michaelni changed the topic of #ffmpeg-devel to: Welcome to the FFmpeg development channel | Questions about using FFmpeg or developing with libav* libs should be asked in #ffmpeg | This channel is publicly logged | FFmpeg 8.0 has been released! | Please read ffmpeg.org/developer.html#Code-of-conduct
BradleyS has quit [Quit: quit]
BradleyS has joined #ffmpeg-devel
lexano has quit [Remote host closed the connection]
lexano has joined #ffmpeg-devel
<fflogger>
[editedticket] michael: Ticket #228 ([ffserver] FFserver: reordering of "<Stream>"s in config file makes difference) updated https://trac.ffmpeg.org/ticket/228#comment:7
<mkver>
Gramner, jamrial: If we changed ff_clear_blocks_avx (in lavc/x86/blockdsp.asm) to zero the xmm register via VEX-encoded xorps, could we then remove the implicit vzeroupper?
<Gramner>
no, at least on some cpus just reading from an ymm register flags the ymm state as dirty which requires a vzeroupper
<mkver>
Then our behavior regarding vzeroupper makes sense (while the behavior of these cpus IMO doesn't).
<mkver>
Can the ymm state also be cleaned by zeroing the upper half of every ymm register manually or does it have to be vzeroupper/vzeroall?
<Gramner>
have to use vzero*
<mkver>
Thanks for your time.
<Gramner>
but a vzeroupper is pretty cheap tbh, so I wouldn't bother going out of the way trying to microoptimize that away