[FFmpeg-devel] [PATCH 1/2] avcodec/x86: move simple_idct to external assembly

[Michael Niedermayer]

On Mon, May 29, 2017 at 09:44:21PM -0400, Ronald S. Bultje wrote:
> Hi,
> 
> On Mon, May 29, 2017 at 5:26 PM, Michael Niedermayer <michael at niedermayer.cc
> > wrote:
> 
> > On Mon, May 29, 2017 at 09:40:49PM +0200, James Darnley wrote:
> > > On 2017-05-29 16:51, James Darnley wrote:
> > > > ---
> > > > Changes:
> > > >  - Changed type of d40000 constant to dwords because it gets used as
> > dwords.
> > > >  - Changed or removed HAVE_MMX_INLINE preprocessor guards.
> > > >  - Added note about conversion from inline.
> > > >  - New file no longer has "2" suffix.
> > > >  - Whitespace (indentation and alignment).
> > > >
> > > >  libavcodec/tests/x86/dct.c     |   2 +-
> > > >  libavcodec/x86/Makefile        |   4 +-
> > > >  libavcodec/x86/idctdsp_init.c  |   4 -
> > > >  libavcodec/x86/simple_idct.asm | 889 ++++++++++++++++++++++++++++++
> > +++++++++
> > > >  libavcodec/x86/simple_idct.c   | 929 ------------------------------
> > -----------
> > > >  5 files changed, 892 insertions(+), 936 deletions(-)
> > > >  create mode 100644 libavcodec/x86/simple_idct.asm
> > > >  delete mode 100644 libavcodec/x86/simple_idct.c
> > >
> > > Ronald queried on IRC about the performance.  The libavcodec/tests/dct
> > > utility reports these numbers
> > >
> > > Yorkfield:
> > >  - inline:   IDCT SIMPLE-MMX: 15715.9 kdct/s
> > >  - external: IDCT SIMPLE-MMX: 15699.9 kdct/s
> > >
> > > Skylake-U:
> > >  - inline:   IDCT SIMPLE-MMX: 11193.3 kdct/s
> > >  - external: IDCT SIMPLE-MMX: 11189.7 kdct/s
> >
> > Its better to benchmark by decoding some videos as the sparsness of
> > the coeffs affects speed
> 
> 
> The vp9 checkasm idct test already tests this (for performance as well as
> coverage purposes), maybe dct-test should include a test for "sub-idcts"
> (including, but not limited to, dc-only) also?

I think dct-test supports dc only and sparse tests, not sure how
relevant these are though, i doubt that they match real world
statistics very well

hmm, but as you mention this, it looks from a quick glance that these
tests are not run in fate

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Those who are too smart to engage in politics are punished by being
governed by those who are dumber. -- Plato 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20170531/77f9681e/attachment.sig>