[FFmpeg-devel] [PATCH 0/14] Vulkan fixes and improvements

[Lynne]

11 Nov 2021, 12:54 by ffmpeg at haasn.xyz:

> On Thu, 11 Nov 2021 09:12:35 +0100 Lynne <dev at lynne.ee> wrote:
>
>> Nov 10, 2021, 08:20 by dev at lynne.ee:
>>
>> > I've extensively fixed and improved our Vulkan codebase, and it's
>> > now up to standard. Since it's a 14-part patchset, I've uploaded
>> > my branch for review here:
>> > https://github.com/cyanreg/FFmpeg/tree/vulkan
>> >
>> > The implementation now follows modern Vulkan API usage practices,
>> > fully takes advantage of all the hardware has to offer, it's faster,
>> > much easier to compile, static linking is possible and most importantly,
>> > it's bug-free and extendable.
>> >
>> > Of interest to anyone who's only interested in public API changes
>> > are patches 2/14 and 6/14. The changes made are backwards compatible,
>> > and will not break API users.
>> >
>> > I've collaborated with Niklas Haas to make our APIs compatible, and
>> > we've both improved our codebases in the process. He'll be submitting
>> > a long-requested vf_libplacebo filter for review as soon as this gets merged.
>> >
>> > I'm also working on a Vulkan hwaccel using the new decode API,
>> > the branch for which can be found here:
>> > https://github.com/cyanreg/FFmpeg/commits/vulkan_decode
>> > Notably, libavfilter/vulkan.c will be moved into libavcodec/vulkan.h,
>> > and will be used to handle both filtering and decoding/encoding.
>> > No API changes will be required for it.
>> >
>> > I've also attached the patchset here.
>> >
>>
>> Ping.
>> libplacebo has already had the changes needed merged.
>>
>
> LGTM overall (except for one comment re: glslang version check).
>
> Will you squash the commits? (If not, there are some minor bug fixes in
> later commits that should probably be squashed onto the first commit)
>

Thanks, fixed all that you mentioned in the github comments.
I won't be squashing the commits. I think they're neatly segmented
into the areas they fix.
I might push this sometime late tonight if there are no further comments.

Could you submit your libplacebo patch too, if you feel like it's good enough?
Some notes about it:
 - Remove the RET redefinition
 - Replace stdbools with ints
 - I think you ought to set `apply_filmgrain` to be enabled by default,
   and rename the description to "Apply film grain side-data", since you
   also implemented support for H2645's film-grain. Would make sense to
   apply film grain if you're already doing processing.