loop), and open-ended splines, with their tangents specified at
their endpoints.
Also change constraint solver matrix size to 1024, on the theory
that a power of two will generate better array indexing, and
replace fabs() with my own function that for some reason is
faster.
[git-p4: depot-paths = "//depot/solvespace/": change = 2055]
point-coincident constraints and the entity- or group-generated
constraints. Also fix a bug where a dragged point was not released
when the mouse pointer left the window.
[git-p4: depot-paths = "//depot/solvespace/": change = 2045]
we're inconsistent (singular Jacobian). That's slow, so we should
provide a library interface to disable it.
[git-p4: depot-paths = "//depot/solvespace/": change = 1946]
little test app that links against it. I still need to polish a few
things, but this is more or less as it should be.
[git-p4: depot-paths = "//depot/solvespace/": change = 1944]
tables from SolveSpace to their own class. This is intended to
simplify use of the constraint solver in a library.
[git-p4: depot-paths = "//depot/solvespace/": change = 1942]
tricky; can't just use the dot product, since that blows up when
you cross pi radians. A gear shift approach, use either sin or cos,
same kind of thing as the 3d-parallel constraint.
And report a NaN constraint as unconverged, of course.
[git-p4: depot-paths = "//depot/solvespace/": change = 1890]
corresponding to each param from the Jacobian, and see if it loses
rank. If it does then that one was important, so it's bound. Then
display a big blue square around its point until the next normal
solve.
[git-p4: depot-paths = "//depot/solvespace/": change = 1887]
not to solve by substitution before rank testing. And report the
unsatisfied constraints when we don't converge.
[git-p4: depot-paths = "//depot/solvespace/": change = 1874]
up more. Also change from stupid linear search lists to sorted
binary search lists, remove a stupid bug where I double-generated
entities, and don't do the triple drawing of entities (since
offsets on the Z buffer were doing the same job already).
[git-p4: depot-paths = "//depot/solvespace/": change = 1776]
and requests to a separate list. It's messy, because I have to make
a deep copy (e.g. of the remap list for the groups, or Expr *
stuff) of some things. Others (e.g. the polygon or mesh) will be
regenerated, so they should be discarded, but they must not get
double-freed.
In any case, works superficially. And fix a few memory leaks
unrelated to this, and remove some dead code.
[git-p4: depot-paths = "//depot/solvespace/": change = 1775]
using the initial numerical guess of whichever vector is already
known to choose our projection planes.
And add a mechanism to defer showing the text window or
regenerating, in order to simplify dependencies on valid actives.
Also yes/no/cancel when about to abandon an unsaved file, and a
bugfix when rotating a rot/trans point.
[git-p4: depot-paths = "//depot/solvespace/": change = 1772]
create a `new' stack-allocated instance of an object; just From,
possibly different versions with different arg types.
[git-p4: depot-paths = "//depot/solvespace/": change = 1763]
and in the case of a singular Jacobian, report which constraints
can be removed to fix it. Also a mechanism to hover and select
entities and constraints from the text window.
[git-p4: depot-paths = "//depot/solvespace/": change = 1746]
lets me reweight the parameters, e.g. to encourage the solver to
make changes to the point being dragged.
[git-p4: depot-paths = "//depot/solvespace/": change = 1719]
we need something to force the points into plane, and the workplane
supplies that), but otherwise straightforward. And add diameter and
equal radius constraints for the arc.
[git-p4: depot-paths = "//depot/solvespace/": change = 1718]
solve the Newton's method iterations in a least squares sense. This
is much less likely to result in numerical disaster; a parameter
will never make a very large change, unless that really is the only
way to satisfy the constraints.
[git-p4: depot-paths = "//depot/solvespace/": change = 1717]
list, and then adding a new entity to that list, and then looking
at that pointer again. Not okay; the add operation might have
forced a realloc. I have to watch for that.
And add a "distance ratio" constraint, plus a new kind of group
that comes with its own workplane. The workplane is not solved for;
it's generated explicitly in terms of elements that are already
solved.
[git-p4: depot-paths = "//depot/solvespace/": change = 1716]
translation; or equivalently, rotation about an arbitrary axis).
Those will be important for step and repeats, and for imported
parts.
Also fix a terrible memory corruption bug: I was freeing the remap
list after I loaded it from the file, but the code that put that
into the SS.group list made only a shallow copy.
[git-p4: depot-paths = "//depot/solvespace/": change = 1715]
constraints. And generate the constraint equations for entities
(e.g., that our unit quaternions have magnitude one). Numerical
troubles there, but it sort of works. Also some stuff to draw
projection lines with projected constraints, and to auto-insert
more constraints as you draw.
[git-p4: depot-paths = "//depot/solvespace/": change = 1711]
vectors", represented by unit quaternions. This permits me to add
circles, where the normal defines the plane of the circle.
Still many things painful. The interface for editing normals is not
so intuitive, and it's not yet clear how I would e.g. export a
circle entity and recreate it properly, since that entity has a
param not associated with a normal or point.
And the transformed points/normals do not yet support rotations.
That will be necessary soon.
[git-p4: depot-paths = "//depot/solvespace/": change = 1705]
the point that's being dragged first, to guarantee that that one
gets the max possible degrees of freedom. The sort code (sort a
list of integers, then apply the permutations by swaps) was more
painful than it should have been.
[git-p4: depot-paths = "//depot/solvespace/": change = 1700]
turned out straightforward, in great part because the planes are
workplanes (6 DOF, represented by a unit quaternion and a point),
and therefore make it easy to get a vector in the plane, as well as
a normal.
And on that subject, replace the previous hack for parallel vector
constraints with a better hack: pivot on the initial numerical
guess, to choose which components of the cross product to drive to
zero. Ugly, but I think that will be as robust as I can get.
[git-p4: depot-paths = "//depot/solvespace/": change = 1699]
or taking partials (constant folding). Also keep a little hash
table to mark with params are used in each equation, in order to
quickly discard trivial partial derivatives. This is solving a
64x64 system in <20 ms. I suspect this is now much faster than
Sketchflat.
Slightly fake situation, though, since substitution solver has not
yet been written, and no partitioning. I'll do those next.
[git-p4: depot-paths = "//depot/solvespace/": change = 1698]
not have much motivation behind them, but they seem to work. And
make sure that we don't solve multiple times without repainting in
between, and tweak the text window a bit more.
[git-p4: depot-paths = "//depot/solvespace/": change = 1696]
erased before redraw, which caused a bit of flicker on show. And
hide debug prints in solver.
[git-p4: depot-paths = "//depot/solvespace/": change = 1693]
other entities. This requires a new point type, for a point that's
defined as a transformation of some other point. All works nicely,
I think. There's ugliness because entities are no longer guaranteed
to have a parent request.
Also speed up display of the text window, by caching brushes
instead of recreating for each character (!), and add a bit more
user interface in the text window.
[git-p4: depot-paths = "//depot/solvespace/": change = 1692]
workplane: a free constraint works in three-space (e.g. true
distance), and a constraint in a workplane works in that plane
(e.g. projected distance). And make the solver go automatically,
though solver itself has lots of pieces missing.
[git-p4: depot-paths = "//depot/solvespace/": change = 1691]
and point-in-plane. These work, but the equation is still stupid,
solving everything at once and not substituting trivial equations.
[git-p4: depot-paths = "//depot/solvespace/": change = 1677]
take the partial derivatives, and run the Newton's method. This
seems to sort of work with a single distance constraint.
[git-p4: depot-paths = "//depot/solvespace/": change = 1675]
Jonathan Westhues