Seventh Coding Week: Cleaning Up Near the Finish Line!#
Hey there!
This was the seventh coding week of my GSoC project with DIPY. This week’s work was less exploratory and research-focused than during previous weeks, but it was still very important. As the project moves closer to its final stages, I focused on cleaning up implementations, improving tests, addressing review feedback, and preparing the remaining pull requests for review.
The main contributions were:
Finalizing the Mutual Information implementation for SyN registration and marking the PR as ready for review.
Opening a new PR to parallelize DIPY’s 2D and 3D image-warping operations.
Cleaning up the affine cross-correlation implementation.
Refining two SynthSeg PRs related to flexible input sizes and brain-mask postprocessing.
Getting the first performance-related contribution of my GSoC project merged into DIPY.
First Performance Contribution Merged#
The week started with some good news: my small Parzen-window optimization,
PR #4071, was merged into DIPY’s
master branch.
As described in my previous GSoC post, the PR removes a redundant cubic B-spline bin evaluation from DIPY’s Parzen-window Mutual Information implementation.
This was the first PR from my GSoC performance work to be merged, so it was a great way to begin the week!
Parallelizing Image Warping#
I also continued the performance work started in PR #4073, which parallelizes displacement-field composition and fixed-point inversion in DIPY’s registration backend.
After reviewing the 2018 OpenMP PR #1613, I identified image warping as another important operation that could benefit from threading. SyN repeatedly warps the static and moving images into a common reference space, so these operations form another relevant part of the registration runtime.
I therefore opened PR #4083, developed in the speed-syn-warp branch.
The PR adds OpenMP parallelization to DIPY’s low-level 2D and 3D image-warping functions, covering both linear and nearest-neighbor interpolation.
The outer rows in 2D and slices in 3D can be processed independently, allowing
the loops to be parallelized using Cython’s prange.
The PR also propagates the existing num_threads configuration through
DiffeomorphicMap.transform(), transform_inverse(), and the image warps
performed internally during SyN registration.
I also ran a small timing experiment on 3D SyN:
Revision and mode |
Time |
NCC |
|---|---|---|
DIPY branch, one thread |
101.249 s |
0.95741 |
DIPY branch, four threads |
76.864 s |
0.95741 |
DIPY |
110.465 s |
0.95741 |
Using four threads reduced runtime while preserving the final NCC.
Reviewing and Updating Existing Pull Requests#
This week, I also had a meeting with my mentors to review the current state of the open pull requests and identify the main changes needed before they can be merged.
Following this review, I focused on updating and cleaning several existing contributions:
PR #4067, adding Mutual Information for SyN registration.
PR #4078, adding cross-correlation to affine registration.
PR #4052, improving the brain mask returned by SynthSeg.
PR #4080, supporting flexible SynthSeg input sizes.
The changes mainly involved simplifying implementations, improving dtype and input validation, refining APIs and documentation, extending test coverage, and addressing specific reviewer feedback. Some of these PRs are now ready for review, while the remaining ones continue to be updated based on the discussion.
Next Week’s Work#
For next week, I plan to focus on bringing the remaining open pull requests to completion by addressing the latest review comments, polishing the implementations, and making sure that the tests and documentation are ready.
I will also begin writing the final GSoC report, summarizing the work completed during the project, the main results, and the current status of each contribution.
Find Me Online#
GitHub: TomasGuija
LinkedIn: Tomás Guija Valiente
Thank you for reading!