GitHub

Migrate from v0.2

v1 is a breaking release. CUDA-specific names remain as deprecated wrappers so an existing NVIDIA workflow can be migrated incrementally.

v0.2v1
cu_transfer_sqrt_arr(n, λ, dx)propagation_grid(n, λ, dx)
cu_transfer(z, n, λ, grid)propagation_kernel(z, λ, grid)
cu_gabor_wavefront(cu(image))gabor_wavefront(image)
cu_phase_retrieval_holo(...)PhaseRetrievalPlan + phase_retrieval
cu_get_reconst_vol(...)ReconstructionPlan + reconstruct
cu_get_reconst_xyprojection(...)xyprojection
cu_get_reconst_vol_and_xyprojection(...)reconstruct_and_projection
cu_dilate(binary)dilate(binary)
CuTransferSqrtPartPropagationGrid
CuTransferPropagationKernel
CuWavefrontWavefront

Minimal rewrite

# v0.2
using ParticleHolography, CUDA
s = cu_transfer_sqrt_arr(n, λ, dx)
front = cu_transfer(-z0, n, λ, s)
step = cu_transfer(-dz, n, λ, s)
wave = cu_gabor_wavefront(cu(image))
volume = cu_get_reconst_vol(wave, front, step, slices)

# v1
using ParticleHolography
backend(:cpu) # or load CUDA/Metal and select it once
grid = propagation_grid(n, λ, dx)
front = propagation_kernel(-z0, λ, grid)
step = propagation_kernel(-dz, λ, grid)
wave = gabor_wavefront(image)
request = ReconstructionRequest(slices; volume=Float32)
plan = ReconstructionPlan(front, step; request)
volume = reconstruct(plan, wave).volume

Behaviour changes

  • CPU-only import no longer loads CUDA, a GPU driver, Makie, or Plots.
  • backend(:cpu), backend(:metal), or backend(:cuda) sets the process-wide default used by later backend-less calls.
  • New reconstruction functions default to unclipped Float32. Deprecated cu_get_* wrappers retain clipped N0f8 defaults where v0.2 did.
  • ReconstructionRequest produces a requested volume and MinIP together in one depth scan and performs a conservative memory preflight.
  • Transfer and filter arrays use FFT-native order. Manually constructed centred arrays must be reordered before wrapping.
  • Combined projection uses unquantised Float32 intensity even if the stored volume is N0f8.
  • Bounding-box overlap is inclusive, UUID generation no longer restarts a fixed random generator, and dictionary deletion is performed safely.
  • Tracking preserves graph branches and rejects duplicate UUIDs across frames.
  • Tests no longer write images or JSON into tracked fixture paths.
  • CUDABackend and MetalBackend implementation types are not exported because CUDA.jl/Metal.jl export names with the same spelling. Use backend(:cuda) and backend(:metal).
  • transfer_sqrt/TransferSqrtPart and transfer/Transfer remain deprecated aliases for propagation_grid/PropagationGrid and propagation_kernel/PropagationKernel.

Package environments

Move CUDA, Metal, and Plots to the application environment that needs them. ParticleHolography lists them as weak dependencies, so a CPU service or docs build need not download accelerator/display stacks.