Currently we have slower path to scroll the image:

• copy entire dirty region from front buffer to back buffer to ensure both buffers have the same content
• scroll the image within back buffer (another copy)
• begin paint (we already have a valid back buffer after scrolling - we don't copy anything in this case)
• paint new content
• end paint and flush

Optimized path to scroll the image:

• gather scrolling deltas and regions which are before begin paint
• in begin paint, prepare new back buffer, but don't copy anything that will be scrolled or painted
• finalize scroll - use the front buffer image to scroll the data to the back buffer (we haven't copied them yet, we'll copy them with applied scrolling deltas)
• paint new content
• end paint and flush

This path prevents redundant copy from front buffer to back buffer the content which will be scrolled and painted (we'll override these pixels anyway).

—

Simple benchmark application: wlscroll.tar.gz

Results on KWin 6.4, 3840x2160, 200% scale, full screen, dual channel DDR5, AMD Ryzen (fixed clock):

Related task for XCB backend scrolling optimization: QTBUG-64414