PERF: Hoist constant tensor-basis SVD out of the per-voxel DTI loop

[hjmjohnson]

DiffusionTensor3DReconstructionImageFilter constructed a vnl_svd of the constant m_TensorBasis once per voxel (~16.7M times for a 256³ DWI) to solve the Stejskal–Tanner equations. The pseudo-inverse (the dual tensor basis) is identical for every voxel, so this precomputes it once in BeforeThreadedGenerateData and applies it per voxel as a matrix-vector product. Since vnl_svd::solve(X) is pinverse()·X, the reconstructed tensors are unchanged. Removing the per-voxel SVD also frees the loop of the thread-unsafe netlib dsvdc call, so the stale single-thread warning is dropped.

Verification

itkDiffusionTensor3DReconstructionImageFilterTest passes (it reconstructs and checks the output tensor at voxel (3,3,3) against expected values at 1e-4). A mutation check (scaling the precomputed pseudo-inverse by 1.5) makes that test fail, confirming it exercises the changed path.

Note: the unit test exercises the AddGradientImage (many-images) path and the shared precompute; the SetGradientImage (single-image) path uses the byte-identical loop edit but has no direct unit-test coverage today.

[greptile-apps]

Greptile Summary

Precomputes the pseudo-inverse of the constant m_TensorBasis once in ComputeTensorBasis() and replaces the per-voxel vnl_svd::solve() call with a matrix-vector multiply, removing a redundant SVD construction and the associated netlib/dsvdc thread-safety constraint. A new test covers the previously-untested SetGradientImage path.

• The mathematical equivalence (pinverse(A)·b == svd(A).solve(b)) is correct, and both the many-images and single-image code paths are updated symmetrically.
• The Doxygen \warning block in the header still instructs users to call SetNumberOfWorkUnits(1) for the now-resolved thread-safety issue; this stale documentation should be removed alongside the inline .hxx comment that was dropped.
• Two new in-source comments exceed the ITK prose-budget limit of one line ≤ 100 chars and should be trimmed or removed.

Confidence Score: 4/5

The core optimization is mathematically correct and both code paths are updated symmetrically, but the user-facing Doxygen warning still tells callers to disable multithreading — defeating the change's purpose for anyone who reads the docs.

The stale \warning block in the public class documentation still instructs callers to call SetNumberOfWorkUnits(1) for a thread-safety concern that this PR resolves. Any user who follows that advice will silently lose the multithreading benefit. The rest of the change — SVD hoist, per-voxel matvec, new test — is correct and complete.

Modules/Filtering/DiffusionTensorImage/include/itkDiffusionTensor3DReconstructionImageFilter.h — the \warning Doxygen block (lines 109–116) and the new member comment (lines 330–332) both need attention.

Important Files Changed

Filename	Overview
Modules/Filtering/DiffusionTensorImage/include/itkDiffusionTensor3DReconstructionImageFilter.h	Adds m_TensorBasisInverse member with a 2-line comment that exceeds the prose budget; the Doxygen \warning about single-threading was not removed, leaving user-visible documentation contradicting the fix.
Modules/Filtering/DiffusionTensorImage/include/itkDiffusionTensor3DReconstructionImageFilter.hxx	Hoists vnl_svd::pinverse() from the per-voxel loop into ComputeTensorBasis(); both code paths updated symmetrically and math is correct. New 3-line comment is prose-budget-over-limit but otherwise the logic is sound.
Modules/Filtering/DiffusionTensorImage/test/itkDiffusionTensor3DReconstructionSingleImageTest.cxx	New test exercising the SetGradientImage (single-VectorImage) path and cross-checking it against AddGradientImage output; sanity checks are appropriate and the 1e-8 tolerance is correct for double precision.
Modules/Filtering/DiffusionTensorImage/test/CMakeLists.txt	Correctly registers the new test with itk_add_test; no issues.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant U as User code
    participant F as Filter::Update()
    participant B as BeforeThreadedGenerateData
    participant C as ComputeTensorBasis
    participant T as DynamicThreadedGenerateData (N threads)

    U->>F: Update()
    F->>B: BeforeThreadedGenerateData()
    B->>C: ComputeTensorBasis()
    Note over C: Builds m_BMatrix, m_TensorBasis
    C->>C: "m_TensorBasisInverse = vnl_svd(m_TensorBasis).pinverse()"
    Note over C: Single SVD computed ONCE (was per-voxel)
    C-->>B: return
    B-->>F: return
    F->>T: spawn N threads
    loop per voxel
        T->>T: "D = m_TensorBasisInverse * B  (matvec only, thread-safe)"
    end
    T-->>F: results merged
    F-->>U: output tensor image

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant U as User code
    participant F as Filter::Update()
    participant B as BeforeThreadedGenerateData
    participant C as ComputeTensorBasis
    participant T as DynamicThreadedGenerateData (N threads)

    U->>F: Update()
    F->>B: BeforeThreadedGenerateData()
    B->>C: ComputeTensorBasis()
    Note over C: Builds m_BMatrix, m_TensorBasis
    C->>C: "m_TensorBasisInverse = vnl_svd(m_TensorBasis).pinverse()"
    Note over C: Single SVD computed ONCE (was per-voxel)
    C-->>B: return
    B-->>F: return
    F->>T: spawn N threads
    loop per voxel
        T->>T: "D = m_TensorBasisInverse * B  (matvec only, thread-safe)"
    end
    T-->>F: results merged
    F-->>U: output tensor image

Loading

Comments Outside Diff (1)

1. Modules/Filtering/DiffusionTensorImage/include/itkDiffusionTensor3DReconstructionImageFilter.h, line 109-116 (link)

   Stale \warning block contradicts the PR's goal

   The \warning telling users to call SetNumberOfWorkUnits(1) because of the netlib/dsvdc thread-safety issue is now wrong. This PR fixes the root cause by moving the SVD into BeforeThreadedGenerateData(), so any user who reads this warning and acts on it will run single-threaded and lose the entire performance benefit. The .hxx inline comment was removed, but the user-visible Doxygen block (the one that renders in IDE tooltips and generated docs) was not. It should be removed.

_{Reviews (1): Last reviewed commit: "ENH: Add single-image-path test for DTI ..." | Re-trigger Greptile}

[thewtex]

greptile has style suggestions