[FFmpeg-devel] [PATCH 09/13] avcodec/svq1dec: clear MMX state after MB decode loop

[u-9iep at aetey.se]

On Mon, Oct 24, 2016 at 09:53:22PM +0200, Henrik Gramner wrote:
> The decision to issue emms manually instead of after every MMX
> function was a deliberate decision. I'd hardly call it "misdesigned"
> to make SIMD code twice as fast at the cost of technically abusing the

It would be nice to look at a benchmarking comparison, to be able to
see the actual practical performance gain of the decision not to follow
the ABI.

My expectation is among others that using MMX in a compliant way would
remain remarkably beneficial compared to no SIMD.

In a perfect world the user might be offered a build time choice:
either lose N% of the performance or take the consequences of not
following the ABI. Apparently, different users/packagers would make
different choices, but the crucial thing would be knowing the N.

Of course, offering to disable MMX _is_ such a choice, but having instead
an option of "somewhat slower, compliant MMX" would be much nicer.

my 2c
Rune