On Tue, Jan 10, 2012 at 1:16 PM, Christophe Gisquet
<[email protected]> wrote:
> 2012/1/10 Jason Garrett-Glaser <[email protected]>:
>> You don't need separate MMX and SSE macros.  Just use movh.
>
> OK.
>
>> This code is needlessly redundant right now.
>
> Do you mean I could avoid duplication of that code elsewhere, or in
> that code (which would be achieved by using movh)

I just mean that it has the same thing written multiple times, right here.

> Can I name them (I'm beginning with x86asm stuff) or do you mean just
> leaving a comment at the start of the macro?

Leave a comment like this:

; dst, src1, src2, tmp
MACRONAME 4
[...]

With possibly more information if you need to describe one.

>> +    cmp        r6, 0      ; are both multiple of 2^9?
>>
>> Does this happen often?  Is it worth optimizing for?  Comment about this.
>
> It did happen fairly often on my test sequences, but maybe that was
> luck. In the above formula, refdist is:
> int refdist = GET_PTS_DIFF(r->next_pts, r->last_pts)
> so it is not completely obvious it is always a power of 2.
>
> However, it was worth it: the *_TIMER results went down by (very
> approximately) 10%.
>
> I guess I can indeed add this to the commit message to make it more obvious.

I think it needs to be a comment in the code.  The relationship to the
refdist formula is not very obvious at first glance.

>> Consider whether the pmac* instructions might be usable here (xop).
>
> I haver never worked with them, and no mean to test them. From this:
> http://support.amd.com/us/Embedded_TechDocs/43479.pdf
> it seems those vpmac* instr may indeed improve things. Though I have
> no further clue.

I think someone has a bulldozer that you can ssh to, but obviously if
you can't find one don't bother.

Jason
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to