ENH: Add ASV-over-native-ITK performance benchmark harness

[hjmjohnson]

Adds an airspeed-velocity (asv) harness that drives the existing native C++ *Benchmark executables for longitudinal performance-regression tracking. Intended for use from ITK CI — companion change to InsightSoftwareConsortium/ITK#612 (replacing the retired Azure macos-11 job).

Also hardens CopyIterationBenchmark against a SIGSEGV exposed by ITK commit 1d87efa52, so the benchmark runs cleanly across the entire ITK v5.3 → current main range.

Architecture (ASV harness)

• environment_type: "existing" keeps ASV out of the ITK build loop. The consuming CI workflow builds ITK (base + head) and points the shim at each build via ITK_BENCHMARK_BIN.
• Each ASV benchmark is a track_* returning seconds parsed from the jsonxx output of HighPriorityRealTimeProbesCollector, bypassing ASV's wall-clock instrumentation of the Python callable.
• repo: "itk-repo" — the caller places a symlink named itk-repo at the harness root pointing at the ITK clone. This lets ASV validate --set-commit-hash values against git, so result labels track ITK history rather than this repository's history.

Contents

• asv.conf.json — ASV configuration
• python/itk_perf_shim/ — subprocess shim parsing jsonxx timings
• benchmarks/{core,filtering,segmentation}.py — 11 prototype benchmarks
• README-asv.md — design notes and smoke-test recipe (validated)
• examples/Core/itkCopyIterationBenchmark.cxx — harden VectorImage pixel construction

CopyIterationBenchmark hardening

Previously, CreateAndInitializeImage seeded pixels via it.Set(static_cast<PixelType>(count)). For scalar / FixedArray pixels this is a value cast, but for VectorImage — whose PixelType is itk::VariableLengthVector<T> — it invokes the VLV length constructor: VLV<float>(count) yields a vector of size count, not a 3-component vector with value count. The first iteration produces a length-0 VLV; ITK's DefaultVectorPixelAccessor::Set then copies NumberOfComponentsPerPixel elements from that source's internal data pointer, which is nullptr on ITK ≥ 1d87efa52 → SIGSEGV.

This commit introduces a PixelFiller<T> trait specialized for VariableLengthVector<T> that constructs a correctly-sized VLV and Fills it with count. The scalar / FixedArray path is unchanged. The benchmark now produces well-defined output across the full ITK bisect range regardless of the internal representation of VariableLengthVector(0).

A companion ITK-side fix is proposed in InsightSoftwareConsortium/ITK#6087 — defensive null-guard in the pixel accessor so any other caller that triggers the same class of bug no longer crashes.

Scope

Harness covers Core (2), Filtering (5), and Segmentation (4) benchmarks — all smoke-validated end-to-end. Registration and Resample benchmarks are intentionally omitted; their C++ sources hit unrelated ITK-level build issues (NrrdIO airFloatQNaN link, Resample -Wmaybe-uninitialized) that are outside the scope of this PR. Adding them later is a small registry + benchmarks/ edit.

Relation to the prior wasm-asv effort

Supersedes the stalled wasm-asv direction on @thewtex's fork (last touched 2024-05-29), which introduced a pnpm+emscripten+wasi dependency surface inappropriate for a pure regression-detection harness.

Smoke-test sample output

asv compare against two labeled runs of the same binary (same-machine jitter only):

| Before [c6b6b608] | After [3a8d2fa6] | Ratio | Benchmark                                         |
|-------------------|------------------|-------|---------------------------------------------------|
| 10.5ms            | 10.2ms           |  0.98 | core.CoreSuite.track_copy_iteration               |
| 4.98ms            | 4.61ms           |  0.93 | core.CoreSuite.track_vector_iteration             |
| 5.39ms            | 5.09ms           |  0.94 | filtering.FilteringSuite.track_binary_add         |
| 48.3ms            | 47.3ms           |  0.98 | filtering.FilteringSuite.track_gradient_magnitude |
| 321ms             | 312ms            |  0.97 | filtering.FilteringSuite.track_median             |

Test plan

• asv check --python=same discovers all benchmarks cleanly
• End-to-end asv run against a local ITK v5.4.5 build (79/79 tests pass, validated 2026-04-19)
• End-to-end run against current upstream main HEAD: CopyIterationBenchmark runs cleanly with the hardening (was SIGSEGV before this commit)
• Probe-JSON parse keys verified against real HighPriorityRealTimeProbesCollector output
• Consuming ITK PR's GHA workflow lands a green run

[dzenanz]

Sounds great, and looks good on a glance. I can hardly wait to see this fully integrated into our CI!

[thewtex]

🚀 🐦