========================================================
This is an advanced setting just for allowing increased choices to power users that have problems with a given sketch and want to
test different flavours of DogLeg algorithm.
This commit does not change the default behaviour of FreeCAD. It is only intended to give more options to power users.
The advanced solver configuration is extended to support three different Gauss-newton steps for DogLeg:
FullPivLU => h_gn = Jx.fullPivLu().solve(-fx);
LeastNormFullPivLU => h_gn = Jx.adjoint()*(Jx*Jx.adjoint()).fullPivLu().solve(-fx);
LeastNormLdlt => h_gn = Jx.adjoint()*(Jx*Jx.adjoint()).ldlt().solve(-fx);
This setting is applied only to DogLeg. It is applied to DogLeg as normal or redundant solver, if DogLeg is the selected solver.
Selecting a solver different from DogLeg for both normal and redundant disables the setting.
We have been told:
https://forum.kde.org/viewtopic.php?f=74&t=129439#p346104
that our default Gauss-Newton step in DogLeg may not be adequate in general (we generally deal with underconstraint systems
unless we have a fully constraint sketch, and even then it is many times overconstraint at least for redundant solving).
We have been told that maybe these LeastNorm options are more suitable for us (performance set aside). This enables you as power
user to test if it works fine with FreeCAD.
===================================================
This commit sets the sketch size multiplier OFF by default (so the number of iterations DogLeg/BGFS/LM is not sketch size dependent).
In complicated sketches having a high number of parameters, with sketch multiplier on, the number of iterations is extremely high (100 parameters*100 iterations => 10000 iterations).
The idea of disabling this comes from tests performed by DeepSOIC and from my own experience using the Sketcher. In general
sketch multiplier makes FreeCAD unresposive (very high amount of iterations, not a real freeze) in big sketches so that users
end up killing the application. This is preventing the users from taking appropriate action, Developers from getting the information of the failure and
users angry.
The idea is that even for complicated sketchs N iterations (100 by default) should be enough to converge, if it is ever going to converge. Experience will tell us
if we have to increase this number in the range [100-300]. 100 iterations in complicated dossiers is in my experience generally under 30 seconds.
N.B.: This commit does not change the defaults stored in your computer, so if you have the sketcher multiplier on, the advanced solver dialog will still enforce this local setting. You
may disable it or click the "defaults" button to disable the sketcher multiplier.
===================================================================
It allows to export the c++ code to create a subsystem with the same information as the one solved using LM/DL/BGFS.
In this commit the functionality is disabled (for production).
To enable the functionality uncomment this line in planegcs/Constraints.h:
//#define _GCS_EXTRACT_SOLVER_SUBSYSTEM_
When enabled, upon solving with LM/DL/BGFS, the c++ code to generate the subsystem is added to a subsystem.txt that is created in the FreeCAD
execution directory.
Note that the file is created in append mode, so it will append all normal/redundant solvings until the file is deleted.
The resulting code can be directly pasted into a project similar to:
https://github.com/abdullahtahiriyo/Eigen_LUPiv_Convergence
Such a project only has libeigen as external dependency.
=====================================
When the solver converged (but did not succeed) or when the solver succeded but the solution is not OCC-valid, no error message was shown in the solver messages dialog.
=======================================
The SparseQR is set as default method. However, in accordance with the previous commit, only if a new enough Eigen library is present will SparseQR be executed.
If the library is old and SparseQR can not be reliably used, FC will automatically fall back to use DenseQR.
This ensures that users that have a new enough library > 3.2.2 will benefit from the performance increase, while the users that use older library will still have
the same performance as in legacy FC.
==========================================
For Eigen>=3.2.2 SparseQR is compiled in
For older versions SparseQR is not compiled in and if the option is selected FC gives a warning in the console and falls back to DenseQR.
This ensures that a user can get a sketch solved and work normally even if SparseQR is selected and the Eigen version is too old to support SparseQR.
================================================================
A new parameter added to the form in order to control how low a value should be
to be considered zero (how high a value shall be to be accounted for in rank calculation)
Debug for QR pivot threshold is also added.
=============================================
EigenQR branch 3.2 with debug code fails an assertion. The result is generally ok if we disable the assertions, however it
eventually leads to memory leakage.
This commit reenables Eigen's assertions when in debug mode, and sets dense QR as default QR algoritm until Eigen's issue is solved.
==================================================
- ternary operator usage was making BFGS not to iterate
- Amount of debug information for BFGS increased
===============================================
- Increased solver debug
- Changing default values from float to string, as float has "only" a precision of 12 decimals.
- Default values as macros
===========================================
- Improvement Debug added to redundant solving in Iteration Mode, to show when just one iteration solved the system
- Fix wrong QR information on empty sketch
==================================================================================
The solver has been adapted to use Eigen's SparseQR QR decomposition algorithm. The original
Dense QR implementation is maintained and can be selected using the Advanced Control TaskBox (see below).
The use of SparseQR provides over an order of magnitude improvement in solving time in complex sketches due to
the Sparse nature of the Jacobian matrix of the system of equations.
The solver advanced control is a new TaskBox in the Sketcher that allows to select which algorithms are to be used for
the different solving operations and tweak its parameters. It is not intended to be a user control, but means to debug
solving problems and improve the algorithms and their configuration.
This commit also introduces multithread support for Eigen. Currently it is only limited to products and does not provide
a substantial speed improvement. It is expected to have more multithreaded operations in Eigen in the future.
As a bonus, the TaskBoxes in the Taskbar of the Sketcher remember the last state (collapsed or deployed).
Fixes a bug where an arc, ellipse, or arc-of-ellipse, being reversed in
XY plane, behaved badly in sketcher (see forum thread "Sketch: how to
handle reversed external arcs?"
http://forum.freecadweb.org/viewtopic.php?f=10&t=9130 ).
Also fixes a problem with rotated arcs (see forum thread "Rotating Arc
in Sketcher"
http://forum.freecadweb.org/viewtopic.php?f=22&t=9145#p74262 ).
This is done by adding an emulation flag to a few methods in
Part::GeomXXX, which makes the shape to pretend being non-reversed
(CCW). This causes endpoints of reversed arcs of circles lineked as
external geometry to swap, causing broken sketches sometimes.
