[FFmpeg-devel] [PATCH] libavfilter/scale: Populate ow/oh when using 0 as w/h

[Michael Niedermayer]

On Tue, Jun 13, 2017 at 12:10:07AM -0400, Kevin Mark wrote:
> On Mon, Jun 12, 2017 at 9:42 PM, Michael Niedermayer
> <michael at niedermayer.cc> wrote:
> > why is there a cast at all ?
> 
> The cast is there because if you run this:
> 
> ffmpeg -frames:v 5 -filter_complex
> "sws_flags=+accurate_rnd+bitexact;testsrc=size=320x240
> [main];testsrc=size=640x360 [ref];[main][ref]
> scale2ref=0:print(ow/641) [main][ref];[ref] nullsink" -map "[main]"
> -flags +bitexact -fflags +bitexact -f md5 -
> 
> This works just fine without a cast for ow. 0 == 0 is true so we set
> it to 640. But for oh, the print() shows that ow/641 is 0.998440. When
> it is truncated from a double to an integer (eval_h = res) it becomes
> 0. But in our comparison, 0.998440 == 0 is false so in this case
> eval_h will be truncated to 0 which is exactly the behavior we're
> trying to correct. Adding that cast resolves the issue. 0.998440 == 0
> is false but (int) 0.998440 == 0 is true.
> 
> For the extra cast I was talking about consider this:
> 
> ffmpeg -frames:v 5 -filter_complex
> "sws_flags=+accurate_rnd+bitexact;testsrc=size=320x240
> [main];testsrc=size=640x360 [ref];[main][ref]
> scale2ref=500/6:print(print(ow)*5) [main][ref];[ref] nullsink" -map
> "[main]" -flags +bitexact -fflags +bitexact -f md5 -
> 
> That will print() 83.333333 and then 416.666667. A user might
> (reasonably, in my opinion) expect that the ow value (or oh) is always
> an integer. With the extra cast you'll see 83.000000 and 415.000000
> printed. 83.333333 truncates to 83 so no (noticeable) change for ow
> but 416.666667 does not truncate to 415 so this is an example of a
> place where the lack of truncation for ow/oh does change the outcome.

ok, makes sense
i agree the 2nd cast seems a good idea


> 
> I hope this clears it up. Perhaps that code should just be entirely
> refactored to be a little more clear?
> 
> Thanks,
> Kevin
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

You can kill me, but you cannot change the truth.
-------------- 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/20170614/6cab8bfd/attachment.sig>