On Sat, Aug 4, 2018 at 6:51 AM Jean-Baptiste Mardelle <j...@kdenlive.org>
wrote:
> On 03.08.2018 19:46, Dan Dennedy wrote:
>
> On Fri, Aug 3, 2018 at 10:23 AM Jean-Baptiste Mardelle <j...@kdenlive.org>
> wrote:
>
>> Hi all,
>>
>> I recently realized that the color producer always creates an alpha
>
> mask, and uses the rgba24a format by default.
>>
>
> Typically, it will receive a format from the downstream consumer services,
> and it will only choose this default if movit is used or mlt_image_none is
> received.
>
>
> Hi Dan,
>
> Thanks for your answer.
>
>
>> In fact, unless the color has an alpha value under 255, it is
>> unnecessary and we could simply use an rgb24 format, saving some memory.
>>
>> I am mostly interested in this because the qtblend transition checks if
>> a producer has an alpha channel before deciding whether a compositing is
>> required. Having a useless alpha channel on the color producer causes
>> unnecessary compositing and slowdowns in this case. Would you be ok to
>> change the color producer to use rgb24 or rgb24a formats depending on
>> the alpha value ?
>>
>>
> Only in the truly default case. If qtblend requests mlt_image_rgb24a and
> the color producer returns mlt_image_rgb24, then the image conversion
> filter will automatically convert it. That is slightly less efficient than
> directly returning what is requested.
>
>
> Ok, I realize I didn't correctly analyze the situation. You are totally
> right that in most cases, the format will be determined by the consumer
> services.
> In fact, the optimization would be to only create the alpha mask if the
> consumer requests an rgba format or if the alpha is less than 255.
>
> something like:
>
> + uint8_t *alpha = NULL;
> + if (color.a < 255 || *format == mlt_image_rgb24a) {
> + alpha = mlt_pool_alloc( alpha_size );
> + (...)
>
>
> If the consumer requests yuv4222 and the color has no transparency, it
> doesn't make sense to create an alpha mask ?
>
>
You are probably correct, and the above change is safe. I will give it a
try.
>
> For frei0r.cairoblend, I check every pixel's alpha until it sees one that
> is not 255. See transition_frei0r.c:is_opaque().
>
>
>
> Sure but it seems to me that simply checking for the presence of an alpha
> mask would be much faster than checking every pixel of a frame..
>
True, and I could add that check as well before checking every pixel.
>
> Thanks & regards
>
>
> Jean-Baptiste
>
>
>
>
>
>> I can provide a patch if you think this is a welcome change.
>>
>>
>> Best regards
>>
>> Jean-Baptiste
>>
>>
>
------------------------------------------------------------------------------
Check out the vibrant tech community on one of the world's most
engaging tech sites, Slashdot.org! http://sdm.link/slashdot
_______________________________________________
Mlt-devel mailing list
Mlt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mlt-devel
