[FFmpeg-devel] [PATCH 4/9] avfilter/af_flanger: use rint instead of floor hack

[Paul B Mahol]

On 12/2/15, Ganesh Ajjanagadde <gajjanagadde at gmail.com> wrote:
> On Wed, Dec 2, 2015 at 1:18 AM, Paul B Mahol <onemda at gmail.com> wrote:
>> On 12/2/15, Ganesh Ajjanagadde <gajjanagadde at gmail.com> wrote:
>>> Signed-off-by: Ganesh Ajjanagadde <gajjanagadde at gmail.com>
>>> ---
>>>  libavfilter/af_flanger.c | 2 +-
>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/libavfilter/af_flanger.c b/libavfilter/af_flanger.c
>>> index f8ec830..a92367c 100644
>>> --- a/libavfilter/af_flanger.c
>>> +++ b/libavfilter/af_flanger.c
>>> @@ -130,7 +130,7 @@ static int config_input(AVFilterLink *inlink)
>>>          return AVERROR(ENOMEM);
>>>
>>>      ff_generate_wave_table(s->wave_shape, AV_SAMPLE_FMT_FLT, s->lfo,
>>> s->lfo_length,
>>> -                           floor(s->delay_min * inlink->sample_rate +
>>> 0.5),
>>> +                           rint(s->delay_min * inlink->sample_rate),
>>>                             s->max_samples - 2., 3 * M_PI_2);
>>>
>>>      return av_samples_alloc_array_and_samples(&s->delay_buffer, NULL,
>>> --
>>> 2.6.2
>>>
>>> _______________________________________________
>>> ffmpeg-devel mailing list
>>> ffmpeg-devel at ffmpeg.org
>>> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>>>
>>
>> Have you checked that output is same?
>
> Well if is not, rint is more accurate than floor, this is the whole
> point of the patch. What I can tell is that FATE passes.
>
> One can craft input so that floor(x + 0.5) is not identical to
> rint(x), and that is the point of these patches - to be more accurate
> when we can be. A simple example: what happens at half-integers, i.e
> 1.5? Then, floor always returns the next up, e.g 2.0, while rint(x)
> rounds to the nearest even integer in accord with IEEE-754. This is
> done to reduce rounding biases on floating point numbers - think of a
> large number of half integer samples, the floor hack results in
> consistent upward bias, the rint (or llrint, lrint more generally)
> avoids this.
>
> I care about technical purity of filters; you seem to care about
> copying it over from some other place and matching another filter
> exactly, regardless of the quality of such filters. In that case, I
> think FFmpeg's monolithic filter design needs to be reconsidered; we
> should allow seamless integration of external filters. These two goals
> are at odds with each other, and I will always personally prefer the
> first, since it actually allows greater flexibility for improvements.
> Ultimately, I am not a maintainer for these things and I have no say
> on the matter or personal interest in it.
>

On 2nd look, patch should be fine.