Optimizing Performance#

Freezing planes with computationally expensive attribute access#

Custom planes may implement stateful or otherwise custom attributes that are dynamically computed at runtime. Although these attributes remain fixed during a propagation, they may be unnecessarily recalculated each time they are accessed (for example, when performing a broadband propagation). To mitigate this performance imapct, it can be beneficial to operate on a copy of the plane where its dynamic attributes are cached or “frozen”. A plane’s freeze() method makes this simple:

frozen_pupil = pupil.freeze()

for wl in wavelengths:
    w = lentil.Wavefront(wl)
    w = w * frozen_pupil  # cached OPD value is used here
    w = lentil.propagate_dft(w, pixelscale=5e-6, shape=(64,64), oversample=5)
    img += w.intensity

Using Wavefront.insert() to reduce array allocation during broadband propagation#

Each time wavefront.field or wavefront.intensity is accessed, a new Numpy array of zeros with shape = wavefront.shape is allocated. It is possible to avoid repeatedly allocating large arrays of zeros when accumulating the result of a broadband propagation by using Wavefront.insert() instead. This can result in significant performance gains, particularly when wavefront.shape is large.

The above example can be rewritten to use Wavefront.insert() instead:

for wl in wavelengths:
    w = lentil.Wavefront(wl)
    w = w * pupil
    w = lentil.propagate_dft(w, pixelscale=5e-6, shape=(64,64),
                             oversample=5)
    img = w.insert(img)

Selecting sensible propagation parameters#

Performing numerical propagation is almost always the most expensive part of any model. Even the most efficient and streamlined models will experience high memory usage and slow performance when working with large arrays and many wavelengths. Understanding how the propagation method’s parameters influence accuracy and speed will allow you to identify appropriate values for your specific needs.

`propagate_dft()`#

Parameter	Usage	Performance considerations
`wave`, `weight`	Arrays representing wavelength and the corresponding weight of each wavelength in the propagation.	Propagation time scales linearly with the number of wavelengths.
`shape`	Shape of output plane.	The output plane size has minimal impact on propagation performance unless it is astronomically large. `shape` should be set to ensure all data is adequately captured by the output plane.
`prop_shape`	Shape of propagation plane.	Propagation time scales quadtatically with `prop_shape`. `prop_shape` should be chosen to capture the data extent necessary to accurately model all required diffraction effects.
`oversample`	Number of times to oversample the output and propagation planes.	Propagation time scales quadratically with `oversample`. For accuracy, `oversample` should be selected to ensure propagations are Nyquist sampled, but there is typically no benefit in selecting larger values.

Strategies for increasing runtime performance#

Perform fewer propagations by increasing wavelength sampling#

Selecting an appropriate wavelength sampling depends on many factors. Any opportunity to reduce the number of propagations required will increase overall model performance. Two patterns for more coarse wavelength sampling are provided here

Use appropriate plane sampling#

Planes should be sized to ensure the smallest spatial features of interest are adequately sampled. Small features (e.g. secondary mirror supports, segment gaps) should be represented by at leat two samples.