Skip to content

tomoxide v0.2.0

Choose a tag to compare

View all tags
@physwkim physwkim released this 30 Jun 02:19
· 255 commits to main since this release
v0.2.0
9d08c87

tomoxide v0.2.0

tomoxide is a Rust tomographic reconstruction toolkit combining the algorithmic
breadth of tomopy (CPU libtomo) with the GPU-accelerated streaming
reconstruction of tomocupy (CUDA), behind a single tri-backend abstraction:
CPU / CUDA / wgpu.

Highlights

  • Full CUDA streaming engine. Device-resident streaming reconstructors for
    the analytic methods (FBP, linerec, fourierrec) and lprec, with an async
    H2D∥compute∥D2H pipeline and memory-aware chunking that lifts the
    large-volume GPU ceiling. tomoxide now beats tomocupy end-to-end on every
    analytic method on an RTX 5000 Ada.
  • Multi-GPU. Per-slice and fused analytic reconstruction across all GPUs,
    plus a multi-GPU z-shard fan-out for the streaming recon CLI. 4-GPU 4096²
    head-to-head now favors tomoxide.
  • Half precision (fp16). End-to-end fp16 analytic path including
    out-of-core fp16 Fbp/Linerec and device-resident fp16 fourierrec streaming;
    fourierrec 2048² single-GPU went 9.83 → 2.70 s, ahead of tomocupy. Enable with
    --dtype float16.
  • On-device stripe removal. Titarenko, Fourier-Wavelet, and Vo all-stripe
    removal run inside the streaming raw path on the GPU.
  • Portable wgpu backend now usable at scale. Cleared the
    65535-workgroup dispatch cap and the WebGL downlevel buffer limits; wgpu
    reconstructs 512²/1024²/2048² volumes and matches the CPU reference at
    Pearson 1.0 (a correct, portable fallback — not a perf path).
  • New algorithms / preprocessing. Vector tomography (tomopy vector.c,
    bit-exact), beam-hardening correction (xraylib, beam-hardening feature), and
    SIFT-based center finding (OpenCV, sift-center feature).
  • Single-crate workspace. The nine library crates were consolidated into one
    tomoxide crate.

Cross-backend convention note

The CUDA analytic streaming kernels emit each slice with tomocupy's handedness:
a vertical (y-axis) flip plus a per-algorithm scale (4/π for fbp/linerec, ≈4·n²
for fourierrec; lprec and gridrec keep the CPU/tomopy orientation and scale 1).
The reconstruction is numerically identical to the CPU/wgpu path once the
convention is undone (Pearson 1.0). This is documented in
docs/ARCHITECTURE.md §4.1 and pinned by
tests/cuda_cpu_convention_parity.rs.

Install

[dependencies]
tomoxide = "0.2.0"

CLI:

cargo install tomoxide-cli

GPU features are opt-in (cuda, gpu-wgpu); see the README for the feature
matrix and benchmark tables.

See CHANGELOG.md for the complete list of changes.

Assets 2
Loading