Convolution perf notes #2

Perf boosts

• NaN treatment #3 NaN treatment
• Kernel normalization treatment #4 Kernel normalization treatment
• Investigate Cython routines for uneccessary persisting python (code that does NOT generate C).
  c.f. generated Cyhton inspection (E.g. cython file.pyx -a)
• Investigate linear memory increase of convolution.convolve()  #6 Investigate possible mem leaks as indicated from Convolution Notes #1 results.
• Use C division to prevent requiring the GIL during loop #5 Python vs C division, the GIL, and DASK
• convolution.convolve() Cython code hoist optimizations #7 manually hoist computations
• look into this.
• Investigate direct convolution in C, NO (P|C)ython #9 Direct C investigation

Other Issues

• convolve() param boundary has default 'fill' whilst doc states default is None. Since the default for fill_value = 0. the two are effectively the same. However, if the user accidentally changes the fill_value this value will be used as boundary=None is NOT the actual default and the internal routine convolve<N>d_boundary_fill() is called and not convolve<N>d _boundary_none(). c.f. Direct conv boundary default - None or fill #10
• The doc for convolve() states that an exception will be raised for nan_treatment='interpolate' if the kernel sums to zero and therefore cannot be normalized. This is not actually true, no exception is raised. Note: not yet tested, code inspection only.