For example when creating a BorderImage which source is 33x33 pixels and
setting the borders to 16 (one pixel in the center gets stretched/repeated
to fill the center area), the rendering performance is horrible with Repeat
and Round modes.

Also the CPU usage gets really high. On my desktop machine, rotating a
1000x1000 sized BorderImage like the one described above consumes 1% of CPU
when using Stretch tile mode and 90% when using Repeat or Round modes.

If implementing the Repeat and Round modes by multiplying the geometry
is the way to go then at least it should be properly documented that
repeating really small source image areas to fill large areas is very
expensive.