[FFmpeg-devel] [PATCH] fate/filter-video.mak: do not use bit-exact check for dnn_processing

[Martin Storsjö]

On Sat, 18 Jan 2020, Guo, Yejun wrote:

>
>
>> -----Original Message-----
>> From: ffmpeg-devel [mailto:ffmpeg-devel-bounces at ffmpeg.org] On Behalf Of
>> Carl Eugen Hoyos
>> Sent: Saturday, January 18, 2020 7:06 AM
>> To: FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
>> Subject: Re: [FFmpeg-devel] [PATCH] fate/filter-video.mak: do not use bit-exact
>> check for dnn_processing
>>
>> Am Fr., 17. Jan. 2020 um 15:37 Uhr schrieb Guo, Yejun <yejun.guo at intel.com>:
>>>
>>> the reason is that the tested models are in float format
>>
>> The two reference files are ~1MB together: Do we want / need that?
>> (No opinion here.)
>
> how about just keeping one test (for format rgb24) for vf_dnn_processing? The rgb format might be the most popular format for vf_dnn_processing.
>
> And the test has two materials:
> 1)  fate-suite/dnn_processing/dnn_rgb_processing.model (can be renamed from halve_first_channel_float.model, 184 byte now)
> 2)  fate-suite/filter-reference/dnn_rgb_processing.raw (can be renamed from dnn_processing-halve_first_channel_float.raw, 446KB now)
>
> When we support more models (the most effort is to add more dnn layers in native mode),
> I can update both dnn_rgb_processing.model (the file size will be increased bit by bit) and dnn_rgb_processing.raw (the file size will not change much).

Keep in mind that ideally, you shouldn't be changing the reference files 
in the separate samples directory incrementially; ideally they should be 
fairly static.

Reducing the number of tests, if you need large test references, sounds 
sensible to me.

// Martin