Well, simply disabling the preview does come with a few downsides, such as no longer seeing a blurry picture when zooming, before the sharp render shows up. Not really a long-term solution.
But as it turns out, there was already an option in the code for using a smaller preview. It had a few bugs, though, if you actually enabled it. I fixed those bugs, and here are the performance numbers:
And now:
- Linux/OpenCL Default: 0.150 s
A bit more than twice as fast as before. (Although some 30% or so were gained by compiling on my own machine, either because the current git version is faster than 3.0.0, or because local compilation sped it up.)
It used to be that the preview took about as long as the image. Now the preview takes a tiny fraction of the time, leaving the image to finish much faster.
I’ll open an issue on Github and see if I accidentally broke something here, or if that’s actually useful for someone.