-
Notifications
You must be signed in to change notification settings - Fork 85
Timing
Data obtained using a 3.6 GHz 10-Core Intel Core i9, Python 3.8.5 and Numpy 1.19.1.
Base package timing to create a rotation matrix, and to convert between SO(3) and SE(3) NumPy matrices
transforms.rotx: 3.6 μs
transforms.trotx: 5 μs
transforms.t2r: 0.469 μs
transforms.r2t: 0.987 μs
The spatial math class adds overhead
SE3.Rx: 10.9 μs
largely in the instance constructor process, we see that the constructor adds 5.9 μs to the cost of an SE(3) rotation matrix base.trotx
.
A chunk of this is attributable to the cost of calling getvector
to process the possible list of values and the constructor creating time
getvector(x): 1.46 μs
SE3(): 2.93 μs
The effect of value checking can be seen when creating an SE3
instance from a NumPy matrix value
SE3(T1): 36.4 μs
SE3(T1 check=False): 1.35 μs
Multiplication using the overloaded class operator vs native NumPy also shows some constructor overhead
SE3 *: 4.72 μs
4x4 @: 1.17 μs
T1 * T2 (R, t): 2.2 μs
but the NumPy @
operator if faster than coding the multiplication in terms of the R and t partitions of the SE(3) matrix.
Inverse also shows some constructor overhead
SE3.inv: 7.24 μs
base.trinv: 4.15 μs
np.linalg.inv: 4.65 μs
but it is interesting to note that exploiting the structure of the SE(3) matrix makes trinv()
~10% faster than the general matrix inverse in NumPy.
To build an SE(3) matrix it is fastest to
T = np.zeros((4,4))
T[3,3] = 1
T[:3,:3] = R
T[:3,3] = t
np.eye(4)
is slower even though the bottom right element doesn't need to be explicitly set. np.pad
or stack
is way slower. np.empty
is faster than np.zeros
but we more than lose the advantage having to initialise the bottom row.