PI: optimize _decode_png_prediction

[exiledkingcc]

this commit makes tests/test_page.py::test_image_new_property cost from 8.98s to 5.69s on my machine, about x1.5 faster.

[codecov]

Codecov Report

Patch coverage: 100.00% and project coverage change: +0.01% 🎉

Comparison is base (f9d25e6) 94.23% compared to head (18ad99c) 94.24%.

❗ Current head 18ad99c differs from pull request most recent head 19429ba. Consider uploading reports for the commit 19429ba to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2068      +/-   ##
==========================================
+ Coverage   94.23%   94.24%   +0.01%     
==========================================
  Files          41       41              
  Lines        7340     7355      +15     
  Branches     1445     1448       +3     
==========================================
+ Hits         6917     6932      +15     
  Misses        263      263              
  Partials      160      160              

Files Changed	Coverage Δ
pypdf/filters.py	94.53% <100.00%> (+0.15%)	⬆️

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[pubpub-zz]

unless I'm wrong, this filter is only for images: wouldn't it be more efficient to use pillow ? your opinion?

[exiledkingcc]

i can not find any thing like this in Pillow. do you have any idea how to do it with pillow ?

[MartinThoma]

Very nice work 👍

On my machine it's:

OLD: 22s - 25s
NEW: 14s - 15s
=> 1.78x to 1.47x

The paeth_predictor can be deprecated, but I'll take care of that.

[MartinThoma]

I'm updating the test code a bit ( #2072 ), but this improvement will almost certainly go into the release this weekend :-)

[MartinThoma]

I would like to add a new benchmark for this as well (before I merge, so that we can see the difference).

The current benchmarks are not really helpful. Lets see if such a big change is visible.

[pubpub-zz]

i can not find any thing like this in Pillow. do you have any idea how to do it with pillow ?

I had a quick review and found no solution.😳

[MartinThoma]

https://py-pdf.github.io/pypdf/dev/bench/ -> look at the very bottom :-)

This shows iterations per second, hence higher is better. Still, I'm a bit disappointed that the magnitude there is way lower than what @exiledkingcc and me observed.

[exiledkingcc]

there is a huge mount of math calculations in the method, which is Python's weeknees. besides, it costs a lot of time for pillow to save png file.