tomoxide v0.3.0

tomoxide v0.3.0

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.

This is a filter correctness / convention release: the CUDA analytic path
now matches tomocupy in absolute amplitude and ramp shape, and the default
filter follows tomocupy. It also adds GPU laminography.

Behaviour changes — read before upgrading

• CUDA analytic amplitude halved. In 0.2.0 every CUDA analytic method
  (fbp / linerec / fourierrec / lprec / laminography, f32 + fp16) emitted output
  2× larger than tomocupy. 0.3.0 fixes the filter normalization so CUDA
  matches tomocupy in absolute amplitude. The CPU/wgpu path still matches
  tomopy. If you compared CUDA output against absolute units or against 0.2.0
  CUDA results, expect a factor-of-2 change.
• Default FBP filter is now parzen (was ramp), matching tomocupy's
  default. Default reconstructions are smoother; set filter_name = FilterName::Ramp to restore the sharp ramp.

Highlights

• CUDA analytic output matches tomocupy bit-for-bit. Beyond the amplitude
  fix, the CUDA filter ramp is ported to tomocupy's exact degree-12 _wint
  quadrature shape, closing the residual ~1% straight-line-ramp gap. On a
  synthetic 180° dataset, CUDA fbp/linerec/fourierrec now match tomocupy at
  scale 1.00000, Pearson 1.000000.
• Per-backend ramp by reference. Each backend now builds the ramp of the
  library it ports: CPU/wgpu use tomopy's linear ramp, CUDA uses tomocupy's
  _wint quadrature ramp (backend::RampShape). Apodization, padding, clamp,
  DC doubling and FFT layout stay shared in make_fbp_filter.
• GPU laminography. recon --lamino_angle runs analytic linerec with a
  tilted rotation axis on CUDA, verified against tomocupy on real leaf data
  (Pearson 0.99997).

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. As of 0.3.0 the scales are
2/π for fbp/linerec and ≈2·n² for fourierrec (half their 0.2.0 values, from
the amplitude fix), ½ for lprec, and 1 for gridrec (which keeps the
CPU/tomopy orientation and scale). After undoing the flip and scale, CUDA
matches the CPU/wgpu path very closely; a deterministic ~0.6% residual remains
from the tomocupy-_wint vs tomopy-linear ramp shapes (Pearson ≈1.0). This is
documented in docs/ARCHITECTURE.md §4.1 and pinned by
tests/cuda_cpu_convention_parity.rs.

Install

[dependencies]
tomoxide = "0.3.0"

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.