[FFmpeg-devel] [PATCH v2] avcodec/nvenc: Include NVENC SDK header

Andreas Cadhalpun andreas.cadhalpun at googlemail.com
Sat Dec 12 11:25:20 CET 2015


On 12.12.2015 10:50, Hendrik Leppkes wrote:
> On Sat, Dec 12, 2015 at 10:39 AM, Andreas Cadhalpun
> <andreas.cadhalpun at googlemail.com> wrote:
>> On 12.12.2015 01:46, Philip Langdale wrote:
>>> On 2015-12-12 00:03, Andreas Cadhalpun wrote:
>>>> On 11.12.2015 09:41, Carl Eugen Hoyos wrote:
>>>>> My point is that so far several people have said that if nvenc
>>>>> is a system library there is no issue (and I fully agree). I
>>>>> didn't see a mail (and even less a patch with a commit message
>>>>> that says so) that claims nvenc is a system library (only that
>>>>> it "should" be one).
>>>>
>>>> So let's ask: Is someone here who claims that nvenc is a system
>>>> library and can explain why?
>>>
>>> I'm not going to claim it's a system library.
>>
>> Interesting...
>>
>>> I'm, instead, going to
>>> ask why we're having this conversation about nvenc, when the qsx/mfx
>>> situation is exactly the same.
>>
>> We have this conversation, because someone sent a patch to enable it
>> by default, together with including the header and removing the
>> 'die_license_disabled nonfree nvenc' line.
>>
>>> The functionality is provided by a
>>> proprietary set of modules that are part of the intel driver on windows
>>> and a separate (almost undiscoverable) download on linux (actually,
>>> that's worse than nvenc where the functionality is shipped with the
>>> driver in both cases). The only structural difference is that ffmpeg
>>> links against a wrapper library for mfx and dlopens in the nvenc
>>> case, but because of your following statement, that cannot make any
>>> difference.
>>
>> Since this requires the mfx wrapper to link, it is not enabled by
>> default. As the license situation seems similar, it might be a good idea
>> to add a 'die_license_disabled nonfree libmfx' line. But these don't have
>> any effect on the legal situation anyway, they are just a help
>> for our users.
>>
>>>>> I am glad we agree that there is no difference (license-wise) if
>>>>> a library is linked statically, dynamically or via dynamic
>>>>> loading;-)
>>>>
>>>> There is that, at least. ;-)
>>>
>>> Oh, and do you know what's funny - I just realised that the primary ffmpeg
>>> code base is LGPL and not GPL, so this whole conversation is slighlty
>>> pointless.
>>
>> No, it's not, because the LGPL and GPL are very similar in terms of the
>> requirements about distributing object code of (L)GPL-ed source code.
>>
>>> Combining ffmpeg with proprietary libraries is covered under section 6 and
>>> section 7,
>>
>> These sections only cover "work that uses the Library" (defined in section 5),
>> not the Library itself.
>>
>>> so even if building the nvenc codec is considered to combine
>>> ffmpeg with nvenc in this sense, it would be acceptable. The key requirement
>>> is that the LGPL covered parts can be rebuilt and modified as desired, and
>>> that is certainly true.
>>>
>>> These sections are generally thought of as enabling a larger proprietary
>>> program to pull in an LGPL library, but the language is symmetric.
>>
>> No, see section 4:
>> "You may copy and distribute the Library (or a portion or derivative of it,
>> under Section 2) in object code or executable form under the terms of Sections
>> 1 and 2 above provided that you accompany it with the complete corresponding
>> machine-readable source code"
>>
>>> Note that I actually don't believe with have a GPL problem here,
>>
>> Why?
>>
>>> but as a step forward, if we can all agree that the nvenc codec is a valid
>>> part of an lgpl build of ffmpeg, that's a step forward.
>>
>> I don't agree with that interpretation, see above explanation.
>>
> 
> We should just add an exception into the license to explicitly allow
> using it with the NVIDIA CUDA library and be done with this debate for
> ever.

That would be an option.

> You know that Open-Source has failed when the project itself is
> arguing days and days for including a feature on license reasons that
> any closed-source app would just write, enable and offer to its users
> without a second thought.

Please try to take a step back.
The "nvenc" feature is already included in FFmpeg, just not enabled by
default.

Best regards,
Andreas



More information about the ffmpeg-devel mailing list