Web lists-archives.com

ITP: vspline -- generic C++ code for uniform b-splines, remap functions




Package: wnpp
Severity: wishlist
Owner: "Kay F. Jahnke" <kfjahnke@xxxxxxxxx>
User: debian-science@xxxxxxxxxxxxxxxx
Usertags: field..mathematics

* Package name    : vspline
  Version         : 0.1.1
  Upstream Author : Kay F. Jahnke <kfjahnke@xxxxxxxxx>
* URL             : https://bitbucket.org/kfj/vspline
* License         : EXPAT
  Programming Lang: C++11
  Description     : generic C++ code for uniform b-splines, remap functions

vspline is a header-only generic C++ library trying to address all
aspects of uniform b-splines. This includes b-spline prefiltering
(conversion of the source data into b-spline coefficients), evaluation
of the spline, and mass evaluation of data with remap-like functions.

I am the developer of this software, and members of the debian science
team have  suggested I should get involved with packaging. The code is
now ready to be packaged, there is a repository at alioth already with
packaging information:

https://anonscm.debian.org/cgit/debian-science/packages/vspline.git

I intend to do the packaging.

I do seek for a mentor/sponsor.

Here's my writeup about vspline:

vspline can create uniform b-splines of

 -  real data types and their aggregates [*]
 -  coming in strided memory
 -  with a reasonable selection of boundary conditions
 -  used in either an implicit or an explicit scheme of extrapolation
 -  arbitrary spline order (up to 24)
 -  arbitrary dimensions of the spline
 -  with multithreaded code
 -  optionally using the CPU's vector units

on the evaluation side it provides

 -  evaluation of the spline at point locations in the defined range
 -  evaluation of the spline's derivatives
 -  mapping of arbitrary coordinates into the defined range
 -  evaluation of nD arrays of coordinates ('remap' function)
 -  target coordinate-fed remap function ('index_remap')
 -  functor-based remap, aka 'transform' function
 -  functor-based 'apply' function

[*] I use 'aggregate' here in the sense of 'several of the same type',
    rather than 'several of any type'

while there is no lack of freely available b-spline code, vspline's aim
is to provide very fast prefiltering and evaluation by exploiting
multithreading and (optionally) the CPU's vector units. multithreading
is done with a tailor-made multithreading implementation using a thread
pool, while vectorization is done generically by using Vc. vspline's
remap routines bring these multithreading and vectorization capabilities
to bear when processing/generating mass data. The remapping routines can
produce interpolated values for nD arrays of coordinates or, more
generally, operate with functors which allow for the processing of
arbitrary value-generating pipelines. vspline's current main focus is on
image processing, but since the code is dimension-agnostic, it can
handle volume data as well, and there are specializations for 1D data, too.

vspline relies on vigra for data handling; vigra offers a convenient
zero-overhead nD array 'view' type, which can easily adopt regular nD
data by passing their shape and strides. It also offers efficient
handling of small aggregates of uniform type (like pixels) which vspline
also relies on.

I use the code in vspline in it's companion project pv, which is a
panoramic image viewer, available from

https://bitbucket.org/kfj/pv

vspline is probably beta stage, but due to the steady use in pv it is
well-tuned especially for image processing. The code has stabilized
nicely, so I think that presenting it to debian is appropriate. The very
low version number is due to the recent adoption of a debian-friendly
tagging scheme. Nevertheless I'd say that vspline is still experimental
- I only have access to a limited set of systems to experiment with, and
there are quite a few heuristics in the code which are probably
suboptimal on some targets. I'd welcome others to experiment with the
code and share their results to improve vspline.

When I started working on b-splines a few years ago, I worked a lot with
libeinspline, which is also available as a debian package, but is coded
in C and only offers cubic b-splines in up to three dimensions. I wanted
a wider scope and a modern code base in C++, and I wanted to use the
signal processing approach to b-splines rather than the linear algebra
approach used in libeinspline, so I set out on vspline. vigra, which I
use for data handling, also offers b-spline code, but I felt more
comfortable implementing the maths myself to fine-tune weight
generation, multithreading and vectorization, and also to make the
spline degree a run-time parameter rather than a template parameter.

Another source for b-spline code is opencv, which also offers a
remap function using b-spline interpolation, as does scipy. I made an
attempt to extend the concept of remapping. The classical remap function
takes an array of coordinates and produces an equally-shaped array of
interpolated values. This can be seen as a specialization of an 'apply'
or 'transform' function where the coordinate value is transformed into
the interpolated value at this location. vspline provides multithreaded, vectorized apply and transform routines and implements remapping as a specialization.

With regards
Kay F. Jahnke