On 06/09/17 19:14, wm4 wrote:
> On Wed, 6 Sep 2017 18:55:46 +0100
> Mark Thompson <[email protected]> wrote:
> 
>> On 06/09/17 17:29, wm4 wrote:
>>> On Wed, 6 Sep 2017 16:53:04 +0100
>>> Mark Thompson <[email protected]> wrote:
>>>   
>>>> On 06/09/17 15:37, wm4 wrote:  
>>>>> On Tue,  5 Sep 2017 23:59:31 +0100
>>>>> Mark Thompson <[email protected]> wrote:
>>>>>     
>>>>>> ---
>>>>>>  doc/APIchanges       |  3 +++
>>>>>>  libavcodec/avcodec.h | 23 +++++++++++++++++++++++
>>>>>>  2 files changed, 26 insertions(+)
>>>>>>
>>>>>> diff --git a/doc/APIchanges b/doc/APIchanges
>>>>>> index 6f70f3c96..f21dc4db0 100644
>>>>>> --- a/doc/APIchanges
>>>>>> +++ b/doc/APIchanges
>>>>>> @@ -14,6 +14,9 @@ libavutil:     2017-03-23
>>>>>>  API changes, most recent first:
>>>>>>  
>>>>>>  2017-xx-xx - xxxxxxx - lavc 58.x+1.0 - avcodec.h
>>>>>> +  Add AVCodecHWConfig and AVCodec.hw_configs.
>>>>>> +
>>>>>> +2017-xx-xx - xxxxxxx - lavc 58.x+1.0 - avcodec.h
>>>>>>    Add AV_HWACCEL_FLAG_REUSE_CONTEXT.
>>>>>>  
>>>>>>  2017-xx-xx - xxxxxxx - lavfi 7.x+1.0 - avfilter.h
>>>>>> diff --git a/libavcodec/avcodec.h b/libavcodec/avcodec.h
>>>>>> index 89e432d16..4f6b76203 100644
>>>>>> --- a/libavcodec/avcodec.h
>>>>>> +++ b/libavcodec/avcodec.h
>>>>>> @@ -35,6 +35,7 @@
>>>>>>  #include "libavutil/cpu.h"
>>>>>>  #include "libavutil/dict.h"
>>>>>>  #include "libavutil/frame.h"
>>>>>> +#include "libavutil/hwcontext.h"
>>>>>>  #include "libavutil/log.h"
>>>>>>  #include "libavutil/pixfmt.h"
>>>>>>  #include "libavutil/rational.h"
>>>>>> @@ -2773,6 +2774,19 @@ typedef struct AVProfile {
>>>>>>      const char *name; ///< short name for the profile
>>>>>>  } AVProfile;
>>>>>>  
>>>>>> +typedef struct AVCodecHWConfig {
>>>>>> +    /**
>>>>>> +     * A device type which can be supplied to the codec.
>>>>>> +     */
>>>>>> +    enum AVHWDeviceType device_type;
>>>>>> +    /**
>>>>>> +     * The matching hardware pixel format which the codec can produce.
>>>>>> +     *
>>>>>> +     * Set to AV_PIX_FMT_NONE if no hardware pixel format is available.
>>>>>> +     */
>>>>>> +    enum AVPixelFormat pix_fmt;
>>>>>> +} AVCodecHWConfig;
>>>>>> +
>>>>>>  typedef struct AVCodecDefault AVCodecDefault;
>>>>>>  
>>>>>>  struct AVSubtitle;
>>>>>> @@ -2808,6 +2822,15 @@ typedef struct AVCodec {
>>>>>>      const AVClass *priv_class;              ///< AVClass for the 
>>>>>> private context
>>>>>>      const AVProfile *profiles;              ///< array of recognized 
>>>>>> profiles, or NULL if unknown, array is terminated by {FF_PROFILE_UNKNOWN}
>>>>>>  
>>>>>> +    /**
>>>>>> +     * Array of hardware configurations supported by the codec, or NULL
>>>>>> +     * if no hardware supported.
>>>>>> +     *
>>>>>> +     * The array is terminated by a configuration with a device_type of
>>>>>> +     * AV_HW_DEVICE_TYPE_NONE.
>>>>>> +     */
>>>>>> +    const AVCodecHWConfig *hw_configs;
>>>>>> +
>>>>>>      /*****************************************************************
>>>>>>       * No fields below this line are part of the public API. They
>>>>>>       * may not be used outside of libavcodec and can be changed and    
>>>>>
>>>>> What if a decoder provides hardware decoding, but has no associated
>>>>> AVHWDeviceType? (Such as mmal.)    
>>>>
>>>> They are in the pix_fmts list.  What other information do you want?  
>>>
>>> They don't necessarily output a hw format. For example, crystalhd
>>> (ffmpeg only) doesn't. It would also not be very uniform.  
>>
>> If it doesn't require any additional setup and it outputs software frames 
>> then it's indistinguishable from a normal software decoder; therefore, treat 
>> it as such.
>>
>> MMAL is harder - I'm only treating the real hwaccels here because they now 
>> have a consistent form we can make use of.
>>
>> I haven't really thought about this properly, but how about something like:
>>
>> enum AVCodecHWConfigMethod {
>>   AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE = 1,
>>   AV_CODEC_HW_CONFIG_METHOD_HW_FRAMES = 2,
>>   AV_CODEC_HW_CONFIG_METHOD_LEGACY_SHIT = 4,
>>   AV_CODEC_HW_CONFIG_METHOD_MAGIC = 8,
>> };
>>
>> struct AVCodecHWConfig {
>>   int methods;
>>   enum AVHWDeviceType device_type;
>>   enum AVPixelFormat pix_fmt;
>> };
>>
>> Then you would label MMAL as MAGIC, because it can just output the MMAL 
>> frames without additional setup.  VAAPI/VDPAU would be HW_DEVICE | HW_FRAMES 
>> | LEGACY_SHIT.  QSV would be HW_FRAMES | LEGACY_SHIT.
> 
> That sounds OK too. Maybe we should include the information whether it's
> a hwaccel, i.e. something that has an actual software decoder under the
> hood, and that dynamic fallback/resuming HW decoding is supported?

I was thinking that was implied by not-MAGIC, but it isn't.  Yeah, that flag 
would help.  Does it actually need to be externally-visible, though?

> What is LEGACY_SHIT exactly?

"This has some arcane rules that can't be expressed here, go read the code" - 
the old hwaccels where you have to put structure X in hwaccel_context and call 
special function Y before you do anything.

>> Is there then sufficient information for ff_get_format() to look at this 
>> list to decide on whether it needs an AVHWAccel, and therefore eliminate 
>> dummy hwaccels?
> 
> I suspect it is. I wonder why this check is even being done at all.

With the needs_hwaccel flag then yes.
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to