Drawing many annotations is very slow

[Azeirah]

I tried converting my private diary which has over 250 pages of hand-written text. It takes well over two minutes on my high-end PC to convert this file to PDFs.

The primary bottleneck lies in drawing the line segments, see the "update_annotations" call below.

See this code: https://github.com/lucasrla/remarks/blob/master/remarks/conversion/drawing.py#L178

[Azeirah]

Luckily, I already found a way to optimize this heavy bottleneck, and it runs waayyy faster.

My theory was that the "update_annot" call has a lot of overhead, so my idea was to batch the calls.

How the code works right now:

• Every time the user has drawn a line on the page
  • Set up a new fitz annotation https://github.com/lucasrla/remarks/blob/master/remarks/conversion/drawing.py#L169
  • Configure the annotation to look like the ReMarkable tool used by the user
  • Finalize the annotation by calling .update()

If you have 1000 lines on the page (which is not strange for when you have hand-drawn pages like in a diary), there will be 1000 calls to add_ink_annot and annot.update(). This is very slow due to the overhead.

My alternative idea is to batch the lines per tool configuration, so if for example the user drew 960 lines with the "pen" tool with width=medium and 40 with the "pen" tool with width=large, then there will only be two calls to add_ink_annot and annot.update():

• Batch every line the user has drawn with the same tool-configuration (taking into account tool-type, stroke_width, color and opacity)
  • Set up a new fitz annotation per batch
  • Configure the tool
  • Finalize by calling update

I simulated this approach without the per-tool batching, so I just defaulted all lines to the same pen configuration and the performance improved drastically! This is a very promising speedup.

[Azeirah]

This one is ran with the actual batching:

Performance difference between the simulated batching and the real batching is negligible, which is great.