BUGS

• Distance Estimator
• Glitch Correction
• Parsers
• Other

Distance Estimator

• overflows double beyond ~1e-145
• goes wrong beyond ~1e-55
  • test case down-the-rabbit-hole.mm (too light)
• not possible to switch between --de and --no-de at runtime
• occasional NaNs with fp32 at low zoom
  • test case examples/mm/fp32-nan-de.mm
• --no-de gives visible seams with tiled rendering
  
  
  

Glitch Correction

• bad glitches without --no-approx
  • test case zoom into emndl_perturbation_test_2013-06-14.mm
• warn: delta(z)_0 == 0, rendering will probably fail
  • all pixels have the same delta(z)_0, insufficient precision in double
  • leads to bad images
• --no-glitch gives minor visible seams with tiled rendering

Parsers

• parse_ppar_corners() uses a fixed precision (256 bits)
• parsers assume ASCII (or other 8bit encoding)
  • UTF-8 is a multi-byte encoding where some bytes of multi-byte characters might clash with ASCII. (possibly false: see http://htmlpurifier.org/docs/enduser-utf8.html#migrate-variablewidth)
• parsers don't check success of mpfr_set_str()
• parsers should use strtod() instead of atof() to be able to detect errors
• fractint .gif parser doesn't read extended precision metadata

Other

• if image is 100% interior, it will never finish rendering. possible algorithm to check this case before starting rendering:
  • first exclude problematic corner points
    • boundary points exactly representable as binary floating point
    • perhaps just exclude real axis and +/- i, are there others?
    • maybe perturb points if this is the case (make rectangle 0.5px smaller?)
  • for each corner, check if it is interior
    • if all corners are interior to the same component
      • if shape estimate of component is cardioid (non-convex)
        • (*) calculate interior distance estimates for each point
        • if there are any regions not excluded by the distance estimates
          • compute the interior|exterior distance estimate for a point in the middle of such a regions
          • if interior and same component, exclude another region and goto (*)
          • otherwise rectangle has exterior points and it's safe to render
      • else rectangle is 100% interior to a convex circular bulb might be easier to check a region in a previous frame, at least for --interactive and --zoom. for --tile perhaps render a small preview of the view to check interior tiles. --timeout mitigates the problem somewhat.
• render.c has a hardcoded maximum period (100000) for finding the first reference
• render.c has a hardcoded maximum pass count (1000) for number of references
• test suite assumes example parameters are in directories matching extension
• need a way to turn view radius into zoom frame count
• need a way to render a subsection of a zoom sequence
  • perhaps start radius / end radius, which are rounded outwards to 2^n values
• timeout calculations are broken if a render is stopped and restarted with ctrl-Z and later fg
• extra/zoom needlessly uses double-size textures
• extra/zoom needlessly uses two passes for cubic interpolation
• extra/zoom needs to have window visible or it pauses rendering
• --slice 0 saves unneeded reference orbit in fpxx_step.c, wasting system memory