#include <BSplSLib.hxx>
Public Member Functions | |
void * | operator new (size_t, void *anAddress) |
void * | operator new (size_t size) |
void | operator delete (void *anAddress) |
Static Public Member Functions | |
static Standard_EXPORT void | RationalDerivative (const Standard_Integer UDeg, const Standard_Integer VDeg, const Standard_Integer N, const Standard_Integer M, Standard_Real &Ders, Standard_Real &RDers, const Standard_Boolean All=Standard_True) |
Computes the derivatives of a ratio of two-variables functions x(u,v) / w(u,v) at orders <N,M>, x(u,v) is a vector in dimension <3>. <ders> is an array containing the values of the input derivatives from 0 to Min(<N>,<UDeg>), 0 to Min(<M>,<VDeg>). For orders higher than <UDeg,VDeg> the input derivatives are assumed to be 0. The <ders> is a 2d array and the dimension of the lines is always (<vdeg>+1) * (<3>+1), even if <n> is smaller than <udeg> (the derivatives higher than <n> are not used). Content of <ders> : x(i,j)[k] means : the composant k of x derivated (i) times in u and (j) times in v. ... First line ... x[1],x[2],...,x[3],w x(0,1)[1],...,x(0,1)[3],w(1,0) ... x(0,VDeg)[1],...,x(0,VDeg)[3],w(0,VDeg) ... Then second line ... x(1,0)[1],...,x(1,0)[3],w(1,0) x(1,1)[1],...,x(1,1)[3],w(1,1) ... x(1,VDeg)[1],...,x(1,VDeg)[3],w(1,VDeg) ... ... Last line ... x(UDeg,0)[1],...,x(UDeg,0)[3],w(UDeg,0) x(UDeg,1)[1],...,x(UDeg,1)[3],w(UDeg,1) ... x(Udeg,VDeg)[1],...,x(UDeg,VDeg)[3],w(Udeg,VDeg) If <all> is false, only the derivative at order <N,M> is computed. <rders> is an array of length 3 which will contain the result : x(1)/w , x(2)/w , ... derivated <n> <m> times If <all> is true multiples derivatives are computed. All the derivatives (i,j) with 0 <= i+j <= Max(N,M) are computed. <rders> is an array of length 3 * (<n>+1) * (<m>+1) which will contains : x(1)/w , x(2)/w , ... x(1)/w , x(2)/w , ... derivated <0,1> times x(1)/w , x(2)/w , ... derivated <0,2> times ... x(1)/w , x(2)/w , ... derivated <0,N> times . | |
static Standard_EXPORT void | D0 (const Standard_Real U, const Standard_Real V, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, gp_Pnt &P) |
static Standard_EXPORT void | D1 (const Standard_Real U, const Standard_Real V, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer Degree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, gp_Pnt &P, gp_Vec &Vu, gp_Vec &Vv) |
static Standard_EXPORT void | D2 (const Standard_Real U, const Standard_Real V, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, gp_Pnt &P, gp_Vec &Vu, gp_Vec &Vv, gp_Vec &Vuu, gp_Vec &Vvv, gp_Vec &Vuv) |
static Standard_EXPORT void | D3 (const Standard_Real U, const Standard_Real V, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, gp_Pnt &P, gp_Vec &Vu, gp_Vec &Vv, gp_Vec &Vuu, gp_Vec &Vvv, gp_Vec &Vuv, gp_Vec &Vuuu, gp_Vec &Vvvv, gp_Vec &Vuuv, gp_Vec &Vuvv) |
static Standard_EXPORT void | DN (const Standard_Real U, const Standard_Real V, const Standard_Integer Nu, const Standard_Integer Nv, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, gp_Vec &Vn) |
static Standard_EXPORT void | Iso (const Standard_Real Param, const Standard_Boolean IsU, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &Knots, const TColStd_Array1OfInteger &Mults, const Standard_Integer Degree, const Standard_Boolean Periodic, TColgp_Array1OfPnt &CPoles, TColStd_Array1OfReal &CWeights) |
Computes the poles and weights of an isoparametric curve at parameter <param> (UIso if <isu> is True, VIso else). . | |
static Standard_EXPORT void | Reverse (TColgp_Array2OfPnt &Poles, const Standard_Integer Last, const Standard_Boolean UDirection) |
Reverses the array of poles. Last is the Index of the new first Row( Col) of Poles. On a non periodic surface Last is Poles.Upper(). On a periodic curve last is (number of flat knots - degree - 1) or (sum of multiplicities(but for the last) + degree - 1) . | |
static Standard_EXPORT void | HomogeneousD0 (const Standard_Real U, const Standard_Real V, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, Standard_Real &W, gp_Pnt &P) |
Makes an homogeneous evaluation of Poles and Weights any and returns in P the Numerator value and in W the Denominator value if Weights are present otherwise returns 1.0e0 . | |
static Standard_EXPORT void | HomogeneousD1 (const Standard_Real U, const Standard_Real V, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, gp_Pnt &N, gp_Vec &Nu, gp_Vec &Nv, Standard_Real &D, Standard_Real &Du, Standard_Real &Dv) |
Makes an homogeneous evaluation of Poles and Weights any and returns in P the Numerator value and in W the Denominator value if Weights are present otherwise returns 1.0e0 . | |
static Standard_EXPORT void | Reverse (TColStd_Array2OfReal &Weights, const Standard_Integer Last, const Standard_Boolean UDirection) |
Reverses the array of weights. . | |
static Standard_EXPORT Standard_Boolean | IsRational (const TColStd_Array2OfReal &Weights, const Standard_Integer I1, const Standard_Integer I2, const Standard_Integer J1, const Standard_Integer J2, const Standard_Real Epsilon=0.0) |
Returns False if all the weights of the array <weights> in the area [I1,I2] * [J1,J2] are identic. Epsilon is used for comparing weights. If Epsilon is 0. the Epsilon of the first weight is used. . | |
static Standard_EXPORT void | SetPoles (const TColgp_Array2OfPnt &Poles, TColStd_Array1OfReal &FP, const Standard_Boolean UDirection) |
Copy in FP the coordinates of the poles. . | |
static Standard_EXPORT void | SetPoles (const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, TColStd_Array1OfReal &FP, const Standard_Boolean UDirection) |
Copy in FP the coordinates of the poles. . | |
static Standard_EXPORT void | GetPoles (const TColStd_Array1OfReal &FP, TColgp_Array2OfPnt &Poles, const Standard_Boolean UDirection) |
Get from FP the coordinates of the poles. . | |
static Standard_EXPORT void | GetPoles (const TColStd_Array1OfReal &FP, TColgp_Array2OfPnt &Poles, TColStd_Array2OfReal &Weights, const Standard_Boolean UDirection) |
Get from FP the coordinates of the poles. . | |
static Standard_EXPORT void | MovePoint (const Standard_Real U, const Standard_Real V, const gp_Vec &Displ, const Standard_Integer UIndex1, const Standard_Integer UIndex2, const Standard_Integer VIndex1, const Standard_Integer VIndex2, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean Rational, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UFlatKnots, const TColStd_Array1OfReal &VFlatKnots, Standard_Integer &UFirstIndex, Standard_Integer &ULastIndex, Standard_Integer &VFirstIndex, Standard_Integer &VLastIndex, TColgp_Array2OfPnt &NewPoles) |
Find the new poles which allows an old point (with a given u,v as parameters) to reach a new position UIndex1,UIndex2 indicate the range of poles we can move for U (1, UNbPoles-1) or (2, UNbPoles) -> no constraint for one side in U (2, UNbPoles-1) -> the ends are enforced for U don't enter (1,NbPoles) and (1,VNbPoles) -> error: rigid move if problem in BSplineBasis calculation, no change for the curve and UFirstIndex, VLastIndex = 0 VFirstIndex, VLastIndex = 0 . | |
static Standard_EXPORT void | InsertKnots (const Standard_Boolean UDirection, const Standard_Integer Degree, const Standard_Boolean Periodic, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &Knots, const TColStd_Array1OfInteger &Mults, const TColStd_Array1OfReal &AddKnots, const TColStd_Array1OfInteger &AddMults, TColgp_Array2OfPnt &NewPoles, TColStd_Array2OfReal &NewWeights, TColStd_Array1OfReal &NewKnots, TColStd_Array1OfInteger &NewMults, const Standard_Real Epsilon, const Standard_Boolean Add=Standard_True) |
static Standard_EXPORT Standard_Boolean | RemoveKnot (const Standard_Boolean UDirection, const Standard_Integer Index, const Standard_Integer Mult, const Standard_Integer Degree, const Standard_Boolean Periodic, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &Knots, const TColStd_Array1OfInteger &Mults, TColgp_Array2OfPnt &NewPoles, TColStd_Array2OfReal &NewWeights, TColStd_Array1OfReal &NewKnots, TColStd_Array1OfInteger &NewMults, const Standard_Real Tolerance) |
static Standard_EXPORT void | IncreaseDegree (const Standard_Boolean UDirection, const Standard_Integer Degree, const Standard_Integer NewDegree, const Standard_Boolean Periodic, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &Knots, const TColStd_Array1OfInteger &Mults, TColgp_Array2OfPnt &NewPoles, TColStd_Array2OfReal &NewWeights, TColStd_Array1OfReal &NewKnots, TColStd_Array1OfInteger &NewMults) |
static Standard_EXPORT void | Unperiodize (const Standard_Boolean UDirection, const Standard_Integer Degree, const TColStd_Array1OfInteger &Mults, const TColStd_Array1OfReal &Knots, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, TColStd_Array1OfInteger &NewMults, TColStd_Array1OfReal &NewKnots, TColgp_Array2OfPnt &NewPoles, TColStd_Array2OfReal &NewWeights) |
static TColStd_Array2OfReal & | NoWeights () |
Used as argument for a non rational curve. . | |
static Standard_EXPORT void | BuildCache (const Standard_Real U, const Standard_Real V, const Standard_Real USpanDomain, const Standard_Real VSpanDomain, const Standard_Boolean UPeriodicFlag, const Standard_Boolean VPeriodicFlag, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Integer UIndex, const Standard_Integer VIndex, const TColStd_Array1OfReal &UFlatKnots, const TColStd_Array1OfReal &VFlatKnots, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, TColgp_Array2OfPnt &CachePoles, TColStd_Array2OfReal &CacheWeights) |
Perform the evaluation of the Taylor expansion of the Bspline normalized between 0 and 1. If rational computes the homogeneous Taylor expension for the numerator and stores it in CachePoles . | |
static Standard_EXPORT void | CacheD0 (const Standard_Real U, const Standard_Real V, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Real UCacheParameter, const Standard_Real VCacheParameter, const Standard_Real USpanLenght, const Standard_Real VSpanLength, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, gp_Pnt &Point) |
Perform the evaluation of the of the cache the parameter must be normalized between the 0 and 1 for the span. The Cache must be valid when calling this routine. Geom Package will insure that. and then multiplies by the weights this just evaluates the current point the CacheParameter is where the Cache was constructed the SpanLength is to normalize the polynomial in the cache to avoid bad conditioning effects . | |
static void | CoefsD0 (const Standard_Real U, const Standard_Real V, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, gp_Pnt &Point) |
Calls CacheD0 for Bezier Surfaces Arrays computed with the method PolesCoefficients. Warning: To be used for BezierSurfaces ONLY!!! . | |
static Standard_EXPORT void | CacheD1 (const Standard_Real U, const Standard_Real V, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Real UCacheParameter, const Standard_Real VCacheParameter, const Standard_Real USpanLenght, const Standard_Real VSpanLength, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, gp_Pnt &Point, gp_Vec &VecU, gp_Vec &VecV) |
Perform the evaluation of the of the cache the parameter must be normalized between the 0 and 1 for the span. The Cache must be valid when calling this routine. Geom Package will insure that. and then multiplies by the weights this just evaluates the current point the CacheParameter is where the Cache was constructed the SpanLength is to normalize the polynomial in the cache to avoid bad conditioning effects . | |
static void | CoefsD1 (const Standard_Real U, const Standard_Real V, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, gp_Pnt &Point, gp_Vec &VecU, gp_Vec &VecV) |
Calls CacheD0 for Bezier Surfaces Arrays computed with the method PolesCoefficients. Warning: To be used for BezierSurfaces ONLY!!! . | |
static Standard_EXPORT void | CacheD2 (const Standard_Real U, const Standard_Real V, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Real UCacheParameter, const Standard_Real VCacheParameter, const Standard_Real USpanLenght, const Standard_Real VSpanLength, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, gp_Pnt &Point, gp_Vec &VecU, gp_Vec &VecV, gp_Vec &VecUU, gp_Vec &VecUV, gp_Vec &VecVV) |
Perform the evaluation of the of the cache the parameter must be normalized between the 0 and 1 for the span. The Cache must be valid when calling this routine. Geom Package will insure that. and then multiplies by the weights this just evaluates the current point the CacheParameter is where the Cache was constructed the SpanLength is to normalize the polynomial in the cache to avoid bad conditioning effects . | |
static void | CoefsD2 (const Standard_Real U, const Standard_Real V, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, gp_Pnt &Point, gp_Vec &VecU, gp_Vec &VecV, gp_Vec &VecUU, gp_Vec &VecUV, gp_Vec &VecVV) |
Calls CacheD0 for Bezier Surfaces Arrays computed with the method PolesCoefficients. Warning: To be used for BezierSurfaces ONLY!!! . | |
static void | PolesCoefficients (const TColgp_Array2OfPnt &Poles, TColgp_Array2OfPnt &CachePoles) |
Warning! To be used for BezierSurfaces ONLY!!! . | |
static Standard_EXPORT void | PolesCoefficients (const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, TColgp_Array2OfPnt &CachePoles, TColStd_Array2OfReal &CacheWeights) |
Encapsulation of BuildCache to perform the evaluation of the Taylor expansion for beziersurfaces at parameters 0.,0.; Warning: To be used for BezierSurfaces ONLY!!! . | |
static Standard_EXPORT void | Resolution (const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UKnots, const TColStd_Array1OfReal &VKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const Standard_Integer UDegree, const Standard_Integer VDegree, const Standard_Boolean URat, const Standard_Boolean VRat, const Standard_Boolean UPer, const Standard_Boolean VPer, const Standard_Real Tolerance3D, Standard_Real &UTolerance, Standard_Real &VTolerance) |
Given a tolerance in 3D space returns two tolerances, one in U one in V such that for all (u1,v1) and (u0,v0) in the domain of the surface f(u,v) we have : | u1 - u0 | < UTolerance and | v1 - v0 | < VTolerance we have |f (u1,v1) - f (u0,v0)| < Tolerance3D . | |
static Standard_EXPORT void | Interpolate (const Standard_Integer UDegree, const Standard_Integer VDegree, const TColStd_Array1OfReal &UFlatKnots, const TColStd_Array1OfReal &VFlatKnots, const TColStd_Array1OfReal &UParameters, const TColStd_Array1OfReal &VParameters, TColgp_Array2OfPnt &Poles, TColStd_Array2OfReal &Weights, Standard_Integer &InversionProblem) |
Performs the interpolation of the data points given in the Poles array in the form [1,...,RL][1,...,RC][1...PolesDimension] . The ColLength CL and the Length of UParameters must be the same. The length of VFlatKnots is VDegree + CL + 1. The RowLength RL and the Length of VParameters must be the same. The length of VFlatKnots is Degree + RL + 1. Warning: the method used to do that interpolation is gauss elimination WITHOUT pivoting. Thus if the diagonal is not dominant there is no guarantee that the algorithm will work. Nevertheless for Cubic interpolation at knots or interpolation at Scheonberg points the method will work. The InversionProblem will report 0 if there was no problem else it will give the index of the faulty pivot . | |
static Standard_EXPORT void | Interpolate (const Standard_Integer UDegree, const Standard_Integer VDegree, const TColStd_Array1OfReal &UFlatKnots, const TColStd_Array1OfReal &VFlatKnots, const TColStd_Array1OfReal &UParameters, const TColStd_Array1OfReal &VParameters, TColgp_Array2OfPnt &Poles, Standard_Integer &InversionProblem) |
Performs the interpolation of the data points given in the Poles array. The ColLength CL and the Length of UParameters must be the same. The length of VFlatKnots is VDegree + CL + 1. The RowLength RL and the Length of VParameters must be the same. The length of VFlatKnots is Degree + RL + 1. Warning: the method used to do that interpolation is gauss elimination WITHOUT pivoting. Thus if the diagonal is not dominant there is no guarantee that the algorithm will work. Nevertheless for Cubic interpolation at knots or interpolation at Scheonberg points the method will work. The InversionProblem will report 0 if there was no problem else it will give the index of the faulty pivot . | |
static Standard_EXPORT void | FunctionMultiply (const BSplSLib_EvaluatorFunction &Function, const Standard_Integer UBSplineDegree, const Standard_Integer VBSplineDegree, const TColStd_Array1OfReal &UBSplineKnots, const TColStd_Array1OfReal &VBSplineKnots, const TColStd_Array1OfInteger &UMults, const TColStd_Array1OfInteger &VMults, const TColgp_Array2OfPnt &Poles, const TColStd_Array2OfReal &Weights, const TColStd_Array1OfReal &UFlatKnots, const TColStd_Array1OfReal &VFlatKnots, const Standard_Integer UNewDegree, const Standard_Integer VNewDegree, TColgp_Array2OfPnt &NewNumerator, TColStd_Array2OfReal &NewDenominator, Standard_Integer &Status) |
this will multiply a given BSpline numerator N(u,v) and denominator D(u,v) defined by its U/VBSplineDegree and U/VBSplineKnots, and U/VMults. Its Poles and Weights are arrays which are coded as array2 of the form [1..UNumPoles][1..VNumPoles] by a function a(u,v) which is assumed to satisfy the following : 1. a(u,v) * N(u,v) and a(u,v) * D(u,v) is a polynomial BSpline that can be expressed exactly as a BSpline of degree U/VNewDegree on the knots U/VFlatKnots 2. the range of a(u,v) is the same as the range of N(u,v) or D(u,v) ---Warning: it is the caller's responsability to insure that conditions 1. and 2. above are satisfied : no check whatsoever is made in this method -- Status will return 0 if OK else it will return the pivot index -- of the matrix that was inverted to compute the multiplied -- BSpline : the method used is interpolation at Schoenenberg -- points of a(u,v)* N(u,v) and a(u,v) * D(u,v) Status will return 0 if OK else it will return the pivot index of the matrix that was inverted to compute the multiplied BSpline : the method used is interpolation at Schoenenberg points of a(u,v)*F(u,v) -- |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|