"Ronald S. Bultje" <rsbultje@...> writes: > Hi, > > On Tue, Jul 31, 2012 at 3:17 PM, Diego Biurrun <diego@...> wrote: > [..] > > The most important thing is missing: why? mmxext is a somewhat more official name. -- -- Måns Rullgård mans@...
On Wed, 1 Aug 2012, Diego Biurrun wrote: > diff --git a/libavutil/x86/x86util.asm b/libavutil/x86/x86util.asm > index 6911a9c..f358257 100644 > --- a/libavutil/x86/x86util.asm > +++ b/libavutil/x86/x86util.asm > <at> <at> -306,6 +306,23 <at> <at> > %endif > %endmacro > > +%macro PSWAPD 2 > +%if cpuflag(sse) > + pshufw %1, %2, 0x4e That's mmx2. The macro was called sse because it was in a float function which thus didn't have a mmx2 version. > +%elif cpuflag(3dnowext) > + pswapd %1, %2 > +%else > +%ifidn %1, %2 > + movd [r0+12], %1 > + punpckhdq %1, [r0+8] > +%else > + movq %1, %2 > + psrlq %1, 32 > + punpckldq %1, %2 > +%endif > +%endif(Continue reading)
Hi, On Tue, Jul 31, 2012 at 3:18 PM, Diego Biurrun <diego@...> wrote: > diff --git a/libavutil/x86/x86util.asm b/libavutil/x86/x86util.asm [..] > +%macro PSWAPD 2 [..] > +%ifidn %1, %2 > + movd [r0+12], %1 > + punpckhdq %1, [r0+8] > +%else No. No, no, no, no, no. No. Ronald
On 07/31/2012 06:17 PM, Diego Biurrun wrote: > -%macro BUTTERFLY_SSE 4 > +%macro BUTTERFLY 4 > +%if cpuflag(avx) > + vsubps %4, %1, %2 > + vaddps %2, %2, %1 > + vmulps %1, %4, %3 > +%elif cpuflag(sse) > movaps %4, %1 > subps %1, %2 > addps %2, %4 > mulps %1, %3 > +%endif > %endmacro The avx/sse abstraction in x86inc.asm can do all that in 1 version without the if/else. -Justin
On Wed, Aug 01, 2012 at 12:17:24AM +0200, Diego Biurrun wrote: > > [PATCH 45/45] x86: fft_mmx: Refactor INTERL macro with cpuflags > > This one is still broken with 3dnow. I am mystified. Forgot to say, I ran this through fate with the following combinations of arches and cpuflags: x86_32: mmx, 3dnow, 3dnowext x86_64: mmx, mmxext, sse, sse2, ssse3 More testing is always welcome. Diego
Hi, On Tue, Jul 31, 2012 at 3:17 PM, Diego Biurrun <diego@...> wrote: > --- > libavcodec/x86/vp8dsp.asm | 143 +++++++++++++++++++++------------------------ > 1 files changed, 66 insertions(+), 77 deletions(-) This is silly, the two codepaths share not a single instruction. Also, this isn't a port to cpuflags, all of vp8dsp already uses cpuflags. It ports a macro to use cpuflags: %macro version_ssse3 .. %endmacro %macro version_mmx2 .. %endmacro becomes %macro version %if cpuflag(ssse3) .. %else .. %endif %endmacro(Continue reading)
On Wed, Aug 01, 2012 at 12:00:58AM +0100, Måns Rullgård wrote: > "Ronald S. Bultje" <rsbultje@...> writes: > > On Tue, Jul 31, 2012 at 3:17 PM, Diego Biurrun <diego@...> wrote: > > [..] > > > > The most important thing is missing: why? > > mmxext is a somewhat more official name. Yes, see my introductory email for the series. I had to unify in one direction or the other, so I picked the one that appeared more sensible, i.e. the more official names. Diego
Hi, On Tue, Jul 31, 2012 at 3:17 PM, Diego Biurrun <diego@...> wrote: > --- > libavcodec/x86/proresdsp.asm | 6 +++--- > 1 files changed, 3 insertions(+), 3 deletions(-) > > diff --git a/libavcodec/x86/proresdsp.asm b/libavcodec/x86/proresdsp.asm > index 93759bc..91e0800 100644 > --- a/libavcodec/x86/proresdsp.asm > +++ b/libavcodec/x86/proresdsp.asm > <at> <at> -404,12 +404,12 <at> <at> cglobal prores_idct_put_10, 4, 4, %1 > RET > %endmacro > > -%macro SIGNEXTEND 2-3 ; dstlow, dsthigh, tmp > -%if cpuflag(sse4) > +%macro SIGNEXTEND 2-3 > +%if cpuflag(sse4) ; dstlow, dsthigh > movhlps %2, %1 > pmovsxwd %1, %1 > pmovsxwd %2, %2 > -%else ; sse2 > +%elif cpuflag(sse2) ; dstlow, dsthigh, tmp > pxor %3, %3 > pcmpgtw %3, %1 > mova %2, %1 > -- > 1.7.2.5(Continue reading)
On Tue, Jul 31, 2012 at 03:46:55PM -0700, Ronald S. Bultje wrote:
> On Tue, Jul 31, 2012 at 3:43 PM, Diego Biurrun <diego <at> biurrun.de> wrote:
> > On Wed, Jul 25, 2012 at 08:30:19PM -0700, Ronald S. Bultje wrote:
> >>
> >> --- a/libavcodec/dct-test.c
> >> +++ b/libavcodec/dct-test.c
> >> <at> <at> -85,7 +85,7 <at> <at> static const struct algo fdct_tab[] = {
> >> { "IJG-AAN-INT", ff_fdct_ifast, SCALE_PERM },
> >> { "IJG-LLM-INT", ff_jpeg_fdct_islow_8, NO_PERM },
> >>
> >> -#if HAVE_MMX
> >> +#if HAVE_MMX && HAVE_INLINE_ASM
> >> { "MMX", ff_fdct_mmx, NO_PERM, AV_CPU_FLAG_MMX },
> >> { "MMX2", ff_fdct_mmx2, NO_PERM, AV_CPU_FLAG_MMX2 },
> >> { "SSE2", ff_fdct_sse2, NO_PERM, AV_CPU_FLAG_SSE2 },
> >
> > Drop this controversial hunk from the patch for now, the rest is OK.
>
> Can we please leave structure of asm init functions out of this
> discussion? I'm leaving it exactly as it was before, so if you want to
> change that, it should be discussed separately and done on its own
> merit, not within this discussion.
Sure, leave the init part as-is, drop the hunk above and we're good
to go.
Diego
_______________________________________________
libav-devel mailing list
libav-devel <at> libav.org
(Continue reading)
Hi, On Tue, Jul 31, 2012 at 3:17 PM, Diego Biurrun <diego@...> wrote: > --- > libavcodec/x86/h264_intrapred_10bit.asm | 307 ++++++++++++++++--------------- > 1 files changed, 158 insertions(+), 149 deletions(-) OK. Ronald
RSS Feed703 | |
|---|---|
968 | |
1313 | |
950 | |
1841 | |
1370 | |
1012 | |
2464 | |
1630 | |
2208 | |
2329 | |
1535 | |
1675 | |
1678 | |
2082 | |
2504 | |
2291 | |
2793 | |
1742 | |
2078 | |
1510 | |
1107 | |
1779 | |
1863 | |
2014 | |
2195 | |
464 |
