,
through the built-in serialization facility (i.e. 3DM files).
-- Raytracing BREPs --
See Abert et al. (raytracing 06 paper: Direct and Fast Ray Tracing of NURBS
Surfaces).
We use a two-dimensional root-finding technique: we represent the ray
as two orthogonal planes (the intersection of the planes includes the
ray), and then find the root of an equation that represents the
gradient of the distance from the point to the intersection of the ray
planes. When this gradient becomes zero (we found a root), we've also
found the uv parameters for the intersection point.
Newton iteration is used, mostly since it is simple, displays
quadratic convergence when using good guesses, and is amenable to
acceleration using SIMD instructions.
Evaluation of the surface and its derivatives is done by the openNURBS
library at this point. The Abert paper gives some information on how
to do the evaluation using SIMD instructions (needed for speeding
things up).
A simple SIMD vector class has been implemented (see vector.h)
supporting both SSE2 and FPU vector operations. This is currently only
lightly used, since no optimization work has taken place yet
(correctness before optimization).
After intersection, we need to trim the surface. Every edge of a BREP
is part of a loop "within" a face. Loops define boundaries on
faces. In our cube example above, the four edges of a each face
comprise a single loop (in this case an outer boundary). The surface
may be defined as an infinite plane, but this outer boundary loop
limits it to the area enclosed within those edges. As you may imagine,
a face may have more than one loop and these additional loops will
always be internal. All loops can also be considered trims, although
this term seems to be reserved for actual geometric curves that are
parameterized within the domain of an individual surface.
Properly ray tracing a BREP is difficult (thus, the obvious
explanation why this primitive is incomplete). There are several facts
about BREPs that result in problematic situations during ray tracing:
BREPs are not like implicit solid geometry: there is no nice equation
to simply solve (a lot of numerical techniques are used); surfaces,
curves and point geometry may not be aligned; there can be gaps
between two mated surfaces... and the list goes on.
Since it's possible to miss a surface but *need* a hit (i.e. it hit an
edge but passed between surfaces that did not mate up or overlap), we
need to do edge checks: at some point, we find out how far an
intersection point or a ray is from some set of edges
-- Current Capabilities & Limitations --
Developing a BRL-CAD primitive requires a minimum of 4 basic
capabilities: reading from a geom db, writing to a geom db, providing
a plot of the primitive (for MGED display), and handling shot
intersection, which involves finding all intersections with the
primitive and returning the results as a list of "segments" (these
segments will later be used by BRL-CAD to do its boolean weaving).
This primitive currently handles the first 3 capabilities fine, but
has problems with the intersections (the most important part!)
Issues:
* bad acne (possibly caused by missing surfaces? duplicate
intersections (making an odd number of hits along the ray), trimming
errors?
* problems with trimming (not completely sure the bezier clipping is
correct)
* possible problems with tolerances (been working with a very small
(~2mm) object that has a moderate number of faces/edges)
* no optimization, bad algorithms: bounding boxes (subsurface bounding
boxes) should contain correct metadata concerning the need for
trimming within the box, also a list of edges touched by the box for
more efficient edge checking, evaluation optimization, etc...
The current issue seems to be that we're missing surfaces altogether or
getting extra intersections! or something funky is happening during
trimming and edge checking. (BTW Pro/E seems to output overlapping
surfaces, relying on the outer boundary loop to trim it - this implies
that it should be quite difficult to actually miss both surfaces at an
edge (since would have to get at least one intersection if they
overlap at the edge!)
[[ some pictures may be useful here ]]
See all instances of "XXX" in the code (some may be out of date, but
if something was fishy or I was being stupid/lazy/ignorant or some
other bit of code was causing a problem, I tried to mark it with XXX.
-- Conceptual/Implementation Issues --
model-space curve pullback to surface domain for trimming (introduces
possible errors while trimming (i.e. at an edge it's possible both
surfaces can be trimmed because of inaccuracies of sampling)...
multiple intersections found
handling edge/surface grazing consistently
bezier conversion of nurbs curves for trimming
subsurface bounding boxes: when subdividing a surface domain and
testing for flatness, hard to create a perfect bounding box around the
subdomain (i.e. such that all points on the surface lie within the
bounding box.) For the most part this may not be a problem, but
oblique shots cause problems here (an image would help). Either way,
the problem of better fitting bounding boxes (without just arbitrarily
increasing the size which would seriously affect performance) needs to
be considered... currently attempting to sample an additional 5 points
(instead of just the corners). This should serve to handle the current
problem.
Close-up axis aligned rendering of a cylinder from the circular ends
produces a halo of spurious points. Ugh!
Optimize for small objects (there are some fixed size adjustments in
the bounding box/subsurface code).
-- IGES converter --
A significant bit of time was spent writing the skeleton for a new
IGES converter in order to get non-trivial BREP geometry from an
external package into BRL-CAD for testing purposes (have been using
the piston head part from one of the Pro/E tutorial models).
I believe this converter can be made production-ready with some work:
* polish options for output
* handle assemblies and their proper mapping to BRL-CAD
* handle naming? (don't know if IGES carries names, or if Pro/E adds
them)
* handle units properly!
* tolerances when converting are still flaky (fix it) (cause BREP
validity problems). Trims endpoint are not within zero tolerance
(1e-12), so they "don't match" but they are very very close
(1e-11)... so need to go through and call ON_Brep::CloseTrimGap() on
each pair of trims in a loop.
* Run lots of test cases!!!