v0.3.2

wgpu-nufft 0.3.2

A Julia package; see CHANGELOG.md.

Added:

• WgpuNufft (julia/WgpuNufft): a Julia package over the C interface, with
  one-call transforms nufft1d1 to nufft3d3, which keep their plans, and a
  Plan with setpts! and execute for batches and any dimension, on Float64
  and Float32 arrays. Options (backend, precision, mode order, threads,
  upsampling) are keywords; library errors are thrown as WgpuNufftError.

Known limitations:

• The MATLAB interface compiles and links but has not yet run in MATLAB.
• On DX12 with FXC, Df64 plans take minutes to create; on Windows, prefer
  Vulkan for Df64 plans.
• 7D and 8D plans are practical only on small grids at tolerances near
  1e-3.
• Metal has not been tested.

v0.3.1

wgpu-nufft 0.3.1

C, Fortran, and MATLAB interfaces, and a first transform of a session as
quick as in 0.2.0; see CHANGELOG.md.

Added:

• wgpu-nufft-c (capi/): a C library declared by capi/include/wgpu_nufft.h.
  Plans of type 1, 2 and 3 in one to eight dimensions take host arrays, set
  their points once and run batches; one-call functions (wgpu_nufft1d1 to
  wgpu_nufft3d3) keep their plans for reuse. Double-precision arrays run in
  native f64 where the GPU supports it and in Df64 elsewhere, single-
  precision arrays in f32, and plans fall back to the CPU backend without a
  usable GPU.
• A Fortran module (fortran/wgpu_nufft.f90) over the C library, tested
  with gfortran and Intel ifx.
• A MATLAB interface (matlab/): one-call functions wgpunufft1d1 to
  wgpunufft3d3, the wgpunufft_plan class, and a MEX gateway for MATLAB
  R2018a or later, which build_wgpu_nufft compiles.

Fixed:

• Type-1 plans compile their spreading shaders when an execution first
  needs them, for the bin geometry it uses. 0.3.0 compiled two variants at
  plan creation, which made the first transform of a session slower than in
  0.2.0.

Known limitations:

• The MATLAB interface compiles and links but has not yet run in MATLAB.
• On DX12 with FXC, Df64 plans take minutes to create; on Windows, prefer
  Vulkan for Df64 plans.
• 7D and 8D plans are practical only on small grids at tolerances near
  1e-3.
• Metal has not been tested.

v0.3.0

wgpu-nufft 0.3.0

GPU plans in up to eight dimensions, and one deterministic spreader for
every type-1 plan, faster in every rank and for sparse and clustered
points; see CHANGELOG.md.

Highlights:

• Type 1, 2 and 3 GPU plans in six to eight dimensions, in all three
  precisions; the 4D and 5D paths are rewritten.
• Every type-1 plan spreads through one rank-generic block spreader with
  bitwise-repeatable sums. It bins points by the first cell of their kernel
  support, chooses its bins per execution from the point density, and
  splits crowded blocks across many workgroups.
• Large bins sort several times faster, and F64 kernel weights evaluate
  faster.
• F64 and Df64 type-2 plans in one to three dimensions interpolate binned
  points.
• F32 plans reject tolerances their rounding would miss
  (NufftError::F32ToleranceUnreachable).
• A standalone browser build, wgpu_web.js, is attached to this release.

Fixed:

• The rank-generic spreader compiles with FXC on DX12.

Known limitations:

• On DX12 with FXC, Df64 plans take minutes to create, mostly in the Df64
  FFT shaders; on Windows, prefer Vulkan for Df64 plans.
• 7D and 8D plans are practical only on small grids at tolerances near
  1e-3; see "Higher dimensions" in the README.
• Metal has not been tested.