[FFmpeg-devel] [PATCH] fate/pixfmt: disable dithering in the scale filter

[James Almer]

On 11/9/2024 1:19 PM, Michael Niedermayer wrote:
> On Sat, Nov 09, 2024 at 11:05:32AM -0300, James Almer wrote:
>> On 11/8/2024 9:03 PM, Michael Niedermayer wrote:
>>> On Fri, Nov 08, 2024 at 04:56:38PM -0300, James Almer wrote:
>>>> On 11/8/2024 5:15 AM, Martin Storsjö wrote:
>>>>> On Thu, 7 Nov 2024, James Almer wrote:
>>>>>
>>>>>> Should fix fate failures across different systems.
>>>>>>
>>>>>> Signed-off-by: James Almer <jamrial at gmail.com>
>>>>>> ---
>>>>>> This only hides the bug in the dither path for unscaled planar yuv
>>>>>> 8bit to
>>>>>> hbd and vice-versa.
>>>>>> Someone more familiar with the code should check what exactly is
>>>>>> going on.
>>>>>
>>>>> LGTM, thanks! Good job in hunting this one down - it seems odd as the
>>>>> exact output value seems stable on each machine (and valgrind doesn't
>>>>> find any nondeterminism), but the output is different on most machines.
>>>>
>>>> The only tests that fail are those using the DITHER_COPY method in
>>>> planarCopyWrapper() when dither is used, and the only difference compared to
>>>> the no dither path is accessing the "dithers" static const array defined in
>>>> the same file.
>>>> Maybe compilers are doing something weird when aligning it, or accessing it?
>>>
>>> Not sure i fully understand but the brute force way to debug this is maybe
>>> just add a printf() to dump every value then diff between to see where differences
>>> start.
>>
>> I can't even reproduce the issue locally, so...
> 
> is this a threads >1 issue ?

Doesn't look like, it happens on fate clients running the default THREADS=1.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature.asc
Type: application/pgp-signature
Size: 495 bytes
Desc: OpenPGP digital signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20241109/7c1f6e87/attachment.sig>