TODO.txt

TODO LIST

- In demo: Enable custom transform even when "haar" is selected (=> modify compute_filters() which returns too son with hlen = 2)
- max_lev is log2(N/(hlen-1)) !
- Better mechanism for custom wavelets. Should allow > 40 coeffs
- Modify inverse haar (and related) to avoid extra memcpy
- Nonseparable SWT re-computes the inverse filters at each inversion (for memory requirements). See if there is a workaround.
- Update filters coefficients in filters.cpp, to be compatible with PyWavelets.
- ISWT:
    - Either return all app coeffs (can be interesing for multi-resolution)
    - Or only store appcoeffs at last scale (for eg. Matlab only uses the last appcoeff for iswt)
- Exceptions to fail gracefully (especially from the Python side)
- More threshold types
- Compute norm related to threshold (eg. soft/L1, group-lasso, etc)
- User-defined target CUDA device
- Documentation/Howtos
- Set properly the variable for LIB_DEST/RUNTIME_DEST
- Define static constexpr filters, and define the max filter size as a
  constexpr variable.

-- DONE :

- All transform work for "odd" sizes
- Refactoring :
    - Create a structure containing informations on WT. Used by Wavelets() class, but also by common functions (to avoid kilometers-long functions signature)
    - Not all files in same folder (test, src, doc, build)
    - Remove useless files (profiling) and put i/o in "test"
    - No more #include "*.cu" !
- TYPE template to support both double and simple precision