[FFmpeg-devel] [RFC] swscale regression tests

[Ramiro Polla]

On Tue, Jul 20, 2010 at 12:23 PM, Vitor Sessak <vitor1001 at gmail.com> wrote:
> On 07/20/2010 06:02 AM, Ramiro Polla wrote:
>> I have attached an attempt at making swscale regression tests. This
>> method creates one file for each input pixel format (currently 42) and
>> compresses the output with bzip2. I have not tested with different
>> platforms (x86_64 linux only yet), and IIRC it doesn't give the same
>> result on all platforms. This is just a request for comments whether
>> this approach is good. I have also removed the fast bilinear test
>> since it works differently on many platforms and would make the
>> regression tests more complex (like keeping one reference file for
>> each platform).
>>
>> The output bzip2ed files add 280Kb of binary data to repository (5.6Mb
>> uncompressed). As for disk usage (with 'du'), it takes 340Kb (5.8Mb
>> uncompressed).
>
> As far as I understood your patch, it will check swscale with a large
> variety with flags and evaluate the SSE for each combination. Wouldn't it
> make more sense in a first patch to check if output is bit-exact when using
> the flag +bitexact (it would involve evaluating a MD5, not just a SSE)?

Yes, I think it would make more sense. I think adding a switch (-md5
or -sse) to swscale-test would be nice. The libavutil functions could
be used to compute it (seems better than outputting to stdout and
running through md5sum, since this way we can get a string for each
test).