[FFmpeg-devel] [PATCH] avfilter/vf_ssim: align temp size

[Ronald S. Bultje]

Hi,

On Wed, Aug 2, 2017 at 8:45 AM, Tobias Rapp <t.rapp at noa-archive.com> wrote:

> On 02.08.2017 14:04, Ronald S. Bultje wrote:
>
>> On Wed, Aug 2, 2017 at 3:10 AM, Tobias Rapp <t.rapp at noa-archive.com>
>> wrote:
>>
>>> On 01.08.2017 17:01, Muhammad Faiz wrote:
>>>
>>>> Fix Ticket6519.
>>>>
>>>> Signed-off-by: Muhammad Faiz <mfcc64 at gmail.com>
>>>> ---
>>>>  libavfilter/vf_ssim.c | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/libavfilter/vf_ssim.c b/libavfilter/vf_ssim.c
>>>> index c3c204268f..dfd276e015 100644
>>>> --- a/libavfilter/vf_ssim.c
>>>> +++ b/libavfilter/vf_ssim.c
>>>> @@ -402,7 +402,7 @@ static int config_input_ref(AVFilterLink *inlink)
>>>>      for (i = 0; i < s->nb_components; i++)
>>>>          s->coefs[i] = (double) s->planeheight[i] * s->planewidth[i] /
>>>> sum;
>>>>
>>>> -    s->temp = av_malloc_array((2 * inlink->w + 12), sizeof(*s->temp) *
>>>> (1 + (desc->comp[0].depth > 8)));
>>>> +    s->temp = av_malloc_array(FFALIGN(2 * inlink->w + 12, 64),
>>>> sizeof(*s->temp) * (1 + (desc->comp[0].depth > 8)));
>>>>      if (!s->temp)
>>>>          return AVERROR(ENOMEM);
>>>>      s->max = (1 << desc->comp[0].depth) - 1;
>>>>
>>>>
>>>> I confirm that the command doesn't crash anymore and the reports of
>>> "invalid read/write" in Valgrind are gone. However there are still some
>>> "use of uninitialized value" reports remaining. Maybe use
>>> av_mallocz_array?
>>>
>>
>>
>> Wait wait, before we do that, which values are uninitialized and what are
>> they used for?
>>
>
> Reads into s->temp seem to use uninitialized values on x86-64 when
> vf_ssim.asm routines are used (-cpuflags all), see attached Valgrind
> report. When vf_ssim.asm is not used (-cpuflags 0) no "use of uninitialized
> value" is reported.
>
> The difference of the reported SSIM scores between my asm and non-asm test
> runs was <= 10^-6.


Which function is causing the warnings? Isit dsp->ssim_4x4_line() or
dsp->ssim_end_line()? (My gut feeling tells me it's ssim_4x4_line(), but I
don't understand why, since the result should be unused in
ssim_end_line().) It's really important to understand the source and
implication of these warnings before we silence them - they may be causing
actual inaccuracies in the result, which we don't want.

Ronald