Re: [Mlt-devel] Webvfx producer transparency

[Dan D. <da...@de...>]

On Fri, Jun 21, 2013 at 11:31 AM, Dan Dennedy <da...@de...> wrote:
> On Tue, Jun 18, 2013 at 6:10 AM, j-b-m <j-...@us...> wrote:
>> I still have a problem when trying to use these producers in Kdenlive because
>> in EffectsImpl::render(double time, Image* renderImage), the
>> QMetaObject::invokeMethod(this, "renderInvokable" (...) sometimes freezes.
>
> What I have noticed while trying to integrate the webvfx filter into
> Shotcut is that if I call Consumer::stop() then it deadlocks because
> it it called on the main Qt thread, inside mlt_consumer_close() it
> waits for the render thread to finish, but the webvfx renderInvokable
> call is waiting on the main Qt thread to process the slot via blocking
> queued connection. I have been tinkering with workarounds that I am
> not so happy with thus far. What I discovered is that if I disable the
> filter, wait a little for the currently blocked render to complete,
> then I can call stop because the consumer render thread is not asking
> webvfx to render anything anymore. I have successfully performed that
> wait using nanosleep() or a QTimer::singleShot() within a method that
> I have split into 2 parts around the call to singleShot. However, I
> would rather have something that works within MLT. I am thinking the
> call to renderInvokable should not be a blocking queued connection.
> Rather, perhaps it should wait on a condition after the invocation -
> one that can be signaled from the actual render or from mlt_consumer
> prior to joining the render thread.
>

JB, I have committed changes to MLT and my WebVfx fork that implements
the last sentence above. I have successfully tested it in my working
copy of Shotcut. It does, however, require that the app sets a
"consumer" data property on the webvfx server to the mlt_consumer
because webvfx uses this to setup an event listener on a stop event in
order to cancel a pending render. I might be a very awkward traversal
for the service to discover the end consumer itself and perhaps
impossible with things like multi consumer. Indeed, I just discovered
I still have a problem with sdl_preview. :-\

--
+-DRD-+