ENH: Umbrella — stage VNL → Eigen numerics migration for ITKv7
Spun out of the Q5 decision on #6230. Maintainer sentiment there favored reviving the long-dormant 2018 plan to migrate ITK's internal numerics from VNL to Eigen; this issue stages that work as an incremental, multi-release deprecation rather than a single break.
Why now / background
The VNL → Eigen aspiration dates to the 2018 internal-Eigen adoption (Discourse #1429) and was reaffirmed on #6230 (Q5) by @thewtex ("+1 for replacing VNL"), @hjmjohnson, and conditionally @dzenanz ("a whole lot of work; if someone is willing to do it"). @phcerdan noted the main hard dependency is VNL's FFT layer, with candidate replacements pocketfft / ducc.
Recent groundwork already reduced VNL surface: vnl_random → std:: random variants.
Proposed staging (incremental, value-tolerant)
vnl_svd → Eigen SVD (first step; @hjmjohnson). Expect numerical drift from algorithmic differences — fortify affected tests with tolerance-based comparison rather than bit-identical baselines.- Inventory remaining
vnl_* linear-algebra call sites (matrix/vector, eigenvalue, least-squares).
- FFT layer — evaluate pocketfft/ducc as
vnl_fft replacement (most permissive licensing vs FFTW).
- Deprecation cycle for public accessors (
GetVnlVector → GetInternalVector, per @phcerdan 2018) spanning ≥2 releases.
Each step is a separate PR; this issue tracks the set.
Blocked-by / depends-on: the #6230 architectural commitment that Eigen stays TU-local and out of public ITK API holds for any new Eigen usage introduced here.
ENH: Umbrella — stage VNL → Eigen numerics migration for ITKv7
Spun out of the Q5 decision on #6230. Maintainer sentiment there favored reviving the long-dormant 2018 plan to migrate ITK's internal numerics from VNL to Eigen; this issue stages that work as an incremental, multi-release deprecation rather than a single break.
Why now / background
The VNL → Eigen aspiration dates to the 2018 internal-Eigen adoption (Discourse #1429) and was reaffirmed on #6230 (Q5) by @thewtex ("+1 for replacing VNL"), @hjmjohnson, and conditionally @dzenanz ("a whole lot of work; if someone is willing to do it"). @phcerdan noted the main hard dependency is VNL's FFT layer, with candidate replacements pocketfft / ducc.
Recent groundwork already reduced VNL surface:
vnl_random→std::random variants.Proposed staging (incremental, value-tolerant)
vnl_svd→ Eigen SVD (first step; @hjmjohnson). Expect numerical drift from algorithmic differences — fortify affected tests with tolerance-based comparison rather than bit-identical baselines.vnl_*linear-algebra call sites (matrix/vector, eigenvalue, least-squares).vnl_fftreplacement (most permissive licensing vs FFTW).GetVnlVector→GetInternalVector, per @phcerdan 2018) spanning ≥2 releases.Each step is a separate PR; this issue tracks the set.
Blocked-by / depends-on: the #6230 architectural commitment that Eigen stays TU-local and out of public ITK API holds for any new Eigen usage introduced here.