[FFmpeg-devel] [PATCH] lavc/x86/videodsp: Drop MMX usage
Ronald S. Bultje
rsbultje at gmail.com
Tue Nov 12 21:27:12 EET 2024
Hi,
On Tue, Nov 12, 2024 at 2:14 PM Frank Plowman <post at frankplowman.com> wrote:
> Remove the MMX versions of these functions and modify the SSE
> implementations to avoid using MMX registers.
>
> Signed-off-by: Frank Plowman <post at frankplowman.com>
> ---
> This wasn't wholly straightforward as the existing SSE implementation did
> not only use SSE but rather a blend of SSE and MMX. I would appreciate
> some review, as I am not particularly familiar with x86 assembly. Of
> particular interest are the changes to {READ,WRITE}_NUM_BYTES. The new
> implementation does not make economic use of the XMM registers like the
> old one did, but it still does not exhaust them -- was this more of a
> concern when this was originally written?
>
I wrote the original SIMD code (IIRC), and back then MMX usage was
acceptable. Nowadays, it carries performance penalties with it and is
deprecated, so the baseline assumptions are just very different.
Your changes look reasonable, thanks for working on this. OK from me.
Ronald
More information about the ffmpeg-devel
mailing list