Tracker: cupyx.scipy.interpolate

Description

This is an issue to keep track of the current improvement work to the cupyx.scipy.interpolate module. The order of the checkboxes denote the planned order and priority of implementation of each module. This plan was devised by @ev-br and is subject to change and discussion.

First phase

• BSpline → https://github.com/cupy/cupy/pull/7086

• RBFInterpolator → https://github.com/cupy/cupy/pull/7190

• make_interp_spline → https://github.com/cupy/cupy/pull/7195

• RegularGridInterpolator: methods linear and nearest → https://github.com/cupy/cupy/pull/7197

• PPoly → https://github.com/cupy/cupy/pull/7204

• make_interp_spline(..., bc_type="periodic") → https://github.com/cupy/cupy/pull/7206

• CubicHermiteSpline: Required by pchip/akima/CubicSpline → https://github.com/cupy/cupy/pull/7242

• PchipInterpolator: aliases (pchip, pchip_interpolate) → https://github.com/cupy/cupy/pull/7255

• Akima1DInterpolator

Second phase (details subject to change)

• RegularGridInterpolator: methods cubic, quintic, slinear, pchip
• Interpn: alias to RegularGridInterpolator
• CubicBSpline
• splantider, splder → https://github.com/cupy/cupy/pull/7086
• BPoly
• NdPPoly

---

QHull-dependent modules

• make_lsq_spline
• interp1d: @ev-br states that it could be ignored (maybe?)

“Odds and ends”

• BarycentricInterpolator → https://github.com/cupy/cupy/pull/6958
• barycentric_interpolate → https://github.com/cupy/cupy/pull/6958
• KroghInterpolator → https://github.com/cupy/cupy/pull/6990
• krogh_interpolate → https://github.com/cupy/cupy/pull/6990
• approximate_taylor_polynomial
• lagrange
• pade

Problematic modules and functions

QHull-dependent

• `CloughTocher2DInterpolator
• LinearNDInterpolator
• NearestNDInterpolator
• griddata: a wrapper over the above three
• RBFInterpolator(..., neighbors=not None): Depends on KD-Tree

FITPACK code (Black-box Fortran code)

• BivariateSpline
• LSQBivariateSpline
• LSQSphereBivariateSpline
• RectBivariateSpline
• RectSphereBivariateSpline
• SmoothBivariateSpline
• SmoothSphereBivariateSpline
• UnivariateSpline
• InterpolatedUnivariateSpline
• LSQUnivariateSpline
• bisplev
• bisplrep
• insert
• spalde
• splev
• splint
• splprep
• splrep
• sproot

---

Deprecated APIs

• Interp2d
• Rbf

New APIs (not available in SciPy main)

• make_smooth_spline
• NdBSpline

Please feel free to discuss any changes to this roadmap/exploration.

cc @rgommers @ev-br

Additional Information

No response

[ev-br]

I guess I’d need to explain a bit the ordering here. This is indeed opinionated 😃 and is very much up to a discussion.

First phase: once these are all done, cupyx.interpolate can be declared MVP: most common usage is possible. The first phase work can be done partly asynchronously:

• make_interp_spline is a factory which constructs BSpline instances from data.
• PPoly is a base class (similar to BSpline) needed for actual interpolators. CubicHermiteSpline is a PPoly subclass, which is then the base for pchip, akima and CubicSpline. The hard part is I guess PPoly, as it has cython code / needs C kernels. The rest is pure python.
• CubicSpline is listed in the 2nd phase mainly because its functionality is a subset of make_interp_spline. Since the work on BSpline is already quite far, I don’t see a reason to not finish it up. (Did I mention that the list is opinionated?)
• RGI is independent
• RBF is independent

So the first phase seems to have four lines of work which can be done in parallel: BSpline/make_interp_spline, PPoly and subclasses, RGI, RBF. In principle, once can pospone the BSpline work and concentrate on PPoly instead if there is a strong feel for it.

The second phase:

There is a small typo (which propagates from the original where I made this typo): BPoly, NdPPoly, CubicSpline and interp1d are not Qhull-based.

BPoly is very similar to PPoly but less used (Mainly superseded by CubicHermiteSpline), NdPPoly is underused, so can be left for later. interp1d is legacy.

All these can be pushed to later, after the QHull based stuff is done. And that needs qhull or a cuda analog. Is there a triangulation library for CUDA we can depend on?

The last big question is what to do with FITPACK based smoothing splines. Hopefully this question can be postponed 😃.

[rgommers]

If I wanted to contribute to this work, how would I do it? Do I just pick something from the list without a check mark or a link to a pull request, and start working on it?

You can just pick something and start digging in, CuPy welcomes PRs from everyone and it doesn’t require pre-assigning a topic. However, given that this tracker is currently quite actively being worked on, I would suggest that once you’ve looked into a particular API for a little bit and think it’s a good fit for you, to post here that you’re tackling it - just to avoid double work.