vspline-dev - 1.0.2-1 main

vspline aims to be as fast as possible, yet mathematically precise.
It's main focus is interpolation of bulk data like sounds, images and
volumes. vspline can handle
.
- splines over real and integer data types and their aggregates:
- all '*xel' data, arbitrary number of channels (template argument)
- single, double precision and long doubles supported (template argument)
- several boundary conditions (mirror, reflect, periodic, natural)
- spline degree up to 45 (runtime argument)
- arbitrary dimensionality of the spline (template argument)
- specialized code for 1D data
- multithreaded code (pthread)
- using the CPU's vector units if possible (like SSE, AVX/2)
- processes nD arrays with arbitrary shapes and strides
.
On the evaluation side it provides
.
- evaluation of the spline at point locations in the defined range
- evaluation of vectorized arguments with SIMD code
- optional use of Vc for explicit SIMD code using intrinsics
- evaluation of the spline's derivatives
- factory functions to create evaluation functors
- specialized code for degrees 0 and 1 (nearest neighbour and n-linear)
- mapping of arbitrary coordinates into the defined range
- evaluation of nD arrays of coordinates ('remap' function)
- discrete-coordinate-fed remap function ('index_remap')
- generalized functor-based 'apply' and 'transform' functions
- restoration of the original data from the coefficients
.
To produce maximum performance, vspline has a fair amount of collateral code,
and some of this code may be helpful beyond vspline:
.
- multithreading with a thread pool
- application of fundamental and SIMD functors to nD arrays
- functional constructs using vspline::unary_functor
- nD forward-backward n-pole recursive filtering
- nD separable convolution
- efficient access to the b-spline basis functions
- precise precalculated constants (made with GNU GSL, BLAS and GNU GMP)
- many examples, ample comments
.
vspline uses a three-step approach to splines: the first step is coefficient
generation (including prefilering), the second step is generation of specific
functors providing evaluation, the third is application of these functors
to n-dimensional data.
data handling is done with vigra data types, using vigra::MultiArrayView
for handling strided nD arrays, and vigra::TinyVector for small aggregates.
vigra::MultiArrayView is a thin wrapper combining an array's address,
dimension, data type, stride and shape, it's easy to use it to 'package'
raw data. vspline handles vector formation from interleaved data, peeling
and re-interleaving automatically.
vspline optionally uses Vc. Without Vc present, it triggers the compiler's
autovectorization by producing deliberately vector-friendly inner loops.
vspline has been under development for several years and is extensively
tested for scope coverage, reliability, precision and performance.