[FFmpeg-devel] [PATCH] Dsputilize some functions from APE decode 2/2 - SSE2
Michael Niedermayer
michaelni
Tue Jul 8 14:18:25 CEST 2008
On Tue, Jul 08, 2008 at 01:48:21PM +0300, Kostya wrote:
> On Tue, Jul 08, 2008 at 12:07:49AM +0200, Michael Niedermayer wrote:
> > On Mon, Jul 07, 2008 at 09:21:07PM +0300, Kostya wrote:
[...]
> > also the loop will likely significantly benefit from being unrolled once
>
> len is declared as multiple of 8, and loop handles 8 elements
is it ever used with %16 != 0 len ?
[...]
> > > + "add $16, %0 \n\t"
> > > + "add $16, %1 \n\t"
> > > + "sub $16, %3 \n\t"
> > > + "jnz 1b \n\t"
> > > + "movd %%xmm7, %2 \n\t"
> > > + : "+r"(v1), "+r"(v2), "=r"(res), "+r"(order)
> >
> > > + : "m"(sh)
> >
> > should be in a register
>
> why? psrad takes either (x)mm register, immediate value or memory
> for input.
memory is likely slower than a register
[...]
> +static void add_int16_sse2(int16_t * v1, int16_t * v2, int order)
> +{
> + x86_reg o = order - 8;
> + asm volatile(
> + "1: \n\t"
> + "movdqu (%1,%2,2), %%xmm0 \n\t"
> + "paddw (%0,%2,2), %%xmm0 \n\t"
> + "movdqa %%xmm0, (%0,%2,2) \n\t"
> + "sub $8, %2 \n\t"
> + "jge 1b \n\t"
> + : "+r"(v1), "+r"(v2), "+r"(o)
> + );
accessing arrays from end to start is likely slower than start to end
(if you want it i want to see benchmarks of it against the equivalent
forward code)
also (a,b) might be faster than (a,b,2)
a add %2, %2 before the loop would avoid it
[...]
--
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
Complexity theory is the science of finding the exact solution to an
approximation. Benchmarking OTOH is finding an approximation of the exact
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20080708/5a957f2f/attachment.pgp>
More information about the ffmpeg-devel
mailing list