On Sat, Sep 23, 2023 at 11:32:28PM +0200, Paul B Mahol wrote: > On Sat, Sep 23, 2023 at 11:02 PM Michael Niedermayer > wrote: > > > From: Paul B Mahol > > > > Instead use float one by default for sample rate conversions. > > The s16p internal transfer format produces visible and hearable > > quantization artifacts. > > > > Signed-off-by: Paul B Mahol > > > > for S8 we continue to use S16 as it should have enough precision > > Fate is adjusted so bitexactness is maintained between mips/arm/x86 > > if more tests became bit-inexact on some platform, the same change > > can be done to them > > > > The use of higher precision and float intermediates inevitably > > leads to more differences between platforms. > > > > Could add S64 precision path, but would be slower than float, probably. > LGTM. will apply later without 2 lines (bugs noticed on rerunning with distclean) thx [...] -- Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB "You are 36 times more likely to die in a bathtub than at the hands of a terrorist. Also, you are 2.5 times more likely to become a president and 2 times more likely to become an astronaut, than to die in a terrorist attack." -- Thoughty2