Files
create/src/Mod/Arch/Dice3DS/util.py
Uwe 04570f81cc [Arch] remove unneeded checks for Python 3
since we only support Python >=3, these checks can go
2022-08-04 01:32:06 +02:00

275 lines
9.3 KiB
Python

# util.py
"""Utitily function for Dice3DS.
Defines some routines for calculating normals and transforming points.
"""
import numpy
# xrange is not available in python3
xrange = range
# Can push numpy.float64 (or even numpy.float80) into this if you
# would like to use higher precision when calculating; results will be
# converted back to numpy.float32
_calc_precision_type = numpy.float32
def translate_points(pointarray,matrix):
"""Translate points in pointarray by the given matrix.
tpointarray = translate_points(pointarray,matrix)
Takes array of points and a homogeneous (4D) transformation
matrix in exactly the same form in which they appear in the
3DS DOM.
Returns a pointarray with the points transformed by the matrix.
"""
n = len(pointarray)
pt = numpy.ones((n,4),_calc_precision_type)
pt[:,:3] = pointarray
tpt = numpy.transpose(numpy.dot(matrix,numpy.transpose(pt)))
return numpy.asarray(tpt[:,:3]/tpt[:,3:4],numpy.float32)
def calculate_normals_no_smoothing(pointarray,facearray,smarray=None):
"""Calculate normals all perpendicular to the faces.
points,norms = calculate_normals_no_smoothing(
pointarray,facearray,smarray=None)
Takes an array of points and faces in exactly the same form in
which they appear in the 3DS DOM. It accepts a smoothing array,
but ignores it.
Returns a numpy.array of points, one per row, and a
numpy.array of the corresponding normals. The points are
returned as a list of consecutive triangles; the first three rows
make up the first triangle, the second three rows make up the
second triangle, and so on.
The normal vectors are determined by calculating the normal to
each face. There is no smoothing.
"""
# prepare to calculate normals. define some arrays
m = len(facearray)
fnorms = numpy.empty((m*3,3),_calc_precision_type)
points = pointarray[facearray.ravel()]
# calculate normals for each face
A = numpy.asarray(pointarray[facearray[:,0]],_calc_precision_type)
B = numpy.asarray(pointarray[facearray[:,1]],_calc_precision_type)
C = numpy.asarray(pointarray[facearray[:,2]],_calc_precision_type)
b = A - C
c = B - A
fnorms[2::3,0] = c[:,2]*b[:,1]-c[:,1]*b[:,2]
fnorms[2::3,1] = c[:,0]*b[:,2]-c[:,2]*b[:,0]
fnorms[2::3,2] = c[:,1]*b[:,0]-c[:,0]*b[:,1]
a = fnorms[2::3]
q = numpy.maximum(numpy.sqrt(numpy.sum(a*a,axis=1)),1e-10)
q = q[:,numpy.newaxis]
a /= q
fnorms[0::3] = fnorms[1::3] = fnorms[2::3]
# we're done
return points, numpy.asarray(fnorms,numpy.float32)
def calculate_normals_by_cross_product(pointarray,facearray,smarray):
"""Calculate normals by smoothing, weighting by cross-product.
points,norms = calculate_normals_by_cross_product(
pointarray,facearray,smarray)
Takes an array of points, faces, and a smoothing group in exactly
the same form in which they appear in the 3DS DOM.
Returns a numpy.array of points, one per row, and a numpy.array of
the corresponding normals. The points are returned as a list of
consecutive triangles; the first three rows make up the first
triangle, the second three rows make up the second triangle, and
so on.
To calculate the normal of a given vertex on a given face, this
function averages the normal vector for all faces which have share
that vertex and a smoothing group.
The normals being averaged are weighted by the cross-product used
to obtain the face's normal, which is proportional to the area of
the face.
"""
# prepare to calculate normals. define some arrays
m = len(facearray)
rnorms = numpy.zeros((m*3,3),_calc_precision_type)
fnorms = numpy.zeros((m*3,3),_calc_precision_type)
points = pointarray[facearray.ravel()]
exarray = numpy.zeros(3*m,numpy.uint32)
if smarray is not None:
exarray[0::3] = exarray[1::3] = exarray[2::3] = smarray
# calculate scaled normals (according to angle subtended)
A = numpy.asarray(pointarray[facearray[:,0]],_calc_precision_type)
B = numpy.asarray(pointarray[facearray[:,1]],_calc_precision_type)
C = numpy.asarray(pointarray[facearray[:,2]],_calc_precision_type)
a = C - B
b = A - C
c = B - A
rnorms[0::3,0] = c[:,2]*b[:,1]-c[:,1]*b[:,2]
rnorms[0::3,1] = c[:,0]*b[:,2]-c[:,2]*b[:,0]
rnorms[0::3,2] = c[:,1]*b[:,0]-c[:,0]*b[:,1]
rnorms[1::3,0] = a[:,2]*c[:,1]-a[:,1]*c[:,2]
rnorms[1::3,1] = a[:,0]*c[:,2]-a[:,2]*c[:,0]
rnorms[1::3,2] = a[:,1]*c[:,0]-a[:,0]*c[:,1]
rnorms[2::3,0] = b[:,2]*a[:,1]-b[:,1]*a[:,2]
rnorms[2::3,1] = b[:,0]*a[:,2]-b[:,2]*a[:,0]
rnorms[2::3,2] = b[:,1]*a[:,0]-b[:,0]*a[:,1]
# normalize vectors according to passed in smoothing group
lex = numpy.lexsort(numpy.transpose(points))
brs = numpy.nonzero(
numpy.any(points[lex[1:],:]-points[lex[:-1],:],axis=1))[0]+1
lslice = numpy.empty((len(brs)+1,),numpy.int)
lslice[0] = 0
lslice[1:] = brs
rslice = numpy.empty((len(brs)+1,),numpy.int)
rslice[:-1] = brs
rslice[-1] = 3*m
for i in xrange(len(brs)+1):
rgroup = lex[lslice[i]:rslice[i]]
xgroup = exarray[rgroup]
normpat = numpy.logical_or(
numpy.bitwise_and.outer(xgroup,xgroup),
numpy.eye(len(xgroup)))
fnorms[rgroup,:] = numpy.dot(normpat,rnorms[rgroup,:])
q = numpy.sum(fnorms*fnorms,axis=1)
qnz = numpy.nonzero(q)[0]
lq = 1.0 / numpy.sqrt(q[qnz])
fnt = numpy.transpose(fnorms)
fnt[:,qnz] *= lq
# we're done
return points, numpy.asarray(fnorms,numpy.float32)
def calculate_normals_by_angle_subtended(pointarray,facearray,smarray):
"""Calculate normals by smoothing, weighting by angle subtended.
points,norms = calculate_normals_by_angle_subtended(
pointarray,facearray,smarray)
Takes an array of points, faces, and a smoothing group in exactly
the same form in which they appear in the 3DS DOM.
Returns a numpy.array of points, one per row, and a numpy.array of
the corresponding normals. The points are returned as a list of
consecutive triangles; the first three rows make up the first
triangle, the second three rows make up the second triangle, and
so on.
To calculate the normal of a given vertex on a given face, this
function averages the normal vector for all faces which have share
that vertex, and a smoothing group.
The normals being averaged are weighted by the angle subtended.
"""
# prepare to calculate normals. define some arrays
m = len(facearray)
rnorms = numpy.zeros((m*3,3),_calc_precision_type)
fnorms = numpy.zeros((m*3,3),_calc_precision_type)
points = pointarray[facearray.ravel()]
exarray = numpy.zeros(3*m,numpy.uint32)
if smarray is not None:
exarray[0::3] = exarray[1::3] = exarray[2::3] = smarray
# weed out degenerate triangles first
# unlike cross-product, angle subtended blows up on degeneracy
A = numpy.asarray(pointarray[facearray[:,0]],_calc_precision_type)
B = numpy.asarray(pointarray[facearray[:,1]],_calc_precision_type)
C = numpy.asarray(pointarray[facearray[:,2]],_calc_precision_type)
a = C - B
b = A - C
c = B - A
p = numpy.zeros((len(facearray),3),_calc_precision_type)
p[:,0] = c[:,2]*b[:,1]-c[:,1]*b[:,2]
p[:,1] = c[:,0]*b[:,2]-c[:,2]*b[:,0]
p[:,2] = c[:,1]*b[:,0]-c[:,0]*b[:,1]
aa = numpy.sum(a*a,axis=1)
bb = numpy.sum(b*b,axis=1)
cc = numpy.sum(c*c,axis=1)
pp = numpy.sum(p*p,axis=1)
ndg = numpy.nonzero(numpy.logical_and.reduce((aa,bb,cc,pp)))[0]
# calculate scaled normals (according to angle subtended)
p = p[ndg]
la = numpy.sqrt(aa[ndg])
lb = numpy.sqrt(bb[ndg])
lc = numpy.sqrt(cc[ndg])
lp = numpy.sqrt(pp[ndg])
sinA = numpy.clip(lp/lb/lc,-1.0,1.0)
sinB = numpy.clip(lp/la/lc,-1.0,1.0)
sinC = numpy.clip(lp/la/lb,-1.0,1.0)
sinA2 = sinA*sinA
sinB2 = sinB*sinB
sinC2 = sinC*sinC
angA = numpy.arcsin(sinA)
angB = numpy.arcsin(sinB)
angC = numpy.arcsin(sinC)
angA = numpy.where(sinA2 > sinB2 + sinC2, numpy.pi - angA, angA)
angB = numpy.where(sinB2 > sinA2 + sinC2, numpy.pi - angB, angB)
angC = numpy.where(sinC2 > sinA2 + sinB2, numpy.pi - angC, angC)
rnorms[0::3][ndg] = p*(angA/lp)[:,numpy.newaxis]
rnorms[1::3][ndg] = p*(angB/lp)[:,numpy.newaxis]
rnorms[2::3][ndg] = p*(angC/lp)[:,numpy.newaxis]
# normalize vectors according to passed in smoothing group
lex = numpy.lexsort(numpy.transpose(points))
brs = numpy.nonzero(
numpy.any(points[lex[1:],:]-points[lex[:-1],:],axis=1))[0]+1
lslice = numpy.empty((len(brs)+1,),numpy.int)
lslice[0] = 0
lslice[1:] = brs
rslice = numpy.empty((len(brs)+1,),numpy.int)
rslice[:-1] = brs
rslice[-1] = 3*m
for i in xrange(len(brs)+1):
rgroup = lex[lslice[i]:rslice[i]]
xgroup = exarray[rgroup]
normpat = numpy.logical_or(
numpy.bitwise_and.outer(xgroup,xgroup),
numpy.eye(len(xgroup)))
fnorms[rgroup,:] = numpy.dot(normpat,rnorms[rgroup,:])
q = numpy.sum(fnorms*fnorms,axis=1)
qnz = numpy.nonzero(q)[0]
lq = 1.0 / numpy.sqrt(q[qnz])
fnt = numpy.transpose(fnorms)
fnt[:,qnz] *= lq
# we're done
return points, numpy.asarray(fnorms,numpy.float32)