On Sun, Jan 08, 2023 at 11:47:42PM +0200, Adrian Bunk wrote: > On Sun, Jan 08, 2023 at 10:11:46PM +0100, Andreas Bombe wrote: > >... > > i386 only, where some > > floats don't match exactly, which I'm still investigating. > >... > > That's "better" results due to x87 excess precison. > > The patch below fixes the tests for me.
Thanks very much, I suspected something in the direction of x87 80-bit floats messing with this, but I didn't know an easy way to fix that. Also fixed the qs1dsearch failures, but I'm leaning towards this still being a gcc misoptimization issue. Just unfortunate that this might make the code slower and the library has no autodetection of SIMD units.
