Skip to content

TIGRE v3.3.0

Latest

Choose a tag to compare

@AnderBiguri AnderBiguri released this 21 Nov 16:01
a3e2221

TIGRE 3.3.0: release notes (final, 21 Nov 2026)

The text pasted into the GitHub Release v3.3.0 (title "TIGRE 3.3.0"). Also useful for the website.

This release is mostly about the GPU code. It brings:

  • a faster and more accurate Siddon projector, now faster than ASTRA's forward projector on a single GPU;
  • PyTorch and CuPy GPU arrays passed straight to TIGRE;
  • faster backprojection, exact for rotated detectors;
  • parallel-beam operators and PICCS on several GPUs and on images larger than GPU memory;
  • many fixes for multi-GPU, low-memory and split cases.

Highlights

  • Siddon projector. Faster and more accurate:
    • cone-beam kernel time 1.4–1.9× faster on an RTX 4070, 5.5–6.5× on a Quadro RTX 6000;
    • parallel beam 1.7–2.0× and 4.4–4.8× faster;
    • the single-precision drift along long rays, up to ~2e-3, is gone (now ~1e-5), and every ray is consistent, also when the image is split;
    • on a single GPU, faster than the ASTRA Toolbox's forward projector: up to 1.9× in our benchmarks (256³–1024³, RTX 4070 and Quadro RTX 6000).
    • The technical note docs/technical-notes/siddon/siddon_note.pdf explains why, and has the comparison with ASTRA.
  • GPU arrays. tigre.Ax / tigre.Atb and the PyTorch bindings take PyTorch CUDA tensors and CuPy arrays, and return the same kind of array on the same GPU, without copying through host memory. In demo d25_Pytorch.py (256³, 360 angles, RTX 4070), Ax takes 0.034 s directly against 0.082 s through the host.
  • Faster backprojection. Cone-beam Atb kernels: matched 0.65–0.73 of its previous time, FDK weights 0.86–0.91 (RTX 4070, Quadro RTX 6000). matched is now exact for rotated detectors and tilted rotation axes. Ax and Atb with NumPy arrays also drop a needless check that cost ~0.1 s per call at 512³.
  • Parallel beam Ax and Atb use all the GPUs in gpuids, and split the image when it doesn't fit (#286, #295).
  • PICCS shares minTV's multi-GPU code. It uses gpuids, several GPUs, and images larger than GPU memory, and no longer returns NaN when the gradient is zero. AwminTV now runs on the same code too (results unchanged).
  • Reproducible on several GPUs. minTV, AwminTV, minPICCS and tvdenoise give the same result on 1 or N GPUs, and tvdenoise the same result whether or not the image is split.
  • Robustness:
    • a CUDA error raises RuntimeError in Python instead of ending the process, and TIGRE frees what the failed call held;
    • clear "out of memory" errors, and splitting that works down to very little free memory;
    • no more texture-object leak in long reconstructions (#617);
    • 64-bit sizes for very large problems;
    • pinned (page-locked) input arrays work;
    • minTV, AwminTV and minPICCS no longer read outside the caller's arrays when a piece of the image is thinner than its halo (#899), and no longer lose a little (pinned) host memory on every call.

Results that change

Check these if you compare against results from 3.2:

  • tvdenoise / im3DDenoise, and minTV / AwminTV (the TV step of the POCS solvers): the first and last z-slice now use the same boundary as x and y. Results change there, and a little further in.
  • minTV / AwminTV: the gradient norm is summed in double precision. This changes rounding, and the POCS solvers amplify it.
  • Siddon Ax: more accurate, by up to ~2e-3 relative on large images.
  • Atb with matched weights: now exact for rotated detectors (rotDetector) and tilted rotation axes, where it was off by 1.5–7%; results change there by up to several %, and by ~1e-4 for unrotated detectors.
  • PyTorch bindings: At and the gradient of A use the matched backprojector. Pass backprojection_type="FDK" for the old behaviour.
  • CTnoise.add / addCTnoise: new noise on every call. Before, every call added the same noise. Use seed= / 'Seed' to fix it; np.random.seed / rng make scripts reproducible.
  • "GPU heavily used by another program" is a warning, not an error.

Install

TIGRE is not on PyPI. Install with pip install git+https://github.com/CERN/TIGRE.git@v3.3.0, or use the wheels attached to this release: Python 3.11–3.14, Linux manylinux_2_28, glibc 2.28+.

For developers

  • Rare-edge-case tests are marked extended and not run by default (pytest -m extended).
  • TIGRE_CUDA_ARCH=native compiles for your GPU only, for much faster development builds.
  • TIGRE_BACKEND selects the GPU backend.

Thanks

  • Jeff Daily (AMD): the analysis and approach for #856, and the ROCm port (#747).
  • Einse57: found #891 and #899 while porting the TV code to SYCL (#885).
  • Everyone who reported issues.

The full list of changes is in [changelog.md](https://github.com/CERN/TIGRE/blob/master/changelog.md).

What's Changed

  • Build wheels for Python 3.14: cibuildwheel 2.23.2 -> 4.3.5 by @AnderBiguri in #827
  • docs: Sphinx site skeleton and GitHub Pages deployment (phase 1) by @AnderBiguri in #839
  • docs: install guides, generated citing page, papers using TIGRE (phase 2) by @AnderBiguri in #841
  • CI: retry the CUDA package install in wheel builds; update actions to Node 24 by @AnderBiguri in #840
  • build: disable pinned memory with TIGRE_NO_PINNED_MEMORY env var by @AnderBiguri in #842
  • docs: 137 more papers using TIGRE, verified from citation contexts by @AnderBiguri in #843
  • docs: user guide (phase 3) by @AnderBiguri in #844
  • Docs/gallery by @AnderBiguri in #845
  • Docs/api by @AnderBiguri in #846
  • docs: Demos and Gallery, PICCS family, geometry figure, logo by @AnderBiguri in #847
  • docs: more Gallery images, geometry figure in the repo by @AnderBiguri in #848
  • docs: logo in README, interoperability and compatibility pages, small… by @AnderBiguri in #849
  • sample_loader: find head.mat where pip installs data_files by @Einse57 in #850
  • Fix texture object leak in Atb (#617) by @AnderBiguri in #851
  • Raise Python exceptions on CUDA errors instead of exit(1); busy GPU i… by @AnderBiguri in #854
  • Overlap the multi-GPU TV halo exchange across GPU boundaries by @jeffdaily in #809
  • Python: Ax/Atb take PyTorch CUDA tensors and CuPy arrays, on their GPU by @AnderBiguri in #857
  • Fix __CUDART_VERSION typo in TV/PICCS warp reduction kernels by @hashimoto-koh in #862
  • TV: split correctly with little free memory; no more endless tvdenois… by @AnderBiguri in #864
  • Ax: wait for each batch's copy-out before reusing its buffer; test #860 by @AnderBiguri in #861
  • Tests: fix three Windows failures that are not TIGRE bugs (import order, sum() guard, TV low-memory fill) by @AnderBiguri in #866
  • TV functions: release everything on error, as Ax/Atb do by @AnderBiguri in #870
  • Use 64-bit sizes where 32-bit integer products wrapped on large problems by @AnderBiguri in #868
  • Siddon (cone and parallel): consistent, more accurate and faster traversal (fixes #859) by @AnderBiguri in #865
  • tvdenoise split path: zero the right buffer, wait for copy-back (replaces #806, partly #863) by @AnderBiguri in #872
  • tvdenoise: split results equal the unsplit ones (fixes #863) by @AnderBiguri in #873
  • Ax/Atb: don't trim the caller's GpuIds (fixes #855) by @AnderBiguri in #871
  • Ax/Atb with very little GPU memory: clear "out of memory", and split until the texture fits by @AnderBiguri in #875
  • docs: Siddon note, parallel-beam timing on Linux by @AnderBiguri in #884
  • tests: rare edge cases and stress tests marked extended, not run by default by @AnderBiguri in #886
  • Build: choose the GPU backend (TIGRE_BACKEND, Compile -backend) or detect it from the installed compiler by @AnderBiguri in #889
  • pyproject: license as an SPDX expression; setup.py: drop a stray print by @AnderBiguri in #890
  • minTV/AwminTV: add the gradient norm in double across GPUs (#856) by @AnderBiguri in #887
  • Parallel beam: Ax and Atb on several GPUs, and split when the image does not fit (#286, #295) by @AnderBiguri in #888
  • tvdenoise: size the multi-GPU halo host buffer for buffer_length slices (#891) by @AnderBiguri in #892
  • TV functions: no empty pieces when sharing few z slices between GPUs by @AnderBiguri in #893
  • Build: TIGRE_CUDA_ARCH to compile for fewer GPU architectures (faster dev builds) by @AnderBiguri in #897
  • Noise: new seed every call; gpuids and clean-up in AddNoise; pinned arrays in cone Ax/Atb by @AnderBiguri in #894
  • Demo d25: PyTorch GPU tensors straight into tigre.Ax/Atb, with a timing by @AnderBiguri in #895
  • PICCS on the GPU code shared with minTV; z boundary like x/y in minTV/AwminTV by @AnderBiguri in #896
  • TV functions: halos stop at the image's ends (#899); free host memory on success by @AnderBiguri in #900
  • Faster backprojection; exact matched weight for rotated detectors; no isreal check in Ax/Atb by @AnderBiguri in #901
  • Release 3.3.0 by @AnderBiguri in #898

New Contributors

Full Changelog: v3.2.0...v3.3.0