Reduced technical debt from original adaptation of pure ptyhon algorithm, which should make future development easier. Refined the gpu PIV-processing functions for better resource-use efficiency, and speed. PIV results should be more accurate at the subpixel level. PIV should be more robust near edges of masks/imagesfinements to the gpu_process module targeting increased readability, resource-use efficiency, and speed.

Main changes:

• Broke-up the old F-structure from widim into smaller arrays to make development easier.
• Remove dependence on Cython, since acceleration can be done completely by CUDA and numpy routines.
• Increase speed of PIV processingSpeed has been increased by GPU-accelerating remaining cpu-bound functions and using faster alternatives to some logic--time to complete a 10-image-pair benchmark with 5 iterations and minimum windows size of 8 was reduced by ~75%.
• Add validation module with validation routines. This can be used independently of the processing module.
• MakLarge functions with high number of degrees of freedom object-orientedhave broken into smaller, more specialized functions.
• Add test modules covering gpu functions. There are unit tests and regression tests.
• Create gpu_smoothn module containingn smoothn logic that is more readable, and which will allow for GPU-acceleration in the future.
• Put all gpu modules into their package to be a clean add-on to base python pacakage.
• GPU-accelerated functions are broken across multiple modules in a single package added on top of OpenPIV.
• Test using CUDA 12.

Improved functionality:

• Change specification of window sizes and iterations into one argument, which is most clear to users.
• Validate user inputs to processing functions using Python descriptors.
• Add centroid and parabolic subpixel estimation methodsAdd centriod and parabolic subpixel estimation methods.
• Validate user inputs to processing functions using Python descriptors
• Change specification of window sizes and iterations into one argument, which is most clear to users.
• put all gpu modules into their package to be a clean add-on to base python pacakage.

Other changes:

• Refactor variables names to be more clear.
• GPU-accelerated functions are broken across multiple modules.
• Mask images before doing PIV.
• Include subpixel and centroid subpixel methods.
• Masked points are not considered in the validation or as data for smoothing.
• Replace vectors with median vectors.
• Add input checks to PIV parametersmove gpu_extended_search_area since it wraps negligible logic and remains unused.
• Update README.
• Add input checks to PIV parameters.
• Include subpixel and centroid subpixel methods.
• Consolidate validation gpu kernels into fewer kernels.
• Move kernel compilation out of functions so they don't compile more than once.
• Add ability to center vector field on image domain. This is now default behaviour.
• Log processing using logging module status rather than printing to console. This cleans up output when using multiple processes or wrappers.
• Add gpu function to interpolate from masked data.
• Add GPU cosine transform to gpu_smoothn module.
• Add methods to free gpu data held by classes.
• Use only positive values in subpixel estimation.
• Estimate strain next to masked points and edges using first-order differencing.

Performance:

• Consolidate validation gpu kernels into fewer kernels.
• Optimize GPU kernels doing vector replacement, subpixel approximation, window deformation, masking, gradient approximation, validation.
• Simplify logic in all gpu kernels and remove as many switch cases as possible.
• Move kernel compilation out of functions so they don't compile more than once.
• Eliminate switching between multiple contexts for single process.
• In most cases, create GPU arrays on the device itself to minimize transfers between CPU and GPU.
• Simplify logic for vector replacement to avoid interpolating twice in one iteration.

Other changes:

• RefactorUse private/public variable/function/method naming.
• Optimize GPU kernels doing vector replacement, subpixel approximation, window deformation, masking, gradient approximation, validation.
• Fix bugs in gradient approximation, suxpixel approximation, validation, smoothing, and other gpu kernels.
• SImplify logic in all gpu kernels and remove as many switch cases as possible.
• Add gpu function to interpolate from masked data.
• In most cases, create GPU arrays on the device itself to minimize transfers between CPU and GPU.
• Add GPU cosine transform to gpu_smoothn module.
• Add methods to free gpu data
• Use python data types for non-numpy-array varaiables names to be more clear.
• Remove gpu_extended_search_area since it wrapped negligible logic and remaine.
• Use only positive values in subpixel estimation.
• Estimate strain next to masked points and edges
• Simplify logic for vector replacement to avoid uintested.
• Update README install instrucrpolating twice in one iterations.
• Use private/public variable/function/method naming.
• Fix bugs in gradient approximation, suxpixel approximation, validation, smoothing, and other gpu kernels.
• Use python data types for non-numpy-array variables.
  Replace vectors with median vectors
• Separate public from private methods
• Test on the newest CUDA toolkit version