[FFmpeg-devel] [PATCH 1/4] vp8: Add hwaccel hooks

Mark Thompson sw at jkqxz.net
Mon Feb 20 13:20:48 EET 2017

On 20/02/17 10:44, Michael Niedermayer wrote:
> On Mon, Feb 20, 2017 at 10:23:27AM +0000, Mark Thompson wrote:
>> On 20/02/17 02:35, Michael Niedermayer wrote:
>>> On Sun, Feb 19, 2017 at 09:29:33PM +0000, Mark Thompson wrote:
>>>> On 19/02/17 21:04, Ronald S. Bultje wrote:
>>>>> Hi,
>>>>> On Sun, Feb 19, 2017 at 12:23 PM, Mark Thompson <sw at jkqxz.net> wrote:
>>>>>> diff --git a/libavcodec/webp.c b/libavcodec/webp.c
>>>>> [..]
>>>>>> +        avctx->get_format = webp_get_format;
>>>>> Docs say:
>>>>> "decoding: Set by user, if not set the native format will be chosen."
>>>>> So I don't think decoders are supposed to set this.
>>>> The webp decoder does not support get_format.  I suppose the user could technically store something there and then read it back, so save and restore the value across the relevant calls?
>>> This is quite ugly, why do you want to do that ?
>>> get_format is set by the user
>>> the get_format API requires the function to choose one of the caller
>>> provided formats and it can choose any.
>>> I dont know what your function does but its different from the API.
>>> It smells very much like a hack ...
>>> The one situation in which you can set get_format from libavcodec is
>>> when there is a seperate codec context you created, that is that you
>>> are its user.
>>> can you explain why this code messes with avctx->get_format ?
>>> and for example doesnt change the code that calls get_format so that
>>> it passes a correct pixel format list which then by any get_format()
>>> results in a correct format ?
>>> or am i missing something ?
>> The current WebP decoder calls the VP8 decoder /using the same AVCodecContext/.
> so they are kind of the same decoder
>> Previously the format was set directly on the VP8Context, but now that the VP8 decoder uses ff_get_format() that initial value is not used.  This change adds a get_format() callback on the AVCodecContext used by the VP8 decoder to set the format appropriately.
> But this is semantically wrong, the formats supported by the decoder
> implementation are choosen by the code calling get_format().
> the get_format callback chooses amongth these formats depending on the
> users preferrance/usefullness.

Yes.  It's a hack to make the VP8 decoder allocate a non-native frame format (YUVA420P) which is compatible with its native format (YUV420P) for any use it makes of it.  This is a hack which already exists, we are just moving it to a different place now that the VP8 decoder uses ff_get_format().

>> Now, because that is the same AVCodecContext as the one supplied by the user, the user could themselves have stored something into the get_format field (even though it isn't used by the WebP decoder) and expect to be able to read it back correctly later.  I think this would be madness, but is I suppose technically allowed; saving and restoring the value across the VP8 calls would fix that case.
> Why do you try to misuse the API ?

To make the minimal modifications to existing hacks for fate to pass for WebP.  I do not care at all about WebP, but the existing hacks abusing the VP8 decoder block the other changes.

> i mean i think we agree that this violates the API. Making sure it
> works doesnt solve that it violates the API. And anyone working on
> the API would likely assume that code conforms to the API documentation.
> -> the developer would think one thing and the code would do another
> thats a recipe for creating bugs.
> I think the code calling *get_format() should pass the correct list
> of formats to the callback and not some other part override the users
> get_format calback to work around that the list passed is wrong.

The user's get_format callback is never used.  (There is only one output format possibility for each input, so it wouldn't do anything anyway.)

> am i missing something ?
> is what i suggested difficult to do or do you see a issue with that
> solution ?

The list passed by the VP8 decoder to ff_get_format() is not wrong - the VP8 decoder does not support alpha so including YUVA420P would be incorrect.

- Mark

More information about the ffmpeg-devel mailing list