Subtitle Overlays and Hardware-Accelerated Playback

This document describes some of the considerations and requirements that led to the current GstVideoOverlayCompositionMeta API which allows attaching of subtitle bitmaps or logos to video buffers.

Background

Subtitles can be muxed in containers or come from an external source.

Subtitles come in many shapes and colours. Usually they are either text-based (incl. 'pango markup'), or bitmap-based (e.g. DVD subtitles and the most common form of DVB subs). Bitmap based subtitles are usually compressed in some way, like some form of run-length encoding.

Subtitles are currently decoded and rendered in subtitle-format-specific overlay elements. These elements have two sink pads (one for raw video and one for the subtitle format in question) and one raw video source pad.

They will take care of synchronising the two input streams, and of decoding and rendering the subtitles on top of the raw video stream.

Digression: one could theoretically have dedicated decoder/render elements that output an AYUV or ARGB image, and then let a videomixer element do the actual overlaying, but this is not very efficient, because it requires us to allocate and blend whole pictures (1920x1080 AYUV = 8MB, 1280x720 AYUV = 3.6MB, 720x576 AYUV = 1.6MB) even if the overlay region is only a small rectangle at the bottom. This wastes memory and CPU. We could do something better by introducing a new format that only encodes the region(s) of interest, but we don't have such a format yet, and are not necessarily keen to rewrite this part of the logic in playbin at this point - and we can't change existing elements' behaviour, so would need to introduce new elements for this.

Playbin supports outputting compressed formats, i.e. it does not force decoding to a raw format, but is happy to output to a non-raw format as long as the sink supports that as well.

In case of certain hardware-accelerated decoding APIs, we will make use of that functionality. However, the decoder will not output a raw video format then, but some kind of hardware/API-specific format (in the caps) and the buffers will reference hardware/API-specific objects that the hardware/API-specific sink will know how to handle.

The Problem

In the case of such hardware-accelerated decoding, the decoder will not output raw pixels that can easily be manipulated. Instead, it will output hardware/API-specific objects that can later be used to render a frame using the same API.

Even if we could transform such a buffer into raw pixels, we most likely would want to avoid that, in order to avoid the need to map the data back into system memory (and then later back to the GPU). It's much better to upload the much smaller encoded data to the GPU/DSP and then leave it there until rendered.

Before GstVideoOverlayComposition playbin only supported subtitles on top of raw decoded video. It would try to find a suitable overlay element from the plugin registry based on the input subtitle caps and the rank. (It is assumed that we will be able to convert any raw video format into any format required by the overlay using a converter such as videoconvert.)

It would not render subtitles if the video sent to the sink is not raw YUV or RGB or if conversions had been disabled by setting the native-video flag on playbin.

Subtitle rendering is considered an important feature. Enabling hardware-accelerated decoding by default should not lead to a major feature regression in this area.

This means that we need to support subtitle rendering on top of non-raw video.

Possible Solutions

The goal is to keep knowledge of the subtitle format within the format-specific GStreamer plugins, and knowledge of any specific video acceleration API to the GStreamer plugins implementing that API. We do not want to make the pango/dvbsuboverlay/dvdspu/kate plugins link to libva/libvdpau/etc. and we do not want to make the vaapi/vdpau plugins link to all of libpango/libkate/libass etc.

Multiple possible solutions come to mind:

  1. backend-specific overlay elements

    e.g. vaapitextoverlay, vdpautextoverlay, vaapidvdspu, vdpaudvdspu, vaapidvbsuboverlay, vdpaudvbsuboverlay, etc.

    This assumes the overlay can be done directly on the backend-specific object passed around.

    The main drawback with this solution is that it leads to a lot of code duplication and may also lead to uncertainty about distributing certain duplicated pieces of code. The code duplication is pretty much unavoidable, since making textoverlay, dvbsuboverlay, dvdspu, kate, assrender, etc. available in form of base classes to derive from is not really an option. Similarly, one would not really want the vaapi/vdpau plugin to depend on a bunch of other libraries such as libpango, libkate, libtiger, libass, etc.

    One could add some new kind of overlay plugin feature though in combination with a generic base class of some sort, but in order to accommodate all the different cases and formats one would end up with quite convoluted/tricky API.

    (Of course there could also be a GstFancyVideoBuffer that provides an abstraction for such video accelerated objects and that could provide an API to add overlays to it in a generic way, but in the end this is just a less generic variant of (c), and it is not clear that there are real benefits to a specialised solution vs. a more generic one).

  2. convert backend-specific object to raw pixels and then overlay

    Even where possible technically, this is most likely very inefficient.

  3. attach the overlay data to the backend-specific video frame buffers in a generic way and do the actual overlaying/blitting later in backend-specific code such as the video sink (or an accelerated encoder/transcoder)

    In this case, the actual overlay rendering (i.e. the actual text rendering or decoding DVD/DVB data into pixels) is done in the subtitle-format-specific GStreamer plugin. All knowledge about the subtitle format is contained in the overlay plugin then, and all knowledge about the video backend in the video backend specific plugin.

    The main question then is how to get the overlay pixels (and we will only deal with pixels here) from the overlay element to the video sink.

    This could be done in multiple ways: One could send custom events downstream with the overlay data, or one could attach the overlay data directly to the video buffers in some way.

    Sending inline events has the advantage that is fairly transparent to any elements between the overlay element and the video sink: if an effects plugin creates a new video buffer for the output, nothing special needs to be done to maintain the subtitle overlay information, since the overlay data is not attached to the buffer. However, it slightly complicates things at the sink, since it would also need to look for the new event in question instead of just processing everything in its buffer render function.

    If one attaches the overlay data to the buffer directly, any element between overlay and video sink that creates a new video buffer would need to be aware of the overlay data attached to it and copy it over to the newly-created buffer.

    One would have to do implement a special kind of new query (e.g. FEATURE query) that is not passed on automatically by gst_pad_query_default() in order to make sure that all elements downstream will handle the attached overlay data. (This is only a problem if we want to also attach overlay data to raw video pixel buffers; for new non-raw types we can just make it mandatory and assume support and be done with it; for existing non-raw types nothing changes anyway if subtitles don't work) (we need to maintain backwards compatibility for existing raw video pipelines like e.g.: ..decoder ! suboverlay ! encoder..)

    Even though slightly more work, attaching the overlay information to buffers seems more intuitive than sending it interleaved as events. And buffers stored or passed around (e.g. via the "last-buffer" property in the sink when doing screenshots via playbin) always contain all the information needed.

  4. create a video/x-raw-*-delta format and use a backend-specific videomixer

    This possibility was hinted at already in the digression in section

    1. It would satisfy the goal of keeping subtitle format knowledge in the subtitle plugins and video backend knowledge in the video backend plugin. It would also add a concept that might be generally useful (think ximagesrc capture with xdamage). However, it would require adding foorender variants of all the existing overlay elements, and changing playbin to that new design, which is somewhat intrusive. And given the general nature of such a new format/API, we would need to take a lot of care to be able to accommodate all possible use cases when designing the API, which makes it considerably more ambitious. Lastly, we would need to write videomixer variants for the various accelerated video backends as well.

Overall (c) appears to be the most promising solution. It is the least intrusive and should be fairly straight-forward to implement with reasonable effort, requiring only small changes to existing elements and requiring no new elements.

Doing the final overlaying in the sink as opposed to a videomixer or overlay in the middle of the pipeline has other advantages:

  • if video frames need to be dropped, e.g. for QoS reasons, we could also skip the actual subtitle overlaying and possibly the decoding/rendering as well, if the implementation and API allows for that to be delayed.

  • the sink often knows the actual size of the window/surface/screen the output video is rendered to. This may make it possible to render the overlay image in a higher resolution than the input video, solving a long standing issue with pixelated subtitles on top of low-resolution videos that are then scaled up in the sink. This would require for the rendering to be delayed of course instead of just attaching an AYUV/ARGB/RGBA blog of pixels to the video buffer in the overlay, but that could all be supported.

  • if the video backend / sink has support for high-quality text rendering (clutter?) we could just pass the text or pango markup to the sink and let it do the rest (this is unlikely to be supported in the general case - text and glyph rendering is hard; also, we don't really want to make up our own text markup system, and pango markup is probably too limited for complex karaoke stuff).

API needed

  1. Representation of subtitle overlays to be rendered

    We need to pass the overlay pixels from the overlay element to the sink somehow. Whatever the exact mechanism, let's assume we pass a refcounted GstVideoOverlayComposition struct or object.

    A composition is made up of one or more overlays/rectangles.

    In the simplest case an overlay rectangle is just a blob of RGBA/ABGR [FIXME?] or AYUV pixels with positioning info and other metadata, and there is only one rectangle to render.

    We're keeping the naming generic ("OverlayFoo" rather than "SubtitleFoo") here, since this might also be handy for other use cases such as e.g. logo overlays or so. It is not designed for full-fledged video stream mixing though.

        // Note: don't mind the exact implementation details, they'll be hidden
        
        // FIXME: might be confusing in 0.11 though since GstXOverlay was
        //        renamed to GstVideoOverlay in 0.11, but not much we can do,
        //        maybe we can rename GstVideoOverlay to something better
        
        struct GstVideoOverlayComposition
        {
            guint                          num_rectangles;
            GstVideoOverlayRectangle    ** rectangles;
        
            /* lowest rectangle sequence number still used by the upstream
             * overlay element. This way a renderer maintaining some kind of
             * rectangles <-> surface cache can know when to free cached
             * surfaces/rectangles. */
            guint                          min_seq_num_used;
        
            /* sequence number for the composition (same series as rectangles) */
            guint                          seq_num;
        }
        
        struct GstVideoOverlayRectangle
        {
            /* Position on video frame and dimension of output rectangle in
             * output frame terms (already adjusted for the PAR of the output
             * frame). x/y can be negative (overlay will be clipped then) */
            gint  x, y;
            guint render_width, render_height;
        
            /* Dimensions of overlay pixels */
            guint width, height, stride;
        
            /* This is the PAR of the overlay pixels */
            guint par_n, par_d;
        
            /* Format of pixels, GST_VIDEO_FORMAT_ARGB on big-endian systems,
             * and BGRA on little-endian systems (i.e. pixels are treated as
             * 32-bit values and alpha is always in the most-significant byte,
             * and blue is in the least-significant byte).
             *
             * FIXME: does anyone actually use AYUV in practice? (we do
             * in our utility function to blend on top of raw video)
             * What about AYUV and endianness? Do we always have [A][Y][U][V]
             * in memory? */
            /* FIXME: maybe use our own enum? */
            GstVideoFormat format;
        
            /* Refcounted blob of memory, no caps or timestamps */
            GstBuffer *pixels;
        
            // FIXME: how to express source like text or pango markup?
            //        (just add source type enum + source buffer with data)
            //
            // FOR 0.10: always send pixel blobs, but attach source data in
            // addition (reason: if downstream changes, we can't renegotiate
            // that properly, if we just do a query of supported formats from
            // the start). Sink will just ignore pixels and use pango markup
            // from source data if it supports that.
            //
            // FOR 0.11: overlay should query formats (pango markup, pixels)
            // supported by downstream and then only send that. We can
            // renegotiate via the reconfigure event.
            //
        
            /* sequence number: useful for backends/renderers/sinks that want
             * to maintain a cache of rectangles <-> surfaces. The value of
             * the min_seq_num_used in the composition tells the renderer which
             * rectangles have expired. */
            guint      seq_num;
        
            /* FIXME: we also need a (private) way to cache converted/scaled
             * pixel blobs */
        }
    
    (a1) Overlay consumer
        API:
    
        How would this work in a video sink that supports scaling of textures:
        
        gst_foo_sink_render () {
          /* assume only one for now */
          if video_buffer has composition:
            composition = video_buffer.get_composition()
        
            for each rectangle in composition:
              if rectangle.source_data_type == PANGO_MARKUP
                actor = text_from_pango_markup (rectangle.get_source_data())
              else
                pixels = rectangle.get_pixels_unscaled (FORMAT_RGBA, ...)
                actor = texture_from_rgba (pixels, ...)
        
              .. position + scale on top of video surface ...
        }
    
    (a2) Overlay producer
        API:
    
        e.g. logo or subpicture overlay: got pixels, stuff into rectangle:
        
         if (logoverlay->cached_composition == NULL) {
           comp = composition_new ();
        
           rect = rectangle_new (format, pixels_buf,
                                 width, height, stride, par_n, par_d,
                                 x, y, render_width, render_height);
        
           /* composition adds its own ref for the rectangle */
           composition_add_rectangle (comp, rect);
           rectangle_unref (rect);
        
           /* buffer adds its own ref for the composition */
           video_buffer_attach_composition (comp);
        
           /* we take ownership of the composition and save it for later */
           logoverlay->cached_composition = comp;
         } else {
           video_buffer_attach_composition (logoverlay->cached_composition);
         }
FIXME: also add some API to modify render position/dimensions of a
rectangle (probably requires creation of new rectangle, unless we
handle writability like with other mini objects).
  1. Fallback overlay rendering/blitting on top of raw video

    Eventually we want to use this overlay mechanism not only for hardware-accelerated video, but also for plain old raw video, either at the sink or in the overlay element directly.

    Apart from the advantages listed earlier in section 3, this allows us to consolidate a lot of overlaying/blitting code that is currently repeated in every single overlay element in one location. This makes it considerably easier to support a whole range of raw video formats out of the box, add SIMD-optimised rendering using ORC, or handle corner cases correctly.

    (Note: side-effect of overlaying raw video at the video sink is that if e.g. a screnshotter gets the last buffer via the last-buffer property of basesink, it would get an image without the subtitles on top. This could probably be fixed by re-implementing the property in GstVideoSink though. Playbin2 could handle this internally as well).

        void
        gst_video_overlay_composition_blend (GstVideoOverlayComposition * comp
                                             GstBuffer                  * video_buf)
        {
          guint n;
        
          g_return_if_fail (gst_buffer_is_writable (video_buf));
          g_return_if_fail (GST_BUFFER_CAPS (video_buf) != NULL);
        
          ... parse video_buffer caps into BlendVideoFormatInfo ...
        
          for each rectangle in the composition: {
        
                 if (gst_video_format_is_yuv (video_buf_format)) {
                   overlay_format = FORMAT_AYUV;
                 } else if (gst_video_format_is_rgb (video_buf_format)) {
                   overlay_format = FORMAT_ARGB;
                 } else {
                   /* FIXME: grayscale? */
                   return;
                 }
        
                 /* this will scale and convert AYUV<->ARGB if needed */
                 pixels = rectangle_get_pixels_scaled (rectangle, overlay_format);
        
                 ... clip output rectangle ...
        
                 __do_blend (video_buf_format, video_buf->data,
                             overlay_format, pixels->data,
                             x, y, width, height, stride);
        
                 gst_buffer_unref (pixels);
          }
        }
  1. Flatten all rectangles in a composition

    We cannot assume that the video backend API can handle any number of rectangle overlays, it's possible that it only supports one single overlay, in which case we need to squash all rectangles into one.

    However, we'll just declare this a corner case for now, and implement it only if someone actually needs it. It's easy to add later API-wise. Might be a bit tricky if we have rectangles with different PARs/formats (e.g. subs and a logo), though we could probably always just use the code from (b) with a fully transparent video buffer to create a flattened overlay buffer.

  2. query support for the new video composition mechanism

    This is handled via GstMeta and an ALLOCATION query - we can simply query whether downstream supports the GstVideoOverlayComposition meta.

    There appears to be no issue with downstream possibly not being linked yet at the time when an overlay would want to do such a query, but we would just have to default to something and update ourselves later on a reconfigure event then.

Other considerations:

  • renderers (overlays or sinks) may be able to handle only ARGB or only AYUV (for most graphics/hw-API it's likely ARGB of some sort, while our blending utility functions will likely want the same colour space as the underlying raw video format, which is usually YUV of some sort). We need to convert where required, and should cache the conversion.

  • renderers may or may not be able to scale the overlay. We need to do the scaling internally if not (simple case: just horizontal scaling to adjust for PAR differences; complex case: both horizontal and vertical scaling, e.g. if subs come from a different source than the video or the video has been rescaled or cropped between overlay element and sink).

  • renderers may be able to generate (possibly scaled) pixels on demand from the original data (e.g. a string or RLE-encoded data). We will ignore this for now, since this functionality can still be added later via API additions. The most interesting case would be to pass a pango markup string, since e.g. clutter can handle that natively.

  • renderers may be able to write data directly on top of the video pixels (instead of creating an intermediary buffer with the overlay which is then blended on top of the actual video frame), e.g. dvdspu, dvbsuboverlay

However, in the interest of simplicity, we should probably ignore the fact that some elements can blend their overlays directly on top of the video (decoding/uncompressing them on the fly), even more so as it's not obvious that it's actually faster to decode the same overlay 70-90 times (say) (ie. ca. 3 seconds of video frames) and then blend it 70-90 times instead of decoding it once into a temporary buffer and then blending it directly from there, possibly SIMD-accelerated. Also, this is only relevant if the video is raw video and not some hardware-acceleration backend object.

And ultimately it is the overlay element that decides whether to do the overlay right there and then or have the sink do it (if supported). It could decide to keep doing the overlay itself for raw video and only use our new API for non-raw video.

  • renderers may want to make sure they only upload the overlay pixels once per rectangle if that rectangle recurs in subsequent frames (as part of the same composition or a different composition), as is likely. This caching of e.g. surfaces needs to be done renderer-side and can be accomplished based on the sequence numbers. The composition contains the lowest sequence number still in use upstream (an overlay element may want to cache created compositions+rectangles as well after all to re-use them for multiple frames), based on that the renderer can expire cached objects. The caching needs to be done renderer-side because attaching renderer-specific objects to the rectangles won't work well given the refcounted nature of rectangles and compositions, making it unpredictable when a rectangle or composition will be freed or from which thread context it will be freed. The renderer-specific objects are likely bound to other types of renderer-specific contexts, and need to be managed in connection with those.

  • composition/rectangles should internally provide a certain degree of thread-safety. Multiple elements (sinks, overlay element) might access or use the same objects from multiple threads at the same time, and it is expected that elements will keep a ref to compositions and rectangles they push downstream for a while, e.g. until the current subtitle composition expires.

Future considerations

  • alternatives: there may be multiple versions/variants of the same subtitle stream. On DVDs, there may be a 4:3 version and a 16:9 version of the same subtitles. We could attach both variants and let the renderer pick the best one for the situation (currently we just use the 16:9 version). With totem, it's ultimately totem that adds the 'black bars' at the top/bottom, so totem also knows if it's got a 4:3 display and can/wants to fit 4:3 subs (which may render on top of the bars) or not, for example.

Misc. FIXMEs

TEST: should these look (roughly) alike (note text distortion) - needs fixing in textoverlay

gst-launch-1.0 \
   videotestsrc ! video/x-raw,width=640,height=480,pixel-aspect-ratio=1/1 \
     ! textoverlay text=Hello font-desc=72 ! xvimagesink \
   videotestsrc ! video/x-raw,width=320,height=480,pixel-aspect-ratio=2/1 \
     ! textoverlay text=Hello font-desc=72 ! xvimagesink \
   videotestsrc ! video/x-raw,width=640,height=240,pixel-aspect-ratio=1/2 \
     ! textoverlay text=Hello font-desc=72 ! xvimagesink

The results of the search are