[FFmpeg-devel] [PATCH] x86/flacdsp: add SSE2 and AVX decorrelate functions

[Clément Bœsch]

On Sun, Nov 02, 2014 at 07:55:35PM -0300, James Almer wrote:
> On 02/11/14 7:43 PM, Clément Bœsch wrote:
> > On Sun, Nov 02, 2014 at 07:31:48PM -0300, James Almer wrote:
> >> Two to four times faster depending on instruction set, block size and channel count.
> >>
> >> Signed-off-by: James Almer <jamrial at gmail.com>
> >> ---
> >> TODO: 16 bits indep for 4, 6 and 8 channels. 24/32 bits indep for 8 channels.
> >>       AVX2 and maybe MMX versions.
> >>       Planar?
> >>
> >>  libavcodec/arm/flacdsp_init_arm.c |   2 +-
> >>  libavcodec/flacdec.c              |   6 +-
> >>  libavcodec/flacdsp.c              |   6 +-
> >>  libavcodec/flacdsp.h              |   6 +-
> >>  libavcodec/flacenc.c              |   2 +-
> >>  libavcodec/x86/flacdsp.asm        | 206 ++++++++++++++++++++++++++++++++++++++
> >>  libavcodec/x86/flacdsp_init.c     |  48 ++++++++-
> >>  7 files changed, 264 insertions(+), 12 deletions(-)
> > [...]
> >> +    mova       m0, [in0q]
> >> +    mova       m1, [in0q+in1q]
> >> +%if %1 > 2
> >> +    mova       m2, [in0q+in2q]
> >> +    mova       m3, [in0q+in3q]
> >> +%if %1 > 4
> >> +    mova       m4, [in0q+in4q]
> >> +    mova       m5, [in0q+in5q]
> >> +%endif
> >> +%endif
> >> +    pslld      m0, m%2
> >> +    pslld      m1, m%2
> >> +%if %1 > 2
> >> +    pslld      m2, m%2
> >> +    pslld      m3, m%2
> >> +%if %1 > 4
> >> +    pslld      m4, m%2
> >> +    pslld      m5, m%2
> >> +%endif
> >> +%endif
> > 
> > Can't you do something like this? (untested)
> >     pslld      m0, [in0q], m%2
> >     %assign i 0
> >     %rep %1
> >     pslld      m%i, [in0q+in%iq], m%2
> >     %assign    i i+1
> >     %endrep
> 
> YASM    libavcodec/x86/flacdsp.o
> D:/MinGW/msys/1.0/ffmpeg/src/libavcodec/x86/flacdsp.asm:271: error: undefined symbol `m' (first use)
> D:/MinGW/msys/1.0/ffmpeg/src/libavcodec/x86/flacdsp.asm:271: error: undefined symbol `i' (first use)
> D:/MinGW/msys/1.0/ffmpeg/src/libavcodec/x86/flacdsp.asm:271: error: undefined symbol `in' (first use)
> D:/MinGW/msys/1.0/ffmpeg/src/libavcodec/x86/flacdsp.asm:271: error: undefined symbol `iq' (first use)
> D:/MinGW/msys/1.0/ffmpeg/src/libavcodec/x86/flacdsp.asm:271: error:  (Each undefined symbol is reported only once.)
> make: *** [libavcodec/x86/flacdsp.o] Error 1
> 
> A %rep like that is only four lines shorter. Do you consider it more readable than the alternative to justify trying 
> to get it working?

Totally up to you, it looked easier to maintain and obvious than several
nested ifdefery.

-- 
Clément B.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 473 bytes
Desc: not available
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20141102/5f919ddc/attachment.asc>