On Wed, Jun 07, 2023 at 07:29:13PM +0200, Paul B Mahol wrote:
> On Wed, Jun 7, 2023 at 6:01 PM Michael Niedermayer <michael@niedermayer.cc>
> wrote:
> 
> > On Wed, Jun 07, 2023 at 05:46:25PM +0200, Paul B Mahol wrote:
> > > On Sat, May 13, 2023 at 3:09 AM Michael Niedermayer <
> > michael@niedermayer.cc>
> > > wrote:
> > >
> > > > On Fri, May 12, 2023 at 07:28:08PM +0200, Paul B Mahol wrote:
> > > > > With fixed alignment requirements.
> > > >
> > > > >  rematrix.c            |   54
> > > > ++++++++++++++++++++++++++++++++++++++++++++++++--
> > > > >  swresample.c          |    5 ++++
> > > > >  swresample_internal.h |    2 +
> > > > >  3 files changed, 59 insertions(+), 2 deletions(-)
> > > > > 3b99c9eb2e2f1f17d1f306e37ddd7107405fede4
> > > > 0001-swresample-reuse-DSP-functions-from-avutil.patch
> > > > > From 771bc1414b737475bc42c7263fd7f21b4d9cc9b7 Mon Sep 17 00:00:00
> > 2001
> > > > > From: Paul B Mahol <onemda@gmail.com>
> > > > > Date: Wed, 10 May 2023 15:41:01 +0200
> > > > > Subject: [PATCH] swresample: reuse DSP functions from avutil
> > > > >
> > > > > Improves generic mixing dramatically.
> > > > >
> > > > > Signed-off-by: Paul B Mahol <onemda@gmail.com>
> > > > > ---
> > > > >  libswresample/rematrix.c            | 54
> > +++++++++++++++++++++++++++--
> > > > >  libswresample/swresample.c          |  5 +++
> > > > >  libswresample/swresample_internal.h |  2 ++
> > > > >  3 files changed, 59 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/libswresample/rematrix.c b/libswresample/rematrix.c
> > > > > index 79e8a43eac..2133b0f90d 100644
> > > > > --- a/libswresample/rematrix.c
> > > > > +++ b/libswresample/rematrix.c
> > > > > @@ -652,7 +652,32 @@ int swri_rematrix(SwrContext *s, AudioData *out,
> > > > AudioData *in, int len, int mus
> > > > >              break;}
> > > > >          default:
> > > > >              if(s->int_sample_fmt == AV_SAMPLE_FMT_FLTP){
> > > > > -                for(i=0; i<len; i++){
> > > > > +                if (out->planar && in->planar)
> > > > > +                    len1 = len & ~15;
> > > > > +                else
> > > > > +                    len1 = 0;
> > > > > +                if ((intptr_t)out->ch[out_i] & 0x1f)
> > > > > +                    len1 = 0;
> > > > > +                for (j = 0; j < s->matrix_ch[out_i][0] && len1 > 0;
> > > > j++) {
> > > > > +                    in_i = s->matrix_ch[out_i][1+j];
> > > > > +                    if ((intptr_t)in->ch[in_i] & 0x1f) {
> > > > > +                        len1 = 0;
> > > > > +                        break;
> > > > > +                    }
> > > > > +                }
> > > >
> > > > Cant this be done outside the "inner" loop ?
> > >
> > >
> > > Sure.
> > >
> > >
> > >
> > > >
> > > >
> > >
> > > > also this produces some new NaN values
> > > >
> > > > @@ -91810,16 +91810,16 @@
> > > >  [e:0.246031 c:-nan max:0.988908] len:  936
> > > >  [e:0.247006 c:-nan max:0.988908] len:  936
> > > >  [e:0.247174 c:-nan max:0.988908] len:  936
> > > > -[e:0.197683 c:0.773693 max:0.825360] len:  936
> > > > -[e:0.192089 c:0.814010 max:0.820662] len:  936
> > > > +[e:0.245992 c:0.031094 max:0.988908] len:  936
> > > > +[e:0.246535 c:0.031025 max:0.988908] len:  936
> > > >  [e:0.013306 c:0.996638 max:0.037320] len:   32 F:  2
> > > >  [e:0.049179 c:0.909927 max:0.081071] len:   32 F:  2
> > > >  [e:0.159079 c:-nan max:0.299026] len:   32 F:  2
> > > >  [e:0.116819 c:-nan max:0.297598] len:   32 F:  2
> > > >  [e:0.159382 c:-nan max:0.299980] len:   32 F:  2
> > > >  [e:0.115993 c:-nan max:0.296648] len:   32 F:  2
> > > > -[e:0.099115 c:0.996999 max:0.189015] len:   32 F:  2
> > > > -[e:0.071657 c:0.998728 max:0.187209] len:   32 F:  2
> > > > +[e:0.159577 c:-nan max:0.299503] len:   32 F:  2
> > > > +[e:0.115367 c:-nan max:0.299503] len:   32 F:  2
> > > >
> > > >
> > > Is that really important to you?
> >
> > The important part is not what the tool displays. But that this
> > points to a worsening of the tested code (or a bug in the tool)
> > The other numbers also seem to worsen by non trivial amounts
> >
> 
> Can you elaborate your reasoning. Otherwise I will not take this into
> serious
> account for very slow library wasting CPU cycles.


if you do care that the output resembles the input then this from
above should be concerning:

> > > > -[e:0.197683 c:0.773693 max:0.825360] len:  936
> > > > -[e:0.192089 c:0.814010 max:0.820662] len:  936
> > > > +[e:0.245992 c:0.031094 max:0.988908] len:  936
> > > > +[e:0.246535 c:0.031025 max:0.988908] len:  936

before there is 0.81 correlation and afterwards there is
0.03 correlation

I mean the tool is telling us, we go kind of from 81% similarity to 3% similarity
if thats not a bug in the tool thats very odd

The NaN could mean that the resampler turned a non zero signal into all zeros
or a all zero signal into non zero. Or maybe it returned NaN directly. Not sure
these are the only ways you get a NaN there
If both signals are all zero the max difference would be 0 and its not.

This tool tries to excercise corner cases, so if it changes theres a good
chance theres either a new bug in a corner case or one fixed. I dont think
this should be ignored.

thx

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

Observe your enemies, for they first find out your faults. -- Antisthenes